annotate src/share/vm/opto/memnode.cpp @ 113:ba764ed4b6f2

6420645: Create a vm that uses compressed oops for up to 32gb heapsizes Summary: Compressed oops in instances, arrays, and headers. Code contributors are coleenp, phh, never, swamyv Reviewed-by: jmasa, kamg, acorn, tbell, kvn, rasbold
author coleenp
date Sun, 13 Apr 2008 17:43:42 -0400
parents de93acbb64fc
children d1a5218d7eaf
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@74 136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM) {
kvn@74 137 // clone the Phi with our address type
kvn@74 138 result = mphi->split_out_instance(t_adr, igvn);
kvn@74 139 } else {
kvn@74 140 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
kvn@74 141 }
kvn@74 142 }
kvn@74 143 return result;
kvn@74 144 }
kvn@74 145
kvn@64 146 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
kvn@64 147 uint alias_idx = phase->C->get_alias_index(tp);
kvn@64 148 Node *mem = mmem;
kvn@64 149 #ifdef ASSERT
kvn@64 150 {
kvn@64 151 // Check that current type is consistent with the alias index used during graph construction
kvn@64 152 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
kvn@64 153 bool consistent = adr_check == NULL || adr_check->empty() ||
kvn@64 154 phase->C->must_alias(adr_check, alias_idx );
kvn@64 155 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
kvn@64 156 if( !consistent && adr_check != NULL && !adr_check->empty() &&
kvn@64 157 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
kvn@64 158 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
kvn@64 159 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
kvn@64 160 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
kvn@64 161 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
kvn@64 162 // don't assert if it is dead code.
kvn@64 163 consistent = true;
kvn@64 164 }
kvn@64 165 if( !consistent ) {
kvn@64 166 st->print("alias_idx==%d, adr_check==", alias_idx);
kvn@64 167 if( adr_check == NULL ) {
kvn@64 168 st->print("NULL");
kvn@64 169 } else {
kvn@64 170 adr_check->dump();
kvn@64 171 }
kvn@64 172 st->cr();
kvn@64 173 print_alias_types();
kvn@64 174 assert(consistent, "adr_check must match alias idx");
kvn@64 175 }
kvn@64 176 }
kvn@64 177 #endif
kvn@64 178 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
kvn@64 179 // means an array I have not precisely typed yet. Do not do any
kvn@64 180 // alias stuff with it any time soon.
kvn@64 181 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 182 if( tp->base() != Type::AnyPtr &&
kvn@64 183 !(tinst &&
kvn@64 184 tinst->klass()->is_java_lang_Object() &&
kvn@64 185 tinst->offset() == Type::OffsetBot) ) {
kvn@64 186 // compress paths and change unreachable cycles to TOP
kvn@64 187 // If not, we can update the input infinitely along a MergeMem cycle
kvn@64 188 // Equivalent code in PhiNode::Ideal
kvn@64 189 Node* m = phase->transform(mmem);
kvn@64 190 // If tranformed to a MergeMem, get the desired slice
kvn@64 191 // Otherwise the returned node represents memory for every slice
kvn@64 192 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
kvn@64 193 // Update input if it is progress over what we have now
kvn@64 194 }
kvn@64 195 return mem;
kvn@64 196 }
kvn@64 197
duke@0 198 //--------------------------Ideal_common---------------------------------------
duke@0 199 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
duke@0 200 // Unhook non-raw memories from complete (macro-expanded) initializations.
duke@0 201 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
duke@0 202 // If our control input is a dead region, kill all below the region
duke@0 203 Node *ctl = in(MemNode::Control);
duke@0 204 if (ctl && remove_dead_region(phase, can_reshape))
duke@0 205 return this;
duke@0 206
duke@0 207 // Ignore if memory is dead, or self-loop
duke@0 208 Node *mem = in(MemNode::Memory);
duke@0 209 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 210 assert( mem != this, "dead loop in MemNode::Ideal" );
duke@0 211
duke@0 212 Node *address = in(MemNode::Address);
duke@0 213 const Type *t_adr = phase->type( address );
duke@0 214 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 215
duke@0 216 // Avoid independent memory operations
duke@0 217 Node* old_mem = mem;
duke@0 218
kvn@36 219 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@36 220 // initializations was removed. After macro-expansion all stores catched
kvn@36 221 // by Initialize node became raw stores and there is no information
kvn@36 222 // which memory slices they modify. So it is unsafe to move any memory
kvn@36 223 // operation above these stores. Also in most cases hooked non-raw memories
kvn@36 224 // were already unhooked by using information from detect_ptr_independence()
kvn@36 225 // and find_previous_store().
duke@0 226
duke@0 227 if (mem->is_MergeMem()) {
duke@0 228 MergeMemNode* mmem = mem->as_MergeMem();
duke@0 229 const TypePtr *tp = t_adr->is_ptr();
kvn@64 230
kvn@64 231 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@0 232 }
duke@0 233
duke@0 234 if (mem != old_mem) {
duke@0 235 set_req(MemNode::Memory, mem);
duke@0 236 return this;
duke@0 237 }
duke@0 238
duke@0 239 // let the subclass continue analyzing...
duke@0 240 return NULL;
duke@0 241 }
duke@0 242
duke@0 243 // Helper function for proving some simple control dominations.
duke@0 244 // Attempt to prove that control input 'dom' dominates (or equals) 'sub'.
duke@0 245 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@0 246 // is not a constant (dominated by the method's StartNode).
duke@0 247 // Used by MemNode::find_previous_store to prove that the
duke@0 248 // control input of a memory operation predates (dominates)
duke@0 249 // an allocation it wants to look past.
duke@0 250 bool MemNode::detect_dominating_control(Node* dom, Node* sub) {
duke@0 251 if (dom == NULL) return false;
duke@0 252 if (dom->is_Proj()) dom = dom->in(0);
duke@0 253 if (dom->is_Start()) return true; // anything inside the method
duke@0 254 if (dom->is_Root()) return true; // dom 'controls' a constant
duke@0 255 int cnt = 20; // detect cycle or too much effort
duke@0 256 while (sub != NULL) { // walk 'sub' up the chain to 'dom'
duke@0 257 if (--cnt < 0) return false; // in a cycle or too complex
duke@0 258 if (sub == dom) return true;
duke@0 259 if (sub->is_Start()) return false;
duke@0 260 if (sub->is_Root()) return false;
duke@0 261 Node* up = sub->in(0);
duke@0 262 if (sub == up && sub->is_Region()) {
duke@0 263 for (uint i = 1; i < sub->req(); i++) {
duke@0 264 Node* in = sub->in(i);
duke@0 265 if (in != NULL && !in->is_top() && in != sub) {
duke@0 266 up = in; break; // take any path on the way up to 'dom'
duke@0 267 }
duke@0 268 }
duke@0 269 }
duke@0 270 if (sub == up) return false; // some kind of tight cycle
duke@0 271 sub = up;
duke@0 272 }
duke@0 273 return false;
duke@0 274 }
duke@0 275
duke@0 276 //---------------------detect_ptr_independence---------------------------------
duke@0 277 // Used by MemNode::find_previous_store to prove that two base
duke@0 278 // pointers are never equal.
duke@0 279 // The pointers are accompanied by their associated allocations,
duke@0 280 // if any, which have been previously discovered by the caller.
duke@0 281 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@0 282 Node* p2, AllocateNode* a2,
duke@0 283 PhaseTransform* phase) {
duke@0 284 // Attempt to prove that these two pointers cannot be aliased.
duke@0 285 // They may both manifestly be allocations, and they should differ.
duke@0 286 // Or, if they are not both allocations, they can be distinct constants.
duke@0 287 // Otherwise, one is an allocation and the other a pre-existing value.
duke@0 288 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@0 289 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@0 290 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@0 291 return (a1 != a2);
duke@0 292 } else if (a1 != NULL) { // one allocation a1
duke@0 293 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
duke@0 294 return detect_dominating_control(p2->in(0), a1->in(0));
duke@0 295 } else { //(a2 != NULL) // one allocation a2
duke@0 296 return detect_dominating_control(p1->in(0), a2->in(0));
duke@0 297 }
duke@0 298 return false;
duke@0 299 }
duke@0 300
duke@0 301
duke@0 302 // The logic for reordering loads and stores uses four steps:
duke@0 303 // (a) Walk carefully past stores and initializations which we
duke@0 304 // can prove are independent of this load.
duke@0 305 // (b) Observe that the next memory state makes an exact match
duke@0 306 // with self (load or store), and locate the relevant store.
duke@0 307 // (c) Ensure that, if we were to wire self directly to the store,
duke@0 308 // the optimizer would fold it up somehow.
duke@0 309 // (d) Do the rewiring, and return, depending on some other part of
duke@0 310 // the optimizer to fold up the load.
duke@0 311 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@0 312 // specific to loads and stores, so they are handled by the callers.
duke@0 313 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@0 314 //
duke@0 315 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@0 316 Node* ctrl = in(MemNode::Control);
duke@0 317 Node* adr = in(MemNode::Address);
duke@0 318 intptr_t offset = 0;
duke@0 319 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 320 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@0 321
duke@0 322 if (offset == Type::OffsetBot)
duke@0 323 return NULL; // cannot unalias unless there are precise offsets
duke@0 324
kvn@74 325 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
kvn@74 326
duke@0 327 intptr_t size_in_bytes = memory_size();
duke@0 328
duke@0 329 Node* mem = in(MemNode::Memory); // start searching here...
duke@0 330
duke@0 331 int cnt = 50; // Cycle limiter
duke@0 332 for (;;) { // While we can dance past unrelated stores...
duke@0 333 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@0 334
duke@0 335 if (mem->is_Store()) {
duke@0 336 Node* st_adr = mem->in(MemNode::Address);
duke@0 337 intptr_t st_offset = 0;
duke@0 338 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@0 339 if (st_base == NULL)
duke@0 340 break; // inscrutable pointer
duke@0 341 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@0 342 const int MAX_STORE = BytesPerLong;
duke@0 343 if (st_offset >= offset + size_in_bytes ||
duke@0 344 st_offset <= offset - MAX_STORE ||
duke@0 345 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@0 346 // Success: The offsets are provably independent.
duke@0 347 // (You may ask, why not just test st_offset != offset and be done?
duke@0 348 // The answer is that stores of different sizes can co-exist
duke@0 349 // in the same sequence of RawMem effects. We sometimes initialize
duke@0 350 // a whole 'tile' of array elements with a single jint or jlong.)
duke@0 351 mem = mem->in(MemNode::Memory);
duke@0 352 continue; // (a) advance through independent store memory
duke@0 353 }
duke@0 354 }
duke@0 355 if (st_base != base &&
duke@0 356 detect_ptr_independence(base, alloc,
duke@0 357 st_base,
duke@0 358 AllocateNode::Ideal_allocation(st_base, phase),
duke@0 359 phase)) {
duke@0 360 // Success: The bases are provably independent.
duke@0 361 mem = mem->in(MemNode::Memory);
duke@0 362 continue; // (a) advance through independent store memory
duke@0 363 }
duke@0 364
duke@0 365 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@0 366 // since we have not managed to prove 'this' and 'mem' independent.
duke@0 367 if (st_base == base && st_offset == offset) {
duke@0 368 return mem; // let caller handle steps (c), (d)
duke@0 369 }
duke@0 370
duke@0 371 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 372 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@0 373 AllocateNode* st_alloc = st_init->allocation();
duke@0 374 if (st_alloc == NULL)
duke@0 375 break; // something degenerated
duke@0 376 bool known_identical = false;
duke@0 377 bool known_independent = false;
duke@0 378 if (alloc == st_alloc)
duke@0 379 known_identical = true;
duke@0 380 else if (alloc != NULL)
duke@0 381 known_independent = true;
duke@0 382 else if (ctrl != NULL &&
duke@0 383 detect_dominating_control(ctrl, st_alloc->in(0)))
duke@0 384 known_independent = true;
duke@0 385
duke@0 386 if (known_independent) {
duke@0 387 // The bases are provably independent: Either they are
duke@0 388 // manifestly distinct allocations, or else the control
duke@0 389 // of this load dominates the store's allocation.
duke@0 390 int alias_idx = phase->C->get_alias_index(adr_type());
duke@0 391 if (alias_idx == Compile::AliasIdxRaw) {
duke@0 392 mem = st_alloc->in(TypeFunc::Memory);
duke@0 393 } else {
duke@0 394 mem = st_init->memory(alias_idx);
duke@0 395 }
duke@0 396 continue; // (a) advance through independent store memory
duke@0 397 }
duke@0 398
duke@0 399 // (b) at this point, if we are not looking at a store initializing
duke@0 400 // the same allocation we are loading from, we lose.
duke@0 401 if (known_identical) {
duke@0 402 // From caller, can_see_stored_value will consult find_captured_store.
duke@0 403 return mem; // let caller handle steps (c), (d)
duke@0 404 }
duke@0 405
kvn@74 406 } else if (addr_t != NULL && addr_t->is_instance_field()) {
kvn@74 407 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
kvn@74 408 if (mem->is_Proj() && mem->in(0)->is_Call()) {
kvn@74 409 CallNode *call = mem->in(0)->as_Call();
kvn@74 410 if (!call->may_modify(addr_t, phase)) {
kvn@74 411 mem = call->in(TypeFunc::Memory);
kvn@74 412 continue; // (a) advance through independent call memory
kvn@74 413 }
kvn@74 414 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
kvn@74 415 mem = mem->in(0)->in(TypeFunc::Memory);
kvn@74 416 continue; // (a) advance through independent MemBar memory
kvn@74 417 } else if (mem->is_MergeMem()) {
kvn@74 418 int alias_idx = phase->C->get_alias_index(adr_type());
kvn@74 419 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@74 420 continue; // (a) advance through independent MergeMem memory
kvn@74 421 }
duke@0 422 }
duke@0 423
duke@0 424 // Unless there is an explicit 'continue', we must bail out here,
duke@0 425 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@0 426 break;
duke@0 427 }
duke@0 428
duke@0 429 return NULL; // bail out
duke@0 430 }
duke@0 431
duke@0 432 //----------------------calculate_adr_type-------------------------------------
duke@0 433 // Helper function. Notices when the given type of address hits top or bottom.
duke@0 434 // Also, asserts a cross-check of the type against the expected address type.
duke@0 435 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@0 436 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@0 437 #ifdef PRODUCT
duke@0 438 cross_check = NULL;
duke@0 439 #else
duke@0 440 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@0 441 #endif
duke@0 442 const TypePtr* tp = t->isa_ptr();
duke@0 443 if (tp == NULL) {
duke@0 444 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@0 445 return TypePtr::BOTTOM; // touches lots of memory
duke@0 446 } else {
duke@0 447 #ifdef ASSERT
duke@0 448 // %%%% [phh] We don't check the alias index if cross_check is
duke@0 449 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@0 450 if (cross_check != NULL &&
duke@0 451 cross_check != TypePtr::BOTTOM &&
duke@0 452 cross_check != TypeRawPtr::BOTTOM) {
duke@0 453 // Recheck the alias index, to see if it has changed (due to a bug).
duke@0 454 Compile* C = Compile::current();
duke@0 455 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@0 456 "must stay in the original alias category");
duke@0 457 // The type of the address must be contained in the adr_type,
duke@0 458 // disregarding "null"-ness.
duke@0 459 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@0 460 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
duke@0 461 assert(cross_check->meet(tp_notnull) == cross_check,
duke@0 462 "real address must not escape from expected memory type");
duke@0 463 }
duke@0 464 #endif
duke@0 465 return tp;
duke@0 466 }
duke@0 467 }
duke@0 468
duke@0 469 //------------------------adr_phi_is_loop_invariant----------------------------
duke@0 470 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@0 471 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@0 472 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@0 473 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@0 474 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@0 475 // with the same data. This implies that any path into the loop already
duke@0 476 // includes such a CastPP, and so the original cast, whatever its input,
duke@0 477 // must be covered by an equivalent cast, with an earlier control input.
duke@0 478 ResourceMark rm;
duke@0 479
duke@0 480 // The loop entry input of the phi should be the unique dominating
duke@0 481 // node for every Phi/CastPP in the loop.
duke@0 482 Unique_Node_List closure;
duke@0 483 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@0 484
duke@0 485 // Add the phi node and the cast to the worklist.
duke@0 486 Unique_Node_List worklist;
duke@0 487 worklist.push(adr_phi);
duke@0 488 if( cast != NULL ){
duke@0 489 if( !cast->is_ConstraintCast() ) return false;
duke@0 490 worklist.push(cast);
duke@0 491 }
duke@0 492
duke@0 493 // Begin recursive walk of phi nodes.
duke@0 494 while( worklist.size() ){
duke@0 495 // Take a node off the worklist
duke@0 496 Node *n = worklist.pop();
duke@0 497 if( !closure.member(n) ){
duke@0 498 // Add it to the closure.
duke@0 499 closure.push(n);
duke@0 500 // Make a sanity check to ensure we don't waste too much time here.
duke@0 501 if( closure.size() > 20) return false;
duke@0 502 // This node is OK if:
duke@0 503 // - it is a cast of an identical value
duke@0 504 // - or it is a phi node (then we add its inputs to the worklist)
duke@0 505 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@0 506 if( n->is_ConstraintCast() ){
duke@0 507 worklist.push(n->in(1));
duke@0 508 } else if( n->is_Phi() ) {
duke@0 509 for( uint i = 1; i < n->req(); i++ ) {
duke@0 510 worklist.push(n->in(i));
duke@0 511 }
duke@0 512 } else {
duke@0 513 return false;
duke@0 514 }
duke@0 515 }
duke@0 516 }
duke@0 517
duke@0 518 // Quit when the worklist is empty, and we've found no offending nodes.
duke@0 519 return true;
duke@0 520 }
duke@0 521
duke@0 522 //------------------------------Ideal_DU_postCCP-------------------------------
duke@0 523 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@0 524 // going away in this pass and we need to make this memory op depend on the
duke@0 525 // gating null check.
duke@0 526
duke@0 527 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@0 528 // some sense; we get to keep around the knowledge that an oop is not-null
duke@0 529 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@0 530 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@0 531 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@0 532 // some of the more trivial cases in the optimizer. Removing more useless
duke@0 533 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@0 534 // up on this approach. CNC 10/20/2000
duke@0 535 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
duke@0 536 Node *ctr = in(MemNode::Control);
duke@0 537 Node *mem = in(MemNode::Memory);
duke@0 538 Node *adr = in(MemNode::Address);
duke@0 539 Node *skipped_cast = NULL;
duke@0 540 // Need a null check? Regular static accesses do not because they are
duke@0 541 // from constant addresses. Array ops are gated by the range check (which
duke@0 542 // always includes a NULL check). Just check field ops.
duke@0 543 if( !ctr ) {
duke@0 544 // Scan upwards for the highest location we can place this memory op.
duke@0 545 while( true ) {
duke@0 546 switch( adr->Opcode() ) {
duke@0 547
duke@0 548 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@0 549 adr = adr->in(AddPNode::Base);
duke@0 550 continue;
duke@0 551
coleenp@113 552 case Op_DecodeN: // No change to NULL-ness, so peek thru
coleenp@113 553 adr = adr->in(1);
coleenp@113 554 continue;
coleenp@113 555
duke@0 556 case Op_CastPP:
duke@0 557 // If the CastPP is useless, just peek on through it.
duke@0 558 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@0 559 // Remember the cast that we've peeked though. If we peek
duke@0 560 // through more than one, then we end up remembering the highest
duke@0 561 // one, that is, if in a loop, the one closest to the top.
duke@0 562 skipped_cast = adr;
duke@0 563 adr = adr->in(1);
duke@0 564 continue;
duke@0 565 }
duke@0 566 // CastPP is going away in this pass! We need this memory op to be
duke@0 567 // control-dependent on the test that is guarding the CastPP.
duke@0 568 ccp->hash_delete(this);
duke@0 569 set_req(MemNode::Control, adr->in(0));
duke@0 570 ccp->hash_insert(this);
duke@0 571 return this;
duke@0 572
duke@0 573 case Op_Phi:
duke@0 574 // Attempt to float above a Phi to some dominating point.
duke@0 575 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@0 576 // If we've already peeked through a Cast (which could have set the
duke@0 577 // control), we can't float above a Phi, because the skipped Cast
duke@0 578 // may not be loop invariant.
duke@0 579 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@0 580 adr = adr->in(1);
duke@0 581 continue;
duke@0 582 }
duke@0 583 }
duke@0 584
duke@0 585 // Intentional fallthrough!
duke@0 586
duke@0 587 // No obvious dominating point. The mem op is pinned below the Phi
duke@0 588 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@0 589 // then the mem op can float, but not indefinitely. It must be pinned
duke@0 590 // behind the controls leading to the Phi.
duke@0 591 case Op_CheckCastPP:
duke@0 592 // These usually stick around to change address type, however a
duke@0 593 // useless one can be elided and we still need to pick up a control edge
duke@0 594 if (adr->in(0) == NULL) {
duke@0 595 // This CheckCastPP node has NO control and is likely useless. But we
duke@0 596 // need check further up the ancestor chain for a control input to keep
duke@0 597 // the node in place. 4959717.
duke@0 598 skipped_cast = adr;
duke@0 599 adr = adr->in(1);
duke@0 600 continue;
duke@0 601 }
duke@0 602 ccp->hash_delete(this);
duke@0 603 set_req(MemNode::Control, adr->in(0));
duke@0 604 ccp->hash_insert(this);
duke@0 605 return this;
duke@0 606
duke@0 607 // List of "safe" opcodes; those that implicitly block the memory
duke@0 608 // op below any null check.
duke@0 609 case Op_CastX2P: // no null checks on native pointers
duke@0 610 case Op_Parm: // 'this' pointer is not null
duke@0 611 case Op_LoadP: // Loading from within a klass
coleenp@113 612 case Op_LoadN: // Loading from within a klass
duke@0 613 case Op_LoadKlass: // Loading from within a klass
duke@0 614 case Op_ConP: // Loading from a klass
duke@0 615 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@0 616 case Op_Con: // Reading from TLS
duke@0 617 case Op_CMoveP: // CMoveP is pinned
duke@0 618 break; // No progress
duke@0 619
duke@0 620 case Op_Proj: // Direct call to an allocation routine
duke@0 621 case Op_SCMemProj: // Memory state from store conditional ops
duke@0 622 #ifdef ASSERT
duke@0 623 {
duke@0 624 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@0 625 const Node* call = adr->in(0);
duke@0 626 if (call->is_CallStaticJava()) {
duke@0 627 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
kvn@64 628 const TypeTuple *r = call_java->tf()->range();
kvn@64 629 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@64 630 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@64 631 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@0 632 // We further presume that this is one of
duke@0 633 // new_instance_Java, new_array_Java, or
duke@0 634 // the like, but do not assert for this.
duke@0 635 } else if (call->is_Allocate()) {
duke@0 636 // similar case to new_instance_Java, etc.
duke@0 637 } else if (!call->is_CallLeaf()) {
duke@0 638 // Projections from fetch_oop (OSR) are allowed as well.
duke@0 639 ShouldNotReachHere();
duke@0 640 }
duke@0 641 }
duke@0 642 #endif
duke@0 643 break;
duke@0 644 default:
duke@0 645 ShouldNotReachHere();
duke@0 646 }
duke@0 647 break;
duke@0 648 }
duke@0 649 }
duke@0 650
duke@0 651 return NULL; // No progress
duke@0 652 }
duke@0 653
duke@0 654
duke@0 655 //=============================================================================
duke@0 656 uint LoadNode::size_of() const { return sizeof(*this); }
duke@0 657 uint LoadNode::cmp( const Node &n ) const
duke@0 658 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@0 659 const Type *LoadNode::bottom_type() const { return _type; }
duke@0 660 uint LoadNode::ideal_reg() const {
duke@0 661 return Matcher::base2reg[_type->base()];
duke@0 662 }
duke@0 663
duke@0 664 #ifndef PRODUCT
duke@0 665 void LoadNode::dump_spec(outputStream *st) const {
duke@0 666 MemNode::dump_spec(st);
duke@0 667 if( !Verbose && !WizardMode ) {
duke@0 668 // standard dump does this in Verbose and WizardMode
duke@0 669 st->print(" #"); _type->dump_on(st);
duke@0 670 }
duke@0 671 }
duke@0 672 #endif
duke@0 673
duke@0 674
duke@0 675 //----------------------------LoadNode::make-----------------------------------
duke@0 676 // Polymorphic factory method:
coleenp@113 677 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
coleenp@113 678 Compile* C = gvn.C;
coleenp@113 679
duke@0 680 // sanity check the alias category against the created node type
duke@0 681 assert(!(adr_type->isa_oopptr() &&
duke@0 682 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@0 683 "use LoadKlassNode instead");
duke@0 684 assert(!(adr_type->isa_aryptr() &&
duke@0 685 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@0 686 "use LoadRangeNode instead");
duke@0 687 switch (bt) {
duke@0 688 case T_BOOLEAN:
duke@0 689 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 690 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 691 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 692 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 693 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
duke@0 694 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
duke@0 695 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
duke@0 696 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
coleenp@113 697 case T_OBJECT:
coleenp@113 698 #ifdef _LP64
coleenp@113 699 if (adr->bottom_type()->is_narrow()) {
coleenp@113 700 const TypeNarrowOop* narrowtype;
coleenp@113 701 if (rt->isa_narrowoop()) {
coleenp@113 702 narrowtype = rt->is_narrowoop();
coleenp@113 703 rt = narrowtype->make_oopptr();
coleenp@113 704 } else {
coleenp@113 705 narrowtype = rt->is_oopptr()->make_narrowoop();
coleenp@113 706 }
coleenp@113 707 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, narrowtype));
coleenp@113 708
coleenp@113 709 return new (C, 2) DecodeNNode(load, rt);
coleenp@113 710 } else
coleenp@113 711 #endif
coleenp@113 712 {
coleenp@113 713 assert(!adr->bottom_type()->is_narrow(), "should have got back a narrow oop");
coleenp@113 714 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
coleenp@113 715 }
duke@0 716 }
duke@0 717 ShouldNotReachHere();
duke@0 718 return (LoadNode*)NULL;
duke@0 719 }
duke@0 720
duke@0 721 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
duke@0 722 bool require_atomic = true;
duke@0 723 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
duke@0 724 }
duke@0 725
duke@0 726
duke@0 727
duke@0 728
duke@0 729 //------------------------------hash-------------------------------------------
duke@0 730 uint LoadNode::hash() const {
duke@0 731 // unroll addition of interesting fields
duke@0 732 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@0 733 }
duke@0 734
duke@0 735 //---------------------------can_see_stored_value------------------------------
duke@0 736 // This routine exists to make sure this set of tests is done the same
duke@0 737 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@0 738 // will change the graph shape in a way which makes memory alive twice at the
duke@0 739 // same time (uses the Oracle model of aliasing), then some
duke@0 740 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@0 741 // of aliasing.
duke@0 742 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@0 743 Node* ld_adr = in(MemNode::Address);
duke@0 744
never@17 745 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
never@17 746 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
never@17 747 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
never@17 748 atp->field() != NULL && !atp->field()->is_volatile()) {
never@17 749 uint alias_idx = atp->index();
never@17 750 bool final = atp->field()->is_final();
never@17 751 Node* result = NULL;
never@17 752 Node* current = st;
never@17 753 // Skip through chains of MemBarNodes checking the MergeMems for
never@17 754 // new states for the slice of this load. Stop once any other
never@17 755 // kind of node is encountered. Loads from final memory can skip
never@17 756 // through any kind of MemBar but normal loads shouldn't skip
never@17 757 // through MemBarAcquire since the could allow them to move out of
never@17 758 // a synchronized region.
never@17 759 while (current->is_Proj()) {
never@17 760 int opc = current->in(0)->Opcode();
never@17 761 if ((final && opc == Op_MemBarAcquire) ||
never@17 762 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
never@17 763 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@17 764 if (mem->is_MergeMem()) {
never@17 765 MergeMemNode* merge = mem->as_MergeMem();
never@17 766 Node* new_st = merge->memory_at(alias_idx);
never@17 767 if (new_st == merge->base_memory()) {
never@17 768 // Keep searching
never@17 769 current = merge->base_memory();
never@17 770 continue;
never@17 771 }
never@17 772 // Save the new memory state for the slice and fall through
never@17 773 // to exit.
never@17 774 result = new_st;
never@17 775 }
never@17 776 }
never@17 777 break;
never@17 778 }
never@17 779 if (result != NULL) {
never@17 780 st = result;
never@17 781 }
never@17 782 }
never@17 783
never@17 784
duke@0 785 // Loop around twice in the case Load -> Initialize -> Store.
duke@0 786 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@0 787 for (int trip = 0; trip <= 1; trip++) {
duke@0 788
duke@0 789 if (st->is_Store()) {
duke@0 790 Node* st_adr = st->in(MemNode::Address);
duke@0 791 if (!phase->eqv(st_adr, ld_adr)) {
duke@0 792 // Try harder before giving up... Match raw and non-raw pointers.
duke@0 793 intptr_t st_off = 0;
duke@0 794 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@0 795 if (alloc == NULL) return NULL;
duke@0 796 intptr_t ld_off = 0;
duke@0 797 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
duke@0 798 if (alloc != allo2) return NULL;
duke@0 799 if (ld_off != st_off) return NULL;
duke@0 800 // At this point we have proven something like this setup:
duke@0 801 // A = Allocate(...)
duke@0 802 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@0 803 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@0 804 // (Actually, we haven't yet proven the Q's are the same.)
duke@0 805 // In other words, we are loading from a casted version of
duke@0 806 // the same pointer-and-offset that we stored to.
duke@0 807 // Thus, we are able to replace L by V.
duke@0 808 }
duke@0 809 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@0 810 if (store_Opcode() != st->Opcode())
duke@0 811 return NULL;
duke@0 812 return st->in(MemNode::ValueIn);
duke@0 813 }
duke@0 814
duke@0 815 intptr_t offset = 0; // scratch
duke@0 816
duke@0 817 // A load from a freshly-created object always returns zero.
duke@0 818 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@0 819 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@0 820 if (st->is_Proj() && st->in(0)->is_Allocate() &&
duke@0 821 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
duke@0 822 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
duke@0 823 // return a zero value for the load's basic type
duke@0 824 // (This is one of the few places where a generic PhaseTransform
duke@0 825 // can create new nodes. Think of it as lazily manifesting
duke@0 826 // virtually pre-existing constants.)
duke@0 827 return phase->zerocon(memory_type());
duke@0 828 }
duke@0 829
duke@0 830 // A load from an initialization barrier can match a captured store.
duke@0 831 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@0 832 InitializeNode* init = st->in(0)->as_Initialize();
duke@0 833 AllocateNode* alloc = init->allocation();
duke@0 834 if (alloc != NULL &&
duke@0 835 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
duke@0 836 // examine a captured store value
duke@0 837 st = init->find_captured_store(offset, memory_size(), phase);
duke@0 838 if (st != NULL)
duke@0 839 continue; // take one more trip around
duke@0 840 }
duke@0 841 }
duke@0 842
duke@0 843 break;
duke@0 844 }
duke@0 845
duke@0 846 return NULL;
duke@0 847 }
duke@0 848
kvn@64 849 //----------------------is_instance_field_load_with_local_phi------------------
kvn@64 850 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@64 851 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
kvn@64 852 in(MemNode::Address)->is_AddP() ) {
kvn@64 853 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@64 854 // Only instances.
kvn@64 855 if( t_oop != NULL && t_oop->is_instance_field() &&
kvn@64 856 t_oop->offset() != Type::OffsetBot &&
kvn@64 857 t_oop->offset() != Type::OffsetTop) {
kvn@64 858 return true;
kvn@64 859 }
kvn@64 860 }
kvn@64 861 return false;
kvn@64 862 }
kvn@64 863
duke@0 864 //------------------------------Identity---------------------------------------
duke@0 865 // Loads are identity if previous store is to same address
duke@0 866 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@0 867 // If the previous store-maker is the right kind of Store, and the store is
duke@0 868 // to the same address, then we are equal to the value stored.
duke@0 869 Node* mem = in(MemNode::Memory);
duke@0 870 Node* value = can_see_stored_value(mem, phase);
duke@0 871 if( value ) {
duke@0 872 // byte, short & char stores truncate naturally.
duke@0 873 // A load has to load the truncated value which requires
duke@0 874 // some sort of masking operation and that requires an
duke@0 875 // Ideal call instead of an Identity call.
duke@0 876 if (memory_size() < BytesPerInt) {
duke@0 877 // If the input to the store does not fit with the load's result type,
duke@0 878 // it must be truncated via an Ideal call.
duke@0 879 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@0 880 return this;
duke@0 881 }
duke@0 882 // (This works even when value is a Con, but LoadNode::Value
duke@0 883 // usually runs first, producing the singleton type of the Con.)
duke@0 884 return value;
duke@0 885 }
kvn@64 886
kvn@64 887 // Search for an existing data phi which was generated before for the same
kvn@64 888 // instance's field to avoid infinite genertion of phis in a loop.
kvn@64 889 Node *region = mem->in(0);
kvn@64 890 if (is_instance_field_load_with_local_phi(region)) {
kvn@64 891 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
kvn@64 892 int this_index = phase->C->get_alias_index(addr_t);
kvn@64 893 int this_offset = addr_t->offset();
kvn@64 894 int this_id = addr_t->is_oopptr()->instance_id();
kvn@64 895 const Type* this_type = bottom_type();
kvn@64 896 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@64 897 Node* phi = region->fast_out(i);
kvn@64 898 if (phi->is_Phi() && phi != mem &&
kvn@64 899 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
kvn@64 900 return phi;
kvn@64 901 }
kvn@64 902 }
kvn@64 903 }
kvn@64 904
duke@0 905 return this;
duke@0 906 }
duke@0 907
never@17 908
never@17 909 // Returns true if the AliasType refers to the field that holds the
never@17 910 // cached box array. Currently only handles the IntegerCache case.
never@17 911 static bool is_autobox_cache(Compile::AliasType* atp) {
never@17 912 if (atp != NULL && atp->field() != NULL) {
never@17 913 ciField* field = atp->field();
never@17 914 ciSymbol* klass = field->holder()->name();
never@17 915 if (field->name() == ciSymbol::cache_field_name() &&
never@17 916 field->holder()->uses_default_loader() &&
never@17 917 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 918 return true;
never@17 919 }
never@17 920 }
never@17 921 return false;
never@17 922 }
never@17 923
never@17 924 // Fetch the base value in the autobox array
never@17 925 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
never@17 926 if (atp != NULL && atp->field() != NULL) {
never@17 927 ciField* field = atp->field();
never@17 928 ciSymbol* klass = field->holder()->name();
never@17 929 if (field->name() == ciSymbol::cache_field_name() &&
never@17 930 field->holder()->uses_default_loader() &&
never@17 931 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 932 assert(field->is_constant(), "what?");
never@17 933 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
never@17 934 // Fetch the box object at the base of the array and get its value
never@17 935 ciInstance* box = array->obj_at(0)->as_instance();
never@17 936 ciInstanceKlass* ik = box->klass()->as_instance_klass();
never@17 937 if (ik->nof_nonstatic_fields() == 1) {
never@17 938 // This should be true nonstatic_field_at requires calling
never@17 939 // nof_nonstatic_fields so check it anyway
never@17 940 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
never@17 941 cache_offset = c.as_int();
never@17 942 }
never@17 943 return true;
never@17 944 }
never@17 945 }
never@17 946 return false;
never@17 947 }
never@17 948
never@17 949 // Returns true if the AliasType refers to the value field of an
never@17 950 // autobox object. Currently only handles Integer.
never@17 951 static bool is_autobox_object(Compile::AliasType* atp) {
never@17 952 if (atp != NULL && atp->field() != NULL) {
never@17 953 ciField* field = atp->field();
never@17 954 ciSymbol* klass = field->holder()->name();
never@17 955 if (field->name() == ciSymbol::value_name() &&
never@17 956 field->holder()->uses_default_loader() &&
never@17 957 klass == ciSymbol::java_lang_Integer()) {
never@17 958 return true;
never@17 959 }
never@17 960 }
never@17 961 return false;
never@17 962 }
never@17 963
never@17 964
never@17 965 // We're loading from an object which has autobox behaviour.
never@17 966 // If this object is result of a valueOf call we'll have a phi
never@17 967 // merging a newly allocated object and a load from the cache.
never@17 968 // We want to replace this load with the original incoming
never@17 969 // argument to the valueOf call.
never@17 970 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
never@17 971 Node* base = in(Address)->in(AddPNode::Base);
never@17 972 if (base->is_Phi() && base->req() == 3) {
never@17 973 AllocateNode* allocation = NULL;
never@17 974 int allocation_index = -1;
never@17 975 int load_index = -1;
never@17 976 for (uint i = 1; i < base->req(); i++) {
never@17 977 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
never@17 978 if (allocation != NULL) {
never@17 979 allocation_index = i;
never@17 980 load_index = 3 - allocation_index;
never@17 981 break;
never@17 982 }
never@17 983 }
never@17 984 LoadNode* load = NULL;
never@17 985 if (allocation != NULL && base->in(load_index)->is_Load()) {
never@17 986 load = base->in(load_index)->as_Load();
never@17 987 }
never@17 988 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
never@17 989 // Push the loads from the phi that comes from valueOf up
never@17 990 // through it to allow elimination of the loads and the recovery
never@17 991 // of the original value.
never@17 992 Node* mem_phi = in(Memory);
never@17 993 Node* offset = in(Address)->in(AddPNode::Offset);
never@17 994
never@17 995 Node* in1 = clone();
never@17 996 Node* in1_addr = in1->in(Address)->clone();
never@17 997 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
never@17 998 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
never@17 999 in1_addr->set_req(AddPNode::Offset, offset);
never@17 1000 in1->set_req(0, base->in(allocation_index));
never@17 1001 in1->set_req(Address, in1_addr);
never@17 1002 in1->set_req(Memory, mem_phi->in(allocation_index));
never@17 1003
never@17 1004 Node* in2 = clone();
never@17 1005 Node* in2_addr = in2->in(Address)->clone();
never@17 1006 in2_addr->set_req(AddPNode::Base, base->in(load_index));
never@17 1007 in2_addr->set_req(AddPNode::Address, base->in(load_index));
never@17 1008 in2_addr->set_req(AddPNode::Offset, offset);
never@17 1009 in2->set_req(0, base->in(load_index));
never@17 1010 in2->set_req(Address, in2_addr);
never@17 1011 in2->set_req(Memory, mem_phi->in(load_index));
never@17 1012
never@17 1013 in1_addr = phase->transform(in1_addr);
never@17 1014 in1 = phase->transform(in1);
never@17 1015 in2_addr = phase->transform(in2_addr);
never@17 1016 in2 = phase->transform(in2);
never@17 1017
never@17 1018 PhiNode* result = PhiNode::make_blank(base->in(0), this);
never@17 1019 result->set_req(allocation_index, in1);
never@17 1020 result->set_req(load_index, in2);
never@17 1021 return result;
never@17 1022 }
never@17 1023 } else if (base->is_Load()) {
never@17 1024 // Eliminate the load of Integer.value for integers from the cache
never@17 1025 // array by deriving the value from the index into the array.
never@17 1026 // Capture the offset of the load and then reverse the computation.
never@17 1027 Node* load_base = base->in(Address)->in(AddPNode::Base);
never@17 1028 if (load_base != NULL) {
never@17 1029 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
never@17 1030 intptr_t cache_offset;
never@17 1031 int shift = -1;
never@17 1032 Node* cache = NULL;
never@17 1033 if (is_autobox_cache(atp)) {
kvn@29 1034 shift = exact_log2(type2aelembytes(T_OBJECT));
never@17 1035 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
never@17 1036 }
never@17 1037 if (cache != NULL && base->in(Address)->is_AddP()) {
never@17 1038 Node* elements[4];
never@17 1039 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@17 1040 int cache_low;
never@17 1041 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
never@17 1042 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
never@17 1043 // Add up all the offsets making of the address of the load
never@17 1044 Node* result = elements[0];
never@17 1045 for (int i = 1; i < count; i++) {
never@17 1046 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
never@17 1047 }
never@17 1048 // Remove the constant offset from the address and then
never@17 1049 // remove the scaling of the offset to recover the original index.
never@17 1050 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
never@17 1051 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
never@17 1052 // Peel the shift off directly but wrap it in a dummy node
never@17 1053 // since Ideal can't return existing nodes
never@17 1054 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
never@17 1055 } else {
never@17 1056 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
never@17 1057 }
never@17 1058 #ifdef _LP64
never@17 1059 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
never@17 1060 #endif
never@17 1061 return result;
never@17 1062 }
never@17 1063 }
never@17 1064 }
never@17 1065 }
never@17 1066 return NULL;
never@17 1067 }
never@17 1068
never@17 1069
duke@0 1070 //------------------------------Ideal------------------------------------------
duke@0 1071 // If the load is from Field memory and the pointer is non-null, we can
duke@0 1072 // zero out the control input.
duke@0 1073 // If the offset is constant and the base is an object allocation,
duke@0 1074 // try to hook me up to the exact initializing store.
duke@0 1075 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1076 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1077 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1078
duke@0 1079 Node* ctrl = in(MemNode::Control);
duke@0 1080 Node* address = in(MemNode::Address);
duke@0 1081
duke@0 1082 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@0 1083 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@0 1084 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@0 1085 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@0 1086 ctrl = ctrl->in(0);
duke@0 1087 set_req(MemNode::Control,ctrl);
duke@0 1088 }
duke@0 1089
duke@0 1090 // Check for useless control edge in some common special cases
duke@0 1091 if (in(MemNode::Control) != NULL) {
duke@0 1092 intptr_t ignore = 0;
duke@0 1093 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
duke@0 1094 if (base != NULL
duke@0 1095 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
duke@0 1096 && detect_dominating_control(base->in(0), phase->C->start())) {
duke@0 1097 // A method-invariant, non-null address (constant or 'this' argument).
duke@0 1098 set_req(MemNode::Control, NULL);
duke@0 1099 }
duke@0 1100 }
duke@0 1101
never@17 1102 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
never@17 1103 Node* base = in(Address)->in(AddPNode::Base);
never@17 1104 if (base != NULL) {
never@17 1105 Compile::AliasType* atp = phase->C->alias_type(adr_type());
never@17 1106 if (is_autobox_object(atp)) {
never@17 1107 Node* result = eliminate_autobox(phase);
never@17 1108 if (result != NULL) return result;
never@17 1109 }
never@17 1110 }
never@17 1111 }
never@17 1112
kvn@74 1113 Node* mem = in(MemNode::Memory);
kvn@74 1114 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@74 1115
kvn@74 1116 if (addr_t != NULL) {
kvn@74 1117 // try to optimize our memory input
kvn@74 1118 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
kvn@74 1119 if (opt_mem != mem) {
kvn@74 1120 set_req(MemNode::Memory, opt_mem);
kvn@74 1121 return this;
kvn@74 1122 }
kvn@74 1123 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@74 1124 if (can_reshape && opt_mem->is_Phi() &&
kvn@74 1125 (t_oop != NULL) && t_oop->is_instance_field()) {
kvn@74 1126 assert(t_oop->offset() != Type::OffsetBot && t_oop->offset() != Type::OffsetTop, "");
kvn@74 1127 Node *region = opt_mem->in(0);
kvn@74 1128 uint cnt = opt_mem->req();
kvn@74 1129 for( uint i = 1; i < cnt; i++ ) {
kvn@74 1130 Node *in = opt_mem->in(i);
kvn@74 1131 if( in == NULL ) {
kvn@74 1132 region = NULL; // Wait stable graph
kvn@74 1133 break;
kvn@74 1134 }
kvn@74 1135 }
kvn@74 1136 if (region != NULL) {
kvn@74 1137 // Check for loop invariant.
kvn@74 1138 if (cnt == 3) {
kvn@74 1139 for( uint i = 1; i < cnt; i++ ) {
kvn@74 1140 Node *in = opt_mem->in(i);
kvn@74 1141 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
kvn@74 1142 if (m == opt_mem) {
kvn@74 1143 set_req(MemNode::Memory, opt_mem->in(cnt - i)); // Skip this phi.
kvn@74 1144 return this;
kvn@74 1145 }
kvn@74 1146 }
kvn@74 1147 }
kvn@74 1148 // Split through Phi (see original code in loopopts.cpp).
kvn@74 1149 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
kvn@82 1150
kvn@82 1151 // Do nothing here if Identity will find a value
kvn@82 1152 // (to avoid infinite chain of value phis generation).
kvn@82 1153 if ( !phase->eqv(this, this->Identity(phase)) )
kvn@82 1154 return NULL;
kvn@82 1155
kvn@74 1156 const Type* this_type = this->bottom_type();
kvn@74 1157 int this_index = phase->C->get_alias_index(addr_t);
kvn@74 1158 int this_offset = addr_t->offset();
kvn@74 1159 int this_iid = addr_t->is_oopptr()->instance_id();
kvn@74 1160 int wins = 0;
kvn@74 1161 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@74 1162 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
kvn@74 1163 for( uint i = 1; i < region->req(); i++ ) {
kvn@74 1164 Node *x;
kvn@74 1165 Node* the_clone = NULL;
kvn@74 1166 if( region->in(i) == phase->C->top() ) {
kvn@74 1167 x = phase->C->top(); // Dead path? Use a dead data op
kvn@74 1168 } else {
kvn@74 1169 x = this->clone(); // Else clone up the data op
kvn@74 1170 the_clone = x; // Remember for possible deletion.
kvn@74 1171 // Alter data node to use pre-phi inputs
kvn@74 1172 if( this->in(0) == region ) {
kvn@74 1173 x->set_req( 0, region->in(i) );
kvn@74 1174 } else {
kvn@74 1175 x->set_req( 0, NULL );
kvn@74 1176 }
kvn@74 1177 for( uint j = 1; j < this->req(); j++ ) {
kvn@74 1178 Node *in = this->in(j);
kvn@74 1179 if( in->is_Phi() && in->in(0) == region )
kvn@74 1180 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
kvn@74 1181 }
kvn@74 1182 }
kvn@74 1183 // Check for a 'win' on some paths
kvn@74 1184 const Type *t = x->Value(igvn);
kvn@74 1185
kvn@74 1186 bool singleton = t->singleton();
kvn@74 1187
kvn@74 1188 // See comments in PhaseIdealLoop::split_thru_phi().
kvn@74 1189 if( singleton && t == Type::TOP ) {
kvn@74 1190 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
kvn@74 1191 }
kvn@74 1192
kvn@74 1193 if( singleton ) {
kvn@74 1194 wins++;
kvn@74 1195 x = igvn->makecon(t);
kvn@74 1196 } else {
kvn@74 1197 // We now call Identity to try to simplify the cloned node.
kvn@74 1198 // Note that some Identity methods call phase->type(this).
kvn@74 1199 // Make sure that the type array is big enough for
kvn@74 1200 // our new node, even though we may throw the node away.
kvn@74 1201 // (This tweaking with igvn only works because x is a new node.)
kvn@74 1202 igvn->set_type(x, t);
kvn@74 1203 Node *y = x->Identity(igvn);
kvn@74 1204 if( y != x ) {
kvn@74 1205 wins++;
kvn@74 1206 x = y;
kvn@74 1207 } else {
kvn@74 1208 y = igvn->hash_find(x);
kvn@74 1209 if( y ) {
kvn@74 1210 wins++;
kvn@74 1211 x = y;
kvn@74 1212 } else {
kvn@74 1213 // Else x is a new node we are keeping
kvn@74 1214 // We do not need register_new_node_with_optimizer
kvn@74 1215 // because set_type has already been called.
kvn@74 1216 igvn->_worklist.push(x);
kvn@74 1217 }
kvn@74 1218 }
kvn@74 1219 }
kvn@74 1220 if (x != the_clone && the_clone != NULL)
kvn@74 1221 igvn->remove_dead_node(the_clone);
kvn@74 1222 phi->set_req(i, x);
kvn@74 1223 }
kvn@74 1224 if( wins > 0 ) {
kvn@74 1225 // Record Phi
kvn@74 1226 igvn->register_new_node_with_optimizer(phi);
kvn@74 1227 return phi;
kvn@74 1228 } else {
kvn@74 1229 igvn->remove_dead_node(phi);
kvn@74 1230 }
kvn@74 1231 }
kvn@74 1232 }
kvn@74 1233 }
kvn@74 1234
duke@0 1235 // Check for prior store with a different base or offset; make Load
duke@0 1236 // independent. Skip through any number of them. Bail out if the stores
duke@0 1237 // are in an endless dead cycle and report no progress. This is a key
duke@0 1238 // transform for Reflection. However, if after skipping through the Stores
duke@0 1239 // we can't then fold up against a prior store do NOT do the transform as
duke@0 1240 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@0 1241 // array memory alive twice: once for the hoisted Load and again after the
duke@0 1242 // bypassed Store. This situation only works if EVERYBODY who does
duke@0 1243 // anti-dependence work knows how to bypass. I.e. we need all
duke@0 1244 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@0 1245 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@0 1246 // fold up, do so.
duke@0 1247 Node* prev_mem = find_previous_store(phase);
duke@0 1248 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1249 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@0 1250 // (c) See if we can fold up on the spot, but don't fold up here.
duke@0 1251 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
duke@0 1252 // just return a prior value, which is done by Identity calls.
duke@0 1253 if (can_see_stored_value(prev_mem, phase)) {
duke@0 1254 // Make ready for step (d):
duke@0 1255 set_req(MemNode::Memory, prev_mem);
duke@0 1256 return this;
duke@0 1257 }
duke@0 1258 }
duke@0 1259
duke@0 1260 return NULL; // No further progress
duke@0 1261 }
duke@0 1262
duke@0 1263 // Helper to recognize certain Klass fields which are invariant across
duke@0 1264 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@0 1265 const Type*
duke@0 1266 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@0 1267 ciKlass* klass) const {
duke@0 1268 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1269 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@0 1270 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@0 1271 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@0 1272 return TypeInt::make(klass->modifier_flags());
duke@0 1273 }
duke@0 1274 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1275 // The field is Klass::_access_flags. Return its (constant) value.
duke@0 1276 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@0 1277 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@0 1278 return TypeInt::make(klass->access_flags());
duke@0 1279 }
duke@0 1280 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1281 // The field is Klass::_layout_helper. Return its constant value if known.
duke@0 1282 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1283 return TypeInt::make(klass->layout_helper());
duke@0 1284 }
duke@0 1285
duke@0 1286 // No match.
duke@0 1287 return NULL;
duke@0 1288 }
duke@0 1289
duke@0 1290 //------------------------------Value-----------------------------------------
duke@0 1291 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@0 1292 // Either input is TOP ==> the result is TOP
duke@0 1293 Node* mem = in(MemNode::Memory);
duke@0 1294 const Type *t1 = phase->type(mem);
duke@0 1295 if (t1 == Type::TOP) return Type::TOP;
duke@0 1296 Node* adr = in(MemNode::Address);
duke@0 1297 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@0 1298 if (tp == NULL || tp->empty()) return Type::TOP;
duke@0 1299 int off = tp->offset();
duke@0 1300 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
duke@0 1301
duke@0 1302 // Try to guess loaded type from pointer type
duke@0 1303 if (tp->base() == Type::AryPtr) {
duke@0 1304 const Type *t = tp->is_aryptr()->elem();
duke@0 1305 // Don't do this for integer types. There is only potential profit if
duke@0 1306 // the element type t is lower than _type; that is, for int types, if _type is
duke@0 1307 // more restrictive than t. This only happens here if one is short and the other
duke@0 1308 // char (both 16 bits), and in those cases we've made an intentional decision
duke@0 1309 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@0 1310 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@0 1311 //
duke@0 1312 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@0 1313 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@0 1314 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@0 1315 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@0 1316 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@0 1317 // In fact, that could have been the original type of p1, and p1 could have
duke@0 1318 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@0 1319 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@0 1320 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
duke@0 1321 && Opcode() != Op_LoadKlass) {
duke@0 1322 // t might actually be lower than _type, if _type is a unique
duke@0 1323 // concrete subclass of abstract class t.
duke@0 1324 // Make sure the reference is not into the header, by comparing
duke@0 1325 // the offset against the offset of the start of the array's data.
duke@0 1326 // Different array types begin at slightly different offsets (12 vs. 16).
duke@0 1327 // We choose T_BYTE as an example base type that is least restrictive
duke@0 1328 // as to alignment, which will therefore produce the smallest
duke@0 1329 // possible base offset.
duke@0 1330 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1331 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
duke@0 1332 const Type* jt = t->join(_type);
duke@0 1333 // In any case, do not allow the join, per se, to empty out the type.
duke@0 1334 if (jt->empty() && !t->empty()) {
duke@0 1335 // This can happen if a interface-typed array narrows to a class type.
duke@0 1336 jt = _type;
duke@0 1337 }
never@17 1338
never@17 1339 if (EliminateAutoBox) {
never@17 1340 // The pointers in the autobox arrays are always non-null
never@17 1341 Node* base = in(Address)->in(AddPNode::Base);
never@17 1342 if (base != NULL) {
never@17 1343 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
never@17 1344 if (is_autobox_cache(atp)) {
never@17 1345 return jt->join(TypePtr::NOTNULL)->is_ptr();
never@17 1346 }
never@17 1347 }
never@17 1348 }
duke@0 1349 return jt;
duke@0 1350 }
duke@0 1351 }
duke@0 1352 } else if (tp->base() == Type::InstPtr) {
duke@0 1353 assert( off != Type::OffsetBot ||
duke@0 1354 // arrays can be cast to Objects
duke@0 1355 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@0 1356 // unsafe field access may not have a constant offset
duke@0 1357 phase->C->has_unsafe_access(),
duke@0 1358 "Field accesses must be precise" );
duke@0 1359 // For oop loads, we expect the _type to be precise
duke@0 1360 } else if (tp->base() == Type::KlassPtr) {
duke@0 1361 assert( off != Type::OffsetBot ||
duke@0 1362 // arrays can be cast to Objects
duke@0 1363 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@0 1364 // also allow array-loading from the primary supertype
duke@0 1365 // array during subtype checks
duke@0 1366 Opcode() == Op_LoadKlass,
duke@0 1367 "Field accesses must be precise" );
duke@0 1368 // For klass/static loads, we expect the _type to be precise
duke@0 1369 }
duke@0 1370
duke@0 1371 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1372 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1373 ciKlass* klass = tkls->klass();
duke@0 1374 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@0 1375 // We are loading a field from a Klass metaobject whose identity
duke@0 1376 // is known at compile time (the type is "exact" or "precise").
duke@0 1377 // Check for fields we know are maintained as constants by the VM.
duke@0 1378 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1379 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@0 1380 // (Folds up type checking code.)
duke@0 1381 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@0 1382 return TypeInt::make(klass->super_check_offset());
duke@0 1383 }
duke@0 1384 // Compute index into primary_supers array
duke@0 1385 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1386 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1387 if( depth < ciKlass::primary_super_limit() ) {
duke@0 1388 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1389 // (Folds up type checking code.)
duke@0 1390 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1391 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1392 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1393 }
duke@0 1394 const Type* aift = load_array_final_field(tkls, klass);
duke@0 1395 if (aift != NULL) return aift;
duke@0 1396 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
duke@0 1397 && klass->is_array_klass()) {
duke@0 1398 // The field is arrayKlass::_component_mirror. Return its (constant) value.
duke@0 1399 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@0 1400 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@0 1401 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@0 1402 }
duke@0 1403 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1404 // The field is Klass::_java_mirror. Return its (constant) value.
duke@0 1405 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@0 1406 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@0 1407 return TypeInstPtr::make(klass->java_mirror());
duke@0 1408 }
duke@0 1409 }
duke@0 1410
duke@0 1411 // We can still check if we are loading from the primary_supers array at a
duke@0 1412 // shallow enough depth. Even though the klass is not exact, entries less
duke@0 1413 // than or equal to its super depth are correct.
duke@0 1414 if (klass->is_loaded() ) {
duke@0 1415 ciType *inner = klass->klass();
duke@0 1416 while( inner->is_obj_array_klass() )
duke@0 1417 inner = inner->as_obj_array_klass()->base_element_type();
duke@0 1418 if( inner->is_instance_klass() &&
duke@0 1419 !inner->as_instance_klass()->flags().is_interface() ) {
duke@0 1420 // Compute index into primary_supers array
duke@0 1421 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1422 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1423 if( depth < ciKlass::primary_super_limit() &&
duke@0 1424 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@0 1425 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1426 // (Folds up type checking code.)
duke@0 1427 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1428 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1429 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1430 }
duke@0 1431 }
duke@0 1432 }
duke@0 1433
duke@0 1434 // If the type is enough to determine that the thing is not an array,
duke@0 1435 // we can give the layout_helper a positive interval type.
duke@0 1436 // This will help short-circuit some reflective code.
duke@0 1437 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
duke@0 1438 && !klass->is_array_klass() // not directly typed as an array
duke@0 1439 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@0 1440 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@0 1441 ) {
duke@0 1442 // Note: When interfaces are reliable, we can narrow the interface
duke@0 1443 // test to (klass != Serializable && klass != Cloneable).
duke@0 1444 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1445 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@0 1446 // The key property of this type is that it folds up tests
duke@0 1447 // for array-ness, since it proves that the layout_helper is positive.
duke@0 1448 // Thus, a generic value like the basic object layout helper works fine.
duke@0 1449 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@0 1450 }
duke@0 1451 }
duke@0 1452
duke@0 1453 // If we are loading from a freshly-allocated object, produce a zero,
duke@0 1454 // if the load is provably beyond the header of the object.
duke@0 1455 // (Also allow a variable load from a fresh array to produce zero.)
duke@0 1456 if (ReduceFieldZeroing) {
duke@0 1457 Node* value = can_see_stored_value(mem,phase);
duke@0 1458 if (value != NULL && value->is_Con())
duke@0 1459 return value->bottom_type();
duke@0 1460 }
duke@0 1461
kvn@64 1462 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 1463 if (tinst != NULL && tinst->is_instance_field()) {
kvn@64 1464 // If we have an instance type and our memory input is the
kvn@64 1465 // programs's initial memory state, there is no matching store,
kvn@64 1466 // so just return a zero of the appropriate type
kvn@64 1467 Node *mem = in(MemNode::Memory);
kvn@64 1468 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@64 1469 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@64 1470 return Type::get_zero_type(_type->basic_type());
kvn@64 1471 }
kvn@64 1472 }
duke@0 1473 return _type;
duke@0 1474 }
duke@0 1475
duke@0 1476 //------------------------------match_edge-------------------------------------
duke@0 1477 // Do we Match on this edge index or not? Match only the address.
duke@0 1478 uint LoadNode::match_edge(uint idx) const {
duke@0 1479 return idx == MemNode::Address;
duke@0 1480 }
duke@0 1481
duke@0 1482 //--------------------------LoadBNode::Ideal--------------------------------------
duke@0 1483 //
duke@0 1484 // If the previous store is to the same address as this load,
duke@0 1485 // and the value stored was larger than a byte, replace this load
duke@0 1486 // with the value stored truncated to a byte. If no truncation is
duke@0 1487 // needed, the replacement is done in LoadNode::Identity().
duke@0 1488 //
duke@0 1489 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1490 Node* mem = in(MemNode::Memory);
duke@0 1491 Node* value = can_see_stored_value(mem,phase);
duke@0 1492 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1493 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
duke@0 1494 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
duke@0 1495 }
duke@0 1496 // Identity call will handle the case where truncation is not needed.
duke@0 1497 return LoadNode::Ideal(phase, can_reshape);
duke@0 1498 }
duke@0 1499
duke@0 1500 //--------------------------LoadCNode::Ideal--------------------------------------
duke@0 1501 //
duke@0 1502 // If the previous store is to the same address as this load,
duke@0 1503 // and the value stored was larger than a char, replace this load
duke@0 1504 // with the value stored truncated to a char. If no truncation is
duke@0 1505 // needed, the replacement is done in LoadNode::Identity().
duke@0 1506 //
duke@0 1507 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1508 Node* mem = in(MemNode::Memory);
duke@0 1509 Node* value = can_see_stored_value(mem,phase);
duke@0 1510 if( value && !phase->type(value)->higher_equal( _type ) )
duke@0 1511 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
duke@0 1512 // Identity call will handle the case where truncation is not needed.
duke@0 1513 return LoadNode::Ideal(phase, can_reshape);
duke@0 1514 }
duke@0 1515
duke@0 1516 //--------------------------LoadSNode::Ideal--------------------------------------
duke@0 1517 //
duke@0 1518 // If the previous store is to the same address as this load,
duke@0 1519 // and the value stored was larger than a short, replace this load
duke@0 1520 // with the value stored truncated to a short. If no truncation is
duke@0 1521 // needed, the replacement is done in LoadNode::Identity().
duke@0 1522 //
duke@0 1523 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1524 Node* mem = in(MemNode::Memory);
duke@0 1525 Node* value = can_see_stored_value(mem,phase);
duke@0 1526 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1527 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
duke@0 1528 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
duke@0 1529 }
duke@0 1530 // Identity call will handle the case where truncation is not needed.
duke@0 1531 return LoadNode::Ideal(phase, can_reshape);
duke@0 1532 }
duke@0 1533
duke@0 1534 //=============================================================================
duke@0 1535 //------------------------------Value------------------------------------------
duke@0 1536 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
duke@0 1537 // Either input is TOP ==> the result is TOP
duke@0 1538 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1539 if (t1 == Type::TOP) return Type::TOP;
duke@0 1540 Node *adr = in(MemNode::Address);
duke@0 1541 const Type *t2 = phase->type( adr );
duke@0 1542 if (t2 == Type::TOP) return Type::TOP;
duke@0 1543 const TypePtr *tp = t2->is_ptr();
duke@0 1544 if (TypePtr::above_centerline(tp->ptr()) ||
duke@0 1545 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@0 1546
duke@0 1547 // Return a more precise klass, if possible
duke@0 1548 const TypeInstPtr *tinst = tp->isa_instptr();
duke@0 1549 if (tinst != NULL) {
duke@0 1550 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@0 1551 int offset = tinst->offset();
duke@0 1552 if (ik == phase->C->env()->Class_klass()
duke@0 1553 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1554 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1555 // We are loading a special hidden field from a Class mirror object,
duke@0 1556 // the field which points to the VM's Klass metaobject.
duke@0 1557 ciType* t = tinst->java_mirror_type();
duke@0 1558 // java_mirror_type returns non-null for compile-time Class constants.
duke@0 1559 if (t != NULL) {
duke@0 1560 // constant oop => constant klass
duke@0 1561 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1562 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@0 1563 }
duke@0 1564 if (!t->is_klass()) {
duke@0 1565 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@0 1566 return TypePtr::NULL_PTR;
duke@0 1567 }
duke@0 1568 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@0 1569 return TypeKlassPtr::make(t->as_klass());
duke@0 1570 }
duke@0 1571 // non-constant mirror, so we can't tell what's going on
duke@0 1572 }
duke@0 1573 if( !ik->is_loaded() )
duke@0 1574 return _type; // Bail out if not loaded
duke@0 1575 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1576 if (tinst->klass_is_exact()) {
duke@0 1577 return TypeKlassPtr::make(ik);
duke@0 1578 }
duke@0 1579 // See if we can become precise: no subklasses and no interface
duke@0 1580 // (Note: We need to support verified interfaces.)
duke@0 1581 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1582 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1583 // Add a dependence; if any subclass added we need to recompile
duke@0 1584 if (!ik->is_final()) {
duke@0 1585 // %%% should use stronger assert_unique_concrete_subtype instead
duke@0 1586 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1587 }
duke@0 1588 // Return precise klass
duke@0 1589 return TypeKlassPtr::make(ik);
duke@0 1590 }
duke@0 1591
duke@0 1592 // Return root of possible klass
duke@0 1593 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@0 1594 }
duke@0 1595 }
duke@0 1596
duke@0 1597 // Check for loading klass from an array
duke@0 1598 const TypeAryPtr *tary = tp->isa_aryptr();
duke@0 1599 if( tary != NULL ) {
duke@0 1600 ciKlass *tary_klass = tary->klass();
duke@0 1601 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@0 1602 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@0 1603 if (tary->klass_is_exact()) {
duke@0 1604 return TypeKlassPtr::make(tary_klass);
duke@0 1605 }
duke@0 1606 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@0 1607 // If the klass is an object array, we defer the question to the
duke@0 1608 // array component klass.
duke@0 1609 if( ak->is_obj_array_klass() ) {
duke@0 1610 assert( ak->is_loaded(), "" );
duke@0 1611 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@0 1612 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@0 1613 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@0 1614 // See if we can become precise: no subklasses and no interface
duke@0 1615 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1616 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1617 // Add a dependence; if any subclass added we need to recompile
duke@0 1618 if (!ik->is_final()) {
duke@0 1619 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1620 }
duke@0 1621 // Return precise array klass
duke@0 1622 return TypeKlassPtr::make(ak);
duke@0 1623 }
duke@0 1624 }
duke@0 1625 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 1626 } else { // Found a type-array?
duke@0 1627 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1628 assert( ak->is_type_array_klass(), "" );
duke@0 1629 return TypeKlassPtr::make(ak); // These are always precise
duke@0 1630 }
duke@0 1631 }
duke@0 1632 }
duke@0 1633
duke@0 1634 // Check for loading klass from an array klass
duke@0 1635 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1636 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1637 ciKlass* klass = tkls->klass();
duke@0 1638 if( !klass->is_loaded() )
duke@0 1639 return _type; // Bail out if not loaded
duke@0 1640 if( klass->is_obj_array_klass() &&
duke@0 1641 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1642 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@0 1643 // // Always returning precise element type is incorrect,
duke@0 1644 // // e.g., element type could be object and array may contain strings
duke@0 1645 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@0 1646
duke@0 1647 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@0 1648 // according to the element type's subclassing.
duke@0 1649 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@0 1650 }
duke@0 1651 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
duke@0 1652 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1653 ciKlass* sup = klass->as_instance_klass()->super();
duke@0 1654 // The field is Klass::_super. Return its (constant) value.
duke@0 1655 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@0 1656 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@0 1657 }
duke@0 1658 }
duke@0 1659
duke@0 1660 // Bailout case
duke@0 1661 return LoadNode::Value(phase);
duke@0 1662 }
duke@0 1663
duke@0 1664 //------------------------------Identity---------------------------------------
duke@0 1665 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@0 1666 // Also feed through the klass in Allocate(...klass...)._klass.
duke@0 1667 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
duke@0 1668 Node* x = LoadNode::Identity(phase);
duke@0 1669 if (x != this) return x;
duke@0 1670
duke@0 1671 // Take apart the address into an oop and and offset.
duke@0 1672 // Return 'this' if we cannot.
duke@0 1673 Node* adr = in(MemNode::Address);
duke@0 1674 intptr_t offset = 0;
duke@0 1675 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1676 if (base == NULL) return this;
duke@0 1677 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@0 1678 if (toop == NULL) return this;
duke@0 1679
duke@0 1680 // We can fetch the klass directly through an AllocateNode.
duke@0 1681 // This works even if the klass is not constant (clone or newArray).
duke@0 1682 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1683 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@0 1684 if (allocated_klass != NULL) {
duke@0 1685 return allocated_klass;
duke@0 1686 }
duke@0 1687 }
duke@0 1688
duke@0 1689 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
duke@0 1690 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
duke@0 1691 // See inline_native_Class_query for occurrences of these patterns.
duke@0 1692 // Java Example: x.getClass().isAssignableFrom(y)
duke@0 1693 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@0 1694 //
duke@0 1695 // This improves reflective code, often making the Class
duke@0 1696 // mirror go completely dead. (Current exception: Class
duke@0 1697 // mirrors may appear in debug info, but we could clean them out by
duke@0 1698 // introducing a new debug info operator for klassOop.java_mirror).
duke@0 1699 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@0 1700 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1701 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1702 // We are loading a special hidden field from a Class mirror,
duke@0 1703 // the field which points to its Klass or arrayKlass metaobject.
duke@0 1704 if (base->is_Load()) {
duke@0 1705 Node* adr2 = base->in(MemNode::Address);
duke@0 1706 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@0 1707 if (tkls != NULL && !tkls->empty()
duke@0 1708 && (tkls->klass()->is_instance_klass() ||
duke@0 1709 tkls->klass()->is_array_klass())
duke@0 1710 && adr2->is_AddP()
duke@0 1711 ) {
duke@0 1712 int mirror_field = Klass::java_mirror_offset_in_bytes();
duke@0 1713 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1714 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
duke@0 1715 }
duke@0 1716 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
duke@0 1717 return adr2->in(AddPNode::Base);
duke@0 1718 }
duke@0 1719 }
duke@0 1720 }
duke@0 1721 }
duke@0 1722
duke@0 1723 return this;
duke@0 1724 }
duke@0 1725
duke@0 1726 //------------------------------Value-----------------------------------------
duke@0 1727 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@0 1728 // Either input is TOP ==> the result is TOP
duke@0 1729 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1730 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1731 Node *adr = in(MemNode::Address);
duke@0 1732 const Type *t2 = phase->type( adr );
duke@0 1733 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1734 const TypePtr *tp = t2->is_ptr();
duke@0 1735 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@0 1736 const TypeAryPtr *tap = tp->isa_aryptr();
duke@0 1737 if( !tap ) return _type;
duke@0 1738 return tap->size();
duke@0 1739 }
duke@0 1740
duke@0 1741 //------------------------------Identity---------------------------------------
duke@0 1742 // Feed through the length in AllocateArray(...length...)._length.
duke@0 1743 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@0 1744 Node* x = LoadINode::Identity(phase);
duke@0 1745 if (x != this) return x;
duke@0 1746
duke@0 1747 // Take apart the address into an oop and and offset.
duke@0 1748 // Return 'this' if we cannot.
duke@0 1749 Node* adr = in(MemNode::Address);
duke@0 1750 intptr_t offset = 0;
duke@0 1751 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1752 if (base == NULL) return this;
duke@0 1753 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@0 1754 if (tary == NULL) return this;
duke@0 1755
duke@0 1756 // We can fetch the length directly through an AllocateArrayNode.
duke@0 1757 // This works even if the length is not constant (clone or newArray).
duke@0 1758 if (offset == arrayOopDesc::length_offset_in_bytes()) {
duke@0 1759 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
duke@0 1760 if (allocated_length != NULL) {
duke@0 1761 return allocated_length;
duke@0 1762 }
duke@0 1763 }
duke@0 1764
duke@0 1765 return this;
duke@0 1766
duke@0 1767 }
duke@0 1768 //=============================================================================
duke@0 1769 //---------------------------StoreNode::make-----------------------------------
duke@0 1770 // Polymorphic factory method:
coleenp@113 1771 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
coleenp@113 1772 Compile* C = gvn.C;
coleenp@113 1773
duke@0 1774 switch (bt) {
duke@0 1775 case T_BOOLEAN:
duke@0 1776 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
duke@0 1777 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
duke@0 1778 case T_CHAR:
duke@0 1779 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
duke@0 1780 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
duke@0 1781 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
duke@0 1782 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
duke@0 1783 case T_ADDRESS:
coleenp@113 1784 case T_OBJECT:
coleenp@113 1785 #ifdef _LP64
coleenp@113 1786 if (adr->bottom_type()->is_narrow() ||
coleenp@113 1787 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
coleenp@113 1788 adr->bottom_type()->isa_rawptr())) {
coleenp@113 1789 const TypePtr* type = val->bottom_type()->is_ptr();
coleenp@113 1790 Node* cp;
coleenp@113 1791 if (type->isa_oopptr()) {
coleenp@113 1792 const TypeNarrowOop* etype = type->is_oopptr()->make_narrowoop();
coleenp@113 1793 cp = gvn.transform(new (C, 2) EncodePNode(val, etype));
coleenp@113 1794 } else if (type == TypePtr::NULL_PTR) {
coleenp@113 1795 cp = gvn.transform(new (C, 1) ConNNode(TypeNarrowOop::NULL_PTR));
coleenp@113 1796 } else {
coleenp@113 1797 ShouldNotReachHere();
coleenp@113 1798 }
coleenp@113 1799 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp);
coleenp@113 1800 } else
coleenp@113 1801 #endif
coleenp@113 1802 {
coleenp@113 1803 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
coleenp@113 1804 }
duke@0 1805 }
duke@0 1806 ShouldNotReachHere();
duke@0 1807 return (StoreNode*)NULL;
duke@0 1808 }
duke@0 1809
duke@0 1810 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
duke@0 1811 bool require_atomic = true;
duke@0 1812 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
duke@0 1813 }
duke@0 1814
duke@0 1815
duke@0 1816 //--------------------------bottom_type----------------------------------------
duke@0 1817 const Type *StoreNode::bottom_type() const {
duke@0 1818 return Type::MEMORY;
duke@0 1819 }
duke@0 1820
duke@0 1821 //------------------------------hash-------------------------------------------
duke@0 1822 uint StoreNode::hash() const {
duke@0 1823 // unroll addition of interesting fields
duke@0 1824 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@0 1825
duke@0 1826 // Since they are not commoned, do not hash them:
duke@0 1827 return NO_HASH;
duke@0 1828 }
duke@0 1829
duke@0 1830 //------------------------------Ideal------------------------------------------
duke@0 1831 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@0 1832 // When a store immediately follows a relevant allocation/initialization,
duke@0 1833 // try to capture it into the initialization, or hoist it above.
duke@0 1834 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1835 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1836 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1837
duke@0 1838 Node* mem = in(MemNode::Memory);
duke@0 1839 Node* address = in(MemNode::Address);
duke@0 1840
duke@0 1841 // Back-to-back stores to same address? Fold em up.
duke@0 1842 // Generally unsafe if I have intervening uses...
duke@0 1843 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
duke@0 1844 // Looking at a dead closed cycle of memory?
duke@0 1845 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@0 1846
duke@0 1847 assert(Opcode() == mem->Opcode() ||
duke@0 1848 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@0 1849 "no mismatched stores, except on raw memory");
duke@0 1850
duke@0 1851 if (mem->outcnt() == 1 && // check for intervening uses
duke@0 1852 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@0 1853 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@0 1854 // For example, 'mem' might be the final state at a conditional return.
duke@0 1855 // Or, 'mem' might be used by some node which is live at the same time
duke@0 1856 // 'this' is live, which might be unschedulable. So, require exactly
duke@0 1857 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@0 1858 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@0 1859 if (can_reshape) { // (%%% is this an anachronism?)
duke@0 1860 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@0 1861 phase->is_IterGVN());
duke@0 1862 } else {
duke@0 1863 // It's OK to do this in the parser, since DU info is always accurate,
duke@0 1864 // and the parser always refers to nodes via SafePointNode maps.
duke@0 1865 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@0 1866 }
duke@0 1867 return this;
duke@0 1868 }
duke@0 1869 }
duke@0 1870
duke@0 1871 // Capture an unaliased, unconditional, simple store into an initializer.
duke@0 1872 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@0 1873 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@0 1874 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 1875 InitializeNode* init = mem->in(0)->as_Initialize();
duke@0 1876 intptr_t offset = init->can_capture_store(this, phase);
duke@0 1877 if (offset > 0) {
duke@0 1878 Node* moved = init->capture_store(this, offset, phase);
duke@0 1879 // If the InitializeNode captured me, it made a raw copy of me,
duke@0 1880 // and I need to disappear.
duke@0 1881 if (moved != NULL) {
duke@0 1882 // %%% hack to ensure that Ideal returns a new node:
duke@0 1883 mem = MergeMemNode::make(phase->C, mem);
duke@0 1884 return mem; // fold me away
duke@0 1885 }
duke@0 1886 }
duke@0 1887 }
duke@0 1888
duke@0 1889 return NULL; // No further progress
duke@0 1890 }
duke@0 1891
duke@0 1892 //------------------------------Value-----------------------------------------
duke@0 1893 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@0 1894 // Either input is TOP ==> the result is TOP
duke@0 1895 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1896 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1897 const Type *t2 = phase->type( in(MemNode::Address) );
duke@0 1898 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1899 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@0 1900 if( t3 == Type::TOP ) return Type::TOP;
duke@0 1901 return Type::MEMORY;
duke@0 1902 }
duke@0 1903
duke@0 1904 //------------------------------Identity---------------------------------------
duke@0 1905 // Remove redundant stores:
duke@0 1906 // Store(m, p, Load(m, p)) changes to m.
duke@0 1907 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@0 1908 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@0 1909 Node* mem = in(MemNode::Memory);
duke@0 1910 Node* adr = in(MemNode::Address);
duke@0 1911 Node* val = in(MemNode::ValueIn);
duke@0 1912
duke@0 1913 // Load then Store? Then the Store is useless
duke@0 1914 if (val->is_Load() &&
duke@0 1915 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
duke@0 1916 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
duke@0 1917 val->as_Load()->store_Opcode() == Opcode()) {
duke@0 1918 return mem;
duke@0 1919 }
duke@0 1920
duke@0 1921 // Two stores in a row of the same value?
duke@0 1922 if (mem->is_Store() &&
duke@0 1923 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
duke@0 1924 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
duke@0 1925 mem->Opcode() == Opcode()) {
duke@0 1926 return mem;
duke@0 1927 }
duke@0 1928
duke@0 1929 // Store of zero anywhere into a freshly-allocated object?
duke@0 1930 // Then the store is useless.
duke@0 1931 // (It must already have been captured by the InitializeNode.)
duke@0 1932 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@0 1933 // a newly allocated object is already all-zeroes everywhere
duke@0 1934 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@0 1935 return mem;
duke@0 1936 }
duke@0 1937
duke@0 1938 // the store may also apply to zero-bits in an earlier object
duke@0 1939 Node* prev_mem = find_previous_store(phase);
duke@0 1940 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1941 if (prev_mem != NULL) {
duke@0 1942 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@0 1943 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@0 1944 // prev_val and val might differ by a cast; it would be good
duke@0 1945 // to keep the more informative of the two.
duke@0 1946 return mem;
duke@0 1947 }
duke@0 1948 }
duke@0 1949 }
duke@0 1950
duke@0 1951 return this;
duke@0 1952 }
duke@0 1953
duke@0 1954 //------------------------------match_edge-------------------------------------
duke@0 1955 // Do we Match on this edge index or not? Match only memory & value
duke@0 1956 uint StoreNode::match_edge(uint idx) const {
duke@0 1957 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@0 1958 }
duke@0 1959
duke@0 1960 //------------------------------cmp--------------------------------------------
duke@0 1961 // Do not common stores up together. They generally have to be split
duke@0 1962 // back up anyways, so do not bother.
duke@0 1963 uint StoreNode::cmp( const Node &n ) const {
duke@0 1964 return (&n == this); // Always fail except on self
duke@0 1965 }
duke@0 1966
duke@0 1967 //------------------------------Ideal_masked_input-----------------------------
duke@0 1968 // Check for a useless mask before a partial-word store
duke@0 1969 // (StoreB ... (AndI valIn conIa) )
duke@0 1970 // If (conIa & mask == mask) this simplifies to
duke@0 1971 // (StoreB ... (valIn) )
duke@0 1972 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@0 1973 Node *val = in(MemNode::ValueIn);
duke@0 1974 if( val->Opcode() == Op_AndI ) {
duke@0 1975 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 1976 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@0 1977 set_req(MemNode::ValueIn, val->in(1));
duke@0 1978 return this;
duke@0 1979 }
duke@0 1980 }
duke@0 1981 return NULL;
duke@0 1982 }
duke@0 1983
duke@0 1984
duke@0 1985 //------------------------------Ideal_sign_extended_input----------------------
duke@0 1986 // Check for useless sign-extension before a partial-word store
duke@0 1987 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@0 1988 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@0 1989 // (StoreB ... (valIn) )
duke@0 1990 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@0 1991 Node *val = in(MemNode::ValueIn);
duke@0 1992 if( val->Opcode() == Op_RShiftI ) {
duke@0 1993 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 1994 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@0 1995 Node *shl = val->in(1);
duke@0 1996 if( shl->Opcode() == Op_LShiftI ) {
duke@0 1997 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@0 1998 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@0 1999 set_req(MemNode::ValueIn, shl->in(1));
duke@0 2000 return this;
duke@0 2001 }
duke@0 2002 }
duke@0 2003 }
duke@0 2004 }
duke@0 2005 return NULL;
duke@0 2006 }
duke@0 2007
duke@0 2008 //------------------------------value_never_loaded-----------------------------------
duke@0 2009 // Determine whether there are any possible loads of the value stored.
duke@0 2010 // For simplicity, we actually check if there are any loads from the
duke@0 2011 // address stored to, not just for loads of the value stored by this node.
duke@0 2012 //
duke@0 2013 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@0 2014 Node *adr = in(Address);
duke@0 2015 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@0 2016 if (adr_oop == NULL)
duke@0 2017 return false;
kvn@64 2018 if (!adr_oop->is_instance_field())
duke@0 2019 return false; // if not a distinct instance, there may be aliases of the address
duke@0 2020 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@0 2021 Node *use = adr->fast_out(i);
duke@0 2022 int opc = use->Opcode();
duke@0 2023 if (use->is_Load() || use->is_LoadStore()) {
duke@0 2024 return false;
duke@0 2025 }
duke@0 2026 }
duke@0 2027 return true;
duke@0 2028 }
duke@0 2029
duke@0 2030 //=============================================================================
duke@0 2031 //------------------------------Ideal------------------------------------------
duke@0 2032 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2033 // we can skip the AND operation. If the store is from a sign-extension
duke@0 2034 // (a left shift, then right shift) we can skip both.
duke@0 2035 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2036 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@0 2037 if( progress != NULL ) return progress;
duke@0 2038
duke@0 2039 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@0 2040 if( progress != NULL ) return progress;
duke@0 2041
duke@0 2042 // Finally check the default case
duke@0 2043 return StoreNode::Ideal(phase, can_reshape);
duke@0 2044 }
duke@0 2045
duke@0 2046 //=============================================================================
duke@0 2047 //------------------------------Ideal------------------------------------------
duke@0 2048 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2049 // we can skip the AND operation
duke@0 2050 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2051 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@0 2052 if( progress != NULL ) return progress;
duke@0 2053
duke@0 2054 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@0 2055 if( progress != NULL ) return progress;
duke@0 2056
duke@0 2057 // Finally check the default case
duke@0 2058 return StoreNode::Ideal(phase, can_reshape);
duke@0 2059 }
duke@0 2060
duke@0 2061 //=============================================================================
duke@0 2062 //------------------------------Identity---------------------------------------
duke@0 2063 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@0 2064 // No need to card mark when storing a null ptr
duke@0 2065 Node* my_store = in(MemNode::OopStore);
duke@0 2066 if (my_store->is_Store()) {
duke@0 2067 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@0 2068 if( t1 == TypePtr::NULL_PTR ) {
duke@0 2069 return in(MemNode::Memory);
duke@0 2070 }
duke@0 2071 }
duke@0 2072 return this;
duke@0 2073 }
duke@0 2074
duke@0 2075 //------------------------------Value-----------------------------------------
duke@0 2076 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@43 2077 // Either input is TOP ==> the result is TOP
kvn@43 2078 const Type *t = phase->type( in(MemNode::Memory) );
kvn@43 2079 if( t == Type::TOP ) return Type::TOP;
kvn@43 2080 t = phase->type( in(MemNode::Address) );
kvn@43 2081 if( t == Type::TOP ) return Type::TOP;
kvn@43 2082 t = phase->type( in(MemNode::ValueIn) );
kvn@43 2083 if( t == Type::TOP ) return Type::TOP;
duke@0 2084 // If extra input is TOP ==> the result is TOP
kvn@43 2085 t = phase->type( in(MemNode::OopStore) );
kvn@43 2086 if( t == Type::TOP ) return Type::TOP;
duke@0 2087
duke@0 2088 return StoreNode::Value( phase );
duke@0 2089 }
duke@0 2090
duke@0 2091
duke@0 2092 //=============================================================================
duke@0 2093 //----------------------------------SCMemProjNode------------------------------
duke@0 2094 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@0 2095 {
duke@0 2096 return bottom_type();
duke@0 2097 }
duke@0 2098
duke@0 2099 //=============================================================================
duke@0 2100 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
duke@0 2101 init_req(MemNode::Control, c );
duke@0 2102 init_req(MemNode::Memory , mem);
duke@0 2103 init_req(MemNode::Address, adr);
duke@0 2104 init_req(MemNode::ValueIn, val);
duke@0 2105 init_req( ExpectedIn, ex );
duke@0 2106 init_class_id(Class_LoadStore);
duke@0 2107
duke@0 2108 }
duke@0 2109
duke@0 2110 //=============================================================================
duke@0 2111 //-------------------------------adr_type--------------------------------------
duke@0 2112 // Do we Match on this edge index or not? Do not match memory
duke@0 2113 const TypePtr* ClearArrayNode::adr_type() const {
duke@0 2114 Node *adr = in(3);
duke@0 2115 return MemNode::calculate_adr_type(adr->bottom_type());
duke@0 2116 }
duke@0 2117
duke@0 2118 //------------------------------match_edge-------------------------------------
duke@0 2119 // Do we Match on this edge index or not? Do not match memory
duke@0 2120 uint ClearArrayNode::match_edge(uint idx) const {
duke@0 2121 return idx > 1;
duke@0 2122 }
duke@0 2123
duke@0 2124 //------------------------------Identity---------------------------------------
duke@0 2125 // Clearing a zero length array does nothing
duke@0 2126 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@68 2127 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@0 2128 }
duke@0 2129
duke@0 2130 //------------------------------Idealize---------------------------------------
duke@0 2131 // Clearing a short array is faster with stores
duke@0 2132 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2133 const int unit = BytesPerLong;
duke@0 2134 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@0 2135 if (!t) return NULL;
duke@0 2136 if (!t->is_con()) return NULL;
duke@0 2137 intptr_t raw_count = t->get_con();
duke@0 2138 intptr_t size = raw_count;
duke@0 2139 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@0 2140 // Clearing nothing uses the Identity call.
duke@0 2141 // Negative clears are possible on dead ClearArrays
duke@0 2142 // (see jck test stmt114.stmt11402.val).
duke@0 2143 if (size <= 0 || size % unit != 0) return NULL;
duke@0 2144 intptr_t count = size / unit;
duke@0 2145 // Length too long; use fast hardware clear
duke@0 2146 if (size > Matcher::init_array_short_size) return NULL;
duke@0 2147 Node *mem = in(1);
duke@0 2148 if( phase->type(mem)==Type::TOP ) return NULL;
duke@0 2149 Node *adr = in(3);
duke@0 2150 const Type* at = phase->type(adr);
duke@0 2151 if( at==Type::TOP ) return NULL;
duke@0 2152 const TypePtr* atp = at->isa_ptr();
duke@0 2153 // adjust atp to be the correct array element address type
duke@0 2154 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@0 2155 else atp = atp->add_offset(Type::OffsetBot);
duke@0 2156 // Get base for derived pointer purposes
duke@0 2157 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@0 2158 Node *base = adr->in(1);
duke@0 2159
duke@0 2160 Node *zero = phase->makecon(TypeLong::ZERO);
duke@0 2161 Node *off = phase->MakeConX(BytesPerLong);
duke@0 2162 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2163 count--;
duke@0 2164 while( count-- ) {
duke@0 2165 mem = phase->transform(mem);
duke@0 2166 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
duke@0 2167 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2168 }
duke@0 2169 return mem;
duke@0 2170 }
duke@0 2171
duke@0 2172 //----------------------------clear_memory-------------------------------------
duke@0 2173 // Generate code to initialize object storage to zero.
duke@0 2174 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2175 intptr_t start_offset,
duke@0 2176 Node* end_offset,
duke@0 2177 PhaseGVN* phase) {
duke@0 2178 Compile* C = phase->C;
duke@0 2179 intptr_t offset = start_offset;
duke@0 2180
duke@0 2181 int unit = BytesPerLong;
duke@0 2182 if ((offset % unit) != 0) {
duke@0 2183 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
duke@0 2184 adr = phase->transform(adr);
duke@0 2185 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2186 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2187 mem = phase->transform(mem);
duke@0 2188 offset += BytesPerInt;
duke@0 2189 }
duke@0 2190 assert((offset % unit) == 0, "");
duke@0 2191
duke@0 2192 // Initialize the remaining stuff, if any, with a ClearArray.
duke@0 2193 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@0 2194 }
duke@0 2195
duke@0 2196 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2197 Node* start_offset,
duke@0 2198 Node* end_offset,
duke@0 2199 PhaseGVN* phase) {
never@68 2200 if (start_offset == end_offset) {
never@68 2201 // nothing to do
never@68 2202 return mem;
never@68 2203 }
never@68 2204
duke@0 2205 Compile* C = phase->C;
duke@0 2206 int unit = BytesPerLong;
duke@0 2207 Node* zbase = start_offset;
duke@0 2208 Node* zend = end_offset;
duke@0 2209
duke@0 2210 // Scale to the unit required by the CPU:
duke@0 2211 if (!Matcher::init_array_count_is_in_bytes) {
duke@0 2212 Node* shift = phase->intcon(exact_log2(unit));
duke@0 2213 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
duke@0 2214 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
duke@0 2215 }
duke@0 2216
duke@0 2217 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
duke@0 2218 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
duke@0 2219
duke@0 2220 // Bulk clear double-words
duke@0 2221 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
duke@0 2222 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
duke@0 2223 return phase->transform(mem);
duke@0 2224 }
duke@0 2225
duke@0 2226 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2227 intptr_t start_offset,
duke@0 2228 intptr_t end_offset,
duke@0 2229 PhaseGVN* phase) {
never@68 2230 if (start_offset == end_offset) {
never@68 2231 // nothing to do
never@68 2232 return mem;
never@68 2233 }
never@68 2234
duke@0 2235 Compile* C = phase->C;
duke@0 2236 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@0 2237 intptr_t done_offset = end_offset;
duke@0 2238 if ((done_offset % BytesPerLong) != 0) {
duke@0 2239 done_offset -= BytesPerInt;
duke@0 2240 }
duke@0 2241 if (done_offset > start_offset) {
duke@0 2242 mem = clear_memory(ctl, mem, dest,
duke@0 2243 start_offset, phase->MakeConX(done_offset), phase);
duke@0 2244 }
duke@0 2245 if (done_offset < end_offset) { // emit the final 32-bit store
duke@0 2246 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@0 2247 adr = phase->transform(adr);
duke@0 2248 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2249 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2250 mem = phase->transform(mem);
duke@0 2251 done_offset += BytesPerInt;
duke@0 2252 }
duke@0 2253 assert(done_offset == end_offset, "");
duke@0 2254 return mem;
duke@0 2255 }
duke@0 2256
duke@0 2257 //=============================================================================
duke@0 2258 // Do we match on this edge? No memory edges
duke@0 2259 uint StrCompNode::match_edge(uint idx) const {
duke@0 2260 return idx == 5 || idx == 6;
duke@0 2261 }
duke@0 2262
duke@0 2263 //------------------------------Ideal------------------------------------------
duke@0 2264 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2265 // control copies
duke@0 2266 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2267 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2268 }
duke@0 2269
duke@0 2270
duke@0 2271 //=============================================================================
duke@0 2272 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@0 2273 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@0 2274 _adr_type(C->get_adr_type(alias_idx))
duke@0 2275 {
duke@0 2276 init_class_id(Class_MemBar);
duke@0 2277 Node* top = C->top();
duke@0 2278 init_req(TypeFunc::I_O,top);
duke@0 2279 init_req(TypeFunc::FramePtr,top);
duke@0 2280 init_req(TypeFunc::ReturnAdr,top);
duke@0 2281 if (precedent != NULL)
duke@0 2282 init_req(TypeFunc::Parms, precedent);
duke@0 2283 }
duke@0 2284
duke@0 2285 //------------------------------cmp--------------------------------------------
duke@0 2286 uint MemBarNode::hash() const { return NO_HASH; }
duke@0 2287 uint MemBarNode::cmp( const Node &n ) const {
duke@0 2288 return (&n == this); // Always fail except on self
duke@0 2289 }
duke@0 2290
duke@0 2291 //------------------------------make-------------------------------------------
duke@0 2292 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@0 2293 int len = Precedent + (pn == NULL? 0: 1);
duke@0 2294 switch (opcode) {
duke@0 2295 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
duke@0 2296 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
duke@0 2297 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
duke@0 2298 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
duke@0 2299 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
duke@0 2300 default: ShouldNotReachHere(); return NULL;
duke@0 2301 }
duke@0 2302 }
duke@0 2303
duke@0 2304 //------------------------------Ideal------------------------------------------
duke@0 2305 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2306 // control copies
duke@0 2307 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 2308 if (remove_dead_region(phase, can_reshape)) return this;
duke@0 2309 return NULL;
duke@0 2310 }
duke@0 2311
duke@0 2312 //------------------------------Value------------------------------------------
duke@0 2313 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@0 2314 if( !in(0) ) return Type::TOP;
duke@0 2315 if( phase->type(in(0)) == Type::TOP )
duke@0 2316 return Type::TOP;
duke@0 2317 return TypeTuple::MEMBAR;
duke@0 2318 }
duke@0 2319
duke@0 2320 //------------------------------match------------------------------------------
duke@0 2321 // Construct projections for memory.
duke@0 2322 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@0 2323 switch (proj->_con) {
duke@0 2324 case TypeFunc::Control:
duke@0 2325 case TypeFunc::Memory:
duke@0 2326 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@0 2327 }
duke@0 2328 ShouldNotReachHere();
duke@0 2329 return NULL;
duke@0 2330 }
duke@0 2331
duke@0 2332 //===========================InitializeNode====================================
duke@0 2333 // SUMMARY:
duke@0 2334 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@0 2335 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@0 2336 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@0 2337 // It can coalesce related raw stores into larger units (called 'tiles').
duke@0 2338 // It can avoid zeroing new storage for memory units which have raw inits.
duke@0 2339 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@0 2340 //
duke@0 2341 // EXAMPLE:
duke@0 2342 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@0 2343 // ctl = incoming control; mem* = incoming memory
duke@0 2344 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@0 2345 // First allocate uninitialized memory and fill in the header:
duke@0 2346 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@0 2347 // ctl := alloc.Control; mem* := alloc.Memory*
duke@0 2348 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@0 2349 // Then initialize to zero the non-header parts of the raw memory block:
duke@0 2350 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@0 2351 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@0 2352 // After the initialize node executes, the object is ready for service:
duke@0 2353 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@0 2354 // Suppose its body is immediately initialized as {1,2}:
duke@0 2355 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2356 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2357 // mem.SLICE(#short[*]) := store2
duke@0 2358 //
duke@0 2359 // DETAILS:
duke@0 2360 // An InitializeNode collects and isolates object initialization after
duke@0 2361 // an AllocateNode and before the next possible safepoint. As a
duke@0 2362 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@0 2363 // down past any safepoint or any publication of the allocation.
duke@0 2364 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@0 2365 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@0 2366 //
duke@0 2367 // The semantics of the InitializeNode include an implicit zeroing of
duke@0 2368 // the new object from object header to the end of the object.
duke@0 2369 // (The object header and end are determined by the AllocateNode.)
duke@0 2370 //
duke@0 2371 // Certain stores may be added as direct inputs to the InitializeNode.
duke@0 2372 // These stores must update raw memory, and they must be to addresses
duke@0 2373 // derived from the raw address produced by AllocateNode, and with
duke@0 2374 // a constant offset. They must be ordered by increasing offset.
duke@0 2375 // The first one is at in(RawStores), the last at in(req()-1).
duke@0 2376 // Unlike most memory operations, they are not linked in a chain,
duke@0 2377 // but are displayed in parallel as users of the rawmem output of
duke@0 2378 // the allocation.
duke@0 2379 //
duke@0 2380 // (See comments in InitializeNode::capture_store, which continue
duke@0 2381 // the example given above.)
duke@0 2382 //
duke@0 2383 // When the associated Allocate is macro-expanded, the InitializeNode
duke@0 2384 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@0 2385 // may also be created at that point to represent any required zeroing.
duke@0 2386 // The InitializeNode is then marked 'complete', prohibiting further
duke@0 2387 // capturing of nearby memory operations.
duke@0 2388 //
duke@0 2389 // During macro-expansion, all captured initializations which store
duke@0 2390 // constant values of 32 bits or smaller are coalesced (if advantagous)
duke@0 2391 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@0 2392 // initialized in fewer memory operations. Memory words which are
duke@0 2393 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@0 2394 // by explicit stores of zero. (The code shape happens to do all
duke@0 2395 // zeroing first, then all other stores, with both sequences occurring
duke@0 2396 // in order of ascending offsets.)
duke@0 2397 //
duke@0 2398 // Alternatively, code may be inserted between an AllocateNode and its
duke@0 2399 // InitializeNode, to perform arbitrary initialization of the new object.
duke@0 2400 // E.g., the object copying intrinsics insert complex data transfers here.
duke@0 2401 // The initialization must then be marked as 'complete' disable the
duke@0 2402 // built-in zeroing semantics and the collection of initializing stores.
duke@0 2403 //
duke@0 2404 // While an InitializeNode is incomplete, reads from the memory state
duke@0 2405 // produced by it are optimizable if they match the control edge and
duke@0 2406 // new oop address associated with the allocation/initialization.
duke@0 2407 // They return a stored value (if the offset matches) or else zero.
duke@0 2408 // A write to the memory state, if it matches control and address,
duke@0 2409 // and if it is to a constant offset, may be 'captured' by the
duke@0 2410 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@0 2411 // inside the initialization, to the raw oop produced by the allocation.
duke@0 2412 // Operations on addresses which are provably distinct (e.g., to
duke@0 2413 // other AllocateNodes) are allowed to bypass the initialization.
duke@0 2414 //
duke@0 2415 // The effect of all this is to consolidate object initialization
duke@0 2416 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@0 2417 // single location, where it can be optimized as a unit.
duke@0 2418 //
duke@0 2419 // Only stores with an offset less than TrackedInitializationLimit words
duke@0 2420 // will be considered for capture by an InitializeNode. This puts a
duke@0 2421 // reasonable limit on the complexity of optimized initializations.
duke@0 2422
duke@0 2423 //---------------------------InitializeNode------------------------------------
duke@0 2424 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
duke@0 2425 : _is_complete(false),
duke@0 2426 MemBarNode(C, adr_type, rawoop)
duke@0 2427 {
duke@0 2428 init_class_id(Class_Initialize);
duke@0 2429
duke@0 2430 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@0 2431 assert(in(RawAddress) == rawoop, "proper init");
duke@0 2432 // Note: allocation() can be NULL, for secondary initialization barriers
duke@0 2433 }
duke@0 2434
duke@0 2435 // Since this node is not matched, it will be processed by the
duke@0 2436 // register allocator. Declare that there are no constraints
duke@0 2437 // on the allocation of the RawAddress edge.
duke@0 2438 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@0 2439 // This edge should be set to top, by the set_complete. But be conservative.
duke@0 2440 if (idx == InitializeNode::RawAddress)
duke@0 2441 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@0 2442 return RegMask::Empty;
duke@0 2443 }
duke@0 2444
duke@0 2445 Node* InitializeNode::memory(uint alias_idx) {
duke@0 2446 Node* mem = in(Memory);
duke@0 2447 if (mem->is_MergeMem()) {
duke@0 2448 return mem->as_MergeMem()->memory_at(alias_idx);
duke@0 2449 } else {
duke@0 2450 // incoming raw memory is not split
duke@0 2451 return mem;
duke@0 2452 }
duke@0 2453 }
duke@0 2454
duke@0 2455 bool InitializeNode::is_non_zero() {
duke@0 2456 if (is_complete()) return false;
duke@0 2457 remove_extra_zeroes();
duke@0 2458 return (req() > RawStores);
duke@0 2459 }
duke@0 2460
duke@0 2461 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@0 2462 assert(!is_complete(), "caller responsibility");
duke@0 2463 _is_complete = true;
duke@0 2464
duke@0 2465 // After this node is complete, it contains a bunch of
duke@0 2466 // raw-memory initializations. There is no need for
duke@0 2467 // it to have anything to do with non-raw memory effects.
duke@0 2468 // Therefore, tell all non-raw users to re-optimize themselves,
duke@0 2469 // after skipping the memory effects of this initialization.
duke@0 2470 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@0 2471 if (igvn) igvn->add_users_to_worklist(this);
duke@0 2472 }
duke@0 2473
duke@0 2474 // convenience function
duke@0 2475 // return false if the init contains any stores already
duke@0 2476 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@0 2477 InitializeNode* init = initialization();
duke@0 2478 if (init == NULL || init->is_complete()) return false;
duke@0 2479 init->remove_extra_zeroes();
duke@0 2480 // for now, if this allocation has already collected any inits, bail:
duke@0 2481 if (init->is_non_zero()) return false;
duke@0 2482 init->set_complete(phase);
duke@0 2483 return true;
duke@0 2484 }
duke@0 2485
duke@0 2486 void InitializeNode::remove_extra_zeroes() {
duke@0 2487 if (req() == RawStores) return;
duke@0 2488 Node* zmem = zero_memory();
duke@0 2489 uint fill = RawStores;
duke@0 2490 for (uint i = fill; i < req(); i++) {
duke@0 2491 Node* n = in(i);
duke@0 2492 if (n->is_top() || n == zmem) continue; // skip
duke@0 2493 if (fill < i) set_req(fill, n); // compact
duke@0 2494 ++fill;
duke@0 2495 }
duke@0 2496 // delete any empty spaces created:
duke@0 2497 while (fill < req()) {
duke@0 2498 del_req(fill);
duke@0 2499 }
duke@0 2500 }
duke@0 2501
duke@0 2502 // Helper for remembering which stores go with which offsets.
duke@0 2503 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@0 2504 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@0 2505 intptr_t offset = -1;
duke@0 2506 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@0 2507 phase, offset);
duke@0 2508 if (base == NULL) return -1; // something is dead,
duke@0 2509 if (offset < 0) return -1; // dead, dead
duke@0 2510 return offset;
duke@0 2511 }
duke@0 2512
duke@0 2513 // Helper for proving that an initialization expression is
duke@0 2514 // "simple enough" to be folded into an object initialization.
duke@0 2515 // Attempts to prove that a store's initial value 'n' can be captured
duke@0 2516 // within the initialization without creating a vicious cycle, such as:
duke@0 2517 // { Foo p = new Foo(); p.next = p; }
duke@0 2518 // True for constants and parameters and small combinations thereof.
duke@0 2519 bool InitializeNode::detect_init_independence(Node* n,
duke@0 2520 bool st_is_pinned,
duke@0 2521 int& count) {
duke@0 2522 if (n == NULL) return true; // (can this really happen?)
duke@0 2523 if (n->is_Proj()) n = n->in(0);
duke@0 2524 if (n == this) return false; // found a cycle
duke@0 2525 if (n->is_Con()) return true;
duke@0 2526 if (n->is_Start()) return true; // params, etc., are OK
duke@0 2527 if (n->is_Root()) return true; // even better
duke@0 2528
duke@0 2529 Node* ctl = n->in(0);
duke@0 2530 if (ctl != NULL && !ctl->is_top()) {
duke@0 2531 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@0 2532 if (ctl == this) return false;
duke@0 2533
duke@0 2534 // If we already know that the enclosing memory op is pinned right after
duke@0 2535 // the init, then any control flow that the store has picked up
duke@0 2536 // must have preceded the init, or else be equal to the init.
duke@0 2537 // Even after loop optimizations (which might change control edges)
duke@0 2538 // a store is never pinned *before* the availability of its inputs.
duke@0 2539 if (!MemNode::detect_dominating_control(ctl, this->in(0)))
duke@0 2540 return false; // failed to prove a good control
duke@0 2541
duke@0 2542 }
duke@0 2543
duke@0 2544 // Check data edges for possible dependencies on 'this'.
duke@0 2545 if ((count += 1) > 20) return false; // complexity limit
duke@0 2546 for (uint i = 1; i < n->req(); i++) {
duke@0 2547 Node* m = n->in(i);
duke@0 2548 if (m == NULL || m == n || m->is_top()) continue;
duke@0 2549 uint first_i = n->find_edge(m);
duke@0 2550 if (i != first_i) continue; // process duplicate edge just once
duke@0 2551 if (!detect_init_independence(m, st_is_pinned, count)) {
duke@0 2552 return false;
duke@0 2553 }
duke@0 2554 }
duke@0 2555
duke@0 2556 return true;
duke@0 2557 }
duke@0 2558
duke@0 2559 // Here are all the checks a Store must pass before it can be moved into
duke@0 2560 // an initialization. Returns zero if a check fails.
duke@0 2561 // On success, returns the (constant) offset to which the store applies,
duke@0 2562 // within the initialized memory.
duke@0 2563 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
duke@0 2564 const int FAIL = 0;
duke@0 2565 if (st->req() != MemNode::ValueIn + 1)
duke@0 2566 return FAIL; // an inscrutable StoreNode (card mark?)
duke@0 2567 Node* ctl = st->in(MemNode::Control);
duke@0 2568 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@0 2569 return FAIL; // must be unconditional after the initialization
duke@0 2570 Node* mem = st->in(MemNode::Memory);
duke@0 2571 if (!(mem->is_Proj() && mem->in(0) == this))
duke@0 2572 return FAIL; // must not be preceded by other stores
duke@0 2573 Node* adr = st->in(MemNode::Address);
duke@0 2574 intptr_t offset;
duke@0 2575 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@0 2576 if (alloc == NULL)
duke@0 2577 return FAIL; // inscrutable address
duke@0 2578 if (alloc != allocation())
duke@0 2579 return FAIL; // wrong allocation! (store needs to float up)
duke@0 2580 Node* val = st->in(MemNode::ValueIn);
duke@0 2581 int complexity_count = 0;
duke@0 2582 if (!detect_init_independence(val, true, complexity_count))
duke@0 2583 return FAIL; // stored value must be 'simple enough'
duke@0 2584
duke@0 2585 return offset; // success
duke@0 2586 }
duke@0 2587
duke@0 2588 // Find the captured store in(i) which corresponds to the range
duke@0 2589 // [start..start+size) in the initialized object.
duke@0 2590 // If there is one, return its index i. If there isn't, return the
duke@0 2591 // negative of the index where it should be inserted.
duke@0 2592 // Return 0 if the queried range overlaps an initialization boundary
duke@0 2593 // or if dead code is encountered.
duke@0 2594 // If size_in_bytes is zero, do not bother with overlap checks.
duke@0 2595 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@0 2596 int size_in_bytes,
duke@0 2597 PhaseTransform* phase) {
duke@0 2598 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@0 2599
duke@0 2600 if (is_complete())
duke@0 2601 return FAIL; // arraycopy got here first; punt
duke@0 2602
duke@0 2603 assert(allocation() != NULL, "must be present");
duke@0 2604
duke@0 2605 // no negatives, no header fields:
coleenp@113 2606 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@0 2607
duke@0 2608 // after a certain size, we bail out on tracking all the stores:
duke@0 2609 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2610 if (start >= ti_limit) return FAIL;
duke@0 2611
duke@0 2612 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@0 2613 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@0 2614
duke@0 2615 Node* st = in(i);
duke@0 2616 intptr_t st_off = get_store_offset(st, phase);
duke@0 2617 if (st_off < 0) {
duke@0 2618 if (st != zero_memory()) {
duke@0 2619 return FAIL; // bail out if there is dead garbage
duke@0 2620 }
duke@0 2621 } else if (st_off > start) {
duke@0 2622 // ...we are done, since stores are ordered
duke@0 2623 if (st_off < start + size_in_bytes) {
duke@0 2624 return FAIL; // the next store overlaps
duke@0 2625 }
duke@0 2626 return -(int)i; // not found; here is where to put it
duke@0 2627 } else if (st_off < start) {
duke@0 2628 if (size_in_bytes != 0 &&
duke@0 2629 start < st_off + MAX_STORE &&
duke@0 2630 start < st_off + st->as_Store()->memory_size()) {
duke@0 2631 return FAIL; // the previous store overlaps
duke@0 2632 }
duke@0 2633 } else {
duke@0 2634 if (size_in_bytes != 0 &&
duke@0 2635 st->as_Store()->memory_size() != size_in_bytes) {
duke@0 2636 return FAIL; // mismatched store size
duke@0 2637 }
duke@0 2638 return i;
duke@0 2639 }
duke@0 2640
duke@0 2641 ++i;
duke@0 2642 }
duke@0 2643 }
duke@0 2644
duke@0 2645 // Look for a captured store which initializes at the offset 'start'
duke@0 2646 // with the given size. If there is no such store, and no other
duke@0 2647 // initialization interferes, then return zero_memory (the memory
duke@0 2648 // projection of the AllocateNode).
duke@0 2649 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@0 2650 PhaseTransform* phase) {
duke@0 2651 assert(stores_are_sane(phase), "");
duke@0 2652 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2653 if (i == 0) {
duke@0 2654 return NULL; // something is dead
duke@0 2655 } else if (i < 0) {
duke@0 2656 return zero_memory(); // just primordial zero bits here
duke@0 2657 } else {
duke@0 2658 Node* st = in(i); // here is the store at this position
duke@0 2659 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@0 2660 return st;
duke@0 2661 }
duke@0 2662 }
duke@0 2663
duke@0 2664 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@0 2665 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@0 2666 PhaseTransform* phase) {
duke@0 2667 Node* addr = in(RawAddress);
duke@0 2668 if (offset != 0) {
duke@0 2669 Compile* C = phase->C;
duke@0 2670 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
duke@0 2671 phase->MakeConX(offset)) );
duke@0 2672 }
duke@0 2673 return addr;
duke@0 2674 }
duke@0 2675
duke@0 2676 // Clone the given store, converting it into a raw store
duke@0 2677 // initializing a field or element of my new object.
duke@0 2678 // Caller is responsible for retiring the original store,
duke@0 2679 // with subsume_node or the like.
duke@0 2680 //
duke@0 2681 // From the example above InitializeNode::InitializeNode,
duke@0 2682 // here are the old stores to be captured:
duke@0 2683 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2684 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2685 //
duke@0 2686 // Here is the changed code; note the extra edges on init:
duke@0 2687 // alloc = (Allocate ...)
duke@0 2688 // rawoop = alloc.RawAddress
duke@0 2689 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@0 2690 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@0 2691 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@0 2692 // rawstore1 rawstore2)
duke@0 2693 //
duke@0 2694 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
duke@0 2695 PhaseTransform* phase) {
duke@0 2696 assert(stores_are_sane(phase), "");
duke@0 2697
duke@0 2698 if (start < 0) return NULL;
duke@0 2699 assert(can_capture_store(st, phase) == start, "sanity");
duke@0 2700
duke@0 2701 Compile* C = phase->C;
duke@0 2702 int size_in_bytes = st->memory_size();
duke@0 2703 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2704 if (i == 0) return NULL; // bail out
duke@0 2705 Node* prev_mem = NULL; // raw memory for the captured store
duke@0 2706 if (i > 0) {
duke@0 2707 prev_mem = in(i); // there is a pre-existing store under this one
duke@0 2708 set_req(i, C->top()); // temporarily disconnect it
duke@0 2709 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@0 2710 } else {
duke@0 2711 i = -i; // no pre-existing store
duke@0 2712 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@0 2713 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@0 2714 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@0 2715 else
duke@0 2716 ins_req(i, C->top()); // build a new edge
duke@0 2717 }
duke@0 2718 Node* new_st = st->clone();
duke@0 2719 new_st->set_req(MemNode::Control, in(Control));
duke@0 2720 new_st->set_req(MemNode::Memory, prev_mem);
duke@0 2721 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@0 2722 new_st = phase->transform(new_st);
duke@0 2723
duke@0 2724 // At this point, new_st might have swallowed a pre-existing store
duke@0 2725 // at the same offset, or perhaps new_st might have disappeared,
duke@0 2726 // if it redundantly stored the same value (or zero to fresh memory).
duke@0 2727
duke@0 2728 // In any case, wire it in:
duke@0 2729 set_req(i, new_st);
duke@0 2730
duke@0 2731 // The caller may now kill the old guy.
duke@0 2732 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@0 2733 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@0 2734 assert(!is_complete(), "");
duke@0 2735 return new_st;
duke@0 2736 }
duke@0 2737
duke@0 2738 static bool store_constant(jlong* tiles, int num_tiles,
duke@0 2739 intptr_t st_off, int st_size,
duke@0 2740 jlong con) {
duke@0 2741 if ((st_off & (st_size-1)) != 0)
duke@0 2742 return false; // strange store offset (assume size==2**N)
duke@0 2743 address addr = (address)tiles + st_off;
duke@0 2744 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@0 2745 switch (st_size) {
duke@0 2746 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@0 2747 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@0 2748 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@0 2749 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@0 2750 default: return false; // strange store size (detect size!=2**N here)
duke@0 2751 }
duke@0 2752 return true; // return success to caller
duke@0 2753 }
duke@0 2754
duke@0 2755 // Coalesce subword constants into int constants and possibly
duke@0 2756 // into long constants. The goal, if the CPU permits,
duke@0 2757 // is to initialize the object with a small number of 64-bit tiles.
duke@0 2758 // Also, convert floating-point constants to bit patterns.
duke@0 2759 // Non-constants are not relevant to this pass.
duke@0 2760 //
duke@0 2761 // In terms of the running example on InitializeNode::InitializeNode
duke@0 2762 // and InitializeNode::capture_store, here is the transformation
duke@0 2763 // of rawstore1 and rawstore2 into rawstore12:
duke@0 2764 // alloc = (Allocate ...)
duke@0 2765 // rawoop = alloc.RawAddress
duke@0 2766 // tile12 = 0x00010002
duke@0 2767 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@0 2768 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@0 2769 //
duke@0 2770 void
duke@0 2771 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@0 2772 Node* size_in_bytes,
duke@0 2773 PhaseGVN* phase) {
duke@0 2774 Compile* C = phase->C;
duke@0 2775
duke@0 2776 assert(stores_are_sane(phase), "");
duke@0 2777 // Note: After this pass, they are not completely sane,
duke@0 2778 // since there may be some overlaps.
duke@0 2779
duke@0 2780 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@0 2781
duke@0 2782 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2783 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@0 2784 size_limit = MIN2(size_limit, ti_limit);
duke@0 2785 size_limit = align_size_up(size_limit, BytesPerLong);
duke@0 2786 int num_tiles = size_limit / BytesPerLong;
duke@0 2787
duke@0 2788 // allocate space for the tile map:
duke@0 2789 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@0 2790 jlong tiles_buf[small_len];
duke@0 2791 Node* nodes_buf[small_len];
duke@0 2792 jlong inits_buf[small_len];
duke@0 2793 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@0 2794 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2795 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@0 2796 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@0 2797 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@0 2798 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2799 // tiles: exact bitwise model of all primitive constants
duke@0 2800 // nodes: last constant-storing node subsumed into the tiles model
duke@0 2801 // inits: which bytes (in each tile) are touched by any initializations
duke@0 2802
duke@0 2803 //// Pass A: Fill in the tile model with any relevant stores.
duke@0 2804
duke@0 2805 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@0 2806 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@0 2807 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@0 2808 Node* zmem = zero_memory(); // initially zero memory state
duke@0 2809 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 2810 Node* st = in(i);
duke@0 2811 intptr_t st_off = get_store_offset(st, phase);
duke@0 2812
duke@0 2813 // Figure out the store's offset and constant value:
duke@0 2814 if (st_off < header_size) continue; //skip (ignore header)
duke@0 2815 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@0 2816 int st_size = st->as_Store()->memory_size();
duke@0 2817 if (st_off + st_size > size_limit) break;
duke@0 2818
duke@0 2819 // Record which bytes are touched, whether by constant or not.
duke@0 2820 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@0 2821 continue; // skip (strange store size)
duke@0 2822
duke@0 2823 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@0 2824 if (!val->singleton()) continue; //skip (non-con store)
duke@0 2825 BasicType type = val->basic_type();
duke@0 2826
duke@0 2827 jlong con = 0;
duke@0 2828 switch (type) {
duke@0 2829 case T_INT: con = val->is_int()->get_con(); break;
duke@0 2830 case T_LONG: con = val->is_long()->get_con(); break;
duke@0 2831 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@0 2832 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@0 2833 default: continue; //skip (odd store type)
duke@0 2834 }
duke@0 2835
duke@0 2836 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@0 2837 st->Opcode() == Op_StoreL) {
duke@0 2838 continue; // This StoreL is already optimal.
duke@0 2839 }
duke@0 2840
duke@0 2841 // Store down the constant.
duke@0 2842 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@0 2843
duke@0 2844 intptr_t j = st_off >> LogBytesPerLong;
duke@0 2845
duke@0 2846 if (type == T_INT && st_size == BytesPerInt
duke@0 2847 && (st_off & BytesPerInt) == BytesPerInt) {
duke@0 2848 jlong lcon = tiles[j];
duke@0 2849 if (!Matcher::isSimpleConstant64(lcon) &&
duke@0 2850 st->Opcode() == Op_StoreI) {
duke@0 2851 // This StoreI is already optimal by itself.
duke@0 2852 jint* intcon = (jint*) &tiles[j];
duke@0 2853 intcon[1] = 0; // undo the store_constant()
duke@0 2854
duke@0 2855 // If the previous store is also optimal by itself, back up and
duke@0 2856 // undo the action of the previous loop iteration... if we can.
duke@0 2857 // But if we can't, just let the previous half take care of itself.
duke@0 2858 st = nodes[j];
duke@0 2859 st_off -= BytesPerInt;
duke@0 2860 con = intcon[0];
duke@0 2861 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@0 2862 assert(st_off >= header_size, "still ignoring header");
duke@0 2863 assert(get_store_offset(st, phase) == st_off, "must be");
duke@0 2864 assert(in(i-1) == zmem, "must be");
duke@0 2865 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@0 2866 assert(con == tcon->is_int()->get_con(), "must be");
duke@0 2867 // Undo the effects of the previous loop trip, which swallowed st:
duke@0 2868 intcon[0] = 0; // undo store_constant()
duke@0 2869 set_req(i-1, st); // undo set_req(i, zmem)
duke@0 2870 nodes[j] = NULL; // undo nodes[j] = st
duke@0 2871 --old_subword; // undo ++old_subword
duke@0 2872 }
duke@0 2873 continue; // This StoreI is already optimal.
duke@0 2874 }
duke@0 2875 }
duke@0 2876
duke@0 2877 // This store is not needed.
duke@0 2878 set_req(i, zmem);
duke@0 2879 nodes[j] = st; // record for the moment
duke@0 2880 if (st_size < BytesPerLong) // something has changed
duke@0 2881 ++old_subword; // includes int/float, but who's counting...
duke@0 2882 else ++old_long;
duke@0 2883 }
duke@0 2884
duke@0 2885 if ((old_subword + old_long) == 0)
duke@0 2886 return; // nothing more to do
duke@0 2887
duke@0 2888 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@0 2889 // Be sure to insert them before overlapping non-constant stores.
duke@0 2890 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@0 2891 for (int j = 0; j < num_tiles; j++) {
duke@0 2892 jlong con = tiles[j];
duke@0 2893 jlong init = inits[j];
duke@0 2894 if (con == 0) continue;
duke@0 2895 jint con0, con1; // split the constant, address-wise
duke@0 2896 jint init0, init1; // split the init map, address-wise
duke@0 2897 { union { jlong con; jint intcon[2]; } u;
duke@0 2898 u.con = con;
duke@0 2899 con0 = u.intcon[0];
duke@0 2900 con1 = u.intcon[1];
duke@0 2901 u.con = init;
duke@0 2902 init0 = u.intcon[0];
duke@0 2903 init1 = u.intcon[1];
duke@0 2904 }
duke@0 2905
duke@0 2906 Node* old = nodes[j];
duke@0 2907 assert(old != NULL, "need the prior store");
duke@0 2908 intptr_t offset = (j * BytesPerLong);
duke@0 2909
duke@0 2910 bool split = !Matcher::isSimpleConstant64(con);
duke@0 2911
duke@0 2912 if (offset < header_size) {
duke@0 2913 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@0 2914 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@0 2915 split = true; // only the second word counts
duke@0 2916 // Example: int a[] = { 42 ... }
duke@0 2917 } else if (con0 == 0 && init0 == -1) {
duke@0 2918 split = true; // first word is covered by full inits
duke@0 2919 // Example: int a[] = { ... foo(), 42 ... }
duke@0 2920 } else if (con1 == 0 && init1 == -1) {
duke@0 2921 split = true; // second word is covered by full inits
duke@0 2922 // Example: int a[] = { ... 42, foo() ... }
duke@0 2923 }
duke@0 2924
duke@0 2925 // Here's a case where init0 is neither 0 nor -1:
duke@0 2926 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@0 2927 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@0 2928 // In this case the tile is not split; it is (jlong)42.
duke@0 2929 // The big tile is stored down, and then the foo() value is inserted.
duke@0 2930 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@0 2931
duke@0 2932 Node* ctl = old->in(MemNode::Control);
duke@0 2933 Node* adr = make_raw_address(offset, phase);
duke@0 2934 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 2935
duke@0 2936 // One or two coalesced stores to plop down.
duke@0 2937 Node* st[2];
duke@0 2938 intptr_t off[2];
duke@0 2939 int nst = 0;
duke@0 2940 if (!split) {
duke@0 2941 ++new_long;
duke@0 2942 off[nst] = offset;
coleenp@113 2943 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 2944 phase->longcon(con), T_LONG);
duke@0 2945 } else {
duke@0 2946 // Omit either if it is a zero.
duke@0 2947 if (con0 != 0) {
duke@0 2948 ++new_int;
duke@0 2949 off[nst] = offset;
coleenp@113 2950 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 2951 phase->intcon(con0), T_INT);
duke@0 2952 }
duke@0 2953 if (con1 != 0) {
duke@0 2954 ++new_int;
duke@0 2955 offset += BytesPerInt;
duke@0 2956 adr = make_raw_address(offset, phase);
duke@0 2957 off[nst] = offset;
coleenp@113 2958 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 2959 phase->intcon(con1), T_INT);
duke@0 2960 }
duke@0 2961 }
duke@0 2962
duke@0 2963 // Insert second store first, then the first before the second.
duke@0 2964 // Insert each one just before any overlapping non-constant stores.
duke@0 2965 while (nst > 0) {
duke@0 2966 Node* st1 = st[--nst];
duke@0 2967 C->copy_node_notes_to(st1, old);
duke@0 2968 st1 = phase->transform(st1);
duke@0 2969 offset = off[nst];
duke@0 2970 assert(offset >= header_size, "do not smash header");
duke@0 2971 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@0 2972 guarantee(ins_idx != 0, "must re-insert constant store");
duke@0 2973 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@0 2974 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@0 2975 set_req(--ins_idx, st1);
duke@0 2976 else
duke@0 2977 ins_req(ins_idx, st1);
duke@0 2978 }
duke@0 2979 }
duke@0 2980
duke@0 2981 if (PrintCompilation && WizardMode)
duke@0 2982 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@0 2983 old_subword, old_long, new_int, new_long);
duke@0 2984 if (C->log() != NULL)
duke@0 2985 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@0 2986 old_subword, old_long, new_int, new_long);
duke@0 2987
duke@0 2988 // Clean up any remaining occurrences of zmem:
duke@0 2989 remove_extra_zeroes();
duke@0 2990 }
duke@0 2991
duke@0 2992 // Explore forward from in(start) to find the first fully initialized
duke@0 2993 // word, and return its offset. Skip groups of subword stores which
duke@0 2994 // together initialize full words. If in(start) is itself part of a
duke@0 2995 // fully initialized word, return the offset of in(start). If there
duke@0 2996 // are no following full-word stores, or if something is fishy, return
duke@0 2997 // a negative value.
duke@0 2998 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@0 2999 int int_map = 0;
duke@0 3000 intptr_t int_map_off = 0;
duke@0 3001 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@0 3002
duke@0 3003 for (uint i = start, limit = req(); i < limit; i++) {
duke@0 3004 Node* st = in(i);
duke@0 3005
duke@0 3006 intptr_t st_off = get_store_offset(st, phase);
duke@0 3007 if (st_off < 0) break; // return conservative answer
duke@0 3008
duke@0 3009 int st_size = st->as_Store()->memory_size();
duke@0 3010 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@0 3011 return st_off; // we found a complete word init
duke@0 3012 }
duke@0 3013
duke@0 3014 // update the map:
duke@0 3015
duke@0 3016 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@0 3017 if (this_int_off != int_map_off) {
duke@0 3018 // reset the map:
duke@0 3019 int_map = 0;
duke@0 3020 int_map_off = this_int_off;
duke@0 3021 }
duke@0 3022
duke@0 3023 int subword_off = st_off - this_int_off;
duke@0 3024 int_map |= right_n_bits(st_size) << subword_off;
duke@0 3025 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@0 3026 return this_int_off; // we found a complete word init
duke@0 3027 }
duke@0 3028
duke@0 3029 // Did this store hit or cross the word boundary?
duke@0 3030 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@0 3031 if (next_int_off == this_int_off + BytesPerInt) {
duke@0 3032 // We passed the current int, without fully initializing it.
duke@0 3033 int_map_off = next_int_off;
duke@0 3034 int_map >>= BytesPerInt;
duke@0 3035 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@0 3036 // We passed the current and next int.
duke@0 3037 return this_int_off + BytesPerInt;
duke@0 3038 }
duke@0 3039 }
duke@0 3040
duke@0 3041 return -1;
duke@0 3042 }
duke@0 3043
duke@0 3044
duke@0 3045 // Called when the associated AllocateNode is expanded into CFG.
duke@0 3046 // At this point, we may perform additional optimizations.
duke@0 3047 // Linearize the stores by ascending offset, to make memory
duke@0 3048 // activity as coherent as possible.
duke@0 3049 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@0 3050 intptr_t header_size,
duke@0 3051 Node* size_in_bytes,
duke@0 3052 PhaseGVN* phase) {
duke@0 3053 assert(!is_complete(), "not already complete");
duke@0 3054 assert(stores_are_sane(phase), "");
duke@0 3055 assert(allocation() != NULL, "must be present");
duke@0 3056
duke@0 3057 remove_extra_zeroes();
duke@0 3058
duke@0 3059 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@0 3060 // reduce instruction count for common initialization patterns
duke@0 3061 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@0 3062
duke@0 3063 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3064 Node* inits = zmem; // accumulating a linearized chain of inits
duke@0 3065 #ifdef ASSERT
coleenp@113 3066 intptr_t first_offset = allocation()->minimum_header_size();
coleenp@113 3067 intptr_t last_init_off = first_offset; // previous init offset
coleenp@113 3068 intptr_t last_init_end = first_offset; // previous init offset+size
coleenp@113 3069 intptr_t last_tile_end = first_offset; // previous tile offset+size
duke@0 3070 #endif
duke@0 3071 intptr_t zeroes_done = header_size;
duke@0 3072
duke@0 3073 bool do_zeroing = true; // we might give up if inits are very sparse
duke@0 3074 int big_init_gaps = 0; // how many large gaps have we seen?
duke@0 3075
duke@0 3076 if (ZeroTLAB) do_zeroing = false;
duke@0 3077 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@0 3078
duke@0 3079 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3080 Node* st = in(i);
duke@0 3081 intptr_t st_off = get_store_offset(st, phase);
duke@0 3082 if (st_off < 0)
duke@0 3083 break; // unknown junk in the inits
duke@0 3084 if (st->in(MemNode::Memory) != zmem)
duke@0 3085 break; // complicated store chains somehow in list
duke@0 3086
duke@0 3087 int st_size = st->as_Store()->memory_size();
duke@0 3088 intptr_t next_init_off = st_off + st_size;
duke@0 3089
duke@0 3090 if (do_zeroing && zeroes_done < next_init_off) {
duke@0 3091 // See if this store needs a zero before it or under it.
duke@0 3092 intptr_t zeroes_needed = st_off;
duke@0 3093
duke@0 3094 if (st_size < BytesPerInt) {
duke@0 3095 // Look for subword stores which only partially initialize words.
duke@0 3096 // If we find some, we must lay down some word-level zeroes first,
duke@0 3097 // underneath the subword stores.
duke@0 3098 //
duke@0 3099 // Examples:
duke@0 3100 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
duke@0 3101 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
duke@0 3102 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
duke@0 3103 //
duke@0 3104 // Note: coalesce_subword_stores may have already done this,
duke@0 3105 // if it was prompted by constant non-zero subword initializers.
duke@0 3106 // But this case can still arise with non-constant stores.
duke@0 3107
duke@0 3108 intptr_t next_full_store = find_next_fullword_store(i, phase);
duke@0 3109
duke@0 3110 // In the examples above:
duke@0 3111 // in(i) p q r s x y z
duke@0 3112 // st_off 12 13 14 15 12 13 14
duke@0 3113 // st_size 1 1 1 1 1 1 1
duke@0 3114 // next_full_s. 12 16 16 16 16 16 16
duke@0 3115 // z's_done 12 16 16 16 12 16 12
duke@0 3116 // z's_needed 12 16 16 16 16 16 16
duke@0 3117 // zsize 0 0 0 0 4 0 4
duke@0 3118 if (next_full_store < 0) {
duke@0 3119 // Conservative tack: Zero to end of current word.
duke@0 3120 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
duke@0 3121 } else {
duke@0 3122 // Zero to beginning of next fully initialized word.
duke@0 3123 // Or, don't zero at all, if we are already in that word.
duke@0 3124 assert(next_full_store >= zeroes_needed, "must go forward");
duke@0 3125 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
duke@0 3126 zeroes_needed = next_full_store;
duke@0 3127 }
duke@0 3128 }
duke@0 3129
duke@0 3130 if (zeroes_needed > zeroes_done) {
duke@0 3131 intptr_t zsize = zeroes_needed - zeroes_done;
duke@0 3132 // Do some incremental zeroing on rawmem, in parallel with inits.
duke@0 3133 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3134 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3135 zeroes_done, zeroes_needed,
duke@0 3136 phase);
duke@0 3137 zeroes_done = zeroes_needed;
duke@0 3138 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
duke@0 3139 do_zeroing = false; // leave the hole, next time
duke@0 3140 }
duke@0 3141 }
duke@0 3142
duke@0 3143 // Collect the store and move on:
duke@0 3144 st->set_req(MemNode::Memory, inits);
duke@0 3145 inits = st; // put it on the linearized chain
duke@0 3146 set_req(i, zmem); // unhook from previous position
duke@0 3147
duke@0 3148 if (zeroes_done == st_off)
duke@0 3149 zeroes_done = next_init_off;
duke@0 3150
duke@0 3151 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
duke@0 3152
duke@0 3153 #ifdef ASSERT
duke@0 3154 // Various order invariants. Weaker than stores_are_sane because
duke@0 3155 // a large constant tile can be filled in by smaller non-constant stores.
duke@0 3156 assert(st_off >= last_init_off, "inits do not reverse");
duke@0 3157 last_init_off = st_off;
duke@0 3158 const Type* val = NULL;
duke@0 3159 if (st_size >= BytesPerInt &&
duke@0 3160 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
duke@0 3161 (int)val->basic_type() < (int)T_OBJECT) {
duke@0 3162 assert(st_off >= last_tile_end, "tiles do not overlap");
duke@0 3163 assert(st_off >= last_init_end, "tiles do not overwrite inits");
duke@0 3164 last_tile_end = MAX2(last_tile_end, next_init_off);
duke@0 3165 } else {
duke@0 3166 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
duke@0 3167 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
duke@0 3168 assert(st_off >= last_init_end, "inits do not overlap");
duke@0 3169 last_init_end = next_init_off; // it's a non-tile
duke@0 3170 }
duke@0 3171 #endif //ASSERT
duke@0 3172 }
duke@0 3173
duke@0 3174 remove_extra_zeroes(); // clear out all the zmems left over
duke@0 3175 add_req(inits);
duke@0 3176
duke@0 3177 if (!ZeroTLAB) {
duke@0 3178 // If anything remains to be zeroed, zero it all now.
duke@0 3179 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3180 // if it is the last unused 4 bytes of an instance, forget about it
duke@0 3181 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
duke@0 3182 if (zeroes_done + BytesPerLong >= size_limit) {
duke@0 3183 assert(allocation() != NULL, "");
duke@0 3184 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
duke@0 3185 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
duke@0 3186 if (zeroes_done == k->layout_helper())
duke@0 3187 zeroes_done = size_limit;
duke@0 3188 }
duke@0 3189 if (zeroes_done < size_limit) {
duke@0 3190 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3191 zeroes_done, size_in_bytes, phase);
duke@0 3192 }
duke@0 3193 }
duke@0 3194
duke@0 3195 set_complete(phase);
duke@0 3196 return rawmem;
duke@0 3197 }
duke@0 3198
duke@0 3199
duke@0 3200 #ifdef ASSERT
duke@0 3201 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
duke@0 3202 if (is_complete())
duke@0 3203 return true; // stores could be anything at this point
coleenp@113 3204 assert(allocation() != NULL, "must be present");
coleenp@113 3205 intptr_t last_off = allocation()->minimum_header_size();
duke@0 3206 for (uint i = InitializeNode::RawStores; i < req(); i++) {
duke@0 3207 Node* st = in(i);
duke@0 3208 intptr_t st_off = get_store_offset(st, phase);
duke@0 3209 if (st_off < 0) continue; // ignore dead garbage
duke@0 3210 if (last_off > st_off) {
duke@0 3211 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
duke@0 3212 this->dump(2);
duke@0 3213 assert(false, "ascending store offsets");
duke@0 3214 return false;
duke@0 3215 }
duke@0 3216 last_off = st_off + st->as_Store()->memory_size();
duke@0 3217 }
duke@0 3218 return true;
duke@0 3219 }
duke@0 3220 #endif //ASSERT
duke@0 3221
duke@0 3222
duke@0 3223
duke@0 3224
duke@0 3225 //============================MergeMemNode=====================================
duke@0 3226 //
duke@0 3227 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
duke@0 3228 // contributing store or call operations. Each contributor provides the memory
duke@0 3229 // state for a particular "alias type" (see Compile::alias_type). For example,
duke@0 3230 // if a MergeMem has an input X for alias category #6, then any memory reference
duke@0 3231 // to alias category #6 may use X as its memory state input, as an exact equivalent
duke@0 3232 // to using the MergeMem as a whole.
duke@0 3233 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
duke@0 3234 //
duke@0 3235 // (Here, the <N> notation gives the index of the relevant adr_type.)
duke@0 3236 //
duke@0 3237 // In one special case (and more cases in the future), alias categories overlap.
duke@0 3238 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
duke@0 3239 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
duke@0 3240 // it is exactly equivalent to that state W:
duke@0 3241 // MergeMem(<Bot>: W) <==> W
duke@0 3242 //
duke@0 3243 // Usually, the merge has more than one input. In that case, where inputs
duke@0 3244 // overlap (i.e., one is Bot), the narrower alias type determines the memory
duke@0 3245 // state for that type, and the wider alias type (Bot) fills in everywhere else:
duke@0 3246 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
duke@0 3247 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
duke@0 3248 //
duke@0 3249 // A merge can take a "wide" memory state as one of its narrow inputs.
duke@0 3250 // This simply means that the merge observes out only the relevant parts of
duke@0 3251 // the wide input. That is, wide memory states arriving at narrow merge inputs
duke@0 3252 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
duke@0 3253 //
duke@0 3254 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
duke@0 3255 // and that memory slices "leak through":
duke@0 3256 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
duke@0 3257 //
duke@0 3258 // But, in such a cascade, repeated memory slices can "block the leak":
duke@0 3259 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
duke@0 3260 //
duke@0 3261 // In the last example, Y is not part of the combined memory state of the
duke@0 3262 // outermost MergeMem. The system must, of course, prevent unschedulable
duke@0 3263 // memory states from arising, so you can be sure that the state Y is somehow
duke@0 3264 // a precursor to state Y'.
duke@0 3265 //
duke@0 3266 //
duke@0 3267 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
duke@0 3268 // of each MergeMemNode array are exactly the numerical alias indexes, including
duke@0 3269 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
duke@0 3270 // Compile::alias_type (and kin) produce and manage these indexes.
duke@0 3271 //
duke@0 3272 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
duke@0 3273 // (Note that this provides quick access to the top node inside MergeMem methods,
duke@0 3274 // without the need to reach out via TLS to Compile::current.)
duke@0 3275 //
duke@0 3276 // As a consequence of what was just described, a MergeMem that represents a full
duke@0 3277 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
duke@0 3278 // containing all alias categories.
duke@0 3279 //
duke@0 3280 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
duke@0 3281 //
duke@0 3282 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
duke@0 3283 // a memory state for the alias type <N>, or else the top node, meaning that
duke@0 3284 // there is no particular input for that alias type. Note that the length of
duke@0 3285 // a MergeMem is variable, and may be extended at any time to accommodate new
duke@0 3286 // memory states at larger alias indexes. When merges grow, they are of course
duke@0 3287 // filled with "top" in the unused in() positions.
duke@0 3288 //
duke@0 3289 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
duke@0 3290 // (Top was chosen because it works smoothly with passes like GCM.)
duke@0 3291 //
duke@0 3292 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
duke@0 3293 // the type of random VM bits like TLS references.) Since it is always the
duke@0 3294 // first non-Bot memory slice, some low-level loops use it to initialize an
duke@0 3295 // index variable: for (i = AliasIdxRaw; i < req(); i++).
duke@0 3296 //
duke@0 3297 //
duke@0 3298 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
duke@0 3299 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
duke@0 3300 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
duke@0 3301 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
duke@0 3302 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
duke@0 3303 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
duke@0 3304 //
duke@0 3305 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
duke@0 3306 // really that different from the other memory inputs. An abbreviation called
duke@0 3307 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
duke@0 3308 //
duke@0 3309 //
duke@0 3310 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
duke@0 3311 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
duke@0 3312 // that "emerges though" the base memory will be marked as excluding the alias types
duke@0 3313 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
duke@0 3314 //
duke@0 3315 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
duke@0 3316 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
duke@0 3317 //
duke@0 3318 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
duke@0 3319 // (It is currently unimplemented.) As you can see, the resulting merge is
duke@0 3320 // actually a disjoint union of memory states, rather than an overlay.
duke@0 3321 //
duke@0 3322
duke@0 3323 //------------------------------MergeMemNode-----------------------------------
duke@0 3324 Node* MergeMemNode::make_empty_memory() {
duke@0 3325 Node* empty_memory = (Node*) Compile::current()->top();
duke@0 3326 assert(empty_memory->is_top(), "correct sentinel identity");
duke@0 3327 return empty_memory;
duke@0 3328 }
duke@0 3329
duke@0 3330 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
duke@0 3331 init_class_id(Class_MergeMem);
duke@0 3332 // all inputs are nullified in Node::Node(int)
duke@0 3333 // set_input(0, NULL); // no control input
duke@0 3334
duke@0 3335 // Initialize the edges uniformly to top, for starters.
duke@0 3336 Node* empty_mem = make_empty_memory();
duke@0 3337 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
duke@0 3338 init_req(i,empty_mem);
duke@0 3339 }
duke@0 3340 assert(empty_memory() == empty_mem, "");
duke@0 3341
duke@0 3342 if( new_base != NULL && new_base->is_MergeMem() ) {
duke@0 3343 MergeMemNode* mdef = new_base->as_MergeMem();
duke@0 3344 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
duke@0 3345 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
duke@0 3346 mms.set_memory(mms.memory2());
duke@0 3347 }
duke@0 3348 assert(base_memory() == mdef->base_memory(), "");
duke@0 3349 } else {
duke@0 3350 set_base_memory(new_base);
duke@0 3351 }
duke@0 3352 }
duke@0 3353
duke@0 3354 // Make a new, untransformed MergeMem with the same base as 'mem'.
duke@0 3355 // If mem is itself a MergeMem, populate the result with the same edges.
duke@0 3356 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
duke@0 3357 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
duke@0 3358 }
duke@0 3359
duke@0 3360 //------------------------------cmp--------------------------------------------
duke@0 3361 uint MergeMemNode::hash() const { return NO_HASH; }
duke@0 3362 uint MergeMemNode::cmp( const Node &n ) const {
duke@0 3363 return (&n == this); // Always fail except on self
duke@0 3364 }
duke@0 3365
duke@0 3366 //------------------------------Identity---------------------------------------
duke@0 3367 Node* MergeMemNode::Identity(PhaseTransform *phase) {
duke@0 3368 // Identity if this merge point does not record any interesting memory
duke@0 3369 // disambiguations.
duke@0 3370 Node* base_mem = base_memory();
duke@0 3371 Node* empty_mem = empty_memory();
duke@0 3372 if (base_mem != empty_mem) { // Memory path is not dead?
duke@0 3373 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3374 Node* mem = in(i);
duke@0 3375 if (mem != empty_mem && mem != base_mem) {
duke@0 3376 return this; // Many memory splits; no change
duke@0 3377 }
duke@0 3378 }
duke@0 3379 }
duke@0 3380 return base_mem; // No memory splits; ID on the one true input
duke@0 3381 }
duke@0 3382
duke@0 3383 //------------------------------Ideal------------------------------------------
duke@0 3384 // This method is invoked recursively on chains of MergeMem nodes
duke@0 3385 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 3386 // Remove chain'd MergeMems
duke@0 3387 //
duke@0 3388 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
duke@0 3389 // relative to the "in(Bot)". Since we are patching both at the same time,
duke@0 3390 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
duke@0 3391 // but rewrite each "in(i)" relative to the new "in(Bot)".
duke@0 3392 Node *progress = NULL;
duke@0 3393
duke@0 3394
duke@0 3395 Node* old_base = base_memory();
duke@0 3396 Node* empty_mem = empty_memory();
duke@0 3397 if (old_base == empty_mem)
duke@0 3398 return NULL; // Dead memory path.
duke@0 3399
duke@0 3400 MergeMemNode* old_mbase;
duke@0 3401 if (old_base != NULL && old_base->is_MergeMem())
duke@0 3402 old_mbase = old_base->as_MergeMem();
duke@0 3403 else
duke@0 3404 old_mbase = NULL;
duke@0 3405 Node* new_base = old_base;
duke@0 3406
duke@0 3407 // simplify stacked MergeMems in base memory
duke@0 3408 if (old_mbase) new_base = old_mbase->base_memory();
duke@0 3409
duke@0 3410 // the base memory might contribute new slices beyond my req()
duke@0 3411 if (old_mbase) grow_to_match(old_mbase);
duke@0 3412
duke@0 3413 // Look carefully at the base node if it is a phi.
duke@0 3414 PhiNode* phi_base;
duke@0 3415 if (new_base != NULL && new_base->is_Phi())
duke@0 3416 phi_base = new_base->as_Phi();
duke@0 3417 else
duke@0 3418 phi_base = NULL;
duke@0 3419
duke@0 3420 Node* phi_reg = NULL;
duke@0 3421 uint phi_len = (uint)-1;
duke@0 3422 if (phi_base != NULL && !phi_base->is_copy()) {
duke@0 3423 // do not examine phi if degraded to a copy
duke@0 3424 phi_reg = phi_base->region();
duke@0 3425 phi_len = phi_base->req();
duke@0 3426 // see if the phi is unfinished
duke@0 3427 for (uint i = 1; i < phi_len; i++) {
duke@0 3428 if (phi_base->in(i) == NULL) {
duke@0 3429 // incomplete phi; do not look at it yet!
duke@0 3430 phi_reg = NULL;
duke@0 3431 phi_len = (uint)-1;
duke@0 3432 break;
duke@0 3433 }
duke@0 3434 }
duke@0 3435 }
duke@0 3436
duke@0 3437 // Note: We do not call verify_sparse on entry, because inputs
duke@0 3438 // can normalize to the base_memory via subsume_node or similar
duke@0 3439 // mechanisms. This method repairs that damage.
duke@0 3440
duke@0 3441 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
duke@0 3442
duke@0 3443 // Look at each slice.
duke@0 3444 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3445 Node* old_in = in(i);
duke@0 3446 // calculate the old memory value
duke@0 3447 Node* old_mem = old_in;
duke@0 3448 if (old_mem == empty_mem) old_mem = old_base;
duke@0 3449 assert(old_mem == memory_at(i), "");
duke@0 3450
duke@0 3451 // maybe update (reslice) the old memory value
duke@0 3452
duke@0 3453 // simplify stacked MergeMems
duke@0 3454 Node* new_mem = old_mem;
duke@0 3455 MergeMemNode* old_mmem;
duke@0 3456 if (old_mem != NULL && old_mem->is_MergeMem())
duke@0 3457 old_mmem = old_mem->as_MergeMem();
duke@0 3458 else
duke@0 3459 old_mmem = NULL;
duke@0 3460 if (old_mmem == this) {
duke@0 3461 // This can happen if loops break up and safepoints disappear.
duke@0 3462 // A merge of BotPtr (default) with a RawPtr memory derived from a
duke@0 3463 // safepoint can be rewritten to a merge of the same BotPtr with
duke@0 3464 // the BotPtr phi coming into the loop. If that phi disappears
duke@0 3465 // also, we can end up with a self-loop of the mergemem.
duke@0 3466 // In general, if loops degenerate and memory effects disappear,
duke@0 3467 // a mergemem can be left looking at itself. This simply means
duke@0 3468 // that the mergemem's default should be used, since there is
duke@0 3469 // no longer any apparent effect on this slice.
duke@0 3470 // Note: If a memory slice is a MergeMem cycle, it is unreachable
duke@0 3471 // from start. Update the input to TOP.
duke@0 3472 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
duke@0 3473 }
duke@0 3474 else if (old_mmem != NULL) {
duke@0 3475 new_mem = old_mmem->memory_at(i);
duke@0 3476 }
duke@0 3477 // else preceeding memory was not a MergeMem
duke@0 3478
duke@0 3479 // replace equivalent phis (unfortunately, they do not GVN together)
duke@0 3480 if (new_mem != NULL && new_mem != new_base &&
duke@0 3481 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
duke@0 3482 if (new_mem->is_Phi()) {
duke@0 3483 PhiNode* phi_mem = new_mem->as_Phi();
duke@0 3484 for (uint i = 1; i < phi_len; i++) {
duke@0 3485 if (phi_base->in(i) != phi_mem->in(i)) {
duke@0 3486 phi_mem = NULL;
duke@0 3487 break;
duke@0 3488 }
duke@0 3489 }
duke@0 3490 if (phi_mem != NULL) {
duke@0 3491 // equivalent phi nodes; revert to the def
duke@0 3492 new_mem = new_base;
duke@0 3493 }
duke@0 3494 }
duke@0 3495 }
duke@0 3496
duke@0 3497 // maybe store down a new value
duke@0 3498 Node* new_in = new_mem;
duke@0 3499 if (new_in == new_base) new_in = empty_mem;
duke@0 3500
duke@0 3501 if (new_in != old_in) {
duke@0 3502 // Warning: Do not combine this "if" with the previous "if"
duke@0 3503 // A memory slice might have be be rewritten even if it is semantically
duke@0 3504 // unchanged, if the base_memory value has changed.
duke@0 3505 set_req(i, new_in);
duke@0 3506 progress = this; // Report progress
duke@0 3507 }
duke@0 3508 }
duke@0 3509
duke@0 3510 if (new_base != old_base) {
duke@0 3511 set_req(Compile::AliasIdxBot, new_base);
duke@0 3512 // Don't use set_base_memory(new_base), because we need to update du.
duke@0 3513 assert(base_memory() == new_base, "");
duke@0 3514 progress = this;
duke@0 3515 }
duke@0 3516
duke@0 3517 if( base_memory() == this ) {
duke@0 3518 // a self cycle indicates this memory path is dead
duke@0 3519 set_req(Compile::AliasIdxBot, empty_mem);
duke@0 3520 }
duke@0 3521
duke@0 3522 // Resolve external cycles by calling Ideal on a MergeMem base_memory
duke@0 3523 // Recursion must occur after the self cycle check above
duke@0 3524 if( base_memory()->is_MergeMem() ) {
duke@0 3525 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
duke@0 3526 Node *m = phase->transform(new_mbase); // Rollup any cycles
duke@0 3527 if( m != NULL && (m->is_top() ||
duke@0 3528 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
duke@0 3529 // propagate rollup of dead cycle to self
duke@0 3530 set_req(Compile::AliasIdxBot, empty_mem);
duke@0 3531 }
duke@0 3532 }
duke@0 3533
duke@0 3534 if( base_memory() == empty_mem ) {
duke@0 3535 progress = this;
duke@0 3536 // Cut inputs during Parse phase only.
duke@0 3537 // During Optimize phase a dead MergeMem node will be subsumed by Top.
duke@0 3538 if( !can_reshape ) {
duke@0 3539 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3540 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
duke@0 3541 }
duke@0 3542 }
duke@0 3543 }
duke@0 3544
duke@0 3545 if( !progress && base_memory()->is_Phi() && can_reshape ) {
duke@0 3546 // Check if PhiNode::Ideal's "Split phis through memory merges"
duke@0 3547 // transform should be attempted. Look for this->phi->this cycle.
duke@0 3548 uint merge_width = req();
duke@0 3549 if (merge_width > Compile::AliasIdxRaw) {
duke@0 3550 PhiNode* phi = base_memory()->as_Phi();
duke@0 3551 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
duke@0 3552 if (phi->in(i) == this) {
duke@0 3553 phase->is_IterGVN()->_worklist.push(phi);
duke@0 3554 break;
duke@0 3555 }
duke@0 3556 }
duke@0 3557 }
duke@0 3558 }
duke@0 3559
kvn@64 3560 assert(progress || verify_sparse(), "please, no dups of base");
duke@0 3561 return progress;
duke@0 3562 }
duke@0 3563
duke@0 3564 //-------------------------set_base_memory-------------------------------------
duke@0 3565 void MergeMemNode::set_base_memory(Node *new_base) {
duke@0 3566 Node* empty_mem = empty_memory();
duke@0 3567 set_req(Compile::AliasIdxBot, new_base);
duke@0 3568 assert(memory_at(req()) == new_base, "must set default memory");
duke@0 3569 // Clear out other occurrences of new_base:
duke@0 3570 if (new_base != empty_mem) {
duke@0 3571 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3572 if (in(i) == new_base) set_req(i, empty_mem);
duke@0 3573 }
duke@0 3574 }
duke@0 3575 }
duke@0 3576
duke@0 3577 //------------------------------out_RegMask------------------------------------
duke@0 3578 const RegMask &MergeMemNode::out_RegMask() const {
duke@0 3579 return RegMask::Empty;
duke@0 3580 }
duke@0 3581
duke@0 3582 //------------------------------dump_spec--------------------------------------
duke@0 3583 #ifndef PRODUCT
duke@0 3584 void MergeMemNode::dump_spec(outputStream *st) const {
duke@0 3585 st->print(" {");
duke@0 3586 Node* base_mem = base_memory();
duke@0 3587 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
duke@0 3588 Node* mem = memory_at(i);
duke@0 3589 if (mem == base_mem) { st->print(" -"); continue; }
duke@0 3590 st->print( " N%d:", mem->_idx );
duke@0 3591 Compile::current()->get_adr_type(i)->dump_on(st);
duke@0 3592 }
duke@0 3593 st->print(" }");
duke@0 3594 }
duke@0 3595 #endif // !PRODUCT
duke@0 3596
duke@0 3597
duke@0 3598 #ifdef ASSERT
duke@0 3599 static bool might_be_same(Node* a, Node* b) {
duke@0 3600 if (a == b) return true;
duke@0 3601 if (!(a->is_Phi() || b->is_Phi())) return false;
duke@0 3602 // phis shift around during optimization
duke@0 3603 return true; // pretty stupid...
duke@0 3604 }
duke@0 3605
duke@0 3606 // verify a narrow slice (either incoming or outgoing)
duke@0 3607 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
duke@0 3608 if (!VerifyAliases) return; // don't bother to verify unless requested
duke@0 3609 if (is_error_reported()) return; // muzzle asserts when debugging an error
duke@0 3610 if (Node::in_dump()) return; // muzzle asserts when printing
duke@0 3611 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
duke@0 3612 assert(n != NULL, "");
duke@0 3613 // Elide intervening MergeMem's
duke@0 3614 while (n->is_MergeMem()) {