annotate src/share/vm/opto/memnode.cpp @ 366:8261ee795323

6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int") Summary: insert CastII nodes to narrow type of load_array_length() node Reviewed-by: never, kvn
author rasbold
date Wed, 17 Sep 2008 08:29:17 -0700
parents ab075d07f1ba
children a1980da045cc
rev   line source
duke@0 1 /*
xdono@196 2 * Copyright 1997-2008 Sun Microsystems, Inc. All Rights Reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
duke@0 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
duke@0 20 * CA 95054 USA or visit www.sun.com if you need additional information or
duke@0 21 * have any questions.
duke@0 22 *
duke@0 23 */
duke@0 24
duke@0 25 // Portions of code courtesy of Clifford Click
duke@0 26
duke@0 27 // Optimization - Graph Style
duke@0 28
duke@0 29 #include "incls/_precompiled.incl"
duke@0 30 #include "incls/_memnode.cpp.incl"
duke@0 31
kvn@74 32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
kvn@74 33
duke@0 34 //=============================================================================
duke@0 35 uint MemNode::size_of() const { return sizeof(*this); }
duke@0 36
duke@0 37 const TypePtr *MemNode::adr_type() const {
duke@0 38 Node* adr = in(Address);
duke@0 39 const TypePtr* cross_check = NULL;
duke@0 40 DEBUG_ONLY(cross_check = _adr_type);
duke@0 41 return calculate_adr_type(adr->bottom_type(), cross_check);
duke@0 42 }
duke@0 43
duke@0 44 #ifndef PRODUCT
duke@0 45 void MemNode::dump_spec(outputStream *st) const {
duke@0 46 if (in(Address) == NULL) return; // node is dead
duke@0 47 #ifndef ASSERT
duke@0 48 // fake the missing field
duke@0 49 const TypePtr* _adr_type = NULL;
duke@0 50 if (in(Address) != NULL)
duke@0 51 _adr_type = in(Address)->bottom_type()->isa_ptr();
duke@0 52 #endif
duke@0 53 dump_adr_type(this, _adr_type, st);
duke@0 54
duke@0 55 Compile* C = Compile::current();
duke@0 56 if( C->alias_type(_adr_type)->is_volatile() )
duke@0 57 st->print(" Volatile!");
duke@0 58 }
duke@0 59
duke@0 60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
duke@0 61 st->print(" @");
duke@0 62 if (adr_type == NULL) {
duke@0 63 st->print("NULL");
duke@0 64 } else {
duke@0 65 adr_type->dump_on(st);
duke@0 66 Compile* C = Compile::current();
duke@0 67 Compile::AliasType* atp = NULL;
duke@0 68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
duke@0 69 if (atp == NULL)
duke@0 70 st->print(", idx=?\?;");
duke@0 71 else if (atp->index() == Compile::AliasIdxBot)
duke@0 72 st->print(", idx=Bot;");
duke@0 73 else if (atp->index() == Compile::AliasIdxTop)
duke@0 74 st->print(", idx=Top;");
duke@0 75 else if (atp->index() == Compile::AliasIdxRaw)
duke@0 76 st->print(", idx=Raw;");
duke@0 77 else {
duke@0 78 ciField* field = atp->field();
duke@0 79 if (field) {
duke@0 80 st->print(", name=");
duke@0 81 field->print_name_on(st);
duke@0 82 }
duke@0 83 st->print(", idx=%d;", atp->index());
duke@0 84 }
duke@0 85 }
duke@0 86 }
duke@0 87
duke@0 88 extern void print_alias_types();
duke@0 89
duke@0 90 #endif
duke@0 91
kvn@74 92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 93 const TypeOopPtr *tinst = t_adr->isa_oopptr();
kvn@223 94 if (tinst == NULL || !tinst->is_known_instance_field())
kvn@74 95 return mchain; // don't try to optimize non-instance types
kvn@74 96 uint instance_id = tinst->instance_id();
kvn@253 97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
kvn@74 98 Node *prev = NULL;
kvn@74 99 Node *result = mchain;
kvn@74 100 while (prev != result) {
kvn@74 101 prev = result;
kvn@253 102 if (result == start_mem)
kvn@253 103 break; // hit one of our sentinals
kvn@74 104 // skip over a call which does not affect this memory slice
kvn@74 105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
kvn@74 106 Node *proj_in = result->in(0);
kvn@253 107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
kvn@253 108 break; // hit one of our sentinals
kvn@253 109 } else if (proj_in->is_Call()) {
kvn@74 110 CallNode *call = proj_in->as_Call();
kvn@74 111 if (!call->may_modify(t_adr, phase)) {
kvn@74 112 result = call->in(TypeFunc::Memory);
kvn@74 113 }
kvn@74 114 } else if (proj_in->is_Initialize()) {
kvn@74 115 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
kvn@74 116 // Stop if this is the initialization for the object instance which
kvn@74 117 // which contains this memory slice, otherwise skip over it.
kvn@74 118 if (alloc != NULL && alloc->_idx != instance_id) {
kvn@74 119 result = proj_in->in(TypeFunc::Memory);
kvn@74 120 }
kvn@74 121 } else if (proj_in->is_MemBar()) {
kvn@74 122 result = proj_in->in(TypeFunc::Memory);
kvn@253 123 } else {
kvn@253 124 assert(false, "unexpected projection");
kvn@74 125 }
kvn@74 126 } else if (result->is_MergeMem()) {
kvn@74 127 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
kvn@74 128 }
kvn@74 129 }
kvn@74 130 return result;
kvn@74 131 }
kvn@74 132
kvn@74 133 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 134 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
kvn@223 135 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
kvn@74 136 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@74 137 Node *result = mchain;
kvn@74 138 result = optimize_simple_memory_chain(result, t_adr, phase);
kvn@74 139 if (is_instance && igvn != NULL && result->is_Phi()) {
kvn@74 140 PhiNode *mphi = result->as_Phi();
kvn@74 141 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
kvn@74 142 const TypePtr *t = mphi->adr_type();
kvn@163 143 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
kvn@223 144 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
kvn@247 145 t->is_oopptr()->cast_to_exactness(true)
kvn@247 146 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
kvn@247 147 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
kvn@74 148 // clone the Phi with our address type
kvn@74 149 result = mphi->split_out_instance(t_adr, igvn);
kvn@74 150 } else {
kvn@74 151 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
kvn@74 152 }
kvn@74 153 }
kvn@74 154 return result;
kvn@74 155 }
kvn@74 156
kvn@64 157 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
kvn@64 158 uint alias_idx = phase->C->get_alias_index(tp);
kvn@64 159 Node *mem = mmem;
kvn@64 160 #ifdef ASSERT
kvn@64 161 {
kvn@64 162 // Check that current type is consistent with the alias index used during graph construction
kvn@64 163 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
kvn@64 164 bool consistent = adr_check == NULL || adr_check->empty() ||
kvn@64 165 phase->C->must_alias(adr_check, alias_idx );
kvn@64 166 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
kvn@64 167 if( !consistent && adr_check != NULL && !adr_check->empty() &&
rasbold@169 168 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
kvn@64 169 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
kvn@64 170 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
kvn@64 171 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
kvn@64 172 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
kvn@64 173 // don't assert if it is dead code.
kvn@64 174 consistent = true;
kvn@64 175 }
kvn@64 176 if( !consistent ) {
kvn@64 177 st->print("alias_idx==%d, adr_check==", alias_idx);
kvn@64 178 if( adr_check == NULL ) {
kvn@64 179 st->print("NULL");
kvn@64 180 } else {
kvn@64 181 adr_check->dump();
kvn@64 182 }
kvn@64 183 st->cr();
kvn@64 184 print_alias_types();
kvn@64 185 assert(consistent, "adr_check must match alias idx");
kvn@64 186 }
kvn@64 187 }
kvn@64 188 #endif
kvn@64 189 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
kvn@64 190 // means an array I have not precisely typed yet. Do not do any
kvn@64 191 // alias stuff with it any time soon.
kvn@64 192 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 193 if( tp->base() != Type::AnyPtr &&
kvn@64 194 !(tinst &&
kvn@64 195 tinst->klass()->is_java_lang_Object() &&
kvn@64 196 tinst->offset() == Type::OffsetBot) ) {
kvn@64 197 // compress paths and change unreachable cycles to TOP
kvn@64 198 // If not, we can update the input infinitely along a MergeMem cycle
kvn@64 199 // Equivalent code in PhiNode::Ideal
kvn@64 200 Node* m = phase->transform(mmem);
kvn@64 201 // If tranformed to a MergeMem, get the desired slice
kvn@64 202 // Otherwise the returned node represents memory for every slice
kvn@64 203 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
kvn@64 204 // Update input if it is progress over what we have now
kvn@64 205 }
kvn@64 206 return mem;
kvn@64 207 }
kvn@64 208
duke@0 209 //--------------------------Ideal_common---------------------------------------
duke@0 210 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
duke@0 211 // Unhook non-raw memories from complete (macro-expanded) initializations.
duke@0 212 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
duke@0 213 // If our control input is a dead region, kill all below the region
duke@0 214 Node *ctl = in(MemNode::Control);
duke@0 215 if (ctl && remove_dead_region(phase, can_reshape))
duke@0 216 return this;
kvn@305 217 ctl = in(MemNode::Control);
kvn@305 218 // Don't bother trying to transform a dead node
kvn@305 219 if( ctl && ctl->is_top() ) return NodeSentinel;
duke@0 220
duke@0 221 // Ignore if memory is dead, or self-loop
duke@0 222 Node *mem = in(MemNode::Memory);
duke@0 223 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 224 assert( mem != this, "dead loop in MemNode::Ideal" );
duke@0 225
duke@0 226 Node *address = in(MemNode::Address);
duke@0 227 const Type *t_adr = phase->type( address );
duke@0 228 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 229
duke@0 230 // Avoid independent memory operations
duke@0 231 Node* old_mem = mem;
duke@0 232
kvn@36 233 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@36 234 // initializations was removed. After macro-expansion all stores catched
kvn@36 235 // by Initialize node became raw stores and there is no information
kvn@36 236 // which memory slices they modify. So it is unsafe to move any memory
kvn@36 237 // operation above these stores. Also in most cases hooked non-raw memories
kvn@36 238 // were already unhooked by using information from detect_ptr_independence()
kvn@36 239 // and find_previous_store().
duke@0 240
duke@0 241 if (mem->is_MergeMem()) {
duke@0 242 MergeMemNode* mmem = mem->as_MergeMem();
duke@0 243 const TypePtr *tp = t_adr->is_ptr();
kvn@64 244
kvn@64 245 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@0 246 }
duke@0 247
duke@0 248 if (mem != old_mem) {
duke@0 249 set_req(MemNode::Memory, mem);
kvn@305 250 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
duke@0 251 return this;
duke@0 252 }
duke@0 253
duke@0 254 // let the subclass continue analyzing...
duke@0 255 return NULL;
duke@0 256 }
duke@0 257
duke@0 258 // Helper function for proving some simple control dominations.
kvn@119 259 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
duke@0 260 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@0 261 // is not a constant (dominated by the method's StartNode).
duke@0 262 // Used by MemNode::find_previous_store to prove that the
duke@0 263 // control input of a memory operation predates (dominates)
duke@0 264 // an allocation it wants to look past.
kvn@119 265 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
kvn@119 266 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
kvn@119 267 return false; // Conservative answer for dead code
kvn@119 268
kvn@193 269 // Check 'dom'. Skip Proj and CatchProj nodes.
kvn@119 270 dom = dom->find_exact_control(dom);
kvn@119 271 if (dom == NULL || dom->is_top())
kvn@119 272 return false; // Conservative answer for dead code
kvn@119 273
kvn@193 274 if (dom == sub) {
kvn@193 275 // For the case when, for example, 'sub' is Initialize and the original
kvn@193 276 // 'dom' is Proj node of the 'sub'.
kvn@193 277 return false;
kvn@193 278 }
kvn@193 279
kvn@155 280 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
kvn@119 281 return true;
kvn@119 282
kvn@119 283 // 'dom' dominates 'sub' if its control edge and control edges
kvn@119 284 // of all its inputs dominate or equal to sub's control edge.
kvn@119 285
kvn@119 286 // Currently 'sub' is either Allocate, Initialize or Start nodes.
kvn@163 287 // Or Region for the check in LoadNode::Ideal();
kvn@163 288 // 'sub' should have sub->in(0) != NULL.
kvn@163 289 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
kvn@163 290 sub->is_Region(), "expecting only these nodes");
kvn@119 291
kvn@119 292 // Get control edge of 'sub'.
kvn@193 293 Node* orig_sub = sub;
kvn@119 294 sub = sub->find_exact_control(sub->in(0));
kvn@119 295 if (sub == NULL || sub->is_top())
kvn@119 296 return false; // Conservative answer for dead code
kvn@119 297
kvn@119 298 assert(sub->is_CFG(), "expecting control");
kvn@119 299
kvn@119 300 if (sub == dom)
kvn@119 301 return true;
kvn@119 302
kvn@119 303 if (sub->is_Start() || sub->is_Root())
kvn@119 304 return false;
kvn@119 305
kvn@119 306 {
kvn@119 307 // Check all control edges of 'dom'.
kvn@119 308
kvn@119 309 ResourceMark rm;
kvn@119 310 Arena* arena = Thread::current()->resource_area();
kvn@119 311 Node_List nlist(arena);
kvn@119 312 Unique_Node_List dom_list(arena);
kvn@119 313
kvn@119 314 dom_list.push(dom);
kvn@119 315 bool only_dominating_controls = false;
kvn@119 316
kvn@119 317 for (uint next = 0; next < dom_list.size(); next++) {
kvn@119 318 Node* n = dom_list.at(next);
kvn@193 319 if (n == orig_sub)
kvn@193 320 return false; // One of dom's inputs dominated by sub.
kvn@119 321 if (!n->is_CFG() && n->pinned()) {
kvn@119 322 // Check only own control edge for pinned non-control nodes.
kvn@119 323 n = n->find_exact_control(n->in(0));
kvn@119 324 if (n == NULL || n->is_top())
kvn@119 325 return false; // Conservative answer for dead code
kvn@119 326 assert(n->is_CFG(), "expecting control");
kvn@193 327 dom_list.push(n);
kvn@193 328 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
kvn@119 329 only_dominating_controls = true;
kvn@119 330 } else if (n->is_CFG()) {
kvn@119 331 if (n->dominates(sub, nlist))
kvn@119 332 only_dominating_controls = true;
kvn@119 333 else
kvn@119 334 return false;
kvn@119 335 } else {
kvn@119 336 // First, own control edge.
kvn@119 337 Node* m = n->find_exact_control(n->in(0));
kvn@155 338 if (m != NULL) {
kvn@155 339 if (m->is_top())
kvn@155 340 return false; // Conservative answer for dead code
kvn@155 341 dom_list.push(m);
kvn@155 342 }
kvn@119 343 // Now, the rest of edges.
kvn@119 344 uint cnt = n->req();
kvn@119 345 for (uint i = 1; i < cnt; i++) {
kvn@119 346 m = n->find_exact_control(n->in(i));
kvn@119 347 if (m == NULL || m->is_top())
kvn@119 348 continue;
kvn@119 349 dom_list.push(m);
duke@0 350 }
duke@0 351 }
duke@0 352 }
kvn@119 353 return only_dominating_controls;
duke@0 354 }
duke@0 355 }
duke@0 356
duke@0 357 //---------------------detect_ptr_independence---------------------------------
duke@0 358 // Used by MemNode::find_previous_store to prove that two base
duke@0 359 // pointers are never equal.
duke@0 360 // The pointers are accompanied by their associated allocations,
duke@0 361 // if any, which have been previously discovered by the caller.
duke@0 362 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@0 363 Node* p2, AllocateNode* a2,
duke@0 364 PhaseTransform* phase) {
duke@0 365 // Attempt to prove that these two pointers cannot be aliased.
duke@0 366 // They may both manifestly be allocations, and they should differ.
duke@0 367 // Or, if they are not both allocations, they can be distinct constants.
duke@0 368 // Otherwise, one is an allocation and the other a pre-existing value.
duke@0 369 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@0 370 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@0 371 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@0 372 return (a1 != a2);
duke@0 373 } else if (a1 != NULL) { // one allocation a1
duke@0 374 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
kvn@119 375 return all_controls_dominate(p2, a1);
duke@0 376 } else { //(a2 != NULL) // one allocation a2
kvn@119 377 return all_controls_dominate(p1, a2);
duke@0 378 }
duke@0 379 return false;
duke@0 380 }
duke@0 381
duke@0 382
duke@0 383 // The logic for reordering loads and stores uses four steps:
duke@0 384 // (a) Walk carefully past stores and initializations which we
duke@0 385 // can prove are independent of this load.
duke@0 386 // (b) Observe that the next memory state makes an exact match
duke@0 387 // with self (load or store), and locate the relevant store.
duke@0 388 // (c) Ensure that, if we were to wire self directly to the store,
duke@0 389 // the optimizer would fold it up somehow.
duke@0 390 // (d) Do the rewiring, and return, depending on some other part of
duke@0 391 // the optimizer to fold up the load.
duke@0 392 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@0 393 // specific to loads and stores, so they are handled by the callers.
duke@0 394 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@0 395 //
duke@0 396 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@0 397 Node* ctrl = in(MemNode::Control);
duke@0 398 Node* adr = in(MemNode::Address);
duke@0 399 intptr_t offset = 0;
duke@0 400 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 401 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@0 402
duke@0 403 if (offset == Type::OffsetBot)
duke@0 404 return NULL; // cannot unalias unless there are precise offsets
duke@0 405
kvn@74 406 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
kvn@74 407
duke@0 408 intptr_t size_in_bytes = memory_size();
duke@0 409
duke@0 410 Node* mem = in(MemNode::Memory); // start searching here...
duke@0 411
duke@0 412 int cnt = 50; // Cycle limiter
duke@0 413 for (;;) { // While we can dance past unrelated stores...
duke@0 414 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@0 415
duke@0 416 if (mem->is_Store()) {
duke@0 417 Node* st_adr = mem->in(MemNode::Address);
duke@0 418 intptr_t st_offset = 0;
duke@0 419 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@0 420 if (st_base == NULL)
duke@0 421 break; // inscrutable pointer
duke@0 422 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@0 423 const int MAX_STORE = BytesPerLong;
duke@0 424 if (st_offset >= offset + size_in_bytes ||
duke@0 425 st_offset <= offset - MAX_STORE ||
duke@0 426 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@0 427 // Success: The offsets are provably independent.
duke@0 428 // (You may ask, why not just test st_offset != offset and be done?
duke@0 429 // The answer is that stores of different sizes can co-exist
duke@0 430 // in the same sequence of RawMem effects. We sometimes initialize
duke@0 431 // a whole 'tile' of array elements with a single jint or jlong.)
duke@0 432 mem = mem->in(MemNode::Memory);
duke@0 433 continue; // (a) advance through independent store memory
duke@0 434 }
duke@0 435 }
duke@0 436 if (st_base != base &&
duke@0 437 detect_ptr_independence(base, alloc,
duke@0 438 st_base,
duke@0 439 AllocateNode::Ideal_allocation(st_base, phase),
duke@0 440 phase)) {
duke@0 441 // Success: The bases are provably independent.
duke@0 442 mem = mem->in(MemNode::Memory);
duke@0 443 continue; // (a) advance through independent store memory
duke@0 444 }
duke@0 445
duke@0 446 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@0 447 // since we have not managed to prove 'this' and 'mem' independent.
duke@0 448 if (st_base == base && st_offset == offset) {
duke@0 449 return mem; // let caller handle steps (c), (d)
duke@0 450 }
duke@0 451
duke@0 452 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 453 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@0 454 AllocateNode* st_alloc = st_init->allocation();
duke@0 455 if (st_alloc == NULL)
duke@0 456 break; // something degenerated
duke@0 457 bool known_identical = false;
duke@0 458 bool known_independent = false;
duke@0 459 if (alloc == st_alloc)
duke@0 460 known_identical = true;
duke@0 461 else if (alloc != NULL)
duke@0 462 known_independent = true;
kvn@119 463 else if (all_controls_dominate(this, st_alloc))
duke@0 464 known_independent = true;
duke@0 465
duke@0 466 if (known_independent) {
duke@0 467 // The bases are provably independent: Either they are
duke@0 468 // manifestly distinct allocations, or else the control
duke@0 469 // of this load dominates the store's allocation.
duke@0 470 int alias_idx = phase->C->get_alias_index(adr_type());
duke@0 471 if (alias_idx == Compile::AliasIdxRaw) {
duke@0 472 mem = st_alloc->in(TypeFunc::Memory);
duke@0 473 } else {
duke@0 474 mem = st_init->memory(alias_idx);
duke@0 475 }
duke@0 476 continue; // (a) advance through independent store memory
duke@0 477 }
duke@0 478
duke@0 479 // (b) at this point, if we are not looking at a store initializing
duke@0 480 // the same allocation we are loading from, we lose.
duke@0 481 if (known_identical) {
duke@0 482 // From caller, can_see_stored_value will consult find_captured_store.
duke@0 483 return mem; // let caller handle steps (c), (d)
duke@0 484 }
duke@0 485
kvn@223 486 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
kvn@74 487 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
kvn@74 488 if (mem->is_Proj() && mem->in(0)->is_Call()) {
kvn@74 489 CallNode *call = mem->in(0)->as_Call();
kvn@74 490 if (!call->may_modify(addr_t, phase)) {
kvn@74 491 mem = call->in(TypeFunc::Memory);
kvn@74 492 continue; // (a) advance through independent call memory
kvn@74 493 }
kvn@74 494 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
kvn@74 495 mem = mem->in(0)->in(TypeFunc::Memory);
kvn@74 496 continue; // (a) advance through independent MemBar memory
kvn@74 497 } else if (mem->is_MergeMem()) {
kvn@74 498 int alias_idx = phase->C->get_alias_index(adr_type());
kvn@74 499 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@74 500 continue; // (a) advance through independent MergeMem memory
kvn@74 501 }
duke@0 502 }
duke@0 503
duke@0 504 // Unless there is an explicit 'continue', we must bail out here,
duke@0 505 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@0 506 break;
duke@0 507 }
duke@0 508
duke@0 509 return NULL; // bail out
duke@0 510 }
duke@0 511
duke@0 512 //----------------------calculate_adr_type-------------------------------------
duke@0 513 // Helper function. Notices when the given type of address hits top or bottom.
duke@0 514 // Also, asserts a cross-check of the type against the expected address type.
duke@0 515 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@0 516 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@0 517 #ifdef PRODUCT
duke@0 518 cross_check = NULL;
duke@0 519 #else
duke@0 520 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@0 521 #endif
duke@0 522 const TypePtr* tp = t->isa_ptr();
duke@0 523 if (tp == NULL) {
duke@0 524 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@0 525 return TypePtr::BOTTOM; // touches lots of memory
duke@0 526 } else {
duke@0 527 #ifdef ASSERT
duke@0 528 // %%%% [phh] We don't check the alias index if cross_check is
duke@0 529 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@0 530 if (cross_check != NULL &&
duke@0 531 cross_check != TypePtr::BOTTOM &&
duke@0 532 cross_check != TypeRawPtr::BOTTOM) {
duke@0 533 // Recheck the alias index, to see if it has changed (due to a bug).
duke@0 534 Compile* C = Compile::current();
duke@0 535 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@0 536 "must stay in the original alias category");
duke@0 537 // The type of the address must be contained in the adr_type,
duke@0 538 // disregarding "null"-ness.
duke@0 539 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@0 540 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
duke@0 541 assert(cross_check->meet(tp_notnull) == cross_check,
duke@0 542 "real address must not escape from expected memory type");
duke@0 543 }
duke@0 544 #endif
duke@0 545 return tp;
duke@0 546 }
duke@0 547 }
duke@0 548
duke@0 549 //------------------------adr_phi_is_loop_invariant----------------------------
duke@0 550 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@0 551 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@0 552 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@0 553 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@0 554 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@0 555 // with the same data. This implies that any path into the loop already
duke@0 556 // includes such a CastPP, and so the original cast, whatever its input,
duke@0 557 // must be covered by an equivalent cast, with an earlier control input.
duke@0 558 ResourceMark rm;
duke@0 559
duke@0 560 // The loop entry input of the phi should be the unique dominating
duke@0 561 // node for every Phi/CastPP in the loop.
duke@0 562 Unique_Node_List closure;
duke@0 563 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@0 564
duke@0 565 // Add the phi node and the cast to the worklist.
duke@0 566 Unique_Node_List worklist;
duke@0 567 worklist.push(adr_phi);
duke@0 568 if( cast != NULL ){
duke@0 569 if( !cast->is_ConstraintCast() ) return false;
duke@0 570 worklist.push(cast);
duke@0 571 }
duke@0 572
duke@0 573 // Begin recursive walk of phi nodes.
duke@0 574 while( worklist.size() ){
duke@0 575 // Take a node off the worklist
duke@0 576 Node *n = worklist.pop();
duke@0 577 if( !closure.member(n) ){
duke@0 578 // Add it to the closure.
duke@0 579 closure.push(n);
duke@0 580 // Make a sanity check to ensure we don't waste too much time here.
duke@0 581 if( closure.size() > 20) return false;
duke@0 582 // This node is OK if:
duke@0 583 // - it is a cast of an identical value
duke@0 584 // - or it is a phi node (then we add its inputs to the worklist)
duke@0 585 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@0 586 if( n->is_ConstraintCast() ){
duke@0 587 worklist.push(n->in(1));
duke@0 588 } else if( n->is_Phi() ) {
duke@0 589 for( uint i = 1; i < n->req(); i++ ) {
duke@0 590 worklist.push(n->in(i));
duke@0 591 }
duke@0 592 } else {
duke@0 593 return false;
duke@0 594 }
duke@0 595 }
duke@0 596 }
duke@0 597
duke@0 598 // Quit when the worklist is empty, and we've found no offending nodes.
duke@0 599 return true;
duke@0 600 }
duke@0 601
duke@0 602 //------------------------------Ideal_DU_postCCP-------------------------------
duke@0 603 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@0 604 // going away in this pass and we need to make this memory op depend on the
duke@0 605 // gating null check.
kvn@163 606 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
kvn@163 607 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
kvn@163 608 }
duke@0 609
duke@0 610 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@0 611 // some sense; we get to keep around the knowledge that an oop is not-null
duke@0 612 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@0 613 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@0 614 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@0 615 // some of the more trivial cases in the optimizer. Removing more useless
duke@0 616 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@0 617 // up on this approach. CNC 10/20/2000
kvn@163 618 // This static method may be called not from MemNode (EncodePNode calls it).
kvn@163 619 // Only the control edge of the node 'n' might be updated.
kvn@163 620 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
duke@0 621 Node *skipped_cast = NULL;
duke@0 622 // Need a null check? Regular static accesses do not because they are
duke@0 623 // from constant addresses. Array ops are gated by the range check (which
duke@0 624 // always includes a NULL check). Just check field ops.
kvn@163 625 if( n->in(MemNode::Control) == NULL ) {
duke@0 626 // Scan upwards for the highest location we can place this memory op.
duke@0 627 while( true ) {
duke@0 628 switch( adr->Opcode() ) {
duke@0 629
duke@0 630 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@0 631 adr = adr->in(AddPNode::Base);
duke@0 632 continue;
duke@0 633
coleenp@113 634 case Op_DecodeN: // No change to NULL-ness, so peek thru
coleenp@113 635 adr = adr->in(1);
coleenp@113 636 continue;
coleenp@113 637
duke@0 638 case Op_CastPP:
duke@0 639 // If the CastPP is useless, just peek on through it.
duke@0 640 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@0 641 // Remember the cast that we've peeked though. If we peek
duke@0 642 // through more than one, then we end up remembering the highest
duke@0 643 // one, that is, if in a loop, the one closest to the top.
duke@0 644 skipped_cast = adr;
duke@0 645 adr = adr->in(1);
duke@0 646 continue;
duke@0 647 }
duke@0 648 // CastPP is going away in this pass! We need this memory op to be
duke@0 649 // control-dependent on the test that is guarding the CastPP.
kvn@163 650 ccp->hash_delete(n);
kvn@163 651 n->set_req(MemNode::Control, adr->in(0));
kvn@163 652 ccp->hash_insert(n);
kvn@163 653 return n;
duke@0 654
duke@0 655 case Op_Phi:
duke@0 656 // Attempt to float above a Phi to some dominating point.
duke@0 657 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@0 658 // If we've already peeked through a Cast (which could have set the
duke@0 659 // control), we can't float above a Phi, because the skipped Cast
duke@0 660 // may not be loop invariant.
duke@0 661 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@0 662 adr = adr->in(1);
duke@0 663 continue;
duke@0 664 }
duke@0 665 }
duke@0 666
duke@0 667 // Intentional fallthrough!
duke@0 668
duke@0 669 // No obvious dominating point. The mem op is pinned below the Phi
duke@0 670 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@0 671 // then the mem op can float, but not indefinitely. It must be pinned
duke@0 672 // behind the controls leading to the Phi.
duke@0 673 case Op_CheckCastPP:
duke@0 674 // These usually stick around to change address type, however a
duke@0 675 // useless one can be elided and we still need to pick up a control edge
duke@0 676 if (adr->in(0) == NULL) {
duke@0 677 // This CheckCastPP node has NO control and is likely useless. But we
duke@0 678 // need check further up the ancestor chain for a control input to keep
duke@0 679 // the node in place. 4959717.
duke@0 680 skipped_cast = adr;
duke@0 681 adr = adr->in(1);
duke@0 682 continue;
duke@0 683 }
kvn@163 684 ccp->hash_delete(n);
kvn@163 685 n->set_req(MemNode::Control, adr->in(0));
kvn@163 686 ccp->hash_insert(n);
kvn@163 687 return n;
duke@0 688
duke@0 689 // List of "safe" opcodes; those that implicitly block the memory
duke@0 690 // op below any null check.
duke@0 691 case Op_CastX2P: // no null checks on native pointers
duke@0 692 case Op_Parm: // 'this' pointer is not null
duke@0 693 case Op_LoadP: // Loading from within a klass
coleenp@113 694 case Op_LoadN: // Loading from within a klass
duke@0 695 case Op_LoadKlass: // Loading from within a klass
kvn@164 696 case Op_LoadNKlass: // Loading from within a klass
duke@0 697 case Op_ConP: // Loading from a klass
kvn@163 698 case Op_ConN: // Loading from a klass
duke@0 699 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@0 700 case Op_Con: // Reading from TLS
duke@0 701 case Op_CMoveP: // CMoveP is pinned
kvn@164 702 case Op_CMoveN: // CMoveN is pinned
duke@0 703 break; // No progress
duke@0 704
duke@0 705 case Op_Proj: // Direct call to an allocation routine
duke@0 706 case Op_SCMemProj: // Memory state from store conditional ops
duke@0 707 #ifdef ASSERT
duke@0 708 {
duke@0 709 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@0 710 const Node* call = adr->in(0);
kvn@163 711 if (call->is_CallJava()) {
kvn@163 712 const CallJavaNode* call_java = call->as_CallJava();
kvn@64 713 const TypeTuple *r = call_java->tf()->range();
kvn@64 714 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@64 715 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@64 716 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@0 717 // We further presume that this is one of
duke@0 718 // new_instance_Java, new_array_Java, or
duke@0 719 // the like, but do not assert for this.
duke@0 720 } else if (call->is_Allocate()) {
duke@0 721 // similar case to new_instance_Java, etc.
duke@0 722 } else if (!call->is_CallLeaf()) {
duke@0 723 // Projections from fetch_oop (OSR) are allowed as well.
duke@0 724 ShouldNotReachHere();
duke@0 725 }
duke@0 726 }
duke@0 727 #endif
duke@0 728 break;
duke@0 729 default:
duke@0 730 ShouldNotReachHere();
duke@0 731 }
duke@0 732 break;
duke@0 733 }
duke@0 734 }
duke@0 735
duke@0 736 return NULL; // No progress
duke@0 737 }
duke@0 738
duke@0 739
duke@0 740 //=============================================================================
duke@0 741 uint LoadNode::size_of() const { return sizeof(*this); }
duke@0 742 uint LoadNode::cmp( const Node &n ) const
duke@0 743 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@0 744 const Type *LoadNode::bottom_type() const { return _type; }
duke@0 745 uint LoadNode::ideal_reg() const {
duke@0 746 return Matcher::base2reg[_type->base()];
duke@0 747 }
duke@0 748
duke@0 749 #ifndef PRODUCT
duke@0 750 void LoadNode::dump_spec(outputStream *st) const {
duke@0 751 MemNode::dump_spec(st);
duke@0 752 if( !Verbose && !WizardMode ) {
duke@0 753 // standard dump does this in Verbose and WizardMode
duke@0 754 st->print(" #"); _type->dump_on(st);
duke@0 755 }
duke@0 756 }
duke@0 757 #endif
duke@0 758
duke@0 759
duke@0 760 //----------------------------LoadNode::make-----------------------------------
duke@0 761 // Polymorphic factory method:
coleenp@113 762 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
coleenp@113 763 Compile* C = gvn.C;
coleenp@113 764
duke@0 765 // sanity check the alias category against the created node type
duke@0 766 assert(!(adr_type->isa_oopptr() &&
duke@0 767 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@0 768 "use LoadKlassNode instead");
duke@0 769 assert(!(adr_type->isa_aryptr() &&
duke@0 770 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@0 771 "use LoadRangeNode instead");
duke@0 772 switch (bt) {
duke@0 773 case T_BOOLEAN:
duke@0 774 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 775 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 776 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 777 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 778 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
duke@0 779 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
duke@0 780 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
duke@0 781 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
coleenp@113 782 case T_OBJECT:
coleenp@113 783 #ifdef _LP64
kvn@163 784 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
kvn@221 785 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
kvn@221 786 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
coleenp@113 787 } else
coleenp@113 788 #endif
kvn@163 789 {
kvn@163 790 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@163 791 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
kvn@163 792 }
duke@0 793 }
duke@0 794 ShouldNotReachHere();
duke@0 795 return (LoadNode*)NULL;
duke@0 796 }
duke@0 797
duke@0 798 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
duke@0 799 bool require_atomic = true;
duke@0 800 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
duke@0 801 }
duke@0 802
duke@0 803
duke@0 804
duke@0 805
duke@0 806 //------------------------------hash-------------------------------------------
duke@0 807 uint LoadNode::hash() const {
duke@0 808 // unroll addition of interesting fields
duke@0 809 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@0 810 }
duke@0 811
duke@0 812 //---------------------------can_see_stored_value------------------------------
duke@0 813 // This routine exists to make sure this set of tests is done the same
duke@0 814 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@0 815 // will change the graph shape in a way which makes memory alive twice at the
duke@0 816 // same time (uses the Oracle model of aliasing), then some
duke@0 817 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@0 818 // of aliasing.
duke@0 819 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@0 820 Node* ld_adr = in(MemNode::Address);
duke@0 821
never@17 822 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
never@17 823 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
never@17 824 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
never@17 825 atp->field() != NULL && !atp->field()->is_volatile()) {
never@17 826 uint alias_idx = atp->index();
never@17 827 bool final = atp->field()->is_final();
never@17 828 Node* result = NULL;
never@17 829 Node* current = st;
never@17 830 // Skip through chains of MemBarNodes checking the MergeMems for
never@17 831 // new states for the slice of this load. Stop once any other
never@17 832 // kind of node is encountered. Loads from final memory can skip
never@17 833 // through any kind of MemBar but normal loads shouldn't skip
never@17 834 // through MemBarAcquire since the could allow them to move out of
never@17 835 // a synchronized region.
never@17 836 while (current->is_Proj()) {
never@17 837 int opc = current->in(0)->Opcode();
never@17 838 if ((final && opc == Op_MemBarAcquire) ||
never@17 839 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
never@17 840 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@17 841 if (mem->is_MergeMem()) {
never@17 842 MergeMemNode* merge = mem->as_MergeMem();
never@17 843 Node* new_st = merge->memory_at(alias_idx);
never@17 844 if (new_st == merge->base_memory()) {
never@17 845 // Keep searching
never@17 846 current = merge->base_memory();
never@17 847 continue;
never@17 848 }
never@17 849 // Save the new memory state for the slice and fall through
never@17 850 // to exit.
never@17 851 result = new_st;
never@17 852 }
never@17 853 }
never@17 854 break;
never@17 855 }
never@17 856 if (result != NULL) {
never@17 857 st = result;
never@17 858 }
never@17 859 }
never@17 860
never@17 861
duke@0 862 // Loop around twice in the case Load -> Initialize -> Store.
duke@0 863 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@0 864 for (int trip = 0; trip <= 1; trip++) {
duke@0 865
duke@0 866 if (st->is_Store()) {
duke@0 867 Node* st_adr = st->in(MemNode::Address);
duke@0 868 if (!phase->eqv(st_adr, ld_adr)) {
duke@0 869 // Try harder before giving up... Match raw and non-raw pointers.
duke@0 870 intptr_t st_off = 0;
duke@0 871 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@0 872 if (alloc == NULL) return NULL;
duke@0 873 intptr_t ld_off = 0;
duke@0 874 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
duke@0 875 if (alloc != allo2) return NULL;
duke@0 876 if (ld_off != st_off) return NULL;
duke@0 877 // At this point we have proven something like this setup:
duke@0 878 // A = Allocate(...)
duke@0 879 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@0 880 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@0 881 // (Actually, we haven't yet proven the Q's are the same.)
duke@0 882 // In other words, we are loading from a casted version of
duke@0 883 // the same pointer-and-offset that we stored to.
duke@0 884 // Thus, we are able to replace L by V.
duke@0 885 }
duke@0 886 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@0 887 if (store_Opcode() != st->Opcode())
duke@0 888 return NULL;
duke@0 889 return st->in(MemNode::ValueIn);
duke@0 890 }
duke@0 891
duke@0 892 intptr_t offset = 0; // scratch
duke@0 893
duke@0 894 // A load from a freshly-created object always returns zero.
duke@0 895 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@0 896 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@0 897 if (st->is_Proj() && st->in(0)->is_Allocate() &&
duke@0 898 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
duke@0 899 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
duke@0 900 // return a zero value for the load's basic type
duke@0 901 // (This is one of the few places where a generic PhaseTransform
duke@0 902 // can create new nodes. Think of it as lazily manifesting
duke@0 903 // virtually pre-existing constants.)
duke@0 904 return phase->zerocon(memory_type());
duke@0 905 }
duke@0 906
duke@0 907 // A load from an initialization barrier can match a captured store.
duke@0 908 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@0 909 InitializeNode* init = st->in(0)->as_Initialize();
duke@0 910 AllocateNode* alloc = init->allocation();
duke@0 911 if (alloc != NULL &&
duke@0 912 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
duke@0 913 // examine a captured store value
duke@0 914 st = init->find_captured_store(offset, memory_size(), phase);
duke@0 915 if (st != NULL)
duke@0 916 continue; // take one more trip around
duke@0 917 }
duke@0 918 }
duke@0 919
duke@0 920 break;
duke@0 921 }
duke@0 922
duke@0 923 return NULL;
duke@0 924 }
duke@0 925
kvn@64 926 //----------------------is_instance_field_load_with_local_phi------------------
kvn@64 927 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@64 928 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
kvn@64 929 in(MemNode::Address)->is_AddP() ) {
kvn@64 930 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@64 931 // Only instances.
kvn@223 932 if( t_oop != NULL && t_oop->is_known_instance_field() &&
kvn@64 933 t_oop->offset() != Type::OffsetBot &&
kvn@64 934 t_oop->offset() != Type::OffsetTop) {
kvn@64 935 return true;
kvn@64 936 }
kvn@64 937 }
kvn@64 938 return false;
kvn@64 939 }
kvn@64 940
duke@0 941 //------------------------------Identity---------------------------------------
duke@0 942 // Loads are identity if previous store is to same address
duke@0 943 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@0 944 // If the previous store-maker is the right kind of Store, and the store is
duke@0 945 // to the same address, then we are equal to the value stored.
duke@0 946 Node* mem = in(MemNode::Memory);
duke@0 947 Node* value = can_see_stored_value(mem, phase);
duke@0 948 if( value ) {
duke@0 949 // byte, short & char stores truncate naturally.
duke@0 950 // A load has to load the truncated value which requires
duke@0 951 // some sort of masking operation and that requires an
duke@0 952 // Ideal call instead of an Identity call.
duke@0 953 if (memory_size() < BytesPerInt) {
duke@0 954 // If the input to the store does not fit with the load's result type,
duke@0 955 // it must be truncated via an Ideal call.
duke@0 956 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@0 957 return this;
duke@0 958 }
duke@0 959 // (This works even when value is a Con, but LoadNode::Value
duke@0 960 // usually runs first, producing the singleton type of the Con.)
duke@0 961 return value;
duke@0 962 }
kvn@64 963
kvn@64 964 // Search for an existing data phi which was generated before for the same
kvn@64 965 // instance's field to avoid infinite genertion of phis in a loop.
kvn@64 966 Node *region = mem->in(0);
kvn@64 967 if (is_instance_field_load_with_local_phi(region)) {
kvn@64 968 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
kvn@64 969 int this_index = phase->C->get_alias_index(addr_t);
kvn@64 970 int this_offset = addr_t->offset();
kvn@64 971 int this_id = addr_t->is_oopptr()->instance_id();
kvn@64 972 const Type* this_type = bottom_type();
kvn@64 973 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@64 974 Node* phi = region->fast_out(i);
kvn@64 975 if (phi->is_Phi() && phi != mem &&
kvn@64 976 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
kvn@64 977 return phi;
kvn@64 978 }
kvn@64 979 }
kvn@64 980 }
kvn@64 981
duke@0 982 return this;
duke@0 983 }
duke@0 984
never@17 985
never@17 986 // Returns true if the AliasType refers to the field that holds the
never@17 987 // cached box array. Currently only handles the IntegerCache case.
never@17 988 static bool is_autobox_cache(Compile::AliasType* atp) {
never@17 989 if (atp != NULL && atp->field() != NULL) {
never@17 990 ciField* field = atp->field();
never@17 991 ciSymbol* klass = field->holder()->name();
never@17 992 if (field->name() == ciSymbol::cache_field_name() &&
never@17 993 field->holder()->uses_default_loader() &&
never@17 994 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 995 return true;
never@17 996 }
never@17 997 }
never@17 998 return false;
never@17 999 }
never@17 1000
never@17 1001 // Fetch the base value in the autobox array
never@17 1002 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
never@17 1003 if (atp != NULL && atp->field() != NULL) {
never@17 1004 ciField* field = atp->field();
never@17 1005 ciSymbol* klass = field->holder()->name();
never@17 1006 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1007 field->holder()->uses_default_loader() &&
never@17 1008 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1009 assert(field->is_constant(), "what?");
never@17 1010 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
never@17 1011 // Fetch the box object at the base of the array and get its value
never@17 1012 ciInstance* box = array->obj_at(0)->as_instance();
never@17 1013 ciInstanceKlass* ik = box->klass()->as_instance_klass();
never@17 1014 if (ik->nof_nonstatic_fields() == 1) {
never@17 1015 // This should be true nonstatic_field_at requires calling
never@17 1016 // nof_nonstatic_fields so check it anyway
never@17 1017 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
never@17 1018 cache_offset = c.as_int();
never@17 1019 }
never@17 1020 return true;
never@17 1021 }
never@17 1022 }
never@17 1023 return false;
never@17 1024 }
never@17 1025
never@17 1026 // Returns true if the AliasType refers to the value field of an
never@17 1027 // autobox object. Currently only handles Integer.
never@17 1028 static bool is_autobox_object(Compile::AliasType* atp) {
never@17 1029 if (atp != NULL && atp->field() != NULL) {
never@17 1030 ciField* field = atp->field();
never@17 1031 ciSymbol* klass = field->holder()->name();
never@17 1032 if (field->name() == ciSymbol::value_name() &&
never@17 1033 field->holder()->uses_default_loader() &&
never@17 1034 klass == ciSymbol::java_lang_Integer()) {
never@17 1035 return true;
never@17 1036 }
never@17 1037 }
never@17 1038 return false;
never@17 1039 }
never@17 1040
never@17 1041
never@17 1042 // We're loading from an object which has autobox behaviour.
never@17 1043 // If this object is result of a valueOf call we'll have a phi
never@17 1044 // merging a newly allocated object and a load from the cache.
never@17 1045 // We want to replace this load with the original incoming
never@17 1046 // argument to the valueOf call.
never@17 1047 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
never@17 1048 Node* base = in(Address)->in(AddPNode::Base);
never@17 1049 if (base->is_Phi() && base->req() == 3) {
never@17 1050 AllocateNode* allocation = NULL;
never@17 1051 int allocation_index = -1;
never@17 1052 int load_index = -1;
never@17 1053 for (uint i = 1; i < base->req(); i++) {
never@17 1054 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
never@17 1055 if (allocation != NULL) {
never@17 1056 allocation_index = i;
never@17 1057 load_index = 3 - allocation_index;
never@17 1058 break;
never@17 1059 }
never@17 1060 }
never@17 1061 LoadNode* load = NULL;
never@17 1062 if (allocation != NULL && base->in(load_index)->is_Load()) {
never@17 1063 load = base->in(load_index)->as_Load();
never@17 1064 }
never@17 1065 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
never@17 1066 // Push the loads from the phi that comes from valueOf up
never@17 1067 // through it to allow elimination of the loads and the recovery
never@17 1068 // of the original value.
never@17 1069 Node* mem_phi = in(Memory);
never@17 1070 Node* offset = in(Address)->in(AddPNode::Offset);
never@17 1071
never@17 1072 Node* in1 = clone();
never@17 1073 Node* in1_addr = in1->in(Address)->clone();
never@17 1074 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
never@17 1075 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
never@17 1076 in1_addr->set_req(AddPNode::Offset, offset);
never@17 1077 in1->set_req(0, base->in(allocation_index));
never@17 1078 in1->set_req(Address, in1_addr);
never@17 1079 in1->set_req(Memory, mem_phi->in(allocation_index));
never@17 1080
never@17 1081 Node* in2 = clone();
never@17 1082 Node* in2_addr = in2->in(Address)->clone();
never@17 1083 in2_addr->set_req(AddPNode::Base, base->in(load_index));
never@17 1084 in2_addr->set_req(AddPNode::Address, base->in(load_index));
never@17 1085 in2_addr->set_req(AddPNode::Offset, offset);
never@17 1086 in2->set_req(0, base->in(load_index));
never@17 1087 in2->set_req(Address, in2_addr);
never@17 1088 in2->set_req(Memory, mem_phi->in(load_index));
never@17 1089
never@17 1090 in1_addr = phase->transform(in1_addr);
never@17 1091 in1 = phase->transform(in1);
never@17 1092 in2_addr = phase->transform(in2_addr);
never@17 1093 in2 = phase->transform(in2);
never@17 1094
never@17 1095 PhiNode* result = PhiNode::make_blank(base->in(0), this);
never@17 1096 result->set_req(allocation_index, in1);
never@17 1097 result->set_req(load_index, in2);
never@17 1098 return result;
never@17 1099 }
never@17 1100 } else if (base->is_Load()) {
never@17 1101 // Eliminate the load of Integer.value for integers from the cache
never@17 1102 // array by deriving the value from the index into the array.
never@17 1103 // Capture the offset of the load and then reverse the computation.
never@17 1104 Node* load_base = base->in(Address)->in(AddPNode::Base);
never@17 1105 if (load_base != NULL) {
never@17 1106 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
never@17 1107 intptr_t cache_offset;
never@17 1108 int shift = -1;
never@17 1109 Node* cache = NULL;
never@17 1110 if (is_autobox_cache(atp)) {
kvn@29 1111 shift = exact_log2(type2aelembytes(T_OBJECT));
never@17 1112 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
never@17 1113 }
never@17 1114 if (cache != NULL && base->in(Address)->is_AddP()) {
never@17 1115 Node* elements[4];
never@17 1116 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@17 1117 int cache_low;
never@17 1118 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
never@17 1119 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
never@17 1120 // Add up all the offsets making of the address of the load
never@17 1121 Node* result = elements[0];
never@17 1122 for (int i = 1; i < count; i++) {
never@17 1123 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
never@17 1124 }
never@17 1125 // Remove the constant offset from the address and then
never@17 1126 // remove the scaling of the offset to recover the original index.
never@17 1127 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
never@17 1128 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
never@17 1129 // Peel the shift off directly but wrap it in a dummy node
never@17 1130 // since Ideal can't return existing nodes
never@17 1131 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
never@17 1132 } else {
never@17 1133 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
never@17 1134 }
never@17 1135 #ifdef _LP64
never@17 1136 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
never@17 1137 #endif
never@17 1138 return result;
never@17 1139 }
never@17 1140 }
never@17 1141 }
never@17 1142 }
never@17 1143 return NULL;
never@17 1144 }
never@17 1145
kvn@163 1146 //------------------------------split_through_phi------------------------------
kvn@163 1147 // Split instance field load through Phi.
kvn@163 1148 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
kvn@163 1149 Node* mem = in(MemNode::Memory);
kvn@163 1150 Node* address = in(MemNode::Address);
kvn@163 1151 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@163 1152 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@163 1153
kvn@163 1154 assert(mem->is_Phi() && (t_oop != NULL) &&
kvn@223 1155 t_oop->is_known_instance_field(), "invalide conditions");
kvn@163 1156
kvn@163 1157 Node *region = mem->in(0);
kvn@163 1158 if (region == NULL) {
kvn@163 1159 return NULL; // Wait stable graph
kvn@163 1160 }
kvn@163 1161 uint cnt = mem->req();
kvn@163 1162 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1163 Node *in = mem->in(i);
kvn@163 1164 if( in == NULL ) {
kvn@163 1165 return NULL; // Wait stable graph
kvn@163 1166 }
kvn@163 1167 }
kvn@163 1168 // Check for loop invariant.
kvn@163 1169 if (cnt == 3) {
kvn@163 1170 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1171 Node *in = mem->in(i);
kvn@163 1172 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
kvn@163 1173 if (m == mem) {
kvn@163 1174 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
kvn@163 1175 return this;
kvn@163 1176 }
kvn@163 1177 }
kvn@163 1178 }
kvn@163 1179 // Split through Phi (see original code in loopopts.cpp).
kvn@163 1180 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
kvn@163 1181
kvn@163 1182 // Do nothing here if Identity will find a value
kvn@163 1183 // (to avoid infinite chain of value phis generation).
kvn@163 1184 if ( !phase->eqv(this, this->Identity(phase)) )
kvn@163 1185 return NULL;
kvn@163 1186
kvn@163 1187 // Skip the split if the region dominates some control edge of the address.
kvn@163 1188 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
kvn@163 1189 return NULL;
kvn@163 1190
kvn@163 1191 const Type* this_type = this->bottom_type();
kvn@163 1192 int this_index = phase->C->get_alias_index(addr_t);
kvn@163 1193 int this_offset = addr_t->offset();
kvn@163 1194 int this_iid = addr_t->is_oopptr()->instance_id();
kvn@163 1195 int wins = 0;
kvn@163 1196 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@163 1197 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
kvn@163 1198 for( uint i = 1; i < region->req(); i++ ) {
kvn@163 1199 Node *x;
kvn@163 1200 Node* the_clone = NULL;
kvn@163 1201 if( region->in(i) == phase->C->top() ) {
kvn@163 1202 x = phase->C->top(); // Dead path? Use a dead data op
kvn@163 1203 } else {
kvn@163 1204 x = this->clone(); // Else clone up the data op
kvn@163 1205 the_clone = x; // Remember for possible deletion.
kvn@163 1206 // Alter data node to use pre-phi inputs
kvn@163 1207 if( this->in(0) == region ) {
kvn@163 1208 x->set_req( 0, region->in(i) );
kvn@163 1209 } else {
kvn@163 1210 x->set_req( 0, NULL );
kvn@163 1211 }
kvn@163 1212 for( uint j = 1; j < this->req(); j++ ) {
kvn@163 1213 Node *in = this->in(j);
kvn@163 1214 if( in->is_Phi() && in->in(0) == region )
kvn@163 1215 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
kvn@163 1216 }
kvn@163 1217 }
kvn@163 1218 // Check for a 'win' on some paths
kvn@163 1219 const Type *t = x->Value(igvn);
kvn@163 1220
kvn@163 1221 bool singleton = t->singleton();
kvn@163 1222
kvn@163 1223 // See comments in PhaseIdealLoop::split_thru_phi().
kvn@163 1224 if( singleton && t == Type::TOP ) {
kvn@163 1225 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
kvn@163 1226 }
kvn@163 1227
kvn@163 1228 if( singleton ) {
kvn@163 1229 wins++;
kvn@163 1230 x = igvn->makecon(t);
kvn@163 1231 } else {
kvn@163 1232 // We now call Identity to try to simplify the cloned node.
kvn@163 1233 // Note that some Identity methods call phase->type(this).
kvn@163 1234 // Make sure that the type array is big enough for
kvn@163 1235 // our new node, even though we may throw the node away.
kvn@163 1236 // (This tweaking with igvn only works because x is a new node.)
kvn@163 1237 igvn->set_type(x, t);
kvn@293 1238 // If x is a TypeNode, capture any more-precise type permanently into Node
kvn@293 1239 // othewise it will be not updated during igvn->transform since
kvn@293 1240 // igvn->type(x) is set to x->Value() already.
kvn@293 1241 x->raise_bottom_type(t);
kvn@163 1242 Node *y = x->Identity(igvn);
kvn@163 1243 if( y != x ) {
kvn@163 1244 wins++;
kvn@163 1245 x = y;
kvn@163 1246 } else {
kvn@163 1247 y = igvn->hash_find(x);
kvn@163 1248 if( y ) {
kvn@163 1249 wins++;
kvn@163 1250 x = y;
kvn@163 1251 } else {
kvn@163 1252 // Else x is a new node we are keeping
kvn@163 1253 // We do not need register_new_node_with_optimizer
kvn@163 1254 // because set_type has already been called.
kvn@163 1255 igvn->_worklist.push(x);
kvn@163 1256 }
kvn@163 1257 }
kvn@163 1258 }
kvn@163 1259 if (x != the_clone && the_clone != NULL)
kvn@163 1260 igvn->remove_dead_node(the_clone);
kvn@163 1261 phi->set_req(i, x);
kvn@163 1262 }
kvn@163 1263 if( wins > 0 ) {
kvn@163 1264 // Record Phi
kvn@163 1265 igvn->register_new_node_with_optimizer(phi);
kvn@163 1266 return phi;
kvn@163 1267 }
kvn@163 1268 igvn->remove_dead_node(phi);
kvn@163 1269 return NULL;
kvn@163 1270 }
never@17 1271
duke@0 1272 //------------------------------Ideal------------------------------------------
duke@0 1273 // If the load is from Field memory and the pointer is non-null, we can
duke@0 1274 // zero out the control input.
duke@0 1275 // If the offset is constant and the base is an object allocation,
duke@0 1276 // try to hook me up to the exact initializing store.
duke@0 1277 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1278 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1279 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1280
duke@0 1281 Node* ctrl = in(MemNode::Control);
duke@0 1282 Node* address = in(MemNode::Address);
duke@0 1283
duke@0 1284 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@0 1285 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@0 1286 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@0 1287 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@0 1288 ctrl = ctrl->in(0);
duke@0 1289 set_req(MemNode::Control,ctrl);
duke@0 1290 }
duke@0 1291
duke@0 1292 // Check for useless control edge in some common special cases
duke@0 1293 if (in(MemNode::Control) != NULL) {
duke@0 1294 intptr_t ignore = 0;
duke@0 1295 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
duke@0 1296 if (base != NULL
duke@0 1297 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
kvn@119 1298 && all_controls_dominate(base, phase->C->start())) {
duke@0 1299 // A method-invariant, non-null address (constant or 'this' argument).
duke@0 1300 set_req(MemNode::Control, NULL);
duke@0 1301 }
duke@0 1302 }
duke@0 1303
never@17 1304 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
never@17 1305 Node* base = in(Address)->in(AddPNode::Base);
never@17 1306 if (base != NULL) {
never@17 1307 Compile::AliasType* atp = phase->C->alias_type(adr_type());
never@17 1308 if (is_autobox_object(atp)) {
never@17 1309 Node* result = eliminate_autobox(phase);
never@17 1310 if (result != NULL) return result;
never@17 1311 }
never@17 1312 }
never@17 1313 }
never@17 1314
kvn@74 1315 Node* mem = in(MemNode::Memory);
kvn@74 1316 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@74 1317
kvn@74 1318 if (addr_t != NULL) {
kvn@74 1319 // try to optimize our memory input
kvn@74 1320 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
kvn@74 1321 if (opt_mem != mem) {
kvn@74 1322 set_req(MemNode::Memory, opt_mem);
kvn@305 1323 if (phase->type( opt_mem ) == Type::TOP) return NULL;
kvn@74 1324 return this;
kvn@74 1325 }
kvn@74 1326 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@74 1327 if (can_reshape && opt_mem->is_Phi() &&
kvn@223 1328 (t_oop != NULL) && t_oop->is_known_instance_field()) {
kvn@163 1329 // Split instance field load through Phi.
kvn@163 1330 Node* result = split_through_phi(phase);
kvn@163 1331 if (result != NULL) return result;
kvn@74 1332 }
kvn@74 1333 }
kvn@74 1334
duke@0 1335 // Check for prior store with a different base or offset; make Load
duke@0 1336 // independent. Skip through any number of them. Bail out if the stores
duke@0 1337 // are in an endless dead cycle and report no progress. This is a key
duke@0 1338 // transform for Reflection. However, if after skipping through the Stores
duke@0 1339 // we can't then fold up against a prior store do NOT do the transform as
duke@0 1340 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@0 1341 // array memory alive twice: once for the hoisted Load and again after the
duke@0 1342 // bypassed Store. This situation only works if EVERYBODY who does
duke@0 1343 // anti-dependence work knows how to bypass. I.e. we need all
duke@0 1344 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@0 1345 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@0 1346 // fold up, do so.
duke@0 1347 Node* prev_mem = find_previous_store(phase);
duke@0 1348 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1349 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@0 1350 // (c) See if we can fold up on the spot, but don't fold up here.
duke@0 1351 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
duke@0 1352 // just return a prior value, which is done by Identity calls.
duke@0 1353 if (can_see_stored_value(prev_mem, phase)) {
duke@0 1354 // Make ready for step (d):
duke@0 1355 set_req(MemNode::Memory, prev_mem);
duke@0 1356 return this;
duke@0 1357 }
duke@0 1358 }
duke@0 1359
duke@0 1360 return NULL; // No further progress
duke@0 1361 }
duke@0 1362
duke@0 1363 // Helper to recognize certain Klass fields which are invariant across
duke@0 1364 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@0 1365 const Type*
duke@0 1366 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@0 1367 ciKlass* klass) const {
duke@0 1368 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1369 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@0 1370 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@0 1371 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@0 1372 return TypeInt::make(klass->modifier_flags());
duke@0 1373 }
duke@0 1374 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1375 // The field is Klass::_access_flags. Return its (constant) value.
duke@0 1376 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@0 1377 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@0 1378 return TypeInt::make(klass->access_flags());
duke@0 1379 }
duke@0 1380 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1381 // The field is Klass::_layout_helper. Return its constant value if known.
duke@0 1382 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1383 return TypeInt::make(klass->layout_helper());
duke@0 1384 }
duke@0 1385
duke@0 1386 // No match.
duke@0 1387 return NULL;
duke@0 1388 }
duke@0 1389
duke@0 1390 //------------------------------Value-----------------------------------------
duke@0 1391 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@0 1392 // Either input is TOP ==> the result is TOP
duke@0 1393 Node* mem = in(MemNode::Memory);
duke@0 1394 const Type *t1 = phase->type(mem);
duke@0 1395 if (t1 == Type::TOP) return Type::TOP;
duke@0 1396 Node* adr = in(MemNode::Address);
duke@0 1397 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@0 1398 if (tp == NULL || tp->empty()) return Type::TOP;
duke@0 1399 int off = tp->offset();
duke@0 1400 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
duke@0 1401
duke@0 1402 // Try to guess loaded type from pointer type
duke@0 1403 if (tp->base() == Type::AryPtr) {
duke@0 1404 const Type *t = tp->is_aryptr()->elem();
duke@0 1405 // Don't do this for integer types. There is only potential profit if
duke@0 1406 // the element type t is lower than _type; that is, for int types, if _type is
duke@0 1407 // more restrictive than t. This only happens here if one is short and the other
duke@0 1408 // char (both 16 bits), and in those cases we've made an intentional decision
duke@0 1409 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@0 1410 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@0 1411 //
duke@0 1412 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@0 1413 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@0 1414 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@0 1415 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@0 1416 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@0 1417 // In fact, that could have been the original type of p1, and p1 could have
duke@0 1418 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@0 1419 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@0 1420 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
kvn@293 1421 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
duke@0 1422 // t might actually be lower than _type, if _type is a unique
duke@0 1423 // concrete subclass of abstract class t.
duke@0 1424 // Make sure the reference is not into the header, by comparing
duke@0 1425 // the offset against the offset of the start of the array's data.
duke@0 1426 // Different array types begin at slightly different offsets (12 vs. 16).
duke@0 1427 // We choose T_BYTE as an example base type that is least restrictive
duke@0 1428 // as to alignment, which will therefore produce the smallest
duke@0 1429 // possible base offset.
duke@0 1430 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1431 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
duke@0 1432 const Type* jt = t->join(_type);
duke@0 1433 // In any case, do not allow the join, per se, to empty out the type.
duke@0 1434 if (jt->empty() && !t->empty()) {
duke@0 1435 // This can happen if a interface-typed array narrows to a class type.
duke@0 1436 jt = _type;
duke@0 1437 }
never@17 1438
never@17 1439 if (EliminateAutoBox) {
never@17 1440 // The pointers in the autobox arrays are always non-null
never@17 1441 Node* base = in(Address)->in(AddPNode::Base);
never@17 1442 if (base != NULL) {
never@17 1443 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
never@17 1444 if (is_autobox_cache(atp)) {
never@17 1445 return jt->join(TypePtr::NOTNULL)->is_ptr();
never@17 1446 }
never@17 1447 }
never@17 1448 }
duke@0 1449 return jt;
duke@0 1450 }
duke@0 1451 }
duke@0 1452 } else if (tp->base() == Type::InstPtr) {
duke@0 1453 assert( off != Type::OffsetBot ||
duke@0 1454 // arrays can be cast to Objects
duke@0 1455 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@0 1456 // unsafe field access may not have a constant offset
duke@0 1457 phase->C->has_unsafe_access(),
duke@0 1458 "Field accesses must be precise" );
duke@0 1459 // For oop loads, we expect the _type to be precise
duke@0 1460 } else if (tp->base() == Type::KlassPtr) {
duke@0 1461 assert( off != Type::OffsetBot ||
duke@0 1462 // arrays can be cast to Objects
duke@0 1463 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@0 1464 // also allow array-loading from the primary supertype
duke@0 1465 // array during subtype checks
duke@0 1466 Opcode() == Op_LoadKlass,
duke@0 1467 "Field accesses must be precise" );
duke@0 1468 // For klass/static loads, we expect the _type to be precise
duke@0 1469 }
duke@0 1470
duke@0 1471 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1472 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1473 ciKlass* klass = tkls->klass();
duke@0 1474 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@0 1475 // We are loading a field from a Klass metaobject whose identity
duke@0 1476 // is known at compile time (the type is "exact" or "precise").
duke@0 1477 // Check for fields we know are maintained as constants by the VM.
duke@0 1478 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1479 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@0 1480 // (Folds up type checking code.)
duke@0 1481 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@0 1482 return TypeInt::make(klass->super_check_offset());
duke@0 1483 }
duke@0 1484 // Compute index into primary_supers array
duke@0 1485 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1486 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1487 if( depth < ciKlass::primary_super_limit() ) {
duke@0 1488 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1489 // (Folds up type checking code.)
duke@0 1490 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1491 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1492 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1493 }
duke@0 1494 const Type* aift = load_array_final_field(tkls, klass);
duke@0 1495 if (aift != NULL) return aift;
duke@0 1496 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
duke@0 1497 && klass->is_array_klass()) {
duke@0 1498 // The field is arrayKlass::_component_mirror. Return its (constant) value.
duke@0 1499 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@0 1500 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@0 1501 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@0 1502 }
duke@0 1503 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1504 // The field is Klass::_java_mirror. Return its (constant) value.
duke@0 1505 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@0 1506 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@0 1507 return TypeInstPtr::make(klass->java_mirror());
duke@0 1508 }
duke@0 1509 }
duke@0 1510
duke@0 1511 // We can still check if we are loading from the primary_supers array at a
duke@0 1512 // shallow enough depth. Even though the klass is not exact, entries less
duke@0 1513 // than or equal to its super depth are correct.
duke@0 1514 if (klass->is_loaded() ) {
duke@0 1515 ciType *inner = klass->klass();
duke@0 1516 while( inner->is_obj_array_klass() )
duke@0 1517 inner = inner->as_obj_array_klass()->base_element_type();
duke@0 1518 if( inner->is_instance_klass() &&
duke@0 1519 !inner->as_instance_klass()->flags().is_interface() ) {
duke@0 1520 // Compute index into primary_supers array
duke@0 1521 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1522 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1523 if( depth < ciKlass::primary_super_limit() &&
duke@0 1524 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@0 1525 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1526 // (Folds up type checking code.)
duke@0 1527 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1528 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1529 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1530 }
duke@0 1531 }
duke@0 1532 }
duke@0 1533
duke@0 1534 // If the type is enough to determine that the thing is not an array,
duke@0 1535 // we can give the layout_helper a positive interval type.
duke@0 1536 // This will help short-circuit some reflective code.
duke@0 1537 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
duke@0 1538 && !klass->is_array_klass() // not directly typed as an array
duke@0 1539 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@0 1540 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@0 1541 ) {
duke@0 1542 // Note: When interfaces are reliable, we can narrow the interface
duke@0 1543 // test to (klass != Serializable && klass != Cloneable).
duke@0 1544 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1545 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@0 1546 // The key property of this type is that it folds up tests
duke@0 1547 // for array-ness, since it proves that the layout_helper is positive.
duke@0 1548 // Thus, a generic value like the basic object layout helper works fine.
duke@0 1549 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@0 1550 }
duke@0 1551 }
duke@0 1552
duke@0 1553 // If we are loading from a freshly-allocated object, produce a zero,
duke@0 1554 // if the load is provably beyond the header of the object.
duke@0 1555 // (Also allow a variable load from a fresh array to produce zero.)
duke@0 1556 if (ReduceFieldZeroing) {
duke@0 1557 Node* value = can_see_stored_value(mem,phase);
duke@0 1558 if (value != NULL && value->is_Con())
duke@0 1559 return value->bottom_type();
duke@0 1560 }
duke@0 1561
kvn@64 1562 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@223 1563 if (tinst != NULL && tinst->is_known_instance_field()) {
kvn@64 1564 // If we have an instance type and our memory input is the
kvn@64 1565 // programs's initial memory state, there is no matching store,
kvn@64 1566 // so just return a zero of the appropriate type
kvn@64 1567 Node *mem = in(MemNode::Memory);
kvn@64 1568 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@64 1569 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@64 1570 return Type::get_zero_type(_type->basic_type());
kvn@64 1571 }
kvn@64 1572 }
duke@0 1573 return _type;
duke@0 1574 }
duke@0 1575
duke@0 1576 //------------------------------match_edge-------------------------------------
duke@0 1577 // Do we Match on this edge index or not? Match only the address.
duke@0 1578 uint LoadNode::match_edge(uint idx) const {
duke@0 1579 return idx == MemNode::Address;
duke@0 1580 }
duke@0 1581
duke@0 1582 //--------------------------LoadBNode::Ideal--------------------------------------
duke@0 1583 //
duke@0 1584 // If the previous store is to the same address as this load,
duke@0 1585 // and the value stored was larger than a byte, replace this load
duke@0 1586 // with the value stored truncated to a byte. If no truncation is
duke@0 1587 // needed, the replacement is done in LoadNode::Identity().
duke@0 1588 //
duke@0 1589 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1590 Node* mem = in(MemNode::Memory);
duke@0 1591 Node* value = can_see_stored_value(mem,phase);
duke@0 1592 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1593 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
duke@0 1594 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
duke@0 1595 }
duke@0 1596 // Identity call will handle the case where truncation is not needed.
duke@0 1597 return LoadNode::Ideal(phase, can_reshape);
duke@0 1598 }
duke@0 1599
duke@0 1600 //--------------------------LoadCNode::Ideal--------------------------------------
duke@0 1601 //
duke@0 1602 // If the previous store is to the same address as this load,
duke@0 1603 // and the value stored was larger than a char, replace this load
duke@0 1604 // with the value stored truncated to a char. If no truncation is
duke@0 1605 // needed, the replacement is done in LoadNode::Identity().
duke@0 1606 //
duke@0 1607 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1608 Node* mem = in(MemNode::Memory);
duke@0 1609 Node* value = can_see_stored_value(mem,phase);
duke@0 1610 if( value && !phase->type(value)->higher_equal( _type ) )
duke@0 1611 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
duke@0 1612 // Identity call will handle the case where truncation is not needed.
duke@0 1613 return LoadNode::Ideal(phase, can_reshape);
duke@0 1614 }
duke@0 1615
duke@0 1616 //--------------------------LoadSNode::Ideal--------------------------------------
duke@0 1617 //
duke@0 1618 // If the previous store is to the same address as this load,
duke@0 1619 // and the value stored was larger than a short, replace this load
duke@0 1620 // with the value stored truncated to a short. If no truncation is
duke@0 1621 // needed, the replacement is done in LoadNode::Identity().
duke@0 1622 //
duke@0 1623 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1624 Node* mem = in(MemNode::Memory);
duke@0 1625 Node* value = can_see_stored_value(mem,phase);
duke@0 1626 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1627 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
duke@0 1628 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
duke@0 1629 }
duke@0 1630 // Identity call will handle the case where truncation is not needed.
duke@0 1631 return LoadNode::Ideal(phase, can_reshape);
duke@0 1632 }
duke@0 1633
duke@0 1634 //=============================================================================
kvn@164 1635 //----------------------------LoadKlassNode::make------------------------------
kvn@164 1636 // Polymorphic factory method:
kvn@164 1637 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
kvn@164 1638 Compile* C = gvn.C;
kvn@164 1639 Node *ctl = NULL;
kvn@164 1640 // sanity check the alias category against the created node type
kvn@164 1641 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
kvn@164 1642 assert(adr_type != NULL, "expecting TypeOopPtr");
kvn@164 1643 #ifdef _LP64
kvn@164 1644 if (adr_type->is_ptr_to_narrowoop()) {
kvn@221 1645 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
kvn@221 1646 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
kvn@168 1647 }
kvn@164 1648 #endif
kvn@168 1649 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@168 1650 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
kvn@164 1651 }
kvn@164 1652
duke@0 1653 //------------------------------Value------------------------------------------
duke@0 1654 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1655 return klass_value_common(phase);
kvn@164 1656 }
kvn@164 1657
kvn@164 1658 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
duke@0 1659 // Either input is TOP ==> the result is TOP
duke@0 1660 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1661 if (t1 == Type::TOP) return Type::TOP;
duke@0 1662 Node *adr = in(MemNode::Address);
duke@0 1663 const Type *t2 = phase->type( adr );
duke@0 1664 if (t2 == Type::TOP) return Type::TOP;
duke@0 1665 const TypePtr *tp = t2->is_ptr();
duke@0 1666 if (TypePtr::above_centerline(tp->ptr()) ||
duke@0 1667 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@0 1668
duke@0 1669 // Return a more precise klass, if possible
duke@0 1670 const TypeInstPtr *tinst = tp->isa_instptr();
duke@0 1671 if (tinst != NULL) {
duke@0 1672 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@0 1673 int offset = tinst->offset();
duke@0 1674 if (ik == phase->C->env()->Class_klass()
duke@0 1675 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1676 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1677 // We are loading a special hidden field from a Class mirror object,
duke@0 1678 // the field which points to the VM's Klass metaobject.
duke@0 1679 ciType* t = tinst->java_mirror_type();
duke@0 1680 // java_mirror_type returns non-null for compile-time Class constants.
duke@0 1681 if (t != NULL) {
duke@0 1682 // constant oop => constant klass
duke@0 1683 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1684 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@0 1685 }
duke@0 1686 if (!t->is_klass()) {
duke@0 1687 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@0 1688 return TypePtr::NULL_PTR;
duke@0 1689 }
duke@0 1690 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@0 1691 return TypeKlassPtr::make(t->as_klass());
duke@0 1692 }
duke@0 1693 // non-constant mirror, so we can't tell what's going on
duke@0 1694 }
duke@0 1695 if( !ik->is_loaded() )
duke@0 1696 return _type; // Bail out if not loaded
duke@0 1697 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1698 if (tinst->klass_is_exact()) {
duke@0 1699 return TypeKlassPtr::make(ik);
duke@0 1700 }
duke@0 1701 // See if we can become precise: no subklasses and no interface
duke@0 1702 // (Note: We need to support verified interfaces.)
duke@0 1703 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1704 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1705 // Add a dependence; if any subclass added we need to recompile
duke@0 1706 if (!ik->is_final()) {
duke@0 1707 // %%% should use stronger assert_unique_concrete_subtype instead
duke@0 1708 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1709 }
duke@0 1710 // Return precise klass
duke@0 1711 return TypeKlassPtr::make(ik);
duke@0 1712 }
duke@0 1713
duke@0 1714 // Return root of possible klass
duke@0 1715 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@0 1716 }
duke@0 1717 }
duke@0 1718
duke@0 1719 // Check for loading klass from an array
duke@0 1720 const TypeAryPtr *tary = tp->isa_aryptr();
duke@0 1721 if( tary != NULL ) {
duke@0 1722 ciKlass *tary_klass = tary->klass();
duke@0 1723 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@0 1724 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@0 1725 if (tary->klass_is_exact()) {
duke@0 1726 return TypeKlassPtr::make(tary_klass);
duke@0 1727 }
duke@0 1728 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@0 1729 // If the klass is an object array, we defer the question to the
duke@0 1730 // array component klass.
duke@0 1731 if( ak->is_obj_array_klass() ) {
duke@0 1732 assert( ak->is_loaded(), "" );
duke@0 1733 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@0 1734 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@0 1735 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@0 1736 // See if we can become precise: no subklasses and no interface
duke@0 1737 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1738 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1739 // Add a dependence; if any subclass added we need to recompile
duke@0 1740 if (!ik->is_final()) {
duke@0 1741 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1742 }
duke@0 1743 // Return precise array klass
duke@0 1744 return TypeKlassPtr::make(ak);
duke@0 1745 }
duke@0 1746 }
duke@0 1747 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 1748 } else { // Found a type-array?
duke@0 1749 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1750 assert( ak->is_type_array_klass(), "" );
duke@0 1751 return TypeKlassPtr::make(ak); // These are always precise
duke@0 1752 }
duke@0 1753 }
duke@0 1754 }
duke@0 1755
duke@0 1756 // Check for loading klass from an array klass
duke@0 1757 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1758 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1759 ciKlass* klass = tkls->klass();
duke@0 1760 if( !klass->is_loaded() )
duke@0 1761 return _type; // Bail out if not loaded
duke@0 1762 if( klass->is_obj_array_klass() &&
duke@0 1763 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1764 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@0 1765 // // Always returning precise element type is incorrect,
duke@0 1766 // // e.g., element type could be object and array may contain strings
duke@0 1767 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@0 1768
duke@0 1769 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@0 1770 // according to the element type's subclassing.
duke@0 1771 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@0 1772 }
duke@0 1773 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
duke@0 1774 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1775 ciKlass* sup = klass->as_instance_klass()->super();
duke@0 1776 // The field is Klass::_super. Return its (constant) value.
duke@0 1777 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@0 1778 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@0 1779 }
duke@0 1780 }
duke@0 1781
duke@0 1782 // Bailout case
duke@0 1783 return LoadNode::Value(phase);
duke@0 1784 }
duke@0 1785
duke@0 1786 //------------------------------Identity---------------------------------------
duke@0 1787 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@0 1788 // Also feed through the klass in Allocate(...klass...)._klass.
duke@0 1789 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1790 return klass_identity_common(phase);
kvn@164 1791 }
kvn@164 1792
kvn@164 1793 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
duke@0 1794 Node* x = LoadNode::Identity(phase);
duke@0 1795 if (x != this) return x;
duke@0 1796
duke@0 1797 // Take apart the address into an oop and and offset.
duke@0 1798 // Return 'this' if we cannot.
duke@0 1799 Node* adr = in(MemNode::Address);
duke@0 1800 intptr_t offset = 0;
duke@0 1801 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1802 if (base == NULL) return this;
duke@0 1803 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@0 1804 if (toop == NULL) return this;
duke@0 1805
duke@0 1806 // We can fetch the klass directly through an AllocateNode.
duke@0 1807 // This works even if the klass is not constant (clone or newArray).
duke@0 1808 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1809 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@0 1810 if (allocated_klass != NULL) {
duke@0 1811 return allocated_klass;
duke@0 1812 }
duke@0 1813 }
duke@0 1814
duke@0 1815 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
duke@0 1816 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
duke@0 1817 // See inline_native_Class_query for occurrences of these patterns.
duke@0 1818 // Java Example: x.getClass().isAssignableFrom(y)
duke@0 1819 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@0 1820 //
duke@0 1821 // This improves reflective code, often making the Class
duke@0 1822 // mirror go completely dead. (Current exception: Class
duke@0 1823 // mirrors may appear in debug info, but we could clean them out by
duke@0 1824 // introducing a new debug info operator for klassOop.java_mirror).
duke@0 1825 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@0 1826 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1827 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1828 // We are loading a special hidden field from a Class mirror,
duke@0 1829 // the field which points to its Klass or arrayKlass metaobject.
duke@0 1830 if (base->is_Load()) {
duke@0 1831 Node* adr2 = base->in(MemNode::Address);
duke@0 1832 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@0 1833 if (tkls != NULL && !tkls->empty()
duke@0 1834 && (tkls->klass()->is_instance_klass() ||
duke@0 1835 tkls->klass()->is_array_klass())
duke@0 1836 && adr2->is_AddP()
duke@0 1837 ) {
duke@0 1838 int mirror_field = Klass::java_mirror_offset_in_bytes();
duke@0 1839 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1840 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
duke@0 1841 }
duke@0 1842 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
duke@0 1843 return adr2->in(AddPNode::Base);
duke@0 1844 }
duke@0 1845 }
duke@0 1846 }
duke@0 1847 }
duke@0 1848
duke@0 1849 return this;
duke@0 1850 }
duke@0 1851
kvn@164 1852
kvn@164 1853 //------------------------------Value------------------------------------------
kvn@164 1854 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1855 const Type *t = klass_value_common(phase);
kvn@221 1856 if (t == Type::TOP)
kvn@221 1857 return t;
kvn@221 1858
kvn@221 1859 return t->make_narrowoop();
kvn@164 1860 }
kvn@164 1861
kvn@164 1862 //------------------------------Identity---------------------------------------
kvn@164 1863 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
kvn@164 1864 // Also feed through the klass in Allocate(...klass...)._klass.
kvn@164 1865 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1866 Node *x = klass_identity_common(phase);
kvn@164 1867
kvn@164 1868 const Type *t = phase->type( x );
kvn@164 1869 if( t == Type::TOP ) return x;
kvn@164 1870 if( t->isa_narrowoop()) return x;
kvn@164 1871
kvn@221 1872 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
kvn@164 1873 }
kvn@164 1874
duke@0 1875 //------------------------------Value-----------------------------------------
duke@0 1876 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@0 1877 // Either input is TOP ==> the result is TOP
duke@0 1878 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1879 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1880 Node *adr = in(MemNode::Address);
duke@0 1881 const Type *t2 = phase->type( adr );
duke@0 1882 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1883 const TypePtr *tp = t2->is_ptr();
duke@0 1884 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@0 1885 const TypeAryPtr *tap = tp->isa_aryptr();
duke@0 1886 if( !tap ) return _type;
duke@0 1887 return tap->size();
duke@0 1888 }
duke@0 1889
rasbold@366 1890 //-------------------------------Ideal---------------------------------------
rasbold@366 1891 // Feed through the length in AllocateArray(...length...)._length.
rasbold@366 1892 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
rasbold@366 1893 Node* p = MemNode::Ideal_common(phase, can_reshape);
rasbold@366 1894 if (p) return (p == NodeSentinel) ? NULL : p;
rasbold@366 1895
rasbold@366 1896 // Take apart the address into an oop and and offset.
rasbold@366 1897 // Return 'this' if we cannot.
rasbold@366 1898 Node* adr = in(MemNode::Address);
rasbold@366 1899 intptr_t offset = 0;
rasbold@366 1900 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
rasbold@366 1901 if (base == NULL) return NULL;
rasbold@366 1902 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
rasbold@366 1903 if (tary == NULL) return NULL;
rasbold@366 1904
rasbold@366 1905 // We can fetch the length directly through an AllocateArrayNode.
rasbold@366 1906 // This works even if the length is not constant (clone or newArray).
rasbold@366 1907 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 1908 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 1909 if (alloc != NULL) {
rasbold@366 1910 Node* allocated_length = alloc->Ideal_length();
rasbold@366 1911 Node* len = alloc->make_ideal_length(tary, phase);
rasbold@366 1912 if (allocated_length != len) {
rasbold@366 1913 // New CastII improves on this.
rasbold@366 1914 return len;
rasbold@366 1915 }
rasbold@366 1916 }
rasbold@366 1917 }
rasbold@366 1918
rasbold@366 1919 return NULL;
rasbold@366 1920 }
rasbold@366 1921
duke@0 1922 //------------------------------Identity---------------------------------------
duke@0 1923 // Feed through the length in AllocateArray(...length...)._length.
duke@0 1924 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@0 1925 Node* x = LoadINode::Identity(phase);
duke@0 1926 if (x != this) return x;
duke@0 1927
duke@0 1928 // Take apart the address into an oop and and offset.
duke@0 1929 // Return 'this' if we cannot.
duke@0 1930 Node* adr = in(MemNode::Address);
duke@0 1931 intptr_t offset = 0;
duke@0 1932 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1933 if (base == NULL) return this;
duke@0 1934 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@0 1935 if (tary == NULL) return this;
duke@0 1936
duke@0 1937 // We can fetch the length directly through an AllocateArrayNode.
duke@0 1938 // This works even if the length is not constant (clone or newArray).
duke@0 1939 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 1940 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 1941 if (alloc != NULL) {
rasbold@366 1942 Node* allocated_length = alloc->Ideal_length();
rasbold@366 1943 // Do not allow make_ideal_length to allocate a CastII node.
rasbold@366 1944 Node* len = alloc->make_ideal_length(tary, phase, false);
rasbold@366 1945 if (allocated_length == len) {
rasbold@366 1946 // Return allocated_length only if it would not be improved by a CastII.
rasbold@366 1947 return allocated_length;
rasbold@366 1948 }
duke@0 1949 }
duke@0 1950 }
duke@0 1951
duke@0 1952 return this;
duke@0 1953
duke@0 1954 }
rasbold@366 1955
duke@0 1956 //=============================================================================
duke@0 1957 //---------------------------StoreNode::make-----------------------------------
duke@0 1958 // Polymorphic factory method:
coleenp@113 1959 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
coleenp@113 1960 Compile* C = gvn.C;
coleenp@113 1961
duke@0 1962 switch (bt) {
duke@0 1963 case T_BOOLEAN:
duke@0 1964 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
duke@0 1965 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
duke@0 1966 case T_CHAR:
duke@0 1967 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
duke@0 1968 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
duke@0 1969 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
duke@0 1970 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
duke@0 1971 case T_ADDRESS:
coleenp@113 1972 case T_OBJECT:
coleenp@113 1973 #ifdef _LP64
kvn@163 1974 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
coleenp@113 1975 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
coleenp@113 1976 adr->bottom_type()->isa_rawptr())) {
kvn@221 1977 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
kvn@221 1978 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
coleenp@113 1979 } else
coleenp@113 1980 #endif
kvn@221 1981 {
kvn@221 1982 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
kvn@221 1983 }
duke@0 1984 }
duke@0 1985 ShouldNotReachHere();
duke@0 1986 return (StoreNode*)NULL;
duke@0 1987 }
duke@0 1988
duke@0 1989 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
duke@0 1990 bool require_atomic = true;
duke@0 1991 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
duke@0 1992 }
duke@0 1993
duke@0 1994
duke@0 1995 //--------------------------bottom_type----------------------------------------
duke@0 1996 const Type *StoreNode::bottom_type() const {
duke@0 1997 return Type::MEMORY;
duke@0 1998 }
duke@0 1999
duke@0 2000 //------------------------------hash-------------------------------------------
duke@0 2001 uint StoreNode::hash() const {
duke@0 2002 // unroll addition of interesting fields
duke@0 2003 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@0 2004
duke@0 2005 // Since they are not commoned, do not hash them:
duke@0 2006 return NO_HASH;
duke@0 2007 }
duke@0 2008
duke@0 2009 //------------------------------Ideal------------------------------------------
duke@0 2010 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@0 2011 // When a store immediately follows a relevant allocation/initialization,
duke@0 2012 // try to capture it into the initialization, or hoist it above.
duke@0 2013 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 2014 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 2015 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 2016
duke@0 2017 Node* mem = in(MemNode::Memory);
duke@0 2018 Node* address = in(MemNode::Address);
duke@0 2019
duke@0 2020 // Back-to-back stores to same address? Fold em up.
duke@0 2021 // Generally unsafe if I have intervening uses...
duke@0 2022 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
duke@0 2023 // Looking at a dead closed cycle of memory?
duke@0 2024 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@0 2025
duke@0 2026 assert(Opcode() == mem->Opcode() ||
duke@0 2027 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@0 2028 "no mismatched stores, except on raw memory");
duke@0 2029
duke@0 2030 if (mem->outcnt() == 1 && // check for intervening uses
duke@0 2031 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@0 2032 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@0 2033 // For example, 'mem' might be the final state at a conditional return.
duke@0 2034 // Or, 'mem' might be used by some node which is live at the same time
duke@0 2035 // 'this' is live, which might be unschedulable. So, require exactly
duke@0 2036 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@0 2037 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@0 2038 if (can_reshape) { // (%%% is this an anachronism?)
duke@0 2039 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@0 2040 phase->is_IterGVN());
duke@0 2041 } else {
duke@0 2042 // It's OK to do this in the parser, since DU info is always accurate,
duke@0 2043 // and the parser always refers to nodes via SafePointNode maps.
duke@0 2044 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@0 2045 }
duke@0 2046 return this;
duke@0 2047 }
duke@0 2048 }
duke@0 2049
duke@0 2050 // Capture an unaliased, unconditional, simple store into an initializer.
duke@0 2051 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@0 2052 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@0 2053 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 2054 InitializeNode* init = mem->in(0)->as_Initialize();
duke@0 2055 intptr_t offset = init->can_capture_store(this, phase);
duke@0 2056 if (offset > 0) {
duke@0 2057 Node* moved = init->capture_store(this, offset, phase);
duke@0 2058 // If the InitializeNode captured me, it made a raw copy of me,
duke@0 2059 // and I need to disappear.
duke@0 2060 if (moved != NULL) {
duke@0 2061 // %%% hack to ensure that Ideal returns a new node:
duke@0 2062 mem = MergeMemNode::make(phase->C, mem);
duke@0 2063 return mem; // fold me away
duke@0 2064 }
duke@0 2065 }
duke@0 2066 }
duke@0 2067
duke@0 2068 return NULL; // No further progress
duke@0 2069 }
duke@0 2070
duke@0 2071 //------------------------------Value-----------------------------------------
duke@0 2072 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@0 2073 // Either input is TOP ==> the result is TOP
duke@0 2074 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 2075 if( t1 == Type::TOP ) return Type::TOP;
duke@0 2076 const Type *t2 = phase->type( in(MemNode::Address) );
duke@0 2077 if( t2 == Type::TOP ) return Type::TOP;
duke@0 2078 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@0 2079 if( t3 == Type::TOP ) return Type::TOP;
duke@0 2080 return Type::MEMORY;
duke@0 2081 }
duke@0 2082
duke@0 2083 //------------------------------Identity---------------------------------------
duke@0 2084 // Remove redundant stores:
duke@0 2085 // Store(m, p, Load(m, p)) changes to m.
duke@0 2086 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@0 2087 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@0 2088 Node* mem = in(MemNode::Memory);
duke@0 2089 Node* adr = in(MemNode::Address);
duke@0 2090 Node* val = in(MemNode::ValueIn);
duke@0 2091
duke@0 2092 // Load then Store? Then the Store is useless
duke@0 2093 if (val->is_Load() &&
duke@0 2094 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
duke@0 2095 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
duke@0 2096 val->as_Load()->store_Opcode() == Opcode()) {
duke@0 2097 return mem;
duke@0 2098 }
duke@0 2099
duke@0 2100 // Two stores in a row of the same value?
duke@0 2101 if (mem->is_Store() &&
duke@0 2102 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
duke@0 2103 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
duke@0 2104 mem->Opcode() == Opcode()) {
duke@0 2105 return mem;
duke@0 2106 }
duke@0 2107
duke@0 2108 // Store of zero anywhere into a freshly-allocated object?
duke@0 2109 // Then the store is useless.
duke@0 2110 // (It must already have been captured by the InitializeNode.)
duke@0 2111 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@0 2112 // a newly allocated object is already all-zeroes everywhere
duke@0 2113 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@0 2114 return mem;
duke@0 2115 }
duke@0 2116
duke@0 2117 // the store may also apply to zero-bits in an earlier object
duke@0 2118 Node* prev_mem = find_previous_store(phase);
duke@0 2119 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 2120 if (prev_mem != NULL) {
duke@0 2121 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@0 2122 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@0 2123 // prev_val and val might differ by a cast; it would be good
duke@0 2124 // to keep the more informative of the two.
duke@0 2125 return mem;
duke@0 2126 }
duke@0 2127 }
duke@0 2128 }
duke@0 2129
duke@0 2130 return this;
duke@0 2131 }
duke@0 2132
duke@0 2133 //------------------------------match_edge-------------------------------------
duke@0 2134 // Do we Match on this edge index or not? Match only memory & value
duke@0 2135 uint StoreNode::match_edge(uint idx) const {
duke@0 2136 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@0 2137 }
duke@0 2138
duke@0 2139 //------------------------------cmp--------------------------------------------
duke@0 2140 // Do not common stores up together. They generally have to be split
duke@0 2141 // back up anyways, so do not bother.
duke@0 2142 uint StoreNode::cmp( const Node &n ) const {
duke@0 2143 return (&n == this); // Always fail except on self
duke@0 2144 }
duke@0 2145
duke@0 2146 //------------------------------Ideal_masked_input-----------------------------
duke@0 2147 // Check for a useless mask before a partial-word store
duke@0 2148 // (StoreB ... (AndI valIn conIa) )
duke@0 2149 // If (conIa & mask == mask) this simplifies to
duke@0 2150 // (StoreB ... (valIn) )
duke@0 2151 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@0 2152 Node *val = in(MemNode::ValueIn);
duke@0 2153 if( val->Opcode() == Op_AndI ) {
duke@0 2154 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2155 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@0 2156 set_req(MemNode::ValueIn, val->in(1));
duke@0 2157 return this;
duke@0 2158 }
duke@0 2159 }
duke@0 2160 return NULL;
duke@0 2161 }
duke@0 2162
duke@0 2163
duke@0 2164 //------------------------------Ideal_sign_extended_input----------------------
duke@0 2165 // Check for useless sign-extension before a partial-word store
duke@0 2166 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@0 2167 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@0 2168 // (StoreB ... (valIn) )
duke@0 2169 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@0 2170 Node *val = in(MemNode::ValueIn);
duke@0 2171 if( val->Opcode() == Op_RShiftI ) {
duke@0 2172 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2173 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@0 2174 Node *shl = val->in(1);
duke@0 2175 if( shl->Opcode() == Op_LShiftI ) {
duke@0 2176 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@0 2177 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@0 2178 set_req(MemNode::ValueIn, shl->in(1));
duke@0 2179 return this;
duke@0 2180 }
duke@0 2181 }
duke@0 2182 }
duke@0 2183 }
duke@0 2184 return NULL;
duke@0 2185 }
duke@0 2186
duke@0 2187 //------------------------------value_never_loaded-----------------------------------
duke@0 2188 // Determine whether there are any possible loads of the value stored.
duke@0 2189 // For simplicity, we actually check if there are any loads from the
duke@0 2190 // address stored to, not just for loads of the value stored by this node.
duke@0 2191 //
duke@0 2192 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@0 2193 Node *adr = in(Address);
duke@0 2194 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@0 2195 if (adr_oop == NULL)
duke@0 2196 return false;
kvn@223 2197 if (!adr_oop->is_known_instance_field())
duke@0 2198 return false; // if not a distinct instance, there may be aliases of the address
duke@0 2199 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@0 2200 Node *use = adr->fast_out(i);
duke@0 2201 int opc = use->Opcode();
duke@0 2202 if (use->is_Load() || use->is_LoadStore()) {
duke@0 2203 return false;
duke@0 2204 }
duke@0 2205 }
duke@0 2206 return true;
duke@0 2207 }
duke@0 2208
duke@0 2209 //=============================================================================
duke@0 2210 //------------------------------Ideal------------------------------------------
duke@0 2211 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2212 // we can skip the AND operation. If the store is from a sign-extension
duke@0 2213 // (a left shift, then right shift) we can skip both.
duke@0 2214 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2215 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@0 2216 if( progress != NULL ) return progress;
duke@0 2217
duke@0 2218 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@0 2219 if( progress != NULL ) return progress;
duke@0 2220
duke@0 2221 // Finally check the default case
duke@0 2222 return StoreNode::Ideal(phase, can_reshape);
duke@0 2223 }
duke@0 2224
duke@0 2225 //=============================================================================
duke@0 2226 //------------------------------Ideal------------------------------------------
duke@0 2227 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2228 // we can skip the AND operation
duke@0 2229 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2230 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@0 2231 if( progress != NULL ) return progress;
duke@0 2232
duke@0 2233 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@0 2234 if( progress != NULL ) return progress;
duke@0 2235
duke@0 2236 // Finally check the default case
duke@0 2237 return StoreNode::Ideal(phase, can_reshape);
duke@0 2238 }
duke@0 2239
duke@0 2240 //=============================================================================
duke@0 2241 //------------------------------Identity---------------------------------------
duke@0 2242 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@0 2243 // No need to card mark when storing a null ptr
duke@0 2244 Node* my_store = in(MemNode::OopStore);
duke@0 2245 if (my_store->is_Store()) {
duke@0 2246 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@0 2247 if( t1 == TypePtr::NULL_PTR ) {
duke@0 2248 return in(MemNode::Memory);
duke@0 2249 }
duke@0 2250 }
duke@0 2251 return this;
duke@0 2252 }
duke@0 2253
duke@0 2254 //------------------------------Value-----------------------------------------
duke@0 2255 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@43 2256 // Either input is TOP ==> the result is TOP
kvn@43 2257 const Type *t = phase->type( in(MemNode::Memory) );
kvn@43 2258 if( t == Type::TOP ) return Type::TOP;
kvn@43 2259 t = phase->type( in(MemNode::Address) );
kvn@43 2260 if( t == Type::TOP ) return Type::TOP;
kvn@43 2261 t = phase->type( in(MemNode::ValueIn) );
kvn@43 2262 if( t == Type::TOP ) return Type::TOP;
duke@0 2263 // If extra input is TOP ==> the result is TOP
kvn@43 2264 t = phase->type( in(MemNode::OopStore) );
kvn@43 2265 if( t == Type::TOP ) return Type::TOP;
duke@0 2266
duke@0 2267 return StoreNode::Value( phase );
duke@0 2268 }
duke@0 2269
duke@0 2270
duke@0 2271 //=============================================================================
duke@0 2272 //----------------------------------SCMemProjNode------------------------------
duke@0 2273 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@0 2274 {
duke@0 2275 return bottom_type();
duke@0 2276 }
duke@0 2277
duke@0 2278 //=============================================================================
duke@0 2279 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
duke@0 2280 init_req(MemNode::Control, c );
duke@0 2281 init_req(MemNode::Memory , mem);
duke@0 2282 init_req(MemNode::Address, adr);
duke@0 2283 init_req(MemNode::ValueIn, val);
duke@0 2284 init_req( ExpectedIn, ex );
duke@0 2285 init_class_id(Class_LoadStore);
duke@0 2286
duke@0 2287 }
duke@0 2288
duke@0 2289 //=============================================================================
duke@0 2290 //-------------------------------adr_type--------------------------------------
duke@0 2291 // Do we Match on this edge index or not? Do not match memory
duke@0 2292 const TypePtr* ClearArrayNode::adr_type() const {
duke@0 2293 Node *adr = in(3);
duke@0 2294 return MemNode::calculate_adr_type(adr->bottom_type());
duke@0 2295 }
duke@0 2296
duke@0 2297 //------------------------------match_edge-------------------------------------
duke@0 2298 // Do we Match on this edge index or not? Do not match memory
duke@0 2299 uint ClearArrayNode::match_edge(uint idx) const {
duke@0 2300 return idx > 1;
duke@0 2301 }
duke@0 2302
duke@0 2303 //------------------------------Identity---------------------------------------
duke@0 2304 // Clearing a zero length array does nothing
duke@0 2305 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@68 2306 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@0 2307 }
duke@0 2308
duke@0 2309 //------------------------------Idealize---------------------------------------
duke@0 2310 // Clearing a short array is faster with stores
duke@0 2311 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2312 const int unit = BytesPerLong;
duke@0 2313 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@0 2314 if (!t) return NULL;
duke@0 2315 if (!t->is_con()) return NULL;
duke@0 2316 intptr_t raw_count = t->get_con();
duke@0 2317 intptr_t size = raw_count;
duke@0 2318 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@0 2319 // Clearing nothing uses the Identity call.
duke@0 2320 // Negative clears are possible on dead ClearArrays
duke@0 2321 // (see jck test stmt114.stmt11402.val).
duke@0 2322 if (size <= 0 || size % unit != 0) return NULL;
duke@0 2323 intptr_t count = size / unit;
duke@0 2324 // Length too long; use fast hardware clear
duke@0 2325 if (size > Matcher::init_array_short_size) return NULL;
duke@0 2326 Node *mem = in(1);
duke@0 2327 if( phase->type(mem)==Type::TOP ) return NULL;
duke@0 2328 Node *adr = in(3);
duke@0 2329 const Type* at = phase->type(adr);
duke@0 2330 if( at==Type::TOP ) return NULL;
duke@0 2331 const TypePtr* atp = at->isa_ptr();
duke@0 2332 // adjust atp to be the correct array element address type
duke@0 2333 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@0 2334 else atp = atp->add_offset(Type::OffsetBot);
duke@0 2335 // Get base for derived pointer purposes
duke@0 2336 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@0 2337 Node *base = adr->in(1);
duke@0 2338
duke@0 2339 Node *zero = phase->makecon(TypeLong::ZERO);
duke@0 2340 Node *off = phase->MakeConX(BytesPerLong);
duke@0 2341 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2342 count--;
duke@0 2343 while( count-- ) {
duke@0 2344 mem = phase->transform(mem);
duke@0 2345 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
duke@0 2346 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2347 }
duke@0 2348 return mem;
duke@0 2349 }
duke@0 2350
duke@0 2351 //----------------------------clear_memory-------------------------------------
duke@0 2352 // Generate code to initialize object storage to zero.
duke@0 2353 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2354 intptr_t start_offset,
duke@0 2355 Node* end_offset,
duke@0 2356 PhaseGVN* phase) {
duke@0 2357 Compile* C = phase->C;
duke@0 2358 intptr_t offset = start_offset;
duke@0 2359
duke@0 2360 int unit = BytesPerLong;
duke@0 2361 if ((offset % unit) != 0) {
duke@0 2362 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
duke@0 2363 adr = phase->transform(adr);
duke@0 2364 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2365 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2366 mem = phase->transform(mem);
duke@0 2367 offset += BytesPerInt;
duke@0 2368 }
duke@0 2369 assert((offset % unit) == 0, "");
duke@0 2370
duke@0 2371 // Initialize the remaining stuff, if any, with a ClearArray.
duke@0 2372 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@0 2373 }
duke@0 2374
duke@0 2375 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2376 Node* start_offset,
duke@0 2377 Node* end_offset,
duke@0 2378 PhaseGVN* phase) {
never@68 2379 if (start_offset == end_offset) {
never@68 2380 // nothing to do
never@68 2381 return mem;
never@68 2382 }
never@68 2383
duke@0 2384 Compile* C = phase->C;
duke@0 2385 int unit = BytesPerLong;
duke@0 2386 Node* zbase = start_offset;
duke@0 2387 Node* zend = end_offset;
duke@0 2388
duke@0 2389 // Scale to the unit required by the CPU:
duke@0 2390 if (!Matcher::init_array_count_is_in_bytes) {
duke@0 2391 Node* shift = phase->intcon(exact_log2(unit));
duke@0 2392 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
duke@0 2393 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
duke@0 2394 }
duke@0 2395
duke@0 2396 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
duke@0 2397 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
duke@0 2398
duke@0 2399 // Bulk clear double-words
duke@0 2400 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
duke@0 2401 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
duke@0 2402 return phase->transform(mem);
duke@0 2403 }
duke@0 2404
duke@0 2405 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2406 intptr_t start_offset,
duke@0 2407 intptr_t end_offset,
duke@0 2408 PhaseGVN* phase) {
never@68 2409 if (start_offset == end_offset) {
never@68 2410 // nothing to do
never@68 2411 return mem;
never@68 2412 }
never@68 2413
duke@0 2414 Compile* C = phase->C;
duke@0 2415 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@0 2416 intptr_t done_offset = end_offset;
duke@0 2417 if ((done_offset % BytesPerLong) != 0) {
duke@0 2418 done_offset -= BytesPerInt;
duke@0 2419 }
duke@0 2420 if (done_offset > start_offset) {
duke@0 2421 mem = clear_memory(ctl, mem, dest,
duke@0 2422 start_offset, phase->MakeConX(done_offset), phase);
duke@0 2423 }
duke@0 2424 if (done_offset < end_offset) { // emit the final 32-bit store
duke@0 2425 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@0 2426 adr = phase->transform(adr);
duke@0 2427 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2428 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2429 mem = phase->transform(mem);
duke@0 2430 done_offset += BytesPerInt;
duke@0 2431 }
duke@0 2432 assert(done_offset == end_offset, "");
duke@0 2433 return mem;
duke@0 2434 }
duke@0 2435
duke@0 2436 //=============================================================================
duke@0 2437 // Do we match on this edge? No memory edges
duke@0 2438 uint StrCompNode::match_edge(uint idx) const {
duke@0 2439 return idx == 5 || idx == 6;
duke@0 2440 }
duke@0 2441
duke@0 2442 //------------------------------Ideal------------------------------------------
duke@0 2443 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2444 // control copies
duke@0 2445 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2446 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2447 }
duke@0 2448
rasbold@169 2449 //------------------------------Ideal------------------------------------------
rasbold@169 2450 // Return a node which is more "ideal" than the current node. Strip out
rasbold@169 2451 // control copies
rasbold@169 2452 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
rasbold@169 2453 return remove_dead_region(phase, can_reshape) ? this : NULL;
rasbold@169 2454 }
rasbold@169 2455
duke@0 2456
duke@0 2457 //=============================================================================
duke@0 2458 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@0 2459 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@0 2460 _adr_type(C->get_adr_type(alias_idx))
duke@0 2461 {
duke@0 2462 init_class_id(Class_MemBar);
duke@0 2463 Node* top = C->top();
duke@0 2464 init_req(TypeFunc::I_O,top);
duke@0 2465 init_req(TypeFunc::FramePtr,top);
duke@0 2466 init_req(TypeFunc::ReturnAdr,top);
duke@0 2467 if (precedent != NULL)
duke@0 2468 init_req(TypeFunc::Parms, precedent);
duke@0 2469 }
duke@0 2470
duke@0 2471 //------------------------------cmp--------------------------------------------
duke@0 2472 uint MemBarNode::hash() const { return NO_HASH; }
duke@0 2473 uint MemBarNode::cmp( const Node &n ) const {
duke@0 2474 return (&n == this); // Always fail except on self
duke@0 2475 }
duke@0 2476
duke@0 2477 //------------------------------make-------------------------------------------
duke@0 2478 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@0 2479 int len = Precedent + (pn == NULL? 0: 1);
duke@0 2480 switch (opcode) {
duke@0 2481 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
duke@0 2482 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
duke@0 2483 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
duke@0 2484 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
duke@0 2485 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
duke@0 2486 default: ShouldNotReachHere(); return NULL;
duke@0 2487 }
duke@0 2488 }
duke@0 2489
duke@0 2490 //------------------------------Ideal------------------------------------------
duke@0 2491 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2492 // control copies
duke@0 2493 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@305 2494 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2495 }
duke@0 2496
duke@0 2497 //------------------------------Value------------------------------------------
duke@0 2498 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@0 2499 if( !in(0) ) return Type::TOP;
duke@0 2500 if( phase->type(in(0)) == Type::TOP )
duke@0 2501 return Type::TOP;
duke@0 2502 return TypeTuple::MEMBAR;
duke@0 2503 }
duke@0 2504
duke@0 2505 //------------------------------match------------------------------------------
duke@0 2506 // Construct projections for memory.
duke@0 2507 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@0 2508 switch (proj->_con) {
duke@0 2509 case TypeFunc::Control:
duke@0 2510 case TypeFunc::Memory:
duke@0 2511 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@0 2512 }
duke@0 2513 ShouldNotReachHere();
duke@0 2514 return NULL;
duke@0 2515 }
duke@0 2516
duke@0 2517 //===========================InitializeNode====================================
duke@0 2518 // SUMMARY:
duke@0 2519 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@0 2520 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@0 2521 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@0 2522 // It can coalesce related raw stores into larger units (called 'tiles').
duke@0 2523 // It can avoid zeroing new storage for memory units which have raw inits.
duke@0 2524 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@0 2525 //
duke@0 2526 // EXAMPLE:
duke@0 2527 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@0 2528 // ctl = incoming control; mem* = incoming memory
duke@0 2529 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@0 2530 // First allocate uninitialized memory and fill in the header:
duke@0 2531 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@0 2532 // ctl := alloc.Control; mem* := alloc.Memory*
duke@0 2533 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@0 2534 // Then initialize to zero the non-header parts of the raw memory block:
duke@0 2535 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@0 2536 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@0 2537 // After the initialize node executes, the object is ready for service:
duke@0 2538 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@0 2539 // Suppose its body is immediately initialized as {1,2}:
duke@0 2540 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2541 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2542 // mem.SLICE(#short[*]) := store2
duke@0 2543 //
duke@0 2544 // DETAILS:
duke@0 2545 // An InitializeNode collects and isolates object initialization after
duke@0 2546 // an AllocateNode and before the next possible safepoint. As a
duke@0 2547 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@0 2548 // down past any safepoint or any publication of the allocation.
duke@0 2549 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@0 2550 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@0 2551 //
duke@0 2552 // The semantics of the InitializeNode include an implicit zeroing of
duke@0 2553 // the new object from object header to the end of the object.
duke@0 2554 // (The object header and end are determined by the AllocateNode.)
duke@0 2555 //
duke@0 2556 // Certain stores may be added as direct inputs to the InitializeNode.
duke@0 2557 // These stores must update raw memory, and they must be to addresses
duke@0 2558 // derived from the raw address produced by AllocateNode, and with
duke@0 2559 // a constant offset. They must be ordered by increasing offset.
duke@0 2560 // The first one is at in(RawStores), the last at in(req()-1).
duke@0 2561 // Unlike most memory operations, they are not linked in a chain,
duke@0 2562 // but are displayed in parallel as users of the rawmem output of
duke@0 2563 // the allocation.
duke@0 2564 //
duke@0 2565 // (See comments in InitializeNode::capture_store, which continue
duke@0 2566 // the example given above.)
duke@0 2567 //
duke@0 2568 // When the associated Allocate is macro-expanded, the InitializeNode
duke@0 2569 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@0 2570 // may also be created at that point to represent any required zeroing.
duke@0 2571 // The InitializeNode is then marked 'complete', prohibiting further
duke@0 2572 // capturing of nearby memory operations.
duke@0 2573 //
duke@0 2574 // During macro-expansion, all captured initializations which store
duke@0 2575 // constant values of 32 bits or smaller are coalesced (if advantagous)
duke@0 2576 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@0 2577 // initialized in fewer memory operations. Memory words which are
duke@0 2578 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@0 2579 // by explicit stores of zero. (The code shape happens to do all
duke@0 2580 // zeroing first, then all other stores, with both sequences occurring
duke@0 2581 // in order of ascending offsets.)
duke@0 2582 //
duke@0 2583 // Alternatively, code may be inserted between an AllocateNode and its
duke@0 2584 // InitializeNode, to perform arbitrary initialization of the new object.
duke@0 2585 // E.g., the object copying intrinsics insert complex data transfers here.
duke@0 2586 // The initialization must then be marked as 'complete' disable the
duke@0 2587 // built-in zeroing semantics and the collection of initializing stores.
duke@0 2588 //
duke@0 2589 // While an InitializeNode is incomplete, reads from the memory state
duke@0 2590 // produced by it are optimizable if they match the control edge and
duke@0 2591 // new oop address associated with the allocation/initialization.
duke@0 2592 // They return a stored value (if the offset matches) or else zero.
duke@0 2593 // A write to the memory state, if it matches control and address,
duke@0 2594 // and if it is to a constant offset, may be 'captured' by the
duke@0 2595 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@0 2596 // inside the initialization, to the raw oop produced by the allocation.
duke@0 2597 // Operations on addresses which are provably distinct (e.g., to
duke@0 2598 // other AllocateNodes) are allowed to bypass the initialization.
duke@0 2599 //
duke@0 2600 // The effect of all this is to consolidate object initialization
duke@0 2601 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@0 2602 // single location, where it can be optimized as a unit.
duke@0 2603 //
duke@0 2604 // Only stores with an offset less than TrackedInitializationLimit words
duke@0 2605 // will be considered for capture by an InitializeNode. This puts a
duke@0 2606 // reasonable limit on the complexity of optimized initializations.
duke@0 2607
duke@0 2608 //---------------------------InitializeNode------------------------------------
duke@0 2609 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
duke@0 2610 : _is_complete(false),
duke@0 2611 MemBarNode(C, adr_type, rawoop)
duke@0 2612 {
duke@0 2613 init_class_id(Class_Initialize);
duke@0 2614
duke@0 2615 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@0 2616 assert(in(RawAddress) == rawoop, "proper init");
duke@0 2617 // Note: allocation() can be NULL, for secondary initialization barriers
duke@0 2618 }
duke@0 2619
duke@0 2620 // Since this node is not matched, it will be processed by the
duke@0 2621 // register allocator. Declare that there are no constraints
duke@0 2622 // on the allocation of the RawAddress edge.
duke@0 2623 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@0 2624 // This edge should be set to top, by the set_complete. But be conservative.
duke@0 2625 if (idx == InitializeNode::RawAddress)
duke@0 2626 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@0 2627 return RegMask::Empty;
duke@0 2628 }
duke@0 2629
duke@0 2630 Node* InitializeNode::memory(uint alias_idx) {
duke@0 2631 Node* mem = in(Memory);
duke@0 2632 if (mem->is_MergeMem()) {
duke@0 2633 return mem->as_MergeMem()->memory_at(alias_idx);
duke@0 2634 } else {
duke@0 2635 // incoming raw memory is not split
duke@0 2636 return mem;
duke@0 2637 }
duke@0 2638 }
duke@0 2639
duke@0 2640 bool InitializeNode::is_non_zero() {
duke@0 2641 if (is_complete()) return false;
duke@0 2642 remove_extra_zeroes();
duke@0 2643 return (req() > RawStores);
duke@0 2644 }
duke@0 2645
duke@0 2646 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@0 2647 assert(!is_complete(), "caller responsibility");
duke@0 2648 _is_complete = true;
duke@0 2649
duke@0 2650 // After this node is complete, it contains a bunch of
duke@0 2651 // raw-memory initializations. There is no need for
duke@0 2652 // it to have anything to do with non-raw memory effects.
duke@0 2653 // Therefore, tell all non-raw users to re-optimize themselves,
duke@0 2654 // after skipping the memory effects of this initialization.
duke@0 2655 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@0 2656 if (igvn) igvn->add_users_to_worklist(this);
duke@0 2657 }
duke@0 2658
duke@0 2659 // convenience function
duke@0 2660 // return false if the init contains any stores already
duke@0 2661 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@0 2662 InitializeNode* init = initialization();
duke@0 2663 if (init == NULL || init->is_complete()) return false;
duke@0 2664 init->remove_extra_zeroes();
duke@0 2665 // for now, if this allocation has already collected any inits, bail:
duke@0 2666 if (init->is_non_zero()) return false;
duke@0 2667 init->set_complete(phase);
duke@0 2668 return true;
duke@0 2669 }
duke@0 2670
duke@0 2671 void InitializeNode::remove_extra_zeroes() {
duke@0 2672 if (req() == RawStores) return;
duke@0 2673 Node* zmem = zero_memory();
duke@0 2674 uint fill = RawStores;
duke@0 2675 for (uint i = fill; i < req(); i++) {
duke@0 2676 Node* n = in(i);
duke@0 2677 if (n->is_top() || n == zmem) continue; // skip
duke@0 2678 if (fill < i) set_req(fill, n); // compact
duke@0 2679 ++fill;
duke@0 2680 }
duke@0 2681 // delete any empty spaces created:
duke@0 2682 while (fill < req()) {
duke@0 2683 del_req(fill);
duke@0 2684 }
duke@0 2685 }
duke@0 2686
duke@0 2687 // Helper for remembering which stores go with which offsets.
duke@0 2688 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@0 2689 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@0 2690 intptr_t offset = -1;
duke@0 2691 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@0 2692 phase, offset);
duke@0 2693 if (base == NULL) return -1; // something is dead,
duke@0 2694 if (offset < 0) return -1; // dead, dead
duke@0 2695 return offset;
duke@0 2696 }
duke@0 2697
duke@0 2698 // Helper for proving that an initialization expression is
duke@0 2699 // "simple enough" to be folded into an object initialization.
duke@0 2700 // Attempts to prove that a store's initial value 'n' can be captured
duke@0 2701 // within the initialization without creating a vicious cycle, such as:
duke@0 2702 // { Foo p = new Foo(); p.next = p; }
duke@0 2703 // True for constants and parameters and small combinations thereof.
duke@0 2704 bool InitializeNode::detect_init_independence(Node* n,
duke@0 2705 bool st_is_pinned,
duke@0 2706 int& count) {
duke@0 2707 if (n == NULL) return true; // (can this really happen?)
duke@0 2708 if (n->is_Proj()) n = n->in(0);
duke@0 2709 if (n == this) return false; // found a cycle
duke@0 2710 if (n->is_Con()) return true;
duke@0 2711 if (n->is_Start()) return true; // params, etc., are OK
duke@0 2712 if (n->is_Root()) return true; // even better
duke@0 2713
duke@0 2714 Node* ctl = n->in(0);
duke@0 2715 if (ctl != NULL && !ctl->is_top()) {
duke@0 2716 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@0 2717 if (ctl == this) return false;
duke@0 2718
duke@0 2719 // If we already know that the enclosing memory op is pinned right after
duke@0 2720 // the init, then any control flow that the store has picked up
duke@0 2721 // must have preceded the init, or else be equal to the init.
duke@0 2722 // Even after loop optimizations (which might change control edges)
duke@0 2723 // a store is never pinned *before* the availability of its inputs.
kvn@119 2724 if (!MemNode::all_controls_dominate(n, this))
duke@0 2725 return false; // failed to prove a good control
duke@0 2726
duke@0 2727 }
duke@0 2728
duke@0 2729 // Check data edges for possible dependencies on 'this'.
duke@0 2730 if ((count += 1) > 20) return false; // complexity limit
duke@0 2731 for (uint i = 1; i < n->req(); i++) {
duke@0 2732 Node* m = n->in(i);
duke@0 2733 if (m == NULL || m == n || m->is_top()) continue;
duke@0 2734 uint first_i = n->find_edge(m);
duke@0 2735 if (i != first_i) continue; // process duplicate edge just once
duke@0 2736 if (!detect_init_independence(m, st_is_pinned, count)) {
duke@0 2737 return false;
duke@0 2738 }
duke@0 2739 }
duke@0 2740
duke@0 2741 return true;
duke@0 2742 }
duke@0 2743
duke@0 2744 // Here are all the checks a Store must pass before it can be moved into
duke@0 2745 // an initialization. Returns zero if a check fails.
duke@0 2746 // On success, returns the (constant) offset to which the store applies,
duke@0 2747 // within the initialized memory.
duke@0 2748 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
duke@0 2749 const int FAIL = 0;
duke@0 2750 if (st->req() != MemNode::ValueIn + 1)
duke@0 2751 return FAIL; // an inscrutable StoreNode (card mark?)
duke@0 2752 Node* ctl = st->in(MemNode::Control);
duke@0 2753 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@0 2754 return FAIL; // must be unconditional after the initialization
duke@0 2755 Node* mem = st->in(MemNode::Memory);
duke@0 2756 if (!(mem->is_Proj() && mem->in(0) == this))
duke@0 2757 return FAIL; // must not be preceded by other stores
duke@0 2758 Node* adr = st->in(MemNode::Address);
duke@0 2759 intptr_t offset;
duke@0 2760 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@0 2761 if (alloc == NULL)
duke@0 2762 return FAIL; // inscrutable address
duke@0 2763 if (alloc != allocation())
duke@0 2764 return FAIL; // wrong allocation! (store needs to float up)
duke@0 2765 Node* val = st->in(MemNode::ValueIn);
duke@0 2766 int complexity_count = 0;
duke@0 2767 if (!detect_init_independence(val, true, complexity_count))
duke@0 2768 return FAIL; // stored value must be 'simple enough'
duke@0 2769
duke@0 2770 return offset; // success
duke@0 2771 }
duke@0 2772
duke@0 2773 // Find the captured store in(i) which corresponds to the range
duke@0 2774 // [start..start+size) in the initialized object.
duke@0 2775 // If there is one, return its index i. If there isn't, return the
duke@0 2776 // negative of the index where it should be inserted.
duke@0 2777 // Return 0 if the queried range overlaps an initialization boundary
duke@0 2778 // or if dead code is encountered.
duke@0 2779 // If size_in_bytes is zero, do not bother with overlap checks.
duke@0 2780 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@0 2781 int size_in_bytes,
duke@0 2782 PhaseTransform* phase) {
duke@0 2783 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@0 2784
duke@0 2785 if (is_complete())
duke@0 2786 return FAIL; // arraycopy got here first; punt
duke@0 2787
duke@0 2788 assert(allocation() != NULL, "must be present");
duke@0 2789
duke@0 2790 // no negatives, no header fields:
coleenp@113 2791 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@0 2792
duke@0 2793 // after a certain size, we bail out on tracking all the stores:
duke@0 2794 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2795 if (start >= ti_limit) return FAIL;
duke@0 2796
duke@0 2797 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@0 2798 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@0 2799
duke@0 2800 Node* st = in(i);
duke@0 2801 intptr_t st_off = get_store_offset(st, phase);
duke@0 2802 if (st_off < 0) {
duke@0 2803 if (st != zero_memory()) {
duke@0 2804 return FAIL; // bail out if there is dead garbage
duke@0 2805 }
duke@0 2806 } else if (st_off > start) {
duke@0 2807 // ...we are done, since stores are ordered
duke@0 2808 if (st_off < start + size_in_bytes) {
duke@0 2809 return FAIL; // the next store overlaps
duke@0 2810 }
duke@0 2811 return -(int)i; // not found; here is where to put it
duke@0 2812 } else if (st_off < start) {
duke@0 2813 if (size_in_bytes != 0 &&
duke@0 2814 start < st_off + MAX_STORE &&
duke@0 2815 start < st_off + st->as_Store()->memory_size()) {
duke@0 2816 return FAIL; // the previous store overlaps
duke@0 2817 }
duke@0 2818 } else {
duke@0 2819 if (size_in_bytes != 0 &&
duke@0 2820 st->as_Store()->memory_size() != size_in_bytes) {
duke@0 2821 return FAIL; // mismatched store size
duke@0 2822 }
duke@0 2823 return i;
duke@0 2824 }
duke@0 2825
duke@0 2826 ++i;
duke@0 2827 }
duke@0 2828 }
duke@0 2829
duke@0 2830 // Look for a captured store which initializes at the offset 'start'
duke@0 2831 // with the given size. If there is no such store, and no other
duke@0 2832 // initialization interferes, then return zero_memory (the memory
duke@0 2833 // projection of the AllocateNode).
duke@0 2834 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@0 2835 PhaseTransform* phase) {
duke@0 2836 assert(stores_are_sane(phase), "");
duke@0 2837 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2838 if (i == 0) {
duke@0 2839 return NULL; // something is dead
duke@0 2840 } else if (i < 0) {
duke@0 2841 return zero_memory(); // just primordial zero bits here
duke@0 2842 } else {
duke@0 2843 Node* st = in(i); // here is the store at this position
duke@0 2844 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@0 2845 return st;
duke@0 2846 }
duke@0 2847 }
duke@0 2848
duke@0 2849 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@0 2850 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@0 2851 PhaseTransform* phase) {
duke@0 2852 Node* addr = in(RawAddress);
duke@0 2853 if (offset != 0) {
duke@0 2854 Compile* C = phase->C;
duke@0 2855 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
duke@0 2856 phase->MakeConX(offset)) );
duke@0 2857 }
duke@0 2858 return addr;
duke@0 2859 }
duke@0 2860
duke@0 2861 // Clone the given store, converting it into a raw store
duke@0 2862 // initializing a field or element of my new object.
duke@0 2863 // Caller is responsible for retiring the original store,
duke@0 2864 // with subsume_node or the like.
duke@0 2865 //
duke@0 2866 // From the example above InitializeNode::InitializeNode,
duke@0 2867 // here are the old stores to be captured:
duke@0 2868 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2869 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2870 //
duke@0 2871 // Here is the changed code; note the extra edges on init:
duke@0 2872 // alloc = (Allocate ...)
duke@0 2873 // rawoop = alloc.RawAddress
duke@0 2874 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@0 2875 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@0 2876 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@0 2877 // rawstore1 rawstore2)
duke@0 2878 //
duke@0 2879 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
duke@0 2880 PhaseTransform* phase) {
duke@0 2881 assert(stores_are_sane(phase), "");
duke@0 2882
duke@0 2883 if (start < 0) return NULL;
duke@0 2884 assert(can_capture_store(st, phase) == start, "sanity");
duke@0 2885
duke@0 2886 Compile* C = phase->C;
duke@0 2887 int size_in_bytes = st->memory_size();
duke@0 2888 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2889 if (i == 0) return NULL; // bail out
duke@0 2890 Node* prev_mem = NULL; // raw memory for the captured store
duke@0 2891 if (i > 0) {
duke@0 2892 prev_mem = in(i); // there is a pre-existing store under this one
duke@0 2893 set_req(i, C->top()); // temporarily disconnect it
duke@0 2894 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@0 2895 } else {
duke@0 2896 i = -i; // no pre-existing store
duke@0 2897 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@0 2898 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@0 2899 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@0 2900 else
duke@0 2901 ins_req(i, C->top()); // build a new edge
duke@0 2902 }
duke@0 2903 Node* new_st = st->clone();
duke@0 2904 new_st->set_req(MemNode::Control, in(Control));
duke@0 2905 new_st->set_req(MemNode::Memory, prev_mem);
duke@0 2906 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@0 2907 new_st = phase->transform(new_st);
duke@0 2908
duke@0 2909 // At this point, new_st might have swallowed a pre-existing store
duke@0 2910 // at the same offset, or perhaps new_st might have disappeared,
duke@0 2911 // if it redundantly stored the same value (or zero to fresh memory).
duke@0 2912
duke@0 2913 // In any case, wire it in:
duke@0 2914 set_req(i, new_st);
duke@0 2915
duke@0 2916 // The caller may now kill the old guy.
duke@0 2917 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@0 2918 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@0 2919 assert(!is_complete(), "");
duke@0 2920 return new_st;
duke@0 2921 }
duke@0 2922
duke@0 2923 static bool store_constant(jlong* tiles, int num_tiles,
duke@0 2924 intptr_t st_off, int st_size,
duke@0 2925 jlong con) {
duke@0 2926 if ((st_off & (st_size-1)) != 0)
duke@0 2927 return false; // strange store offset (assume size==2**N)
duke@0 2928 address addr = (address)tiles + st_off;
duke@0 2929 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@0 2930 switch (st_size) {
duke@0 2931 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@0 2932 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@0 2933 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@0 2934 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@0 2935 default: return false; // strange store size (detect size!=2**N here)
duke@0 2936 }
duke@0 2937 return true; // return success to caller
duke@0 2938 }
duke@0 2939
duke@0 2940 // Coalesce subword constants into int constants and possibly
duke@0 2941 // into long constants. The goal, if the CPU permits,
duke@0 2942 // is to initialize the object with a small number of 64-bit tiles.
duke@0 2943 // Also, convert floating-point constants to bit patterns.
duke@0 2944 // Non-constants are not relevant to this pass.
duke@0 2945 //
duke@0 2946 // In terms of the running example on InitializeNode::InitializeNode
duke@0 2947 // and InitializeNode::capture_store, here is the transformation
duke@0 2948 // of rawstore1 and rawstore2 into rawstore12:
duke@0 2949 // alloc = (Allocate ...)
duke@0 2950 // rawoop = alloc.RawAddress
duke@0 2951 // tile12 = 0x00010002
duke@0 2952 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@0 2953 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@0 2954 //
duke@0 2955 void
duke@0 2956 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@0 2957 Node* size_in_bytes,
duke@0 2958 PhaseGVN* phase) {
duke@0 2959 Compile* C = phase->C;
duke@0 2960
duke@0 2961 assert(stores_are_sane(phase), "");
duke@0 2962 // Note: After this pass, they are not completely sane,
duke@0 2963 // since there may be some overlaps.
duke@0 2964
duke@0 2965 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@0 2966
duke@0 2967 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2968 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@0 2969 size_limit = MIN2(size_limit, ti_limit);
duke@0 2970 size_limit = align_size_up(size_limit, BytesPerLong);
duke@0 2971 int num_tiles = size_limit / BytesPerLong;
duke@0 2972
duke@0 2973 // allocate space for the tile map:
duke@0 2974 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@0 2975 jlong tiles_buf[small_len];
duke@0 2976 Node* nodes_buf[small_len];
duke@0 2977 jlong inits_buf[small_len];
duke@0 2978 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@0 2979 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2980 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@0 2981 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@0 2982 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@0 2983 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2984 // tiles: exact bitwise model of all primitive constants
duke@0 2985 // nodes: last constant-storing node subsumed into the tiles model
duke@0 2986 // inits: which bytes (in each tile) are touched by any initializations
duke@0 2987
duke@0 2988 //// Pass A: Fill in the tile model with any relevant stores.
duke@0 2989
duke@0 2990 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@0 2991 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@0 2992 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@0 2993 Node* zmem = zero_memory(); // initially zero memory state
duke@0 2994 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 2995 Node* st = in(i);
duke@0 2996 intptr_t st_off = get_store_offset(st, phase);
duke@0 2997
duke@0 2998 // Figure out the store's offset and constant value:
duke@0 2999 if (st_off < header_size) continue; //skip (ignore header)
duke@0 3000 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@0 3001 int st_size = st->as_Store()->memory_size();
duke@0 3002 if (st_off + st_size > size_limit) break;
duke@0 3003
duke@0 3004 // Record which bytes are touched, whether by constant or not.
duke@0 3005 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@0 3006 continue; // skip (strange store size)
duke@0 3007
duke@0 3008 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@0 3009 if (!val->singleton()) continue; //skip (non-con store)
duke@0 3010 BasicType type = val->basic_type();
duke@0 3011
duke@0 3012 jlong con = 0;
duke@0 3013 switch (type) {
duke@0 3014 case T_INT: con = val->is_int()->get_con(); break;
duke@0 3015 case T_LONG: con = val->is_long()->get_con(); break;
duke@0 3016 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@0 3017 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@0 3018 default: continue; //skip (odd store type)
duke@0 3019 }
duke@0 3020
duke@0 3021 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@0 3022 st->Opcode() == Op_StoreL) {
duke@0 3023 continue; // This StoreL is already optimal.
duke@0 3024 }
duke@0 3025
duke@0 3026 // Store down the constant.
duke@0 3027 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@0 3028
duke@0 3029 intptr_t j = st_off >> LogBytesPerLong;
duke@0 3030
duke@0 3031 if (type == T_INT && st_size == BytesPerInt
duke@0 3032 && (st_off & BytesPerInt) == BytesPerInt) {
duke@0 3033 jlong lcon = tiles[j];
duke@0 3034 if (!Matcher::isSimpleConstant64(lcon) &&
duke@0 3035 st->Opcode() == Op_StoreI) {
duke@0 3036 // This StoreI is already optimal by itself.
duke@0 3037 jint* intcon = (jint*) &tiles[j];
duke@0 3038 intcon[1] = 0; // undo the store_constant()
duke@0 3039
duke@0 3040 // If the previous store is also optimal by itself, back up and
duke@0 3041 // undo the action of the previous loop iteration... if we can.
duke@0 3042 // But if we can't, just let the previous half take care of itself.
duke@0 3043 st = nodes[j];
duke@0 3044 st_off -= BytesPerInt;
duke@0 3045 con = intcon[0];
duke@0 3046 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@0 3047 assert(st_off >= header_size, "still ignoring header");
duke@0 3048 assert(get_store_offset(st, phase) == st_off, "must be");
duke@0 3049 assert(in(i-1) == zmem, "must be");
duke@0 3050 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@0 3051 assert(con == tcon->is_int()->get_con(), "must be");
duke@0 3052 // Undo the effects of the previous loop trip, which swallowed st:
duke@0 3053 intcon[0] = 0; // undo store_constant()
duke@0 3054 set_req(i-1, st); // undo set_req(i, zmem)
duke@0 3055 nodes[j] = NULL; // undo nodes[j] = st
duke@0 3056 --old_subword; // undo ++old_subword
duke@0 3057 }
duke@0 3058 continue; // This StoreI is already optimal.
duke@0 3059 }
duke@0 3060 }
duke@0 3061
duke@0 3062 // This store is not needed.
duke@0 3063 set_req(i, zmem);
duke@0 3064 nodes[j] = st; // record for the moment
duke@0 3065 if (st_size < BytesPerLong) // something has changed
duke@0 3066 ++old_subword; // includes int/float, but who's counting...
duke@0 3067 else ++old_long;
duke@0 3068 }
duke@0 3069
duke@0 3070 if ((old_subword + old_long) == 0)
duke@0 3071 return; // nothing more to do
duke@0 3072
duke@0 3073 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@0 3074 // Be sure to insert them before overlapping non-constant stores.
duke@0 3075 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@0 3076 for (int j = 0; j < num_tiles; j++) {
duke@0 3077 jlong con = tiles[j];
duke@0 3078 jlong init = inits[j];
duke@0 3079 if (con == 0) continue;
duke@0 3080 jint con0, con1; // split the constant, address-wise
duke@0 3081 jint init0, init1; // split the init map, address-wise
duke@0 3082 { union { jlong con; jint intcon[2]; } u;
duke@0 3083 u.con = con;
duke@0 3084 con0 = u.intcon[0];
duke@0 3085 con1 = u.intcon[1];
duke@0 3086 u.con = init;
duke@0 3087 init0 = u.intcon[0];
duke@0 3088 init1 = u.intcon[1];
duke@0 3089 }
duke@0 3090
duke@0 3091 Node* old = nodes[j];
duke@0 3092 assert(old != NULL, "need the prior store");
duke@0 3093 intptr_t offset = (j * BytesPerLong);
duke@0 3094
duke@0 3095 bool split = !Matcher::isSimpleConstant64(con);
duke@0 3096
duke@0 3097 if (offset < header_size) {
duke@0 3098 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@0 3099 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@0 3100 split = true; // only the second word counts
duke@0 3101 // Example: int a[] = { 42 ... }
duke@0 3102 } else if (con0 == 0 && init0 == -1) {
duke@0 3103 split = true; // first word is covered by full inits
duke@0 3104 // Example: int a[] = { ... foo(), 42 ... }
duke@0 3105 } else if (con1 == 0 && init1 == -1) {
duke@0 3106 split = true; // second word is covered by full inits
duke@0 3107 // Example: int a[] = { ... 42, foo() ... }
duke@0 3108 }
duke@0 3109
duke@0 3110 // Here's a case where init0 is neither 0 nor -1:
duke@0 3111 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@0 3112 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@0 3113 // In this case the tile is not split; it is (jlong)42.
duke@0 3114 // The big tile is stored down, and then the foo() value is inserted.
duke@0 3115 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@0 3116
duke@0 3117 Node* ctl = old->in(MemNode::Control);
duke@0 3118 Node* adr = make_raw_address(offset, phase);
duke@0 3119 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 3120
duke@0 3121 // One or two coalesced stores to plop down.
duke@0 3122 Node* st[2];
duke@0 3123 intptr_t off[2];
duke@0 3124 int nst = 0;
duke@0 3125 if (!split) {
duke@0 3126 ++new_long;
duke@0 3127 off[nst] = offset;
coleenp@113 3128 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3129 phase->longcon(con), T_LONG);
duke@0 3130 } else {
duke@0 3131 // Omit either if it is a zero.
duke@0 3132 if (con0 != 0) {
duke@0 3133 ++new_int;
duke@0 3134 off[nst] = offset;
coleenp@113 3135 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3136 phase->intcon(con0), T_INT);
duke@0 3137 }
duke@0 3138 if (con1 != 0) {
duke@0 3139 ++new_int;
duke@0 3140 offset += BytesPerInt;
duke@0 3141 adr = make_raw_address(offset, phase);
duke@0 3142 off[nst] = offset;
coleenp@113 3143 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3144 phase->intcon(con1), T_INT);
duke@0 3145 }
duke@0 3146 }
duke@0 3147
duke@0 3148 // Insert second store first, then the first before the second.
duke@0 3149 // Insert each one just before any overlapping non-constant stores.
duke@0 3150 while (nst > 0) {
duke@0 3151 Node* st1 = st[--nst];
duke@0 3152 C->copy_node_notes_to(st1, old);
duke@0 3153 st1 = phase->transform(st1);
duke@0 3154 offset = off[nst];
duke@0 3155 assert(offset >= header_size, "do not smash header");
duke@0 3156 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@0 3157 guarantee(ins_idx != 0, "must re-insert constant store");
duke@0 3158 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@0 3159 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@0 3160 set_req(--ins_idx, st1);
duke@0 3161 else
duke@0 3162 ins_req(ins_idx, st1);
duke@0 3163 }
duke@0 3164 }
duke@0 3165
duke@0 3166 if (PrintCompilation && WizardMode)
duke@0 3167 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@0 3168 old_subword, old_long, new_int, new_long);
duke@0 3169 if (C->log() != NULL)
duke@0 3170 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@0 3171 old_subword, old_long, new_int, new_long);
duke@0 3172
duke@0 3173 // Clean up any remaining occurrences of zmem:
duke@0 3174 remove_extra_zeroes();
duke@0 3175 }
duke@0 3176
duke@0 3177 // Explore forward from in(start) to find the first fully initialized
duke@0 3178 // word, and return its offset. Skip groups of subword stores which
duke@0 3179 // together initialize full words. If in(start) is itself part of a
duke@0 3180 // fully initialized word, return the offset of in(start). If there
duke@0 3181 // are no following full-word stores, or if something is fishy, return
duke@0 3182 // a negative value.
duke@0 3183 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@0 3184 int int_map = 0;
duke@0 3185 intptr_t int_map_off = 0;
duke@0 3186 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@0 3187
duke@0 3188 for (uint i = start, limit = req(); i < limit; i++) {
duke@0 3189 Node* st = in(i);
duke@0 3190
duke@0 3191 intptr_t st_off = get_store_offset(st, phase);
duke@0 3192 if (st_off < 0) break; // return conservative answer
duke@0 3193
duke@0 3194 int st_size = st->as_Store()->memory_size();
duke@0 3195 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@0 3196 return st_off; // we found a complete word init
duke@0 3197 }
duke@0 3198
duke@0 3199 // update the map:
duke@0 3200
duke@0 3201 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@0 3202 if (this_int_off != int_map_off) {
duke@0 3203 // reset the map:
duke@0 3204 int_map = 0;
duke@0 3205 int_map_off = this_int_off;
duke@0 3206 }
duke@0 3207
duke@0 3208 int subword_off = st_off - this_int_off;
duke@0 3209 int_map |= right_n_bits(st_size) << subword_off;
duke@0 3210 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@0 3211 return this_int_off; // we found a complete word init
duke@0 3212 }
duke@0 3213
duke@0 3214 // Did this store hit or cross the word boundary?
duke@0 3215 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@0 3216 if (next_int_off == this_int_off + BytesPerInt) {
duke@0 3217 // We passed the current int, without fully initializing it.
duke@0 3218 int_map_off = next_int_off;
duke@0 3219 int_map >>= BytesPerInt;
duke@0 3220 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@0 3221 // We passed the current and next int.
duke@0 3222 return this_int_off + BytesPerInt;
duke@0 3223 }
duke@0 3224 }
duke@0 3225
duke@0 3226 return -1;
duke@0 3227 }
duke@0 3228
duke@0 3229
duke@0 3230 // Called when the associated AllocateNode is expanded into CFG.
duke@0 3231 // At this point, we may perform additional optimizations.
duke@0 3232 // Linearize the stores by ascending offset, to make memory
duke@0 3233 // activity as coherent as possible.
duke@0 3234 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@0 3235 intptr_t header_size,
duke@0 3236 Node* size_in_bytes,
duke@0 3237 PhaseGVN* phase) {
duke@0 3238 assert(!is_complete(), "not already complete");
duke@0 3239 assert(stores_are_sane(phase), "");
duke@0 3240 assert(allocation() != NULL, "must be present");
duke@0 3241
duke@0 3242 remove_extra_zeroes();
duke@0 3243
duke@0 3244 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@0 3245 // reduce instruction count for common initialization patterns
duke@0 3246 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@0 3247
duke@0 3248 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3249 Node* inits = zmem; // accumulating a linearized chain of inits
duke@0 3250 #ifdef ASSERT
coleenp@113 3251 intptr_t first_offset = allocation()->minimum_header_size();
coleenp@113 3252 intptr_t last_init_off = first_offset; // previous init offset
coleenp@113 3253 intptr_t last_init_end = first_offset; // previous init offset+size
coleenp@113 3254 intptr_t last_tile_end = first_offset; // previous tile offset+size
duke@0 3255 #endif
duke@0 3256 intptr_t zeroes_done = header_size;
duke@0 3257
duke@0 3258 bool do_zeroing = true; // we might give up if inits are very sparse
duke@0 3259 int big_init_gaps = 0; // how many large gaps have we seen?
duke@0 3260
duke@0 3261 if (ZeroTLAB) do_zeroing = false;
duke@0 3262 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@0 3263
duke@0 3264 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3265 Node* st = in(i);
duke@0 3266 intptr_t st_off = get_store_offset(st, phase);
duke@0 3267 if (st_off < 0)
duke@0 3268 break; // unknown junk in the inits
duke@0 3269 if (st->in(MemNode::Memory) != zmem)
duke@0 3270 break; // complicated store chains somehow in list
duke@0 3271
duke@0 3272 int st_size = st->as_Store()->memory_size();
duke@0 3273 intptr_t next_init_off = st_off + st_size;
duke@0 3274
duke@0 3275 if (do_zeroing && zeroes_done < next_init_off) {
duke@0 3276 // See if this store needs a zero before it or under it.
duke@0 3277 intptr_t zeroes_needed = st_off;
duke@0 3278
duke@0 3279 if (st_size < BytesPerInt) {
duke@0 3280 // Look for subword stores which only partially initialize words.
duke@0 3281 // If we find some, we must lay down some word-level zeroes first,
duke@0 3282 // underneath the subword stores.
duke@0 3283 //
duke@0 3284 // Examples:
duke@0 3285 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
duke@0 3286 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
duke@0 3287 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
duke@0 3288 //
duke@0 3289 // Note: coalesce_subword_stores may have already done this,
duke@0 3290 // if it was prompted by constant non-zero subword initializers.
duke@0 3291 // But this case can still arise with non-constant stores.
duke@0 3292
duke@0 3293 intptr_t next_full_store = find_next_fullword_store(i, phase);
duke@0 3294
duke@0 3295 // In the examples above:
duke@0 3296 // in(i) p q r s x y z
duke@0 3297 // st_off 12 13 14 15 12 13 14
duke@0 3298 // st_size 1 1 1 1 1 1 1
duke@0 3299 // next_full_s. 12 16 16 16 16 16 16
duke@0 3300 // z's_done 12 16 16 16 12 16 12
duke@0 3301 // z's_needed 12 16 16 16 16 16 16
duke@0 3302 // zsize 0 0 0 0 4 0 4
duke@0 3303 if (next_full_store < 0) {
duke@0 3304 // Conservative tack: Zero to end of current word.
duke@0 3305 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
duke@0 3306 } else {
duke@0 3307 // Zero to beginning of next fully initialized word.
duke@0 3308 // Or, don't zero at all, if we are already in that word.
duke@0 3309 assert(next_full_store >= zeroes_needed, "must go forward");
duke@0 3310 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
duke@0 3311 zeroes_needed = next_full_store;
duke@0 3312 }
duke@0 3313 }
duke@0 3314
duke@0 3315 if (zeroes_needed > zeroes_done) {
duke@0 3316 intptr_t zsize = zeroes_needed - zeroes_done;
duke@0 3317 // Do some incremental zeroing on rawmem, in parallel with inits.
duke@0 3318 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3319 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3320 zeroes_done, zeroes_needed,
duke@0 3321 phase);
duke@0 3322 zeroes_done = zeroes_needed;
duke@0 3323 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
duke@0 3324 do_zeroing = false; // leave the hole, next time
duke@0 3325 }
duke@0 3326 }
duke@0 3327
duke@0 3328 // Collect the store and move on:
duke@0 3329 st->set_req(MemNode::Memory, inits);
duke@0 3330 inits = st; // put it on the linearized chain
duke@0 3331 set_req(i, zmem); // unhook from previous position
duke@0 3332
duke@0 3333 if (zeroes_done == st_off)
duke@0 3334 zeroes_done = next_init_off;
duke@0 3335
duke@0 3336 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
duke@0 3337
duke@0 3338 #ifdef ASSERT
duke@0 3339 // Various order invariants. Weaker than stores_are_sane because
duke@0 3340 // a large constant tile can be filled in by smaller non-constant stores.
duke@0 3341 assert(st_off >= last_init_off, "inits do not reverse");
duke@0 3342 last_init_off = st_off;
duke@0 3343 const Type* val = NULL;
duke@0 3344 if (st_size >= BytesPerInt &&
duke@0 3345 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
duke@0 3346 (int)val->basic_type() < (int)T_OBJECT) {
duke@0 3347 assert(st_off >= last_tile_end, "tiles do not overlap");
duke@0 3348 assert(st_off >= last_init_end, "tiles do not overwrite inits");
duke@0 3349 last_tile_end = MAX2(last_tile_end, next_init_off);
duke@0 3350 } else {
duke@0 3351 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
duke@0 3352 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
duke@0 3353 assert(st_off >= last_init_end, "inits do not overlap");
duke@0 3354 last_init_end = next_init_off; // it's a non-tile
duke@0 3355 }
duke@0 3356 #endif //ASSERT
duke@0 3357 }
duke@0 3358
duke@0 3359 remove_extra_zeroes(); // clear out all the zmems left over
duke@0 3360 add_req(inits);
duke@0 3361
duke@0 3362 if (!ZeroTLAB) {
duke@0 3363 // If anything remains to be zeroed, zero it all now.
duke@0 3364 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3365 // if it is the last unused 4 bytes of an instance, forget about it
duke@0 3366 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
duke@0 3367 if (zeroes_done + BytesPerLong >= size_limit) {
duke@0 3368 assert(allocation() != NULL, "");
duke@0 3369 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
duke@0 3370 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
duke@0 3371 if (zeroes_done == k->layout_helper())
duke@0 3372 zeroes_done = size_limit;
duke@0 3373 }
duke@0 3374 if (zeroes_done < size_limit) {
duke@0 3375 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3376 zeroes_done, size_in_bytes, phase);
duke@0 3377 }
duke@0 3378 }
duke@0 3379
duke@0 3380 set_complete(phase);
duke@0 3381 return rawmem;
duke@0 3382 }
duke@0 3383
duke@0 3384
duke@0 3385 #ifdef ASSERT
duke@0 3386 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
duke@0 3387 if (is_complete())
duke@0 3388 return true; // stores could be anything at this point
coleenp@113 3389 assert(allocation() != NULL, "must be present");
coleenp@113 3390 intptr_t last_off = allocation()->minimum_header_size();
duke@0 3391 for (uint i = InitializeNode::RawStores; i < req(); i++) {
duke@0 3392 Node* st = in(i);
duke@0 3393 intptr_t st_off = get_store_offset(st, phase);
duke@0 3394 if (st_off < 0) continue; // ignore dead garbage
duke@0 3395 if (last_off > st_off) {
duke@0 3396 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
duke@0 3397 this->dump(2);
duke@0 3398 assert(false, "ascending store offsets");
duke@0 3399 return false;
duke@0 3400 }
duke@0 3401 last_off = st_off + st->as_Store()->memory_size();
duke@0 3402 }
duke@0 3403 return true;
duke@0 3404 }
duke@0 3405 #endif //ASSERT
duke@0 3406
duke@0 3407
duke@0 3408
duke@0 3409
duke@0 3410 //============================MergeMemNode=====================================
duke@0 3411 //
duke@0 3412 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
duke@0 3413 // contributing store or call operations. Each contributor provides the memory
duke@0 3414 // state for a particular "alias type" (see Compile::alias_type). For example,
duke@0 3415 // if a MergeMem has an input X for alias category #6, then any memory reference
duke@0 3416 // to alias category #6 may use X as its memory state input, as an exact equivalent
duke@0 3417 // to using the MergeMem as a whole.
duke@0 3418 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
duke@0 3419 //
duke@0 3420 // (Here, the <N> notation gives the index of the relevant adr_type.)
duke@0 3421 //
duke@0 3422 // In one special case (and more cases in the future), alias categories overlap.
duke@0 3423 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
duke@0 3424 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
duke@0 3425 // it is exactly equivalent to that state W:
duke@0 3426 // MergeMem(<Bot>: W) <==> W
duke@0 3427 //
duke@0 3428 // Usually, the merge has more than one input. In that case, where inputs
duke@0 3429 // overlap (i.e., one is Bot), the narrower alias type determines the memory
duke@0 3430 // state for that type, and the wider alias type (Bot) fills in everywhere else:
duke@0 3431 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
duke@0 3432 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
duke@0 3433 //
duke@0 3434 // A merge can take a "wide" memory state as one of its narrow inputs.
duke@0 3435 // This simply means that the merge observes out only the relevant parts of
duke@0 3436 // the wide input. That is, wide memory states arriving at narrow merge inputs
duke@0 3437 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
duke@0 3438 //
duke@0 3439 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
duke@0 3440 // and that memory slices "leak through":
duke@0 3441 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
duke@0 3442 //
duke@0 3443 // But, in such a cascade, repeated memory slices can "block the leak":
duke@0 3444 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
duke@0 3445 //
duke@0 3446 // In the last example, Y is not part of the combined memory state of the
duke@0 3447 // outermost MergeMem. The system must, of course, prevent unschedulable
duke@0 3448 // memory states from arising, so you can be sure that the state Y is somehow
duke@0 3449 // a precursor to state Y'.
duke@0 3450 //
duke@0 3451 //
duke@0 3452 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
duke@0 3453 // of each MergeMemNode array are exactly the numerical alias indexes, including
duke@0 3454 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
duke@0 3455 // Compile::alias_type (and kin) produce and manage these indexes.
duke@0 3456 //
duke@0 3457 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
duke@0 3458 // (Note that this provides quick access to the top node inside MergeMem methods,
duke@0 3459 // without the need to reach out via TLS to Compile::current.)
duke@0 3460 //
duke@0 3461 // As a consequence of what was just described, a MergeMem that represents a full
duke@0 3462 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
duke@0 3463 // containing all alias categories.
duke@0 3464 //
duke@0 3465 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
duke@0 3466 //
duke@0 3467 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
duke@0 3468 // a memory state for the alias type <N>, or else the top node, meaning that
duke@0 3469 // there is no particular input for that alias type. Note that the length of
duke@0 3470 // a MergeMem is variable, and may be extended at any time to accommodate new
duke@0 3471 // memory states at larger alias indexes. When merges grow, they are of course
duke@0 3472 // filled with "top" in the unused in() positions.
duke@0 3473 //
duke@0 3474 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
duke@0 3475 // (Top was chosen because it works smoothly with passes like GCM.)
duke@0 3476 //
duke@0 3477 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
duke@0 3478 // the type of random VM bits like TLS references.) Since it is always the
duke@0 3479 // first non-Bot memory slice, some low-level loops use it to initialize an
duke@0 3480 // index variable: for (i = AliasIdxRaw; i < req(); i++).
duke@0 3481 //
duke@0 3482 //
duke@0 3483 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
duke@0 3484 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
duke@0 3485 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
duke@0 3486 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
duke@0 3487 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
duke@0 3488 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
duke@0 3489 //
duke@0 3490 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
duke@0 3491 // really that different from the other memory inputs. An abbreviation called
duke@0 3492 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
duke@0 3493 //
duke@0 3494 //
duke@0 3495 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
duke@0 3496 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
duke@0 3497 // that "emerges though" the base memory will be marked as excluding the alias types
duke@0 3498 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
duke@0 3499 //
duke@0 3500 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
duke@0 3501 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
duke@0 3502 //
duke@0 3503 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
duke@0 3504 // (It is currently unimplemented.) As you can see, the resulting merge is
duke@0 3505 // actually a disjoint union of memory states, rather than an overlay.
duke@0 3506 //
duke@0 3507
duke@0 3508 //------------------------------MergeMemNode-----------------------------------
duke@0 3509 Node* MergeMemNode::make_empty_memory() {
duke@0 3510 Node* empty_memory = (Node*) Compile::current()->top();
duke@0 3511 assert(empty_memory->is_top(), "correct sentinel identity");
duke@0 3512 return empty_memory;
duke@0 3513 }
duke@0 3514
duke@0 3515 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
duke@0 3516 init_class_id(Class_MergeMem);
duke@0 3517 // all inputs are nullified in Node::Node(int)
duke@0 3518 // set_input(0, NULL); // no control input
duke@0 3519
duke@0 3520 // Initialize the edges uniformly to top, for starters.
duke@0 3521 Node* empty_mem = make_empty_memory();
duke@0 3522 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
duke@0 3523 init_req(i,empty_mem);
duke@0 3524 }
duke@0 3525 assert(empty_memory() == empty_mem, "");
duke@0 3526
duke@0 3527 if( new_base != NULL && new_base->is_MergeMem() ) {
duke@0 3528 MergeMemNode* mdef = new_base->as_MergeMem();
duke@0 3529 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
duke@0 3530 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
duke@0 3531 mms.set_memory(mms.memory2());
duke@0 3532 }
duke@0 3533 assert(base_memory() == mdef->base_memory(), "");
duke@0 3534 } else {
duke@0 3535 set_base_memory(new_base);
duke@0 3536 }
duke@0 3537 }
duke@0 3538
duke@0 3539 // Make a new, untransformed MergeMem with the same base as 'mem'.
duke@0 3540 // If mem is itself a MergeMem, populate the result with the same edges.
duke@0 3541 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
duke@0 3542 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
duke@0 3543 }
duke@0 3544
duke@0 3545 //------------------------------cmp--------------------------------------------
duke@0 3546 uint MergeMemNode::hash() const { return NO_HASH; }
duke@0 3547 uint MergeMemNode::cmp( const Node &n ) const {
duke@0 3548 return (&n == this); // Always fail except on self
duke@0 3549 }
duke@0 3550
duke@0 3551 //------------------------------Identity---------------------------------------
duke@0 3552 Node* MergeMemNode::Identity(PhaseTransform *phase) {
duke@0 3553 // Identity if this merge point does not record any interesting memory
duke@0 3554 // disambiguations.
duke@0 3555 Node* base_mem = base_memory();
duke@0 3556 Node* empty_mem = empty_memory();
duke@0 3557 if (base_mem != empty_mem) { // Memory path is not dead?
duke@0 3558 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3559 Node* mem = in(i);
duke@0 3560 if (mem != empty_mem && mem != base_mem) {
duke@0 3561 return this; // Many memory splits; no change
duke@0 3562 }
duke@0 3563 }
duke@0 3564 }
duke@0 3565 return base_mem; // No memory splits; ID on the one true input
duke@0 3566 }
duke@0 3567
duke@0 3568 //------------------------------Ideal------------------------------------------
duke@0