annotate src/share/vm/opto/memnode.cpp @ 558:3b5ac9e7e6ea

6796746: rename LoadC (char) opcode class to LoadUS (unsigned short) Summary: Renaming LoadC to LoadUS would round up the planned introduction of LoadUB and LoadUI. Reviewed-by: phh, kvn
author twisti
date Mon, 26 Jan 2009 16:22:12 +0100
parents 041fe019d769
children dca06e7f503d
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
kvn@420 230 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@420 231 if( can_reshape && igvn != NULL && igvn->_worklist.member(address) ) {
kvn@420 232 // The address's base and type may change when the address is processed.
kvn@420 233 // Delay this mem node transformation until the address is processed.
kvn@420 234 phase->is_IterGVN()->_worklist.push(this);
kvn@420 235 return NodeSentinel; // caller will return NULL
kvn@420 236 }
kvn@420 237
duke@0 238 // Avoid independent memory operations
duke@0 239 Node* old_mem = mem;
duke@0 240
kvn@36 241 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@36 242 // initializations was removed. After macro-expansion all stores catched
kvn@36 243 // by Initialize node became raw stores and there is no information
kvn@36 244 // which memory slices they modify. So it is unsafe to move any memory
kvn@36 245 // operation above these stores. Also in most cases hooked non-raw memories
kvn@36 246 // were already unhooked by using information from detect_ptr_independence()
kvn@36 247 // and find_previous_store().
duke@0 248
duke@0 249 if (mem->is_MergeMem()) {
duke@0 250 MergeMemNode* mmem = mem->as_MergeMem();
duke@0 251 const TypePtr *tp = t_adr->is_ptr();
kvn@64 252
kvn@64 253 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@0 254 }
duke@0 255
duke@0 256 if (mem != old_mem) {
duke@0 257 set_req(MemNode::Memory, mem);
kvn@305 258 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
duke@0 259 return this;
duke@0 260 }
duke@0 261
duke@0 262 // let the subclass continue analyzing...
duke@0 263 return NULL;
duke@0 264 }
duke@0 265
duke@0 266 // Helper function for proving some simple control dominations.
kvn@119 267 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
duke@0 268 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@0 269 // is not a constant (dominated by the method's StartNode).
duke@0 270 // Used by MemNode::find_previous_store to prove that the
duke@0 271 // control input of a memory operation predates (dominates)
duke@0 272 // an allocation it wants to look past.
kvn@119 273 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
kvn@119 274 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
kvn@119 275 return false; // Conservative answer for dead code
kvn@119 276
kvn@193 277 // Check 'dom'. Skip Proj and CatchProj nodes.
kvn@119 278 dom = dom->find_exact_control(dom);
kvn@119 279 if (dom == NULL || dom->is_top())
kvn@119 280 return false; // Conservative answer for dead code
kvn@119 281
kvn@193 282 if (dom == sub) {
kvn@193 283 // For the case when, for example, 'sub' is Initialize and the original
kvn@193 284 // 'dom' is Proj node of the 'sub'.
kvn@193 285 return false;
kvn@193 286 }
kvn@193 287
kvn@155 288 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
kvn@119 289 return true;
kvn@119 290
kvn@119 291 // 'dom' dominates 'sub' if its control edge and control edges
kvn@119 292 // of all its inputs dominate or equal to sub's control edge.
kvn@119 293
kvn@119 294 // Currently 'sub' is either Allocate, Initialize or Start nodes.
kvn@163 295 // Or Region for the check in LoadNode::Ideal();
kvn@163 296 // 'sub' should have sub->in(0) != NULL.
kvn@163 297 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
kvn@163 298 sub->is_Region(), "expecting only these nodes");
kvn@119 299
kvn@119 300 // Get control edge of 'sub'.
kvn@193 301 Node* orig_sub = sub;
kvn@119 302 sub = sub->find_exact_control(sub->in(0));
kvn@119 303 if (sub == NULL || sub->is_top())
kvn@119 304 return false; // Conservative answer for dead code
kvn@119 305
kvn@119 306 assert(sub->is_CFG(), "expecting control");
kvn@119 307
kvn@119 308 if (sub == dom)
kvn@119 309 return true;
kvn@119 310
kvn@119 311 if (sub->is_Start() || sub->is_Root())
kvn@119 312 return false;
kvn@119 313
kvn@119 314 {
kvn@119 315 // Check all control edges of 'dom'.
kvn@119 316
kvn@119 317 ResourceMark rm;
kvn@119 318 Arena* arena = Thread::current()->resource_area();
kvn@119 319 Node_List nlist(arena);
kvn@119 320 Unique_Node_List dom_list(arena);
kvn@119 321
kvn@119 322 dom_list.push(dom);
kvn@119 323 bool only_dominating_controls = false;
kvn@119 324
kvn@119 325 for (uint next = 0; next < dom_list.size(); next++) {
kvn@119 326 Node* n = dom_list.at(next);
kvn@193 327 if (n == orig_sub)
kvn@193 328 return false; // One of dom's inputs dominated by sub.
kvn@119 329 if (!n->is_CFG() && n->pinned()) {
kvn@119 330 // Check only own control edge for pinned non-control nodes.
kvn@119 331 n = n->find_exact_control(n->in(0));
kvn@119 332 if (n == NULL || n->is_top())
kvn@119 333 return false; // Conservative answer for dead code
kvn@119 334 assert(n->is_CFG(), "expecting control");
kvn@193 335 dom_list.push(n);
kvn@193 336 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
kvn@119 337 only_dominating_controls = true;
kvn@119 338 } else if (n->is_CFG()) {
kvn@119 339 if (n->dominates(sub, nlist))
kvn@119 340 only_dominating_controls = true;
kvn@119 341 else
kvn@119 342 return false;
kvn@119 343 } else {
kvn@119 344 // First, own control edge.
kvn@119 345 Node* m = n->find_exact_control(n->in(0));
kvn@155 346 if (m != NULL) {
kvn@155 347 if (m->is_top())
kvn@155 348 return false; // Conservative answer for dead code
kvn@155 349 dom_list.push(m);
kvn@155 350 }
kvn@119 351 // Now, the rest of edges.
kvn@119 352 uint cnt = n->req();
kvn@119 353 for (uint i = 1; i < cnt; i++) {
kvn@119 354 m = n->find_exact_control(n->in(i));
kvn@119 355 if (m == NULL || m->is_top())
kvn@119 356 continue;
kvn@119 357 dom_list.push(m);
duke@0 358 }
duke@0 359 }
duke@0 360 }
kvn@119 361 return only_dominating_controls;
duke@0 362 }
duke@0 363 }
duke@0 364
duke@0 365 //---------------------detect_ptr_independence---------------------------------
duke@0 366 // Used by MemNode::find_previous_store to prove that two base
duke@0 367 // pointers are never equal.
duke@0 368 // The pointers are accompanied by their associated allocations,
duke@0 369 // if any, which have been previously discovered by the caller.
duke@0 370 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@0 371 Node* p2, AllocateNode* a2,
duke@0 372 PhaseTransform* phase) {
duke@0 373 // Attempt to prove that these two pointers cannot be aliased.
duke@0 374 // They may both manifestly be allocations, and they should differ.
duke@0 375 // Or, if they are not both allocations, they can be distinct constants.
duke@0 376 // Otherwise, one is an allocation and the other a pre-existing value.
duke@0 377 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@0 378 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@0 379 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@0 380 return (a1 != a2);
duke@0 381 } else if (a1 != NULL) { // one allocation a1
duke@0 382 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
kvn@119 383 return all_controls_dominate(p2, a1);
duke@0 384 } else { //(a2 != NULL) // one allocation a2
kvn@119 385 return all_controls_dominate(p1, a2);
duke@0 386 }
duke@0 387 return false;
duke@0 388 }
duke@0 389
duke@0 390
duke@0 391 // The logic for reordering loads and stores uses four steps:
duke@0 392 // (a) Walk carefully past stores and initializations which we
duke@0 393 // can prove are independent of this load.
duke@0 394 // (b) Observe that the next memory state makes an exact match
duke@0 395 // with self (load or store), and locate the relevant store.
duke@0 396 // (c) Ensure that, if we were to wire self directly to the store,
duke@0 397 // the optimizer would fold it up somehow.
duke@0 398 // (d) Do the rewiring, and return, depending on some other part of
duke@0 399 // the optimizer to fold up the load.
duke@0 400 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@0 401 // specific to loads and stores, so they are handled by the callers.
duke@0 402 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@0 403 //
duke@0 404 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@0 405 Node* ctrl = in(MemNode::Control);
duke@0 406 Node* adr = in(MemNode::Address);
duke@0 407 intptr_t offset = 0;
duke@0 408 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 409 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@0 410
duke@0 411 if (offset == Type::OffsetBot)
duke@0 412 return NULL; // cannot unalias unless there are precise offsets
duke@0 413
kvn@74 414 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
kvn@74 415
duke@0 416 intptr_t size_in_bytes = memory_size();
duke@0 417
duke@0 418 Node* mem = in(MemNode::Memory); // start searching here...
duke@0 419
duke@0 420 int cnt = 50; // Cycle limiter
duke@0 421 for (;;) { // While we can dance past unrelated stores...
duke@0 422 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@0 423
duke@0 424 if (mem->is_Store()) {
duke@0 425 Node* st_adr = mem->in(MemNode::Address);
duke@0 426 intptr_t st_offset = 0;
duke@0 427 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@0 428 if (st_base == NULL)
duke@0 429 break; // inscrutable pointer
duke@0 430 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@0 431 const int MAX_STORE = BytesPerLong;
duke@0 432 if (st_offset >= offset + size_in_bytes ||
duke@0 433 st_offset <= offset - MAX_STORE ||
duke@0 434 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@0 435 // Success: The offsets are provably independent.
duke@0 436 // (You may ask, why not just test st_offset != offset and be done?
duke@0 437 // The answer is that stores of different sizes can co-exist
duke@0 438 // in the same sequence of RawMem effects. We sometimes initialize
duke@0 439 // a whole 'tile' of array elements with a single jint or jlong.)
duke@0 440 mem = mem->in(MemNode::Memory);
duke@0 441 continue; // (a) advance through independent store memory
duke@0 442 }
duke@0 443 }
duke@0 444 if (st_base != base &&
duke@0 445 detect_ptr_independence(base, alloc,
duke@0 446 st_base,
duke@0 447 AllocateNode::Ideal_allocation(st_base, phase),
duke@0 448 phase)) {
duke@0 449 // Success: The bases are provably independent.
duke@0 450 mem = mem->in(MemNode::Memory);
duke@0 451 continue; // (a) advance through independent store memory
duke@0 452 }
duke@0 453
duke@0 454 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@0 455 // since we have not managed to prove 'this' and 'mem' independent.
duke@0 456 if (st_base == base && st_offset == offset) {
duke@0 457 return mem; // let caller handle steps (c), (d)
duke@0 458 }
duke@0 459
duke@0 460 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 461 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@0 462 AllocateNode* st_alloc = st_init->allocation();
duke@0 463 if (st_alloc == NULL)
duke@0 464 break; // something degenerated
duke@0 465 bool known_identical = false;
duke@0 466 bool known_independent = false;
duke@0 467 if (alloc == st_alloc)
duke@0 468 known_identical = true;
duke@0 469 else if (alloc != NULL)
duke@0 470 known_independent = true;
kvn@119 471 else if (all_controls_dominate(this, st_alloc))
duke@0 472 known_independent = true;
duke@0 473
duke@0 474 if (known_independent) {
duke@0 475 // The bases are provably independent: Either they are
duke@0 476 // manifestly distinct allocations, or else the control
duke@0 477 // of this load dominates the store's allocation.
duke@0 478 int alias_idx = phase->C->get_alias_index(adr_type());
duke@0 479 if (alias_idx == Compile::AliasIdxRaw) {
duke@0 480 mem = st_alloc->in(TypeFunc::Memory);
duke@0 481 } else {
duke@0 482 mem = st_init->memory(alias_idx);
duke@0 483 }
duke@0 484 continue; // (a) advance through independent store memory
duke@0 485 }
duke@0 486
duke@0 487 // (b) at this point, if we are not looking at a store initializing
duke@0 488 // the same allocation we are loading from, we lose.
duke@0 489 if (known_identical) {
duke@0 490 // From caller, can_see_stored_value will consult find_captured_store.
duke@0 491 return mem; // let caller handle steps (c), (d)
duke@0 492 }
duke@0 493
kvn@223 494 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
kvn@74 495 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
kvn@74 496 if (mem->is_Proj() && mem->in(0)->is_Call()) {
kvn@74 497 CallNode *call = mem->in(0)->as_Call();
kvn@74 498 if (!call->may_modify(addr_t, phase)) {
kvn@74 499 mem = call->in(TypeFunc::Memory);
kvn@74 500 continue; // (a) advance through independent call memory
kvn@74 501 }
kvn@74 502 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
kvn@74 503 mem = mem->in(0)->in(TypeFunc::Memory);
kvn@74 504 continue; // (a) advance through independent MemBar memory
kvn@74 505 } else if (mem->is_MergeMem()) {
kvn@74 506 int alias_idx = phase->C->get_alias_index(adr_type());
kvn@74 507 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@74 508 continue; // (a) advance through independent MergeMem memory
kvn@74 509 }
duke@0 510 }
duke@0 511
duke@0 512 // Unless there is an explicit 'continue', we must bail out here,
duke@0 513 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@0 514 break;
duke@0 515 }
duke@0 516
duke@0 517 return NULL; // bail out
duke@0 518 }
duke@0 519
duke@0 520 //----------------------calculate_adr_type-------------------------------------
duke@0 521 // Helper function. Notices when the given type of address hits top or bottom.
duke@0 522 // Also, asserts a cross-check of the type against the expected address type.
duke@0 523 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@0 524 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@0 525 #ifdef PRODUCT
duke@0 526 cross_check = NULL;
duke@0 527 #else
duke@0 528 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@0 529 #endif
duke@0 530 const TypePtr* tp = t->isa_ptr();
duke@0 531 if (tp == NULL) {
duke@0 532 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@0 533 return TypePtr::BOTTOM; // touches lots of memory
duke@0 534 } else {
duke@0 535 #ifdef ASSERT
duke@0 536 // %%%% [phh] We don't check the alias index if cross_check is
duke@0 537 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@0 538 if (cross_check != NULL &&
duke@0 539 cross_check != TypePtr::BOTTOM &&
duke@0 540 cross_check != TypeRawPtr::BOTTOM) {
duke@0 541 // Recheck the alias index, to see if it has changed (due to a bug).
duke@0 542 Compile* C = Compile::current();
duke@0 543 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@0 544 "must stay in the original alias category");
duke@0 545 // The type of the address must be contained in the adr_type,
duke@0 546 // disregarding "null"-ness.
duke@0 547 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@0 548 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
duke@0 549 assert(cross_check->meet(tp_notnull) == cross_check,
duke@0 550 "real address must not escape from expected memory type");
duke@0 551 }
duke@0 552 #endif
duke@0 553 return tp;
duke@0 554 }
duke@0 555 }
duke@0 556
duke@0 557 //------------------------adr_phi_is_loop_invariant----------------------------
duke@0 558 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@0 559 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@0 560 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@0 561 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@0 562 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@0 563 // with the same data. This implies that any path into the loop already
duke@0 564 // includes such a CastPP, and so the original cast, whatever its input,
duke@0 565 // must be covered by an equivalent cast, with an earlier control input.
duke@0 566 ResourceMark rm;
duke@0 567
duke@0 568 // The loop entry input of the phi should be the unique dominating
duke@0 569 // node for every Phi/CastPP in the loop.
duke@0 570 Unique_Node_List closure;
duke@0 571 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@0 572
duke@0 573 // Add the phi node and the cast to the worklist.
duke@0 574 Unique_Node_List worklist;
duke@0 575 worklist.push(adr_phi);
duke@0 576 if( cast != NULL ){
duke@0 577 if( !cast->is_ConstraintCast() ) return false;
duke@0 578 worklist.push(cast);
duke@0 579 }
duke@0 580
duke@0 581 // Begin recursive walk of phi nodes.
duke@0 582 while( worklist.size() ){
duke@0 583 // Take a node off the worklist
duke@0 584 Node *n = worklist.pop();
duke@0 585 if( !closure.member(n) ){
duke@0 586 // Add it to the closure.
duke@0 587 closure.push(n);
duke@0 588 // Make a sanity check to ensure we don't waste too much time here.
duke@0 589 if( closure.size() > 20) return false;
duke@0 590 // This node is OK if:
duke@0 591 // - it is a cast of an identical value
duke@0 592 // - or it is a phi node (then we add its inputs to the worklist)
duke@0 593 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@0 594 if( n->is_ConstraintCast() ){
duke@0 595 worklist.push(n->in(1));
duke@0 596 } else if( n->is_Phi() ) {
duke@0 597 for( uint i = 1; i < n->req(); i++ ) {
duke@0 598 worklist.push(n->in(i));
duke@0 599 }
duke@0 600 } else {
duke@0 601 return false;
duke@0 602 }
duke@0 603 }
duke@0 604 }
duke@0 605
duke@0 606 // Quit when the worklist is empty, and we've found no offending nodes.
duke@0 607 return true;
duke@0 608 }
duke@0 609
duke@0 610 //------------------------------Ideal_DU_postCCP-------------------------------
duke@0 611 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@0 612 // going away in this pass and we need to make this memory op depend on the
duke@0 613 // gating null check.
kvn@163 614 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
kvn@163 615 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
kvn@163 616 }
duke@0 617
duke@0 618 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@0 619 // some sense; we get to keep around the knowledge that an oop is not-null
duke@0 620 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@0 621 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@0 622 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@0 623 // some of the more trivial cases in the optimizer. Removing more useless
duke@0 624 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@0 625 // up on this approach. CNC 10/20/2000
kvn@163 626 // This static method may be called not from MemNode (EncodePNode calls it).
kvn@163 627 // Only the control edge of the node 'n' might be updated.
kvn@163 628 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
duke@0 629 Node *skipped_cast = NULL;
duke@0 630 // Need a null check? Regular static accesses do not because they are
duke@0 631 // from constant addresses. Array ops are gated by the range check (which
duke@0 632 // always includes a NULL check). Just check field ops.
kvn@163 633 if( n->in(MemNode::Control) == NULL ) {
duke@0 634 // Scan upwards for the highest location we can place this memory op.
duke@0 635 while( true ) {
duke@0 636 switch( adr->Opcode() ) {
duke@0 637
duke@0 638 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@0 639 adr = adr->in(AddPNode::Base);
duke@0 640 continue;
duke@0 641
coleenp@113 642 case Op_DecodeN: // No change to NULL-ness, so peek thru
coleenp@113 643 adr = adr->in(1);
coleenp@113 644 continue;
coleenp@113 645
duke@0 646 case Op_CastPP:
duke@0 647 // If the CastPP is useless, just peek on through it.
duke@0 648 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@0 649 // Remember the cast that we've peeked though. If we peek
duke@0 650 // through more than one, then we end up remembering the highest
duke@0 651 // one, that is, if in a loop, the one closest to the top.
duke@0 652 skipped_cast = adr;
duke@0 653 adr = adr->in(1);
duke@0 654 continue;
duke@0 655 }
duke@0 656 // CastPP is going away in this pass! We need this memory op to be
duke@0 657 // control-dependent on the test that is guarding the CastPP.
kvn@163 658 ccp->hash_delete(n);
kvn@163 659 n->set_req(MemNode::Control, adr->in(0));
kvn@163 660 ccp->hash_insert(n);
kvn@163 661 return n;
duke@0 662
duke@0 663 case Op_Phi:
duke@0 664 // Attempt to float above a Phi to some dominating point.
duke@0 665 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@0 666 // If we've already peeked through a Cast (which could have set the
duke@0 667 // control), we can't float above a Phi, because the skipped Cast
duke@0 668 // may not be loop invariant.
duke@0 669 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@0 670 adr = adr->in(1);
duke@0 671 continue;
duke@0 672 }
duke@0 673 }
duke@0 674
duke@0 675 // Intentional fallthrough!
duke@0 676
duke@0 677 // No obvious dominating point. The mem op is pinned below the Phi
duke@0 678 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@0 679 // then the mem op can float, but not indefinitely. It must be pinned
duke@0 680 // behind the controls leading to the Phi.
duke@0 681 case Op_CheckCastPP:
duke@0 682 // These usually stick around to change address type, however a
duke@0 683 // useless one can be elided and we still need to pick up a control edge
duke@0 684 if (adr->in(0) == NULL) {
duke@0 685 // This CheckCastPP node has NO control and is likely useless. But we
duke@0 686 // need check further up the ancestor chain for a control input to keep
duke@0 687 // the node in place. 4959717.
duke@0 688 skipped_cast = adr;
duke@0 689 adr = adr->in(1);
duke@0 690 continue;
duke@0 691 }
kvn@163 692 ccp->hash_delete(n);
kvn@163 693 n->set_req(MemNode::Control, adr->in(0));
kvn@163 694 ccp->hash_insert(n);
kvn@163 695 return n;
duke@0 696
duke@0 697 // List of "safe" opcodes; those that implicitly block the memory
duke@0 698 // op below any null check.
duke@0 699 case Op_CastX2P: // no null checks on native pointers
duke@0 700 case Op_Parm: // 'this' pointer is not null
duke@0 701 case Op_LoadP: // Loading from within a klass
coleenp@113 702 case Op_LoadN: // Loading from within a klass
duke@0 703 case Op_LoadKlass: // Loading from within a klass
kvn@164 704 case Op_LoadNKlass: // Loading from within a klass
duke@0 705 case Op_ConP: // Loading from a klass
kvn@163 706 case Op_ConN: // Loading from a klass
duke@0 707 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@0 708 case Op_Con: // Reading from TLS
duke@0 709 case Op_CMoveP: // CMoveP is pinned
kvn@164 710 case Op_CMoveN: // CMoveN is pinned
duke@0 711 break; // No progress
duke@0 712
duke@0 713 case Op_Proj: // Direct call to an allocation routine
duke@0 714 case Op_SCMemProj: // Memory state from store conditional ops
duke@0 715 #ifdef ASSERT
duke@0 716 {
duke@0 717 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@0 718 const Node* call = adr->in(0);
kvn@163 719 if (call->is_CallJava()) {
kvn@163 720 const CallJavaNode* call_java = call->as_CallJava();
kvn@64 721 const TypeTuple *r = call_java->tf()->range();
kvn@64 722 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@64 723 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@64 724 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@0 725 // We further presume that this is one of
duke@0 726 // new_instance_Java, new_array_Java, or
duke@0 727 // the like, but do not assert for this.
duke@0 728 } else if (call->is_Allocate()) {
duke@0 729 // similar case to new_instance_Java, etc.
duke@0 730 } else if (!call->is_CallLeaf()) {
duke@0 731 // Projections from fetch_oop (OSR) are allowed as well.
duke@0 732 ShouldNotReachHere();
duke@0 733 }
duke@0 734 }
duke@0 735 #endif
duke@0 736 break;
duke@0 737 default:
duke@0 738 ShouldNotReachHere();
duke@0 739 }
duke@0 740 break;
duke@0 741 }
duke@0 742 }
duke@0 743
duke@0 744 return NULL; // No progress
duke@0 745 }
duke@0 746
duke@0 747
duke@0 748 //=============================================================================
duke@0 749 uint LoadNode::size_of() const { return sizeof(*this); }
duke@0 750 uint LoadNode::cmp( const Node &n ) const
duke@0 751 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@0 752 const Type *LoadNode::bottom_type() const { return _type; }
duke@0 753 uint LoadNode::ideal_reg() const {
duke@0 754 return Matcher::base2reg[_type->base()];
duke@0 755 }
duke@0 756
duke@0 757 #ifndef PRODUCT
duke@0 758 void LoadNode::dump_spec(outputStream *st) const {
duke@0 759 MemNode::dump_spec(st);
duke@0 760 if( !Verbose && !WizardMode ) {
duke@0 761 // standard dump does this in Verbose and WizardMode
duke@0 762 st->print(" #"); _type->dump_on(st);
duke@0 763 }
duke@0 764 }
duke@0 765 #endif
duke@0 766
duke@0 767
duke@0 768 //----------------------------LoadNode::make-----------------------------------
duke@0 769 // Polymorphic factory method:
coleenp@113 770 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
coleenp@113 771 Compile* C = gvn.C;
coleenp@113 772
duke@0 773 // sanity check the alias category against the created node type
duke@0 774 assert(!(adr_type->isa_oopptr() &&
duke@0 775 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@0 776 "use LoadKlassNode instead");
duke@0 777 assert(!(adr_type->isa_aryptr() &&
duke@0 778 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@0 779 "use LoadRangeNode instead");
duke@0 780 switch (bt) {
duke@0 781 case T_BOOLEAN:
twisti@558 782 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 783 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 784 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 785 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 786 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
twisti@558 787 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt );
twisti@558 788 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt );
twisti@558 789 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
coleenp@113 790 case T_OBJECT:
coleenp@113 791 #ifdef _LP64
kvn@163 792 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
kvn@221 793 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
kvn@221 794 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
coleenp@113 795 } else
coleenp@113 796 #endif
kvn@163 797 {
kvn@163 798 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@163 799 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
kvn@163 800 }
duke@0 801 }
duke@0 802 ShouldNotReachHere();
duke@0 803 return (LoadNode*)NULL;
duke@0 804 }
duke@0 805
duke@0 806 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
duke@0 807 bool require_atomic = true;
duke@0 808 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
duke@0 809 }
duke@0 810
duke@0 811
duke@0 812
duke@0 813
duke@0 814 //------------------------------hash-------------------------------------------
duke@0 815 uint LoadNode::hash() const {
duke@0 816 // unroll addition of interesting fields
duke@0 817 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@0 818 }
duke@0 819
duke@0 820 //---------------------------can_see_stored_value------------------------------
duke@0 821 // This routine exists to make sure this set of tests is done the same
duke@0 822 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@0 823 // will change the graph shape in a way which makes memory alive twice at the
duke@0 824 // same time (uses the Oracle model of aliasing), then some
duke@0 825 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@0 826 // of aliasing.
duke@0 827 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@0 828 Node* ld_adr = in(MemNode::Address);
duke@0 829
never@17 830 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
never@17 831 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
never@17 832 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
never@17 833 atp->field() != NULL && !atp->field()->is_volatile()) {
never@17 834 uint alias_idx = atp->index();
never@17 835 bool final = atp->field()->is_final();
never@17 836 Node* result = NULL;
never@17 837 Node* current = st;
never@17 838 // Skip through chains of MemBarNodes checking the MergeMems for
never@17 839 // new states for the slice of this load. Stop once any other
never@17 840 // kind of node is encountered. Loads from final memory can skip
never@17 841 // through any kind of MemBar but normal loads shouldn't skip
never@17 842 // through MemBarAcquire since the could allow them to move out of
never@17 843 // a synchronized region.
never@17 844 while (current->is_Proj()) {
never@17 845 int opc = current->in(0)->Opcode();
never@17 846 if ((final && opc == Op_MemBarAcquire) ||
never@17 847 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
never@17 848 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@17 849 if (mem->is_MergeMem()) {
never@17 850 MergeMemNode* merge = mem->as_MergeMem();
never@17 851 Node* new_st = merge->memory_at(alias_idx);
never@17 852 if (new_st == merge->base_memory()) {
never@17 853 // Keep searching
never@17 854 current = merge->base_memory();
never@17 855 continue;
never@17 856 }
never@17 857 // Save the new memory state for the slice and fall through
never@17 858 // to exit.
never@17 859 result = new_st;
never@17 860 }
never@17 861 }
never@17 862 break;
never@17 863 }
never@17 864 if (result != NULL) {
never@17 865 st = result;
never@17 866 }
never@17 867 }
never@17 868
never@17 869
duke@0 870 // Loop around twice in the case Load -> Initialize -> Store.
duke@0 871 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@0 872 for (int trip = 0; trip <= 1; trip++) {
duke@0 873
duke@0 874 if (st->is_Store()) {
duke@0 875 Node* st_adr = st->in(MemNode::Address);
duke@0 876 if (!phase->eqv(st_adr, ld_adr)) {
duke@0 877 // Try harder before giving up... Match raw and non-raw pointers.
duke@0 878 intptr_t st_off = 0;
duke@0 879 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@0 880 if (alloc == NULL) return NULL;
duke@0 881 intptr_t ld_off = 0;
duke@0 882 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
duke@0 883 if (alloc != allo2) return NULL;
duke@0 884 if (ld_off != st_off) return NULL;
duke@0 885 // At this point we have proven something like this setup:
duke@0 886 // A = Allocate(...)
duke@0 887 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@0 888 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@0 889 // (Actually, we haven't yet proven the Q's are the same.)
duke@0 890 // In other words, we are loading from a casted version of
duke@0 891 // the same pointer-and-offset that we stored to.
duke@0 892 // Thus, we are able to replace L by V.
duke@0 893 }
duke@0 894 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@0 895 if (store_Opcode() != st->Opcode())
duke@0 896 return NULL;
duke@0 897 return st->in(MemNode::ValueIn);
duke@0 898 }
duke@0 899
duke@0 900 intptr_t offset = 0; // scratch
duke@0 901
duke@0 902 // A load from a freshly-created object always returns zero.
duke@0 903 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@0 904 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@0 905 if (st->is_Proj() && st->in(0)->is_Allocate() &&
duke@0 906 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
duke@0 907 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
duke@0 908 // return a zero value for the load's basic type
duke@0 909 // (This is one of the few places where a generic PhaseTransform
duke@0 910 // can create new nodes. Think of it as lazily manifesting
duke@0 911 // virtually pre-existing constants.)
duke@0 912 return phase->zerocon(memory_type());
duke@0 913 }
duke@0 914
duke@0 915 // A load from an initialization barrier can match a captured store.
duke@0 916 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@0 917 InitializeNode* init = st->in(0)->as_Initialize();
duke@0 918 AllocateNode* alloc = init->allocation();
duke@0 919 if (alloc != NULL &&
duke@0 920 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
duke@0 921 // examine a captured store value
duke@0 922 st = init->find_captured_store(offset, memory_size(), phase);
duke@0 923 if (st != NULL)
duke@0 924 continue; // take one more trip around
duke@0 925 }
duke@0 926 }
duke@0 927
duke@0 928 break;
duke@0 929 }
duke@0 930
duke@0 931 return NULL;
duke@0 932 }
duke@0 933
kvn@64 934 //----------------------is_instance_field_load_with_local_phi------------------
kvn@64 935 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@64 936 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
kvn@64 937 in(MemNode::Address)->is_AddP() ) {
kvn@64 938 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@64 939 // Only instances.
kvn@223 940 if( t_oop != NULL && t_oop->is_known_instance_field() &&
kvn@64 941 t_oop->offset() != Type::OffsetBot &&
kvn@64 942 t_oop->offset() != Type::OffsetTop) {
kvn@64 943 return true;
kvn@64 944 }
kvn@64 945 }
kvn@64 946 return false;
kvn@64 947 }
kvn@64 948
duke@0 949 //------------------------------Identity---------------------------------------
duke@0 950 // Loads are identity if previous store is to same address
duke@0 951 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@0 952 // If the previous store-maker is the right kind of Store, and the store is
duke@0 953 // to the same address, then we are equal to the value stored.
duke@0 954 Node* mem = in(MemNode::Memory);
duke@0 955 Node* value = can_see_stored_value(mem, phase);
duke@0 956 if( value ) {
duke@0 957 // byte, short & char stores truncate naturally.
duke@0 958 // A load has to load the truncated value which requires
duke@0 959 // some sort of masking operation and that requires an
duke@0 960 // Ideal call instead of an Identity call.
duke@0 961 if (memory_size() < BytesPerInt) {
duke@0 962 // If the input to the store does not fit with the load's result type,
duke@0 963 // it must be truncated via an Ideal call.
duke@0 964 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@0 965 return this;
duke@0 966 }
duke@0 967 // (This works even when value is a Con, but LoadNode::Value
duke@0 968 // usually runs first, producing the singleton type of the Con.)
duke@0 969 return value;
duke@0 970 }
kvn@64 971
kvn@64 972 // Search for an existing data phi which was generated before for the same
kvn@64 973 // instance's field to avoid infinite genertion of phis in a loop.
kvn@64 974 Node *region = mem->in(0);
kvn@64 975 if (is_instance_field_load_with_local_phi(region)) {
kvn@64 976 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
kvn@64 977 int this_index = phase->C->get_alias_index(addr_t);
kvn@64 978 int this_offset = addr_t->offset();
kvn@64 979 int this_id = addr_t->is_oopptr()->instance_id();
kvn@64 980 const Type* this_type = bottom_type();
kvn@64 981 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@64 982 Node* phi = region->fast_out(i);
kvn@64 983 if (phi->is_Phi() && phi != mem &&
kvn@64 984 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
kvn@64 985 return phi;
kvn@64 986 }
kvn@64 987 }
kvn@64 988 }
kvn@64 989
duke@0 990 return this;
duke@0 991 }
duke@0 992
never@17 993
never@17 994 // Returns true if the AliasType refers to the field that holds the
never@17 995 // cached box array. Currently only handles the IntegerCache case.
never@17 996 static bool is_autobox_cache(Compile::AliasType* atp) {
never@17 997 if (atp != NULL && atp->field() != NULL) {
never@17 998 ciField* field = atp->field();
never@17 999 ciSymbol* klass = field->holder()->name();
never@17 1000 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1001 field->holder()->uses_default_loader() &&
never@17 1002 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1003 return true;
never@17 1004 }
never@17 1005 }
never@17 1006 return false;
never@17 1007 }
never@17 1008
never@17 1009 // Fetch the base value in the autobox array
never@17 1010 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
never@17 1011 if (atp != NULL && atp->field() != NULL) {
never@17 1012 ciField* field = atp->field();
never@17 1013 ciSymbol* klass = field->holder()->name();
never@17 1014 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1015 field->holder()->uses_default_loader() &&
never@17 1016 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1017 assert(field->is_constant(), "what?");
never@17 1018 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
never@17 1019 // Fetch the box object at the base of the array and get its value
never@17 1020 ciInstance* box = array->obj_at(0)->as_instance();
never@17 1021 ciInstanceKlass* ik = box->klass()->as_instance_klass();
never@17 1022 if (ik->nof_nonstatic_fields() == 1) {
never@17 1023 // This should be true nonstatic_field_at requires calling
never@17 1024 // nof_nonstatic_fields so check it anyway
never@17 1025 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
never@17 1026 cache_offset = c.as_int();
never@17 1027 }
never@17 1028 return true;
never@17 1029 }
never@17 1030 }
never@17 1031 return false;
never@17 1032 }
never@17 1033
never@17 1034 // Returns true if the AliasType refers to the value field of an
never@17 1035 // autobox object. Currently only handles Integer.
never@17 1036 static bool is_autobox_object(Compile::AliasType* atp) {
never@17 1037 if (atp != NULL && atp->field() != NULL) {
never@17 1038 ciField* field = atp->field();
never@17 1039 ciSymbol* klass = field->holder()->name();
never@17 1040 if (field->name() == ciSymbol::value_name() &&
never@17 1041 field->holder()->uses_default_loader() &&
never@17 1042 klass == ciSymbol::java_lang_Integer()) {
never@17 1043 return true;
never@17 1044 }
never@17 1045 }
never@17 1046 return false;
never@17 1047 }
never@17 1048
never@17 1049
never@17 1050 // We're loading from an object which has autobox behaviour.
never@17 1051 // If this object is result of a valueOf call we'll have a phi
never@17 1052 // merging a newly allocated object and a load from the cache.
never@17 1053 // We want to replace this load with the original incoming
never@17 1054 // argument to the valueOf call.
never@17 1055 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
never@17 1056 Node* base = in(Address)->in(AddPNode::Base);
never@17 1057 if (base->is_Phi() && base->req() == 3) {
never@17 1058 AllocateNode* allocation = NULL;
never@17 1059 int allocation_index = -1;
never@17 1060 int load_index = -1;
never@17 1061 for (uint i = 1; i < base->req(); i++) {
never@17 1062 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
never@17 1063 if (allocation != NULL) {
never@17 1064 allocation_index = i;
never@17 1065 load_index = 3 - allocation_index;
never@17 1066 break;
never@17 1067 }
never@17 1068 }
never@17 1069 LoadNode* load = NULL;
never@17 1070 if (allocation != NULL && base->in(load_index)->is_Load()) {
never@17 1071 load = base->in(load_index)->as_Load();
never@17 1072 }
never@17 1073 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
never@17 1074 // Push the loads from the phi that comes from valueOf up
never@17 1075 // through it to allow elimination of the loads and the recovery
never@17 1076 // of the original value.
never@17 1077 Node* mem_phi = in(Memory);
never@17 1078 Node* offset = in(Address)->in(AddPNode::Offset);
never@553 1079 Node* region = base->in(0);
never@17 1080
never@17 1081 Node* in1 = clone();
never@17 1082 Node* in1_addr = in1->in(Address)->clone();
never@17 1083 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
never@17 1084 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
never@17 1085 in1_addr->set_req(AddPNode::Offset, offset);
never@553 1086 in1->set_req(0, region->in(allocation_index));
never@17 1087 in1->set_req(Address, in1_addr);
never@17 1088 in1->set_req(Memory, mem_phi->in(allocation_index));
never@17 1089
never@17 1090 Node* in2 = clone();
never@17 1091 Node* in2_addr = in2->in(Address)->clone();
never@17 1092 in2_addr->set_req(AddPNode::Base, base->in(load_index));
never@17 1093 in2_addr->set_req(AddPNode::Address, base->in(load_index));
never@17 1094 in2_addr->set_req(AddPNode::Offset, offset);
never@553 1095 in2->set_req(0, region->in(load_index));
never@17 1096 in2->set_req(Address, in2_addr);
never@17 1097 in2->set_req(Memory, mem_phi->in(load_index));
never@17 1098
never@17 1099 in1_addr = phase->transform(in1_addr);
never@17 1100 in1 = phase->transform(in1);
never@17 1101 in2_addr = phase->transform(in2_addr);
never@17 1102 in2 = phase->transform(in2);
never@17 1103
never@553 1104 PhiNode* result = PhiNode::make_blank(region, this);
never@17 1105 result->set_req(allocation_index, in1);
never@17 1106 result->set_req(load_index, in2);
never@17 1107 return result;
never@17 1108 }
never@17 1109 } else if (base->is_Load()) {
never@17 1110 // Eliminate the load of Integer.value for integers from the cache
never@17 1111 // array by deriving the value from the index into the array.
never@17 1112 // Capture the offset of the load and then reverse the computation.
never@17 1113 Node* load_base = base->in(Address)->in(AddPNode::Base);
never@17 1114 if (load_base != NULL) {
never@17 1115 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
never@17 1116 intptr_t cache_offset;
never@17 1117 int shift = -1;
never@17 1118 Node* cache = NULL;
never@17 1119 if (is_autobox_cache(atp)) {
kvn@29 1120 shift = exact_log2(type2aelembytes(T_OBJECT));
never@17 1121 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
never@17 1122 }
never@17 1123 if (cache != NULL && base->in(Address)->is_AddP()) {
never@17 1124 Node* elements[4];
never@17 1125 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@17 1126 int cache_low;
never@17 1127 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
never@17 1128 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
never@17 1129 // Add up all the offsets making of the address of the load
never@17 1130 Node* result = elements[0];
never@17 1131 for (int i = 1; i < count; i++) {
never@17 1132 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
never@17 1133 }
never@17 1134 // Remove the constant offset from the address and then
never@17 1135 // remove the scaling of the offset to recover the original index.
never@17 1136 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
never@17 1137 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
never@17 1138 // Peel the shift off directly but wrap it in a dummy node
never@17 1139 // since Ideal can't return existing nodes
never@17 1140 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
never@17 1141 } else {
never@17 1142 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
never@17 1143 }
never@17 1144 #ifdef _LP64
never@17 1145 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
never@17 1146 #endif
never@17 1147 return result;
never@17 1148 }
never@17 1149 }
never@17 1150 }
never@17 1151 }
never@17 1152 return NULL;
never@17 1153 }
never@17 1154
kvn@163 1155 //------------------------------split_through_phi------------------------------
kvn@163 1156 // Split instance field load through Phi.
kvn@163 1157 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
kvn@163 1158 Node* mem = in(MemNode::Memory);
kvn@163 1159 Node* address = in(MemNode::Address);
kvn@163 1160 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@163 1161 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@163 1162
kvn@163 1163 assert(mem->is_Phi() && (t_oop != NULL) &&
kvn@223 1164 t_oop->is_known_instance_field(), "invalide conditions");
kvn@163 1165
kvn@163 1166 Node *region = mem->in(0);
kvn@163 1167 if (region == NULL) {
kvn@163 1168 return NULL; // Wait stable graph
kvn@163 1169 }
kvn@163 1170 uint cnt = mem->req();
kvn@163 1171 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1172 Node *in = mem->in(i);
kvn@163 1173 if( in == NULL ) {
kvn@163 1174 return NULL; // Wait stable graph
kvn@163 1175 }
kvn@163 1176 }
kvn@163 1177 // Check for loop invariant.
kvn@163 1178 if (cnt == 3) {
kvn@163 1179 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1180 Node *in = mem->in(i);
kvn@163 1181 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
kvn@163 1182 if (m == mem) {
kvn@163 1183 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
kvn@163 1184 return this;
kvn@163 1185 }
kvn@163 1186 }
kvn@163 1187 }
kvn@163 1188 // Split through Phi (see original code in loopopts.cpp).
kvn@163 1189 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
kvn@163 1190
kvn@163 1191 // Do nothing here if Identity will find a value
kvn@163 1192 // (to avoid infinite chain of value phis generation).
kvn@163 1193 if ( !phase->eqv(this, this->Identity(phase)) )
kvn@163 1194 return NULL;
kvn@163 1195
kvn@163 1196 // Skip the split if the region dominates some control edge of the address.
kvn@163 1197 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
kvn@163 1198 return NULL;
kvn@163 1199
kvn@163 1200 const Type* this_type = this->bottom_type();
kvn@163 1201 int this_index = phase->C->get_alias_index(addr_t);
kvn@163 1202 int this_offset = addr_t->offset();
kvn@163 1203 int this_iid = addr_t->is_oopptr()->instance_id();
kvn@163 1204 int wins = 0;
kvn@163 1205 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@163 1206 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
kvn@163 1207 for( uint i = 1; i < region->req(); i++ ) {
kvn@163 1208 Node *x;
kvn@163 1209 Node* the_clone = NULL;
kvn@163 1210 if( region->in(i) == phase->C->top() ) {
kvn@163 1211 x = phase->C->top(); // Dead path? Use a dead data op
kvn@163 1212 } else {
kvn@163 1213 x = this->clone(); // Else clone up the data op
kvn@163 1214 the_clone = x; // Remember for possible deletion.
kvn@163 1215 // Alter data node to use pre-phi inputs
kvn@163 1216 if( this->in(0) == region ) {
kvn@163 1217 x->set_req( 0, region->in(i) );
kvn@163 1218 } else {
kvn@163 1219 x->set_req( 0, NULL );
kvn@163 1220 }
kvn@163 1221 for( uint j = 1; j < this->req(); j++ ) {
kvn@163 1222 Node *in = this->in(j);
kvn@163 1223 if( in->is_Phi() && in->in(0) == region )
kvn@163 1224 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
kvn@163 1225 }
kvn@163 1226 }
kvn@163 1227 // Check for a 'win' on some paths
kvn@163 1228 const Type *t = x->Value(igvn);
kvn@163 1229
kvn@163 1230 bool singleton = t->singleton();
kvn@163 1231
kvn@163 1232 // See comments in PhaseIdealLoop::split_thru_phi().
kvn@163 1233 if( singleton && t == Type::TOP ) {
kvn@163 1234 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
kvn@163 1235 }
kvn@163 1236
kvn@163 1237 if( singleton ) {
kvn@163 1238 wins++;
kvn@163 1239 x = igvn->makecon(t);
kvn@163 1240 } else {
kvn@163 1241 // We now call Identity to try to simplify the cloned node.
kvn@163 1242 // Note that some Identity methods call phase->type(this).
kvn@163 1243 // Make sure that the type array is big enough for
kvn@163 1244 // our new node, even though we may throw the node away.
kvn@163 1245 // (This tweaking with igvn only works because x is a new node.)
kvn@163 1246 igvn->set_type(x, t);
kvn@293 1247 // If x is a TypeNode, capture any more-precise type permanently into Node
kvn@293 1248 // othewise it will be not updated during igvn->transform since
kvn@293 1249 // igvn->type(x) is set to x->Value() already.
kvn@293 1250 x->raise_bottom_type(t);
kvn@163 1251 Node *y = x->Identity(igvn);
kvn@163 1252 if( y != x ) {
kvn@163 1253 wins++;
kvn@163 1254 x = y;
kvn@163 1255 } else {
kvn@163 1256 y = igvn->hash_find(x);
kvn@163 1257 if( y ) {
kvn@163 1258 wins++;
kvn@163 1259 x = y;
kvn@163 1260 } else {
kvn@163 1261 // Else x is a new node we are keeping
kvn@163 1262 // We do not need register_new_node_with_optimizer
kvn@163 1263 // because set_type has already been called.
kvn@163 1264 igvn->_worklist.push(x);
kvn@163 1265 }
kvn@163 1266 }
kvn@163 1267 }
kvn@163 1268 if (x != the_clone && the_clone != NULL)
kvn@163 1269 igvn->remove_dead_node(the_clone);
kvn@163 1270 phi->set_req(i, x);
kvn@163 1271 }
kvn@163 1272 if( wins > 0 ) {
kvn@163 1273 // Record Phi
kvn@163 1274 igvn->register_new_node_with_optimizer(phi);
kvn@163 1275 return phi;
kvn@163 1276 }
kvn@163 1277 igvn->remove_dead_node(phi);
kvn@163 1278 return NULL;
kvn@163 1279 }
never@17 1280
duke@0 1281 //------------------------------Ideal------------------------------------------
duke@0 1282 // If the load is from Field memory and the pointer is non-null, we can
duke@0 1283 // zero out the control input.
duke@0 1284 // If the offset is constant and the base is an object allocation,
duke@0 1285 // try to hook me up to the exact initializing store.
duke@0 1286 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1287 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1288 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1289
duke@0 1290 Node* ctrl = in(MemNode::Control);
duke@0 1291 Node* address = in(MemNode::Address);
duke@0 1292
duke@0 1293 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@0 1294 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@0 1295 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@0 1296 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@0 1297 ctrl = ctrl->in(0);
duke@0 1298 set_req(MemNode::Control,ctrl);
duke@0 1299 }
duke@0 1300
duke@0 1301 // Check for useless control edge in some common special cases
duke@0 1302 if (in(MemNode::Control) != NULL) {
duke@0 1303 intptr_t ignore = 0;
duke@0 1304 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
duke@0 1305 if (base != NULL
duke@0 1306 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
kvn@119 1307 && all_controls_dominate(base, phase->C->start())) {
duke@0 1308 // A method-invariant, non-null address (constant or 'this' argument).
duke@0 1309 set_req(MemNode::Control, NULL);
duke@0 1310 }
duke@0 1311 }
duke@0 1312
never@17 1313 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
never@17 1314 Node* base = in(Address)->in(AddPNode::Base);
never@17 1315 if (base != NULL) {
never@17 1316 Compile::AliasType* atp = phase->C->alias_type(adr_type());
never@17 1317 if (is_autobox_object(atp)) {
never@17 1318 Node* result = eliminate_autobox(phase);
never@17 1319 if (result != NULL) return result;
never@17 1320 }
never@17 1321 }
never@17 1322 }
never@17 1323
kvn@74 1324 Node* mem = in(MemNode::Memory);
kvn@74 1325 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@74 1326
kvn@74 1327 if (addr_t != NULL) {
kvn@74 1328 // try to optimize our memory input
kvn@74 1329 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
kvn@74 1330 if (opt_mem != mem) {
kvn@74 1331 set_req(MemNode::Memory, opt_mem);
kvn@305 1332 if (phase->type( opt_mem ) == Type::TOP) return NULL;
kvn@74 1333 return this;
kvn@74 1334 }
kvn@74 1335 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@74 1336 if (can_reshape && opt_mem->is_Phi() &&
kvn@223 1337 (t_oop != NULL) && t_oop->is_known_instance_field()) {
kvn@163 1338 // Split instance field load through Phi.
kvn@163 1339 Node* result = split_through_phi(phase);
kvn@163 1340 if (result != NULL) return result;
kvn@74 1341 }
kvn@74 1342 }
kvn@74 1343
duke@0 1344 // Check for prior store with a different base or offset; make Load
duke@0 1345 // independent. Skip through any number of them. Bail out if the stores
duke@0 1346 // are in an endless dead cycle and report no progress. This is a key
duke@0 1347 // transform for Reflection. However, if after skipping through the Stores
duke@0 1348 // we can't then fold up against a prior store do NOT do the transform as
duke@0 1349 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@0 1350 // array memory alive twice: once for the hoisted Load and again after the
duke@0 1351 // bypassed Store. This situation only works if EVERYBODY who does
duke@0 1352 // anti-dependence work knows how to bypass. I.e. we need all
duke@0 1353 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@0 1354 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@0 1355 // fold up, do so.
duke@0 1356 Node* prev_mem = find_previous_store(phase);
duke@0 1357 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1358 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@0 1359 // (c) See if we can fold up on the spot, but don't fold up here.
twisti@558 1360 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
duke@0 1361 // just return a prior value, which is done by Identity calls.
duke@0 1362 if (can_see_stored_value(prev_mem, phase)) {
duke@0 1363 // Make ready for step (d):
duke@0 1364 set_req(MemNode::Memory, prev_mem);
duke@0 1365 return this;
duke@0 1366 }
duke@0 1367 }
duke@0 1368
duke@0 1369 return NULL; // No further progress
duke@0 1370 }
duke@0 1371
duke@0 1372 // Helper to recognize certain Klass fields which are invariant across
duke@0 1373 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@0 1374 const Type*
duke@0 1375 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@0 1376 ciKlass* klass) const {
duke@0 1377 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1378 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@0 1379 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@0 1380 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@0 1381 return TypeInt::make(klass->modifier_flags());
duke@0 1382 }
duke@0 1383 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1384 // The field is Klass::_access_flags. Return its (constant) value.
duke@0 1385 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@0 1386 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@0 1387 return TypeInt::make(klass->access_flags());
duke@0 1388 }
duke@0 1389 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1390 // The field is Klass::_layout_helper. Return its constant value if known.
duke@0 1391 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1392 return TypeInt::make(klass->layout_helper());
duke@0 1393 }
duke@0 1394
duke@0 1395 // No match.
duke@0 1396 return NULL;
duke@0 1397 }
duke@0 1398
duke@0 1399 //------------------------------Value-----------------------------------------
duke@0 1400 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@0 1401 // Either input is TOP ==> the result is TOP
duke@0 1402 Node* mem = in(MemNode::Memory);
duke@0 1403 const Type *t1 = phase->type(mem);
duke@0 1404 if (t1 == Type::TOP) return Type::TOP;
duke@0 1405 Node* adr = in(MemNode::Address);
duke@0 1406 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@0 1407 if (tp == NULL || tp->empty()) return Type::TOP;
duke@0 1408 int off = tp->offset();
duke@0 1409 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
duke@0 1410
duke@0 1411 // Try to guess loaded type from pointer type
duke@0 1412 if (tp->base() == Type::AryPtr) {
duke@0 1413 const Type *t = tp->is_aryptr()->elem();
duke@0 1414 // Don't do this for integer types. There is only potential profit if
duke@0 1415 // the element type t is lower than _type; that is, for int types, if _type is
duke@0 1416 // more restrictive than t. This only happens here if one is short and the other
duke@0 1417 // char (both 16 bits), and in those cases we've made an intentional decision
duke@0 1418 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@0 1419 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@0 1420 //
duke@0 1421 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@0 1422 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@0 1423 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@0 1424 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@0 1425 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@0 1426 // In fact, that could have been the original type of p1, and p1 could have
duke@0 1427 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@0 1428 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@0 1429 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
kvn@293 1430 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
duke@0 1431 // t might actually be lower than _type, if _type is a unique
duke@0 1432 // concrete subclass of abstract class t.
duke@0 1433 // Make sure the reference is not into the header, by comparing
duke@0 1434 // the offset against the offset of the start of the array's data.
duke@0 1435 // Different array types begin at slightly different offsets (12 vs. 16).
duke@0 1436 // We choose T_BYTE as an example base type that is least restrictive
duke@0 1437 // as to alignment, which will therefore produce the smallest
duke@0 1438 // possible base offset.
duke@0 1439 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1440 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
duke@0 1441 const Type* jt = t->join(_type);
duke@0 1442 // In any case, do not allow the join, per se, to empty out the type.
duke@0 1443 if (jt->empty() && !t->empty()) {
duke@0 1444 // This can happen if a interface-typed array narrows to a class type.
duke@0 1445 jt = _type;
duke@0 1446 }
never@17 1447
never@17 1448 if (EliminateAutoBox) {
never@17 1449 // The pointers in the autobox arrays are always non-null
never@17 1450 Node* base = in(Address)->in(AddPNode::Base);
never@17 1451 if (base != NULL) {
never@17 1452 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
never@17 1453 if (is_autobox_cache(atp)) {
never@17 1454 return jt->join(TypePtr::NOTNULL)->is_ptr();
never@17 1455 }
never@17 1456 }
never@17 1457 }
duke@0 1458 return jt;
duke@0 1459 }
duke@0 1460 }
duke@0 1461 } else if (tp->base() == Type::InstPtr) {
duke@0 1462 assert( off != Type::OffsetBot ||
duke@0 1463 // arrays can be cast to Objects
duke@0 1464 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@0 1465 // unsafe field access may not have a constant offset
duke@0 1466 phase->C->has_unsafe_access(),
duke@0 1467 "Field accesses must be precise" );
duke@0 1468 // For oop loads, we expect the _type to be precise
duke@0 1469 } else if (tp->base() == Type::KlassPtr) {
duke@0 1470 assert( off != Type::OffsetBot ||
duke@0 1471 // arrays can be cast to Objects
duke@0 1472 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@0 1473 // also allow array-loading from the primary supertype
duke@0 1474 // array during subtype checks
duke@0 1475 Opcode() == Op_LoadKlass,
duke@0 1476 "Field accesses must be precise" );
duke@0 1477 // For klass/static loads, we expect the _type to be precise
duke@0 1478 }
duke@0 1479
duke@0 1480 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1481 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1482 ciKlass* klass = tkls->klass();
duke@0 1483 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@0 1484 // We are loading a field from a Klass metaobject whose identity
duke@0 1485 // is known at compile time (the type is "exact" or "precise").
duke@0 1486 // Check for fields we know are maintained as constants by the VM.
duke@0 1487 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1488 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@0 1489 // (Folds up type checking code.)
duke@0 1490 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@0 1491 return TypeInt::make(klass->super_check_offset());
duke@0 1492 }
duke@0 1493 // Compute index into primary_supers array
duke@0 1494 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1495 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1496 if( depth < ciKlass::primary_super_limit() ) {
duke@0 1497 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1498 // (Folds up type checking code.)
duke@0 1499 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1500 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1501 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1502 }
duke@0 1503 const Type* aift = load_array_final_field(tkls, klass);
duke@0 1504 if (aift != NULL) return aift;
duke@0 1505 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
duke@0 1506 && klass->is_array_klass()) {
duke@0 1507 // The field is arrayKlass::_component_mirror. Return its (constant) value.
duke@0 1508 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@0 1509 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@0 1510 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@0 1511 }
duke@0 1512 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1513 // The field is Klass::_java_mirror. Return its (constant) value.
duke@0 1514 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@0 1515 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@0 1516 return TypeInstPtr::make(klass->java_mirror());
duke@0 1517 }
duke@0 1518 }
duke@0 1519
duke@0 1520 // We can still check if we are loading from the primary_supers array at a
duke@0 1521 // shallow enough depth. Even though the klass is not exact, entries less
duke@0 1522 // than or equal to its super depth are correct.
duke@0 1523 if (klass->is_loaded() ) {
duke@0 1524 ciType *inner = klass->klass();
duke@0 1525 while( inner->is_obj_array_klass() )
duke@0 1526 inner = inner->as_obj_array_klass()->base_element_type();
duke@0 1527 if( inner->is_instance_klass() &&
duke@0 1528 !inner->as_instance_klass()->flags().is_interface() ) {
duke@0 1529 // Compute index into primary_supers array
duke@0 1530 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1531 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1532 if( depth < ciKlass::primary_super_limit() &&
duke@0 1533 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@0 1534 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1535 // (Folds up type checking code.)
duke@0 1536 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1537 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1538 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1539 }
duke@0 1540 }
duke@0 1541 }
duke@0 1542
duke@0 1543 // If the type is enough to determine that the thing is not an array,
duke@0 1544 // we can give the layout_helper a positive interval type.
duke@0 1545 // This will help short-circuit some reflective code.
duke@0 1546 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
duke@0 1547 && !klass->is_array_klass() // not directly typed as an array
duke@0 1548 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@0 1549 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@0 1550 ) {
duke@0 1551 // Note: When interfaces are reliable, we can narrow the interface
duke@0 1552 // test to (klass != Serializable && klass != Cloneable).
duke@0 1553 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1554 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@0 1555 // The key property of this type is that it folds up tests
duke@0 1556 // for array-ness, since it proves that the layout_helper is positive.
duke@0 1557 // Thus, a generic value like the basic object layout helper works fine.
duke@0 1558 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@0 1559 }
duke@0 1560 }
duke@0 1561
duke@0 1562 // If we are loading from a freshly-allocated object, produce a zero,
duke@0 1563 // if the load is provably beyond the header of the object.
duke@0 1564 // (Also allow a variable load from a fresh array to produce zero.)
duke@0 1565 if (ReduceFieldZeroing) {
duke@0 1566 Node* value = can_see_stored_value(mem,phase);
duke@0 1567 if (value != NULL && value->is_Con())
duke@0 1568 return value->bottom_type();
duke@0 1569 }
duke@0 1570
kvn@64 1571 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@223 1572 if (tinst != NULL && tinst->is_known_instance_field()) {
kvn@64 1573 // If we have an instance type and our memory input is the
kvn@64 1574 // programs's initial memory state, there is no matching store,
kvn@64 1575 // so just return a zero of the appropriate type
kvn@64 1576 Node *mem = in(MemNode::Memory);
kvn@64 1577 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@64 1578 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@64 1579 return Type::get_zero_type(_type->basic_type());
kvn@64 1580 }
kvn@64 1581 }
duke@0 1582 return _type;
duke@0 1583 }
duke@0 1584
duke@0 1585 //------------------------------match_edge-------------------------------------
duke@0 1586 // Do we Match on this edge index or not? Match only the address.
duke@0 1587 uint LoadNode::match_edge(uint idx) const {
duke@0 1588 return idx == MemNode::Address;
duke@0 1589 }
duke@0 1590
duke@0 1591 //--------------------------LoadBNode::Ideal--------------------------------------
duke@0 1592 //
duke@0 1593 // If the previous store is to the same address as this load,
duke@0 1594 // and the value stored was larger than a byte, replace this load
duke@0 1595 // with the value stored truncated to a byte. If no truncation is
duke@0 1596 // needed, the replacement is done in LoadNode::Identity().
duke@0 1597 //
duke@0 1598 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1599 Node* mem = in(MemNode::Memory);
duke@0 1600 Node* value = can_see_stored_value(mem,phase);
duke@0 1601 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1602 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
duke@0 1603 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
duke@0 1604 }
duke@0 1605 // Identity call will handle the case where truncation is not needed.
duke@0 1606 return LoadNode::Ideal(phase, can_reshape);
duke@0 1607 }
duke@0 1608
twisti@558 1609 //--------------------------LoadUSNode::Ideal-------------------------------------
duke@0 1610 //
duke@0 1611 // If the previous store is to the same address as this load,
duke@0 1612 // and the value stored was larger than a char, replace this load
duke@0 1613 // with the value stored truncated to a char. If no truncation is
duke@0 1614 // needed, the replacement is done in LoadNode::Identity().
duke@0 1615 //
twisti@558 1616 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1617 Node* mem = in(MemNode::Memory);
duke@0 1618 Node* value = can_see_stored_value(mem,phase);
duke@0 1619 if( value && !phase->type(value)->higher_equal( _type ) )
duke@0 1620 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
duke@0 1621 // Identity call will handle the case where truncation is not needed.
duke@0 1622 return LoadNode::Ideal(phase, can_reshape);
duke@0 1623 }
duke@0 1624
duke@0 1625 //--------------------------LoadSNode::Ideal--------------------------------------
duke@0 1626 //
duke@0 1627 // If the previous store is to the same address as this load,
duke@0 1628 // and the value stored was larger than a short, replace this load
duke@0 1629 // with the value stored truncated to a short. If no truncation is
duke@0 1630 // needed, the replacement is done in LoadNode::Identity().
duke@0 1631 //
duke@0 1632 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1633 Node* mem = in(MemNode::Memory);
duke@0 1634 Node* value = can_see_stored_value(mem,phase);
duke@0 1635 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1636 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
duke@0 1637 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
duke@0 1638 }
duke@0 1639 // Identity call will handle the case where truncation is not needed.
duke@0 1640 return LoadNode::Ideal(phase, can_reshape);
duke@0 1641 }
duke@0 1642
duke@0 1643 //=============================================================================
kvn@164 1644 //----------------------------LoadKlassNode::make------------------------------
kvn@164 1645 // Polymorphic factory method:
kvn@164 1646 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
kvn@164 1647 Compile* C = gvn.C;
kvn@164 1648 Node *ctl = NULL;
kvn@164 1649 // sanity check the alias category against the created node type
kvn@164 1650 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
kvn@164 1651 assert(adr_type != NULL, "expecting TypeOopPtr");
kvn@164 1652 #ifdef _LP64
kvn@164 1653 if (adr_type->is_ptr_to_narrowoop()) {
kvn@221 1654 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
kvn@221 1655 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
kvn@168 1656 }
kvn@164 1657 #endif
kvn@168 1658 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@168 1659 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
kvn@164 1660 }
kvn@164 1661
duke@0 1662 //------------------------------Value------------------------------------------
duke@0 1663 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1664 return klass_value_common(phase);
kvn@164 1665 }
kvn@164 1666
kvn@164 1667 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
duke@0 1668 // Either input is TOP ==> the result is TOP
duke@0 1669 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1670 if (t1 == Type::TOP) return Type::TOP;
duke@0 1671 Node *adr = in(MemNode::Address);
duke@0 1672 const Type *t2 = phase->type( adr );
duke@0 1673 if (t2 == Type::TOP) return Type::TOP;
duke@0 1674 const TypePtr *tp = t2->is_ptr();
duke@0 1675 if (TypePtr::above_centerline(tp->ptr()) ||
duke@0 1676 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@0 1677
duke@0 1678 // Return a more precise klass, if possible
duke@0 1679 const TypeInstPtr *tinst = tp->isa_instptr();
duke@0 1680 if (tinst != NULL) {
duke@0 1681 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@0 1682 int offset = tinst->offset();
duke@0 1683 if (ik == phase->C->env()->Class_klass()
duke@0 1684 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1685 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1686 // We are loading a special hidden field from a Class mirror object,
duke@0 1687 // the field which points to the VM's Klass metaobject.
duke@0 1688 ciType* t = tinst->java_mirror_type();
duke@0 1689 // java_mirror_type returns non-null for compile-time Class constants.
duke@0 1690 if (t != NULL) {
duke@0 1691 // constant oop => constant klass
duke@0 1692 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1693 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@0 1694 }
duke@0 1695 if (!t->is_klass()) {
duke@0 1696 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@0 1697 return TypePtr::NULL_PTR;
duke@0 1698 }
duke@0 1699 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@0 1700 return TypeKlassPtr::make(t->as_klass());
duke@0 1701 }
duke@0 1702 // non-constant mirror, so we can't tell what's going on
duke@0 1703 }
duke@0 1704 if( !ik->is_loaded() )
duke@0 1705 return _type; // Bail out if not loaded
duke@0 1706 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1707 if (tinst->klass_is_exact()) {
duke@0 1708 return TypeKlassPtr::make(ik);
duke@0 1709 }
duke@0 1710 // See if we can become precise: no subklasses and no interface
duke@0 1711 // (Note: We need to support verified interfaces.)
duke@0 1712 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1713 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1714 // Add a dependence; if any subclass added we need to recompile
duke@0 1715 if (!ik->is_final()) {
duke@0 1716 // %%% should use stronger assert_unique_concrete_subtype instead
duke@0 1717 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1718 }
duke@0 1719 // Return precise klass
duke@0 1720 return TypeKlassPtr::make(ik);
duke@0 1721 }
duke@0 1722
duke@0 1723 // Return root of possible klass
duke@0 1724 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@0 1725 }
duke@0 1726 }
duke@0 1727
duke@0 1728 // Check for loading klass from an array
duke@0 1729 const TypeAryPtr *tary = tp->isa_aryptr();
duke@0 1730 if( tary != NULL ) {
duke@0 1731 ciKlass *tary_klass = tary->klass();
duke@0 1732 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@0 1733 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@0 1734 if (tary->klass_is_exact()) {
duke@0 1735 return TypeKlassPtr::make(tary_klass);
duke@0 1736 }
duke@0 1737 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@0 1738 // If the klass is an object array, we defer the question to the
duke@0 1739 // array component klass.
duke@0 1740 if( ak->is_obj_array_klass() ) {
duke@0 1741 assert( ak->is_loaded(), "" );
duke@0 1742 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@0 1743 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@0 1744 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@0 1745 // See if we can become precise: no subklasses and no interface
duke@0 1746 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1747 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1748 // Add a dependence; if any subclass added we need to recompile
duke@0 1749 if (!ik->is_final()) {
duke@0 1750 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1751 }
duke@0 1752 // Return precise array klass
duke@0 1753 return TypeKlassPtr::make(ak);
duke@0 1754 }
duke@0 1755 }
duke@0 1756 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 1757 } else { // Found a type-array?
duke@0 1758 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1759 assert( ak->is_type_array_klass(), "" );
duke@0 1760 return TypeKlassPtr::make(ak); // These are always precise
duke@0 1761 }
duke@0 1762 }
duke@0 1763 }
duke@0 1764
duke@0 1765 // Check for loading klass from an array klass
duke@0 1766 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1767 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1768 ciKlass* klass = tkls->klass();
duke@0 1769 if( !klass->is_loaded() )
duke@0 1770 return _type; // Bail out if not loaded
duke@0 1771 if( klass->is_obj_array_klass() &&
duke@0 1772 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1773 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@0 1774 // // Always returning precise element type is incorrect,
duke@0 1775 // // e.g., element type could be object and array may contain strings
duke@0 1776 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@0 1777
duke@0 1778 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@0 1779 // according to the element type's subclassing.
duke@0 1780 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@0 1781 }
duke@0 1782 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
duke@0 1783 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1784 ciKlass* sup = klass->as_instance_klass()->super();
duke@0 1785 // The field is Klass::_super. Return its (constant) value.
duke@0 1786 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@0 1787 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@0 1788 }
duke@0 1789 }
duke@0 1790
duke@0 1791 // Bailout case
duke@0 1792 return LoadNode::Value(phase);
duke@0 1793 }
duke@0 1794
duke@0 1795 //------------------------------Identity---------------------------------------
duke@0 1796 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@0 1797 // Also feed through the klass in Allocate(...klass...)._klass.
duke@0 1798 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1799 return klass_identity_common(phase);
kvn@164 1800 }
kvn@164 1801
kvn@164 1802 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
duke@0 1803 Node* x = LoadNode::Identity(phase);
duke@0 1804 if (x != this) return x;
duke@0 1805
duke@0 1806 // Take apart the address into an oop and and offset.
duke@0 1807 // Return 'this' if we cannot.
duke@0 1808 Node* adr = in(MemNode::Address);
duke@0 1809 intptr_t offset = 0;
duke@0 1810 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1811 if (base == NULL) return this;
duke@0 1812 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@0 1813 if (toop == NULL) return this;
duke@0 1814
duke@0 1815 // We can fetch the klass directly through an AllocateNode.
duke@0 1816 // This works even if the klass is not constant (clone or newArray).
duke@0 1817 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1818 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@0 1819 if (allocated_klass != NULL) {
duke@0 1820 return allocated_klass;
duke@0 1821 }
duke@0 1822 }
duke@0 1823
duke@0 1824 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
duke@0 1825 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
duke@0 1826 // See inline_native_Class_query for occurrences of these patterns.
duke@0 1827 // Java Example: x.getClass().isAssignableFrom(y)
duke@0 1828 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@0 1829 //
duke@0 1830 // This improves reflective code, often making the Class
duke@0 1831 // mirror go completely dead. (Current exception: Class
duke@0 1832 // mirrors may appear in debug info, but we could clean them out by
duke@0 1833 // introducing a new debug info operator for klassOop.java_mirror).
duke@0 1834 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@0 1835 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1836 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1837 // We are loading a special hidden field from a Class mirror,
duke@0 1838 // the field which points to its Klass or arrayKlass metaobject.
duke@0 1839 if (base->is_Load()) {
duke@0 1840 Node* adr2 = base->in(MemNode::Address);
duke@0 1841 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@0 1842 if (tkls != NULL && !tkls->empty()
duke@0 1843 && (tkls->klass()->is_instance_klass() ||
duke@0 1844 tkls->klass()->is_array_klass())
duke@0 1845 && adr2->is_AddP()
duke@0 1846 ) {
duke@0 1847 int mirror_field = Klass::java_mirror_offset_in_bytes();
duke@0 1848 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1849 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
duke@0 1850 }
duke@0 1851 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
duke@0 1852 return adr2->in(AddPNode::Base);
duke@0 1853 }
duke@0 1854 }
duke@0 1855 }
duke@0 1856 }
duke@0 1857
duke@0 1858 return this;
duke@0 1859 }
duke@0 1860
kvn@164 1861
kvn@164 1862 //------------------------------Value------------------------------------------
kvn@164 1863 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1864 const Type *t = klass_value_common(phase);
kvn@221 1865 if (t == Type::TOP)
kvn@221 1866 return t;
kvn@221 1867
kvn@221 1868 return t->make_narrowoop();
kvn@164 1869 }
kvn@164 1870
kvn@164 1871 //------------------------------Identity---------------------------------------
kvn@164 1872 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
kvn@164 1873 // Also feed through the klass in Allocate(...klass...)._klass.
kvn@164 1874 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1875 Node *x = klass_identity_common(phase);
kvn@164 1876
kvn@164 1877 const Type *t = phase->type( x );
kvn@164 1878 if( t == Type::TOP ) return x;
kvn@164 1879 if( t->isa_narrowoop()) return x;
kvn@164 1880
kvn@221 1881 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
kvn@164 1882 }
kvn@164 1883
duke@0 1884 //------------------------------Value-----------------------------------------
duke@0 1885 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@0 1886 // Either input is TOP ==> the result is TOP
duke@0 1887 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1888 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1889 Node *adr = in(MemNode::Address);
duke@0 1890 const Type *t2 = phase->type( adr );
duke@0 1891 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1892 const TypePtr *tp = t2->is_ptr();
duke@0 1893 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@0 1894 const TypeAryPtr *tap = tp->isa_aryptr();
duke@0 1895 if( !tap ) return _type;
duke@0 1896 return tap->size();
duke@0 1897 }
duke@0 1898
rasbold@366 1899 //-------------------------------Ideal---------------------------------------
rasbold@366 1900 // Feed through the length in AllocateArray(...length...)._length.
rasbold@366 1901 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
rasbold@366 1902 Node* p = MemNode::Ideal_common(phase, can_reshape);
rasbold@366 1903 if (p) return (p == NodeSentinel) ? NULL : p;
rasbold@366 1904
rasbold@366 1905 // Take apart the address into an oop and and offset.
rasbold@366 1906 // Return 'this' if we cannot.
rasbold@366 1907 Node* adr = in(MemNode::Address);
rasbold@366 1908 intptr_t offset = 0;
rasbold@366 1909 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
rasbold@366 1910 if (base == NULL) return NULL;
rasbold@366 1911 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
rasbold@366 1912 if (tary == NULL) return NULL;
rasbold@366 1913
rasbold@366 1914 // We can fetch the length directly through an AllocateArrayNode.
rasbold@366 1915 // This works even if the length is not constant (clone or newArray).
rasbold@366 1916 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 1917 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 1918 if (alloc != NULL) {
rasbold@366 1919 Node* allocated_length = alloc->Ideal_length();
rasbold@366 1920 Node* len = alloc->make_ideal_length(tary, phase);
rasbold@366 1921 if (allocated_length != len) {
rasbold@366 1922 // New CastII improves on this.
rasbold@366 1923 return len;
rasbold@366 1924 }
rasbold@366 1925 }
rasbold@366 1926 }
rasbold@366 1927
rasbold@366 1928 return NULL;
rasbold@366 1929 }
rasbold@366 1930
duke@0 1931 //------------------------------Identity---------------------------------------
duke@0 1932 // Feed through the length in AllocateArray(...length...)._length.
duke@0 1933 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@0 1934 Node* x = LoadINode::Identity(phase);
duke@0 1935 if (x != this) return x;
duke@0 1936
duke@0 1937 // Take apart the address into an oop and and offset.
duke@0 1938 // Return 'this' if we cannot.
duke@0 1939 Node* adr = in(MemNode::Address);
duke@0 1940 intptr_t offset = 0;
duke@0 1941 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1942 if (base == NULL) return this;
duke@0 1943 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@0 1944 if (tary == NULL) return this;
duke@0 1945
duke@0 1946 // We can fetch the length directly through an AllocateArrayNode.
duke@0 1947 // This works even if the length is not constant (clone or newArray).
duke@0 1948 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 1949 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 1950 if (alloc != NULL) {
rasbold@366 1951 Node* allocated_length = alloc->Ideal_length();
rasbold@366 1952 // Do not allow make_ideal_length to allocate a CastII node.
rasbold@366 1953 Node* len = alloc->make_ideal_length(tary, phase, false);
rasbold@366 1954 if (allocated_length == len) {
rasbold@366 1955 // Return allocated_length only if it would not be improved by a CastII.
rasbold@366 1956 return allocated_length;
rasbold@366 1957 }
duke@0 1958 }
duke@0 1959 }
duke@0 1960
duke@0 1961 return this;
duke@0 1962
duke@0 1963 }
rasbold@366 1964
duke@0 1965 //=============================================================================
duke@0 1966 //---------------------------StoreNode::make-----------------------------------
duke@0 1967 // Polymorphic factory method:
coleenp@113 1968 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
coleenp@113 1969 Compile* C = gvn.C;
coleenp@113 1970
duke@0 1971 switch (bt) {
duke@0 1972 case T_BOOLEAN:
duke@0 1973 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
duke@0 1974 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
duke@0 1975 case T_CHAR:
duke@0 1976 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
duke@0 1977 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
duke@0 1978 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
duke@0 1979 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
duke@0 1980 case T_ADDRESS:
coleenp@113 1981 case T_OBJECT:
coleenp@113 1982 #ifdef _LP64
kvn@163 1983 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
coleenp@113 1984 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
coleenp@113 1985 adr->bottom_type()->isa_rawptr())) {
kvn@221 1986 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
kvn@221 1987 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
coleenp@113 1988 } else
coleenp@113 1989 #endif
kvn@221 1990 {
kvn@221 1991 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
kvn@221 1992 }
duke@0 1993 }
duke@0 1994 ShouldNotReachHere();
duke@0 1995 return (StoreNode*)NULL;
duke@0 1996 }
duke@0 1997
duke@0 1998 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
duke@0 1999 bool require_atomic = true;
duke@0 2000 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
duke@0 2001 }
duke@0 2002
duke@0 2003
duke@0 2004 //--------------------------bottom_type----------------------------------------
duke@0 2005 const Type *StoreNode::bottom_type() const {
duke@0 2006 return Type::MEMORY;
duke@0 2007 }
duke@0 2008
duke@0 2009 //------------------------------hash-------------------------------------------
duke@0 2010 uint StoreNode::hash() const {
duke@0 2011 // unroll addition of interesting fields
duke@0 2012 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@0 2013
duke@0 2014 // Since they are not commoned, do not hash them:
duke@0 2015 return NO_HASH;
duke@0 2016 }
duke@0 2017
duke@0 2018 //------------------------------Ideal------------------------------------------
duke@0 2019 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@0 2020 // When a store immediately follows a relevant allocation/initialization,
duke@0 2021 // try to capture it into the initialization, or hoist it above.
duke@0 2022 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 2023 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 2024 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 2025
duke@0 2026 Node* mem = in(MemNode::Memory);
duke@0 2027 Node* address = in(MemNode::Address);
duke@0 2028
duke@0 2029 // Back-to-back stores to same address? Fold em up.
duke@0 2030 // Generally unsafe if I have intervening uses...
duke@0 2031 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
duke@0 2032 // Looking at a dead closed cycle of memory?
duke@0 2033 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@0 2034
duke@0 2035 assert(Opcode() == mem->Opcode() ||
duke@0 2036 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@0 2037 "no mismatched stores, except on raw memory");
duke@0 2038
duke@0 2039 if (mem->outcnt() == 1 && // check for intervening uses
duke@0 2040 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@0 2041 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@0 2042 // For example, 'mem' might be the final state at a conditional return.
duke@0 2043 // Or, 'mem' might be used by some node which is live at the same time
duke@0 2044 // 'this' is live, which might be unschedulable. So, require exactly
duke@0 2045 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@0 2046 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@0 2047 if (can_reshape) { // (%%% is this an anachronism?)
duke@0 2048 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@0 2049 phase->is_IterGVN());
duke@0 2050 } else {
duke@0 2051 // It's OK to do this in the parser, since DU info is always accurate,
duke@0 2052 // and the parser always refers to nodes via SafePointNode maps.
duke@0 2053 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@0 2054 }
duke@0 2055 return this;
duke@0 2056 }
duke@0 2057 }
duke@0 2058
duke@0 2059 // Capture an unaliased, unconditional, simple store into an initializer.
duke@0 2060 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@0 2061 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@0 2062 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 2063 InitializeNode* init = mem->in(0)->as_Initialize();
duke@0 2064 intptr_t offset = init->can_capture_store(this, phase);
duke@0 2065 if (offset > 0) {
duke@0 2066 Node* moved = init->capture_store(this, offset, phase);
duke@0 2067 // If the InitializeNode captured me, it made a raw copy of me,
duke@0 2068 // and I need to disappear.
duke@0 2069 if (moved != NULL) {
duke@0 2070 // %%% hack to ensure that Ideal returns a new node:
duke@0 2071 mem = MergeMemNode::make(phase->C, mem);
duke@0 2072 return mem; // fold me away
duke@0 2073 }
duke@0 2074 }
duke@0 2075 }
duke@0 2076
duke@0 2077 return NULL; // No further progress
duke@0 2078 }
duke@0 2079
duke@0 2080 //------------------------------Value-----------------------------------------
duke@0 2081 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@0 2082 // Either input is TOP ==> the result is TOP
duke@0 2083 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 2084 if( t1 == Type::TOP ) return Type::TOP;
duke@0 2085 const Type *t2 = phase->type( in(MemNode::Address) );
duke@0 2086 if( t2 == Type::TOP ) return Type::TOP;
duke@0 2087 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@0 2088 if( t3 == Type::TOP ) return Type::TOP;
duke@0 2089 return Type::MEMORY;
duke@0 2090 }
duke@0 2091
duke@0 2092 //------------------------------Identity---------------------------------------
duke@0 2093 // Remove redundant stores:
duke@0 2094 // Store(m, p, Load(m, p)) changes to m.
duke@0 2095 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@0 2096 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@0 2097 Node* mem = in(MemNode::Memory);
duke@0 2098 Node* adr = in(MemNode::Address);
duke@0 2099 Node* val = in(MemNode::ValueIn);
duke@0 2100
duke@0 2101 // Load then Store? Then the Store is useless
duke@0 2102 if (val->is_Load() &&
duke@0 2103 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
duke@0 2104 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
duke@0 2105 val->as_Load()->store_Opcode() == Opcode()) {
duke@0 2106 return mem;
duke@0 2107 }
duke@0 2108
duke@0 2109 // Two stores in a row of the same value?
duke@0 2110 if (mem->is_Store() &&
duke@0 2111 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
duke@0 2112 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
duke@0 2113 mem->Opcode() == Opcode()) {
duke@0 2114 return mem;
duke@0 2115 }
duke@0 2116
duke@0 2117 // Store of zero anywhere into a freshly-allocated object?
duke@0 2118 // Then the store is useless.
duke@0 2119 // (It must already have been captured by the InitializeNode.)
duke@0 2120 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@0 2121 // a newly allocated object is already all-zeroes everywhere
duke@0 2122 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@0 2123 return mem;
duke@0 2124 }
duke@0 2125
duke@0 2126 // the store may also apply to zero-bits in an earlier object
duke@0 2127 Node* prev_mem = find_previous_store(phase);
duke@0 2128 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 2129 if (prev_mem != NULL) {
duke@0 2130 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@0 2131 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@0 2132 // prev_val and val might differ by a cast; it would be good
duke@0 2133 // to keep the more informative of the two.
duke@0 2134 return mem;
duke@0 2135 }
duke@0 2136 }
duke@0 2137 }
duke@0 2138
duke@0 2139 return this;
duke@0 2140 }
duke@0 2141
duke@0 2142 //------------------------------match_edge-------------------------------------
duke@0 2143 // Do we Match on this edge index or not? Match only memory & value
duke@0 2144 uint StoreNode::match_edge(uint idx) const {
duke@0 2145 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@0 2146 }
duke@0 2147
duke@0 2148 //------------------------------cmp--------------------------------------------
duke@0 2149 // Do not common stores up together. They generally have to be split
duke@0 2150 // back up anyways, so do not bother.
duke@0 2151 uint StoreNode::cmp( const Node &n ) const {
duke@0 2152 return (&n == this); // Always fail except on self
duke@0 2153 }
duke@0 2154
duke@0 2155 //------------------------------Ideal_masked_input-----------------------------
duke@0 2156 // Check for a useless mask before a partial-word store
duke@0 2157 // (StoreB ... (AndI valIn conIa) )
duke@0 2158 // If (conIa & mask == mask) this simplifies to
duke@0 2159 // (StoreB ... (valIn) )
duke@0 2160 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@0 2161 Node *val = in(MemNode::ValueIn);
duke@0 2162 if( val->Opcode() == Op_AndI ) {
duke@0 2163 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2164 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@0 2165 set_req(MemNode::ValueIn, val->in(1));
duke@0 2166 return this;
duke@0 2167 }
duke@0 2168 }
duke@0 2169 return NULL;
duke@0 2170 }
duke@0 2171
duke@0 2172
duke@0 2173 //------------------------------Ideal_sign_extended_input----------------------
duke@0 2174 // Check for useless sign-extension before a partial-word store
duke@0 2175 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@0 2176 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@0 2177 // (StoreB ... (valIn) )
duke@0 2178 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@0 2179 Node *val = in(MemNode::ValueIn);
duke@0 2180 if( val->Opcode() == Op_RShiftI ) {
duke@0 2181 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2182 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@0 2183 Node *shl = val->in(1);
duke@0 2184 if( shl->Opcode() == Op_LShiftI ) {
duke@0 2185 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@0 2186 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@0 2187 set_req(MemNode::ValueIn, shl->in(1));
duke@0 2188 return this;
duke@0 2189 }
duke@0 2190 }
duke@0 2191 }
duke@0 2192 }
duke@0 2193 return NULL;
duke@0 2194 }
duke@0 2195
duke@0 2196 //------------------------------value_never_loaded-----------------------------------
duke@0 2197 // Determine whether there are any possible loads of the value stored.
duke@0 2198 // For simplicity, we actually check if there are any loads from the
duke@0 2199 // address stored to, not just for loads of the value stored by this node.
duke@0 2200 //
duke@0 2201 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@0 2202 Node *adr = in(Address);
duke@0 2203 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@0 2204 if (adr_oop == NULL)
duke@0 2205 return false;
kvn@223 2206 if (!adr_oop->is_known_instance_field())
duke@0 2207 return false; // if not a distinct instance, there may be aliases of the address
duke@0 2208 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@0 2209 Node *use = adr->fast_out(i);
duke@0 2210 int opc = use->Opcode();
duke@0 2211 if (use->is_Load() || use->is_LoadStore()) {
duke@0 2212 return false;
duke@0 2213 }
duke@0 2214 }
duke@0 2215 return true;
duke@0 2216 }
duke@0 2217
duke@0 2218 //=============================================================================
duke@0 2219 //------------------------------Ideal------------------------------------------
duke@0 2220 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2221 // we can skip the AND operation. If the store is from a sign-extension
duke@0 2222 // (a left shift, then right shift) we can skip both.
duke@0 2223 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2224 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@0 2225 if( progress != NULL ) return progress;
duke@0 2226
duke@0 2227 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@0 2228 if( progress != NULL ) return progress;
duke@0 2229
duke@0 2230 // Finally check the default case
duke@0 2231 return StoreNode::Ideal(phase, can_reshape);
duke@0 2232 }
duke@0 2233
duke@0 2234 //=============================================================================
duke@0 2235 //------------------------------Ideal------------------------------------------
duke@0 2236 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2237 // we can skip the AND operation
duke@0 2238 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2239 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@0 2240 if( progress != NULL ) return progress;
duke@0 2241
duke@0 2242 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@0 2243 if( progress != NULL ) return progress;
duke@0 2244
duke@0 2245 // Finally check the default case
duke@0 2246 return StoreNode::Ideal(phase, can_reshape);
duke@0 2247 }
duke@0 2248
duke@0 2249 //=============================================================================
duke@0 2250 //------------------------------Identity---------------------------------------
duke@0 2251 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@0 2252 // No need to card mark when storing a null ptr
duke@0 2253 Node* my_store = in(MemNode::OopStore);
duke@0 2254 if (my_store->is_Store()) {
duke@0 2255 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@0 2256 if( t1 == TypePtr::NULL_PTR ) {
duke@0 2257 return in(MemNode::Memory);
duke@0 2258 }
duke@0 2259 }
duke@0 2260 return this;
duke@0 2261 }
duke@0 2262
duke@0 2263 //------------------------------Value-----------------------------------------
duke@0 2264 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@43 2265 // Either input is TOP ==> the result is TOP
kvn@43 2266 const Type *t = phase->type( in(MemNode::Memory) );
kvn@43 2267 if( t == Type::TOP ) return Type::TOP;
kvn@43 2268 t = phase->type( in(MemNode::Address) );
kvn@43 2269 if( t == Type::TOP ) return Type::TOP;
kvn@43 2270 t = phase->type( in(MemNode::ValueIn) );
kvn@43 2271 if( t == Type::TOP ) return Type::TOP;
duke@0 2272 // If extra input is TOP ==> the result is TOP
kvn@43 2273 t = phase->type( in(MemNode::OopStore) );
kvn@43 2274 if( t == Type::TOP ) return Type::TOP;
duke@0 2275
duke@0 2276 return StoreNode::Value( phase );
duke@0 2277 }
duke@0 2278
duke@0 2279
duke@0 2280 //=============================================================================
duke@0 2281 //----------------------------------SCMemProjNode------------------------------
duke@0 2282 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@0 2283 {
duke@0 2284 return bottom_type();
duke@0 2285 }
duke@0 2286
duke@0 2287 //=============================================================================
duke@0 2288 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
duke@0 2289 init_req(MemNode::Control, c );
duke@0 2290 init_req(MemNode::Memory , mem);
duke@0 2291 init_req(MemNode::Address, adr);
duke@0 2292 init_req(MemNode::ValueIn, val);
duke@0 2293 init_req( ExpectedIn, ex );
duke@0 2294 init_class_id(Class_LoadStore);
duke@0 2295
duke@0 2296 }
duke@0 2297
duke@0 2298 //=============================================================================
duke@0 2299 //-------------------------------adr_type--------------------------------------
duke@0 2300 // Do we Match on this edge index or not? Do not match memory
duke@0 2301 const TypePtr* ClearArrayNode::adr_type() const {
duke@0 2302 Node *adr = in(3);
duke@0 2303 return MemNode::calculate_adr_type(adr->bottom_type());
duke@0 2304 }
duke@0 2305
duke@0 2306 //------------------------------match_edge-------------------------------------
duke@0 2307 // Do we Match on this edge index or not? Do not match memory
duke@0 2308 uint ClearArrayNode::match_edge(uint idx) const {
duke@0 2309 return idx > 1;
duke@0 2310 }
duke@0 2311
duke@0 2312 //------------------------------Identity---------------------------------------
duke@0 2313 // Clearing a zero length array does nothing
duke@0 2314 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@68 2315 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@0 2316 }
duke@0 2317
duke@0 2318 //------------------------------Idealize---------------------------------------
duke@0 2319 // Clearing a short array is faster with stores
duke@0 2320 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2321 const int unit = BytesPerLong;
duke@0 2322 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@0 2323 if (!t) return NULL;
duke@0 2324 if (!t->is_con()) return NULL;
duke@0 2325 intptr_t raw_count = t->get_con();
duke@0 2326 intptr_t size = raw_count;
duke@0 2327 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@0 2328 // Clearing nothing uses the Identity call.
duke@0 2329 // Negative clears are possible on dead ClearArrays
duke@0 2330 // (see jck test stmt114.stmt11402.val).
duke@0 2331 if (size <= 0 || size % unit != 0) return NULL;
duke@0 2332 intptr_t count = size / unit;
duke@0 2333 // Length too long; use fast hardware clear
duke@0 2334 if (size > Matcher::init_array_short_size) return NULL;
duke@0 2335 Node *mem = in(1);
duke@0 2336 if( phase->type(mem)==Type::TOP ) return NULL;
duke@0 2337 Node *adr = in(3);
duke@0 2338 const Type* at = phase->type(adr);
duke@0 2339 if( at==Type::TOP ) return NULL;
duke@0 2340 const TypePtr* atp = at->isa_ptr();
duke@0 2341 // adjust atp to be the correct array element address type
duke@0 2342 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@0 2343 else atp = atp->add_offset(Type::OffsetBot);
duke@0 2344 // Get base for derived pointer purposes
duke@0 2345 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@0 2346 Node *base = adr->in(1);
duke@0 2347
duke@0 2348 Node *zero = phase->makecon(TypeLong::ZERO);
duke@0 2349 Node *off = phase->MakeConX(BytesPerLong);
duke@0 2350 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2351 count--;
duke@0 2352 while( count-- ) {
duke@0 2353 mem = phase->transform(mem);
duke@0 2354 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
duke@0 2355 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2356 }
duke@0 2357 return mem;
duke@0 2358 }
duke@0 2359
duke@0 2360 //----------------------------clear_memory-------------------------------------
duke@0 2361 // Generate code to initialize object storage to zero.
duke@0 2362 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2363 intptr_t start_offset,
duke@0 2364 Node* end_offset,
duke@0 2365 PhaseGVN* phase) {
duke@0 2366 Compile* C = phase->C;
duke@0 2367 intptr_t offset = start_offset;
duke@0 2368
duke@0 2369 int unit = BytesPerLong;
duke@0 2370 if ((offset % unit) != 0) {
duke@0 2371 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
duke@0 2372 adr = phase->transform(adr);
duke@0 2373 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2374 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2375 mem = phase->transform(mem);
duke@0 2376 offset += BytesPerInt;
duke@0 2377 }
duke@0 2378 assert((offset % unit) == 0, "");
duke@0 2379
duke@0 2380 // Initialize the remaining stuff, if any, with a ClearArray.
duke@0 2381 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@0 2382 }
duke@0 2383
duke@0 2384 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2385 Node* start_offset,
duke@0 2386 Node* end_offset,
duke@0 2387 PhaseGVN* phase) {
never@68 2388 if (start_offset == end_offset) {
never@68 2389 // nothing to do
never@68 2390 return mem;
never@68 2391 }
never@68 2392
duke@0 2393 Compile* C = phase->C;
duke@0 2394 int unit = BytesPerLong;
duke@0 2395 Node* zbase = start_offset;
duke@0 2396 Node* zend = end_offset;
duke@0 2397
duke@0 2398 // Scale to the unit required by the CPU:
duke@0 2399 if (!Matcher::init_array_count_is_in_bytes) {
duke@0 2400 Node* shift = phase->intcon(exact_log2(unit));
duke@0 2401 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
duke@0 2402 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
duke@0 2403 }
duke@0 2404
duke@0 2405 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
duke@0 2406 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
duke@0 2407
duke@0 2408 // Bulk clear double-words
duke@0 2409 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
duke@0 2410 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
duke@0 2411 return phase->transform(mem);
duke@0 2412 }
duke@0 2413
duke@0 2414 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2415 intptr_t start_offset,
duke@0 2416 intptr_t end_offset,
duke@0 2417 PhaseGVN* phase) {
never@68 2418 if (start_offset == end_offset) {
never@68 2419 // nothing to do
never@68 2420 return mem;
never@68 2421 }
never@68 2422
duke@0 2423 Compile* C = phase->C;
duke@0 2424 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@0 2425 intptr_t done_offset = end_offset;
duke@0 2426 if ((done_offset % BytesPerLong) != 0) {
duke@0 2427 done_offset -= BytesPerInt;
duke@0 2428 }
duke@0 2429 if (done_offset > start_offset) {
duke@0 2430 mem = clear_memory(ctl, mem, dest,
duke@0 2431 start_offset, phase->MakeConX(done_offset), phase);
duke@0 2432 }
duke@0 2433 if (done_offset < end_offset) { // emit the final 32-bit store
duke@0 2434 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@0 2435 adr = phase->transform(adr);
duke@0 2436 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2437 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2438 mem = phase->transform(mem);
duke@0 2439 done_offset += BytesPerInt;
duke@0 2440 }
duke@0 2441 assert(done_offset == end_offset, "");
duke@0 2442 return mem;
duke@0 2443 }
duke@0 2444
duke@0 2445 //=============================================================================
duke@0 2446 // Do we match on this edge? No memory edges
duke@0 2447 uint StrCompNode::match_edge(uint idx) const {
duke@0 2448 return idx == 5 || idx == 6;
duke@0 2449 }
duke@0 2450
duke@0 2451 //------------------------------Ideal------------------------------------------
duke@0 2452 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2453 // control copies
duke@0 2454 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2455 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2456 }
duke@0 2457
rasbold@169 2458 //------------------------------Ideal------------------------------------------
rasbold@169 2459 // Return a node which is more "ideal" than the current node. Strip out
rasbold@169 2460 // control copies
rasbold@169 2461 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
rasbold@169 2462 return remove_dead_region(phase, can_reshape) ? this : NULL;
rasbold@169 2463 }
rasbold@169 2464
duke@0 2465
duke@0 2466 //=============================================================================
duke@0 2467 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@0 2468 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@0 2469 _adr_type(C->get_adr_type(alias_idx))
duke@0 2470 {
duke@0 2471 init_class_id(Class_MemBar);
duke@0 2472 Node* top = C->top();
duke@0 2473 init_req(TypeFunc::I_O,top);
duke@0 2474 init_req(TypeFunc::FramePtr,top);
duke@0 2475 init_req(TypeFunc::ReturnAdr,top);
duke@0 2476 if (precedent != NULL)
duke@0 2477 init_req(TypeFunc::Parms, precedent);
duke@0 2478 }
duke@0 2479
duke@0 2480 //------------------------------cmp--------------------------------------------
duke@0 2481 uint MemBarNode::hash() const { return NO_HASH; }
duke@0 2482 uint MemBarNode::cmp( const Node &n ) const {
duke@0 2483 return (&n == this); // Always fail except on self
duke@0 2484 }
duke@0 2485
duke@0 2486 //------------------------------make-------------------------------------------
duke@0 2487 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@0 2488 int len = Precedent + (pn == NULL? 0: 1);
duke@0 2489 switch (opcode) {
duke@0 2490 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
duke@0 2491 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
duke@0 2492 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
duke@0 2493 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
duke@0 2494 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
duke@0 2495 default: ShouldNotReachHere(); return NULL;
duke@0 2496 }
duke@0 2497 }
duke@0 2498
duke@0 2499 //------------------------------Ideal------------------------------------------
duke@0 2500 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2501 // control copies
duke@0 2502 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@305 2503 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2504 }
duke@0 2505
duke@0 2506 //------------------------------Value------------------------------------------
duke@0 2507 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@0 2508 if( !in(0) ) return Type::TOP;
duke@0 2509 if( phase->type(in(0)) == Type::TOP )
duke@0 2510 return Type::TOP;
duke@0 2511 return TypeTuple::MEMBAR;
duke@0 2512 }
duke@0 2513
duke@0 2514 //------------------------------match------------------------------------------
duke@0 2515 // Construct projections for memory.
duke@0 2516 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@0 2517 switch (proj->_con) {
duke@0 2518 case TypeFunc::Control:
duke@0 2519 case TypeFunc::Memory:
duke@0 2520 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@0 2521 }
duke@0 2522 ShouldNotReachHere();
duke@0 2523 return NULL;
duke@0 2524 }
duke@0 2525
duke@0 2526 //===========================InitializeNode====================================
duke@0 2527 // SUMMARY:
duke@0 2528 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@0 2529 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@0 2530 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@0 2531 // It can coalesce related raw stores into larger units (called 'tiles').
duke@0 2532 // It can avoid zeroing new storage for memory units which have raw inits.
duke@0 2533 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@0 2534 //
duke@0 2535 // EXAMPLE:
duke@0 2536 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@0 2537 // ctl = incoming control; mem* = incoming memory
duke@0 2538 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@0 2539 // First allocate uninitialized memory and fill in the header:
duke@0 2540 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@0 2541 // ctl := alloc.Control; mem* := alloc.Memory*
duke@0 2542 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@0 2543 // Then initialize to zero the non-header parts of the raw memory block:
duke@0 2544 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@0 2545 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@0 2546 // After the initialize node executes, the object is ready for service:
duke@0 2547 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@0 2548 // Suppose its body is immediately initialized as {1,2}:
duke@0 2549 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2550 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2551 // mem.SLICE(#short[*]) := store2
duke@0 2552 //
duke@0 2553 // DETAILS:
duke@0 2554 // An InitializeNode collects and isolates object initialization after
duke@0 2555 // an AllocateNode and before the next possible safepoint. As a
duke@0 2556 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@0 2557 // down past any safepoint or any publication of the allocation.
duke@0 2558 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@0 2559 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@0 2560 //
duke@0 2561 // The semantics of the InitializeNode include an implicit zeroing of
duke@0 2562 // the new object from object header to the end of the object.
duke@0 2563 // (The object header and end are determined by the AllocateNode.)
duke@0 2564 //
duke@0 2565 // Certain stores may be added as direct inputs to the InitializeNode.
duke@0 2566 // These stores must update raw memory, and they must be to addresses
duke@0 2567 // derived from the raw address produced by AllocateNode, and with
duke@0 2568 // a constant offset. They must be ordered by increasing offset.
duke@0 2569 // The first one is at in(RawStores), the last at in(req()-1).
duke@0 2570 // Unlike most memory operations, they are not linked in a chain,
duke@0 2571 // but are displayed in parallel as users of the rawmem output of
duke@0 2572 // the allocation.
duke@0 2573 //
duke@0 2574 // (See comments in InitializeNode::capture_store, which continue
duke@0 2575 // the example given above.)
duke@0 2576 //
duke@0 2577 // When the associated Allocate is macro-expanded, the InitializeNode
duke@0 2578 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@0 2579 // may also be created at that point to represent any required zeroing.
duke@0 2580 // The InitializeNode is then marked 'complete', prohibiting further
duke@0 2581 // capturing of nearby memory operations.
duke@0 2582 //
duke@0 2583 // During macro-expansion, all captured initializations which store
duke@0 2584 // constant values of 32 bits or smaller are coalesced (if advantagous)
duke@0 2585 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@0 2586 // initialized in fewer memory operations. Memory words which are
duke@0 2587 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@0 2588 // by explicit stores of zero. (The code shape happens to do all
duke@0 2589 // zeroing first, then all other stores, with both sequences occurring
duke@0 2590 // in order of ascending offsets.)
duke@0 2591 //
duke@0 2592 // Alternatively, code may be inserted between an AllocateNode and its
duke@0 2593 // InitializeNode, to perform arbitrary initialization of the new object.
duke@0 2594 // E.g., the object copying intrinsics insert complex data transfers here.
duke@0 2595 // The initialization must then be marked as 'complete' disable the
duke@0 2596 // built-in zeroing semantics and the collection of initializing stores.
duke@0 2597 //
duke@0 2598 // While an InitializeNode is incomplete, reads from the memory state
duke@0 2599 // produced by it are optimizable if they match the control edge and
duke@0 2600 // new oop address associated with the allocation/initialization.
duke@0 2601 // They return a stored value (if the offset matches) or else zero.
duke@0 2602 // A write to the memory state, if it matches control and address,
duke@0 2603 // and if it is to a constant offset, may be 'captured' by the
duke@0 2604 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@0 2605 // inside the initialization, to the raw oop produced by the allocation.
duke@0 2606 // Operations on addresses which are provably distinct (e.g., to
duke@0 2607 // other AllocateNodes) are allowed to bypass the initialization.
duke@0 2608 //
duke@0 2609 // The effect of all this is to consolidate object initialization
duke@0 2610 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@0 2611 // single location, where it can be optimized as a unit.
duke@0 2612 //
duke@0 2613 // Only stores with an offset less than TrackedInitializationLimit words
duke@0 2614 // will be considered for capture by an InitializeNode. This puts a
duke@0 2615 // reasonable limit on the complexity of optimized initializations.
duke@0 2616
duke@0 2617 //---------------------------InitializeNode------------------------------------
duke@0 2618 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
duke@0 2619 : _is_complete(false),
duke@0 2620 MemBarNode(C, adr_type, rawoop)
duke@0 2621 {
duke@0 2622 init_class_id(Class_Initialize);
duke@0 2623
duke@0 2624 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@0 2625 assert(in(RawAddress) == rawoop, "proper init");
duke@0 2626 // Note: allocation() can be NULL, for secondary initialization barriers
duke@0 2627 }
duke@0 2628
duke@0 2629 // Since this node is not matched, it will be processed by the
duke@0 2630 // register allocator. Declare that there are no constraints
duke@0 2631 // on the allocation of the RawAddress edge.
duke@0 2632 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@0 2633 // This edge should be set to top, by the set_complete. But be conservative.
duke@0 2634 if (idx == InitializeNode::RawAddress)
duke@0 2635 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@0 2636 return RegMask::Empty;
duke@0 2637 }
duke@0 2638
duke@0 2639 Node* InitializeNode::memory(uint alias_idx) {
duke@0 2640 Node* mem = in(Memory);
duke@0 2641 if (mem->is_MergeMem()) {
duke@0 2642 return mem->as_MergeMem()->memory_at(alias_idx);
duke@0 2643 } else {
duke@0 2644 // incoming raw memory is not split
duke@0 2645 return mem;
duke@0 2646 }
duke@0 2647 }
duke@0 2648
duke@0 2649 bool InitializeNode::is_non_zero() {
duke@0 2650 if (is_complete()) return false;
duke@0 2651 remove_extra_zeroes();
duke@0 2652 return (req() > RawStores);
duke@0 2653 }
duke@0 2654
duke@0 2655 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@0 2656 assert(!is_complete(), "caller responsibility");
duke@0 2657 _is_complete = true;
duke@0 2658
duke@0 2659 // After this node is complete, it contains a bunch of
duke@0 2660 // raw-memory initializations. There is no need for
duke@0 2661 // it to have anything to do with non-raw memory effects.
duke@0 2662 // Therefore, tell all non-raw users to re-optimize themselves,
duke@0 2663 // after skipping the memory effects of this initialization.
duke@0 2664 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@0 2665 if (igvn) igvn->add_users_to_worklist(this);
duke@0 2666 }
duke@0 2667
duke@0 2668 // convenience function
duke@0 2669 // return false if the init contains any stores already
duke@0 2670 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@0 2671 InitializeNode* init = initialization();
duke@0 2672 if (init == NULL || init->is_complete()) return false;
duke@0 2673 init->remove_extra_zeroes();
duke@0 2674 // for now, if this allocation has already collected any inits, bail:
duke@0 2675 if (init->is_non_zero()) return false;
duke@0 2676 init->set_complete(phase);
duke@0 2677 return true;
duke@0 2678 }
duke@0 2679
duke@0 2680 void InitializeNode::remove_extra_zeroes() {
duke@0 2681 if (req() == RawStores) return;
duke@0 2682 Node* zmem = zero_memory();
duke@0 2683 uint fill = RawStores;
duke@0 2684 for (uint i = fill; i < req(); i++) {
duke@0 2685 Node* n = in(i);
duke@0 2686 if (n->is_top() || n == zmem) continue; // skip
duke@0 2687 if (fill < i) set_req(fill, n); // compact
duke@0 2688 ++fill;
duke@0 2689 }
duke@0 2690 // delete any empty spaces created:
duke@0 2691 while (fill < req()) {
duke@0 2692 del_req(fill);
duke@0 2693 }
duke@0 2694 }
duke@0 2695
duke@0 2696 // Helper for remembering which stores go with which offsets.
duke@0 2697 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@0 2698 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@0 2699 intptr_t offset = -1;
duke@0 2700 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@0 2701 phase, offset);
duke@0 2702 if (base == NULL) return -1; // something is dead,
duke@0 2703 if (offset < 0) return -1; // dead, dead
duke@0 2704 return offset;
duke@0 2705 }
duke@0 2706
duke@0 2707 // Helper for proving that an initialization expression is
duke@0 2708 // "simple enough" to be folded into an object initialization.
duke@0 2709 // Attempts to prove that a store's initial value 'n' can be captured
duke@0 2710 // within the initialization without creating a vicious cycle, such as:
duke@0 2711 // { Foo p = new Foo(); p.next = p; }
duke@0 2712 // True for constants and parameters and small combinations thereof.
duke@0 2713 bool InitializeNode::detect_init_independence(Node* n,
duke@0 2714 bool st_is_pinned,
duke@0 2715 int& count) {
duke@0 2716 if (n == NULL) return true; // (can this really happen?)
duke@0 2717 if (n->is_Proj()) n = n->in(0);
duke@0 2718 if (n == this) return false; // found a cycle
duke@0 2719 if (n->is_Con()) return true;
duke@0 2720 if (n->is_Start()) return true; // params, etc., are OK
duke@0 2721 if (n->is_Root()) return true; // even better
duke@0 2722
duke@0 2723 Node* ctl = n->in(0);
duke@0 2724 if (ctl != NULL && !ctl->is_top()) {
duke@0 2725 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@0 2726 if (ctl == this) return false;
duke@0 2727
duke@0 2728 // If we already know that the enclosing memory op is pinned right after
duke@0 2729 // the init, then any control flow that the store has picked up
duke@0 2730 // must have preceded the init, or else be equal to the init.
duke@0 2731 // Even after loop optimizations (which might change control edges)
duke@0 2732 // a store is never pinned *before* the availability of its inputs.
kvn@119 2733 if (!MemNode::all_controls_dominate(n, this))
duke@0 2734 return false; // failed to prove a good control
duke@0 2735
duke@0 2736 }
duke@0 2737
duke@0 2738 // Check data edges for possible dependencies on 'this'.
duke@0 2739 if ((count += 1) > 20) return false; // complexity limit
duke@0 2740 for (uint i = 1; i < n->req(); i++) {
duke@0 2741 Node* m = n->in(i);
duke@0 2742 if (m == NULL || m == n || m->is_top()) continue;
duke@0 2743 uint first_i = n->find_edge(m);
duke@0 2744 if (i != first_i) continue; // process duplicate edge just once
duke@0 2745 if (!detect_init_independence(m, st_is_pinned, count)) {
duke@0 2746 return false;
duke@0 2747 }
duke@0 2748 }
duke@0 2749
duke@0 2750 return true;
duke@0 2751 }
duke@0 2752
duke@0 2753 // Here are all the checks a Store must pass before it can be moved into
duke@0 2754 // an initialization. Returns zero if a check fails.
duke@0 2755 // On success, returns the (constant) offset to which the store applies,
duke@0 2756 // within the initialized memory.
duke@0 2757 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
duke@0 2758 const int FAIL = 0;
duke@0 2759 if (st->req() != MemNode::ValueIn + 1)
duke@0 2760 return FAIL; // an inscrutable StoreNode (card mark?)
duke@0 2761 Node* ctl = st->in(MemNode::Control);
duke@0 2762 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@0 2763 return FAIL; // must be unconditional after the initialization
duke@0 2764 Node* mem = st->in(MemNode::Memory);
duke@0 2765 if (!(mem->is_Proj() && mem->in(0) == this))
duke@0 2766 return FAIL; // must not be preceded by other stores
duke@0 2767 Node* adr = st->in(MemNode::Address);
duke@0 2768 intptr_t offset;
duke@0 2769 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@0 2770 if (alloc == NULL)
duke@0 2771 return FAIL; // inscrutable address
duke@0 2772 if (alloc != allocation())
duke@0 2773 return FAIL; // wrong allocation! (store needs to float up)
duke@0 2774 Node* val = st->in(MemNode::ValueIn);
duke@0 2775 int complexity_count = 0;
duke@0 2776 if (!detect_init_independence(val, true, complexity_count))
duke@0 2777 return FAIL; // stored value must be 'simple enough'
duke@0 2778
duke@0 2779 return offset; // success
duke@0 2780 }
duke@0 2781
duke@0 2782 // Find the captured store in(i) which corresponds to the range
duke@0 2783 // [start..start+size) in the initialized object.
duke@0 2784 // If there is one, return its index i. If there isn't, return the
duke@0 2785 // negative of the index where it should be inserted.
duke@0 2786 // Return 0 if the queried range overlaps an initialization boundary
duke@0 2787 // or if dead code is encountered.
duke@0 2788 // If size_in_bytes is zero, do not bother with overlap checks.
duke@0 2789 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@0 2790 int size_in_bytes,
duke@0 2791 PhaseTransform* phase) {
duke@0 2792 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@0 2793
duke@0 2794 if (is_complete())
duke@0 2795 return FAIL; // arraycopy got here first; punt
duke@0 2796
duke@0 2797 assert(allocation() != NULL, "must be present");
duke@0 2798
duke@0 2799 // no negatives, no header fields:
coleenp@113 2800 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@0 2801
duke@0 2802 // after a certain size, we bail out on tracking all the stores:
duke@0 2803 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2804 if (start >= ti_limit) return FAIL;
duke@0 2805
duke@0 2806 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@0 2807 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@0 2808
duke@0 2809 Node* st = in(i);
duke@0 2810 intptr_t st_off = get_store_offset(st, phase);
duke@0 2811 if (st_off < 0) {
duke@0 2812 if (st != zero_memory()) {
duke@0 2813 return FAIL; // bail out if there is dead garbage
duke@0 2814 }
duke@0 2815 } else if (st_off > start) {
duke@0 2816 // ...we are done, since stores are ordered
duke@0 2817 if (st_off < start + size_in_bytes) {
duke@0 2818 return FAIL; // the next store overlaps
duke@0 2819 }
duke@0 2820 return -(int)i; // not found; here is where to put it
duke@0 2821 } else if (st_off < start) {
duke@0 2822 if (size_in_bytes != 0 &&
duke@0 2823 start < st_off + MAX_STORE &&
duke@0 2824 start < st_off + st->as_Store()->memory_size()) {
duke@0 2825 return FAIL; // the previous store overlaps
duke@0 2826 }
duke@0 2827 } else {
duke@0 2828 if (size_in_bytes != 0 &&
duke@0 2829 st->as_Store()->memory_size() != size_in_bytes) {
duke@0 2830 return FAIL; // mismatched store size
duke@0 2831 }
duke@0 2832 return i;
duke@0 2833 }
duke@0 2834
duke@0 2835 ++i;
duke@0 2836 }
duke@0 2837 }
duke@0 2838
duke@0 2839 // Look for a captured store which initializes at the offset 'start'
duke@0 2840 // with the given size. If there is no such store, and no other
duke@0 2841 // initialization interferes, then return zero_memory (the memory
duke@0 2842 // projection of the AllocateNode).
duke@0 2843 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@0 2844 PhaseTransform* phase) {
duke@0 2845 assert(stores_are_sane(phase), "");
duke@0 2846 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2847 if (i == 0) {
duke@0 2848 return NULL; // something is dead
duke@0 2849 } else if (i < 0) {
duke@0 2850 return zero_memory(); // just primordial zero bits here
duke@0 2851 } else {
duke@0 2852 Node* st = in(i); // here is the store at this position
duke@0 2853 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@0 2854 return st;
duke@0 2855 }
duke@0 2856 }
duke@0 2857
duke@0 2858 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@0 2859 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@0 2860 PhaseTransform* phase) {
duke@0 2861 Node* addr = in(RawAddress);
duke@0 2862 if (offset != 0) {
duke@0 2863 Compile* C = phase->C;
duke@0 2864 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
duke@0 2865 phase->MakeConX(offset)) );
duke@0 2866 }
duke@0 2867 return addr;
duke@0 2868 }
duke@0 2869
duke@0 2870 // Clone the given store, converting it into a raw store
duke@0 2871 // initializing a field or element of my new object.
duke@0 2872 // Caller is responsible for retiring the original store,
duke@0 2873 // with subsume_node or the like.
duke@0 2874 //
duke@0 2875 // From the example above InitializeNode::InitializeNode,
duke@0 2876 // here are the old stores to be captured:
duke@0 2877 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2878 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2879 //
duke@0 2880 // Here is the changed code; note the extra edges on init:
duke@0 2881 // alloc = (Allocate ...)
duke@0 2882 // rawoop = alloc.RawAddress
duke@0 2883 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@0 2884 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@0 2885 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@0 2886 // rawstore1 rawstore2)
duke@0 2887 //
duke@0 2888 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
duke@0 2889 PhaseTransform* phase) {
duke@0 2890 assert(stores_are_sane(phase), "");
duke@0 2891
duke@0 2892 if (start < 0) return NULL;
duke@0 2893 assert(can_capture_store(st, phase) == start, "sanity");
duke@0 2894
duke@0 2895 Compile* C = phase->C;
duke@0 2896 int size_in_bytes = st->memory_size();
duke@0 2897 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2898 if (i == 0) return NULL; // bail out
duke@0 2899 Node* prev_mem = NULL; // raw memory for the captured store
duke@0 2900 if (i > 0) {
duke@0 2901 prev_mem = in(i); // there is a pre-existing store under this one
duke@0 2902 set_req(i, C->top()); // temporarily disconnect it
duke@0 2903 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@0 2904 } else {
duke@0 2905 i = -i; // no pre-existing store
duke@0 2906 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@0 2907 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@0 2908 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@0 2909 else
duke@0 2910 ins_req(i, C->top()); // build a new edge
duke@0 2911 }
duke@0 2912 Node* new_st = st->clone();
duke@0 2913 new_st->set_req(MemNode::Control, in(Control));
duke@0 2914 new_st->set_req(MemNode::Memory, prev_mem);
duke@0 2915 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@0 2916 new_st = phase->transform(new_st);
duke@0 2917
duke@0 2918 // At this point, new_st might have swallowed a pre-existing store
duke@0 2919 // at the same offset, or perhaps new_st might have disappeared,
duke@0 2920 // if it redundantly stored the same value (or zero to fresh memory).
duke@0 2921
duke@0 2922 // In any case, wire it in:
duke@0 2923 set_req(i, new_st);
duke@0 2924
duke@0 2925 // The caller may now kill the old guy.
duke@0 2926 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@0 2927 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@0 2928 assert(!is_complete(), "");
duke@0 2929 return new_st;
duke@0 2930 }
duke@0 2931
duke@0 2932 static bool store_constant(jlong* tiles, int num_tiles,
duke@0 2933 intptr_t st_off, int st_size,
duke@0 2934 jlong con) {
duke@0 2935 if ((st_off & (st_size-1)) != 0)
duke@0 2936 return false; // strange store offset (assume size==2**N)
duke@0 2937 address addr = (address)tiles + st_off;
duke@0 2938 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@0 2939 switch (st_size) {
duke@0 2940 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@0 2941 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@0 2942 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@0 2943 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@0 2944 default: return false; // strange store size (detect size!=2**N here)
duke@0 2945 }
duke@0 2946 return true; // return success to caller
duke@0 2947 }
duke@0 2948
duke@0 2949 // Coalesce subword constants into int constants and possibly
duke@0 2950 // into long constants. The goal, if the CPU permits,
duke@0 2951 // is to initialize the object with a small number of 64-bit tiles.
duke@0 2952 // Also, convert floating-point constants to bit patterns.
duke@0 2953 // Non-constants are not relevant to this pass.
duke@0 2954 //
duke@0 2955 // In terms of the running example on InitializeNode::InitializeNode
duke@0 2956 // and InitializeNode::capture_store, here is the transformation
duke@0 2957 // of rawstore1 and rawstore2 into rawstore12:
duke@0 2958 // alloc = (Allocate ...)
duke@0 2959 // rawoop = alloc.RawAddress
duke@0 2960 // tile12 = 0x00010002
duke@0 2961 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@0 2962 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@0 2963 //
duke@0 2964 void
duke@0 2965 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@0 2966 Node* size_in_bytes,
duke@0 2967 PhaseGVN* phase) {
duke@0 2968 Compile* C = phase->C;
duke@0 2969
duke@0 2970 assert(stores_are_sane(phase), "");
duke@0 2971 // Note: After this pass, they are not completely sane,
duke@0 2972 // since there may be some overlaps.
duke@0 2973
duke@0 2974 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@0 2975
duke@0 2976 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2977 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@0 2978 size_limit = MIN2(size_limit, ti_limit);
duke@0 2979 size_limit = align_size_up(size_limit, BytesPerLong);
duke@0 2980 int num_tiles = size_limit / BytesPerLong;
duke@0 2981
duke@0 2982 // allocate space for the tile map:
duke@0 2983 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@0 2984 jlong tiles_buf[small_len];
duke@0 2985 Node* nodes_buf[small_len];
duke@0 2986 jlong inits_buf[small_len];
duke@0 2987 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@0 2988 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2989 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@0 2990 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@0 2991 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@0 2992 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2993 // tiles: exact bitwise model of all primitive constants
duke@0 2994 // nodes: last constant-storing node subsumed into the tiles model
duke@0 2995 // inits: which bytes (in each tile) are touched by any initializations
duke@0 2996
duke@0 2997 //// Pass A: Fill in the tile model with any relevant stores.
duke@0 2998
duke@0 2999 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@0 3000 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@0 3001 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@0 3002 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3003 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3004 Node* st = in(i);
duke@0 3005 intptr_t st_off = get_store_offset(st, phase);
duke@0 3006
duke@0 3007 // Figure out the store's offset and constant value:
duke@0 3008 if (st_off < header_size) continue; //skip (ignore header)
duke@0 3009 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@0 3010 int st_size = st->as_Store()->memory_size();
duke@0 3011 if (st_off + st_size > size_limit) break;
duke@0 3012
duke@0 3013 // Record which bytes are touched, whether by constant or not.
duke@0 3014 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@0 3015 continue; // skip (strange store size)
duke@0 3016
duke@0 3017 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@0 3018 if (!val->singleton()) continue; //skip (non-con store)
duke@0 3019 BasicType type = val->basic_type();
duke@0 3020
duke@0 3021 jlong con = 0;
duke@0 3022 switch (type) {
duke@0 3023 case T_INT: con = val->is_int()->get_con(); break;
duke@0 3024 case T_LONG: con = val->is_long()->get_con(); break;
duke@0 3025 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@0 3026 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@0 3027 default: continue; //skip (odd store type)
duke@0 3028 }
duke@0 3029
duke@0 3030 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@0 3031 st->Opcode() == Op_StoreL) {
duke@0 3032 continue; // This StoreL is already optimal.
duke@0 3033 }
duke@0 3034
duke@0 3035 // Store down the constant.
duke@0 3036 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@0 3037
duke@0 3038 intptr_t j = st_off >> LogBytesPerLong;
duke@0 3039
duke@0 3040 if (type == T_INT && st_size == BytesPerInt
duke@0 3041 && (st_off & BytesPerInt) == BytesPerInt) {
duke@0 3042 jlong lcon = tiles[j];
duke@0 3043 if (!Matcher::isSimpleConstant64(lcon) &&
duke@0 3044 st->Opcode() == Op_StoreI) {
duke@0 3045 // This StoreI is already optimal by itself.
duke@0 3046 jint* intcon = (jint*) &tiles[j];
duke@0 3047 intcon[1] = 0; // undo the store_constant()
duke@0 3048
duke@0 3049 // If the previous store is also optimal by itself, back up and
duke@0 3050 // undo the action of the previous loop iteration... if we can.
duke@0 3051 // But if we can't, just let the previous half take care of itself.
duke@0 3052 st = nodes[j];
duke@0 3053 st_off -= BytesPerInt;
duke@0 3054 con = intcon[0];
duke@0 3055 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@0 3056 assert(st_off >= header_size, "still ignoring header");
duke@0 3057 assert(get_store_offset(st, phase) == st_off, "must be");
duke@0 3058 assert(in(i-1) == zmem, "must be");
duke@0 3059 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@0 3060 assert(con == tcon->is_int()->get_con(), "must be");
duke@0 3061 // Undo the effects of the previous loop trip, which swallowed st:
duke@0 3062 intcon[0] = 0; // undo store_constant()
duke@0 3063 set_req(i-1, st); // undo set_req(i, zmem)
duke@0 3064 nodes[j] = NULL; // undo nodes[j] = st
duke@0 3065 --old_subword; // undo ++old_subword
duke@0 3066 }
duke@0 3067 continue; // This StoreI is already optimal.
duke@0 3068 }
duke@0 3069 }
duke@0 3070
duke@0 3071 // This store is not needed.
duke@0 3072 set_req(i, zmem);
duke@0 3073 nodes[j] = st; // record for the moment
duke@0 3074 if (st_size < BytesPerLong) // something has changed
duke@0 3075 ++old_subword; // includes int/float, but who's counting...
duke@0 3076 else ++old_long;
duke@0 3077 }
duke@0 3078
duke@0 3079 if ((old_subword + old_long) == 0)
duke@0 3080 return; // nothing more to do
duke@0 3081
duke@0 3082 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@0 3083 // Be sure to insert them before overlapping non-constant stores.
duke@0 3084 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@0 3085 for (int j = 0; j < num_tiles; j++) {
duke@0 3086 jlong con = tiles[j];
duke@0 3087 jlong init = inits[j];
duke@0 3088 if (con == 0) continue;
duke@0 3089 jint con0, con1; // split the constant, address-wise
duke@0 3090 jint init0, init1; // split the init map, address-wise
duke@0 3091 { union { jlong con; jint intcon[2]; } u;
duke@0 3092 u.con = con;
duke@0 3093 con0 = u.intcon[0];
duke@0 3094 con1 = u.intcon[1];
duke@0 3095 u.con = init;
duke@0 3096 init0 = u.intcon[0];
duke@0 3097 init1 = u.intcon[1];
duke@0 3098 }
duke@0 3099
duke@0 3100 Node* old = nodes[j];
duke@0 3101 assert(old != NULL, "need the prior store");
duke@0 3102 intptr_t offset = (j * BytesPerLong);
duke@0 3103
duke@0 3104 bool split = !Matcher::isSimpleConstant64(con);
duke@0 3105
duke@0 3106 if (offset < header_size) {
duke@0 3107 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@0 3108 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@0 3109 split = true; // only the second word counts
duke@0 3110 // Example: int a[] = { 42 ... }
duke@0 3111 } else if (con0 == 0 && init0 == -1) {
duke@0 3112 split = true; // first word is covered by full inits
duke@0 3113 // Example: int a[] = { ... foo(), 42 ... }
duke@0 3114 } else if (con1 == 0 && init1 == -1) {
duke@0 3115 split = true; // second word is covered by full inits
duke@0 3116 // Example: int a[] = { ... 42, foo() ... }
duke@0 3117 }
duke@0 3118
duke@0 3119 // Here's a case where init0 is neither 0 nor -1:
duke@0 3120 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@0 3121 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@0 3122 // In this case the tile is not split; it is (jlong)42.
duke@0 3123 // The big tile is stored down, and then the foo() value is inserted.
duke@0 3124 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@0 3125
duke@0 3126 Node* ctl = old->in(MemNode::Control);
duke@0 3127 Node* adr = make_raw_address(offset, phase);
duke@0 3128 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 3129
duke@0 3130 // One or two coalesced stores to plop down.
duke@0 3131 Node* st[2];
duke@0 3132 intptr_t off[2];
duke@0 3133 int nst = 0;
duke@0 3134 if (!split) {
duke@0 3135 ++new_long;
duke@0 3136 off[nst] = offset;
coleenp@113 3137 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3138 phase->longcon(con), T_LONG);
duke@0 3139 } else {
duke@0 3140 // Omit either if it is a zero.
duke@0 3141 if (con0 != 0) {
duke@0 3142 ++new_int;
duke@0 3143 off[nst] = offset;
coleenp@113 3144 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3145 phase->intcon(con0), T_INT);
duke@0 3146 }
duke@0 3147 if (con1 != 0) {
duke@0 3148 ++new_int;
duke@0 3149 offset += BytesPerInt;
duke@0 3150 adr = make_raw_address(offset, phase);
duke@0 3151 off[nst] = offset;
coleenp@113 3152 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3153 phase->intcon(con1), T_INT);
duke@0 3154 }
duke@0 3155 }
duke@0 3156
duke@0 3157 // Insert second store first, then the first before the second.
duke@0 3158 // Insert each one just before any overlapping non-constant stores.
duke@0 3159 while (nst > 0) {
duke@0 3160 Node* st1 = st[--nst];
duke@0 3161 C->copy_node_notes_to(st1, old);
duke@0 3162 st1 = phase->transform(st1);
duke@0 3163 offset = off[nst];
duke@0 3164 assert(offset >= header_size, "do not smash header");
duke@0 3165 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@0 3166 guarantee(ins_idx != 0, "must re-insert constant store");
duke@0 3167 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@0 3168 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@0 3169 set_req(--ins_idx, st1);
duke@0 3170 else
duke@0 3171 ins_req(ins_idx, st1);
duke@0 3172 }
duke@0 3173 }
duke@0 3174
duke@0 3175 if (PrintCompilation && WizardMode)
duke@0 3176 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@0 3177 old_subword, old_long, new_int, new_long);
duke@0 3178 if (C->log() != NULL)
duke@0 3179 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@0 3180 old_subword, old_long, new_int, new_long);
duke@0 3181
duke@0 3182 // Clean up any remaining occurrences of zmem:
duke@0 3183 remove_extra_zeroes();
duke@0 3184 }
duke@0 3185
duke@0 3186 // Explore forward from in(start) to find the first fully initialized
duke@0 3187 // word, and return its offset. Skip groups of subword stores which
duke@0 3188 // together initialize full words. If in(start) is itself part of a
duke@0 3189 // fully initialized word, return the offset of in(start). If there
duke@0 3190 // are no following full-word stores, or if something is fishy, return
duke@0 3191 // a negative value.
duke@0 3192 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@0 3193 int int_map = 0;
duke@0 3194 intptr_t int_map_off = 0;
duke@0 3195 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@0 3196
duke@0 3197 for (uint i = start, limit = req(); i < limit; i++) {
duke@0 3198 Node* st = in(i);
duke@0 3199
duke@0 3200 intptr_t st_off = get_store_offset(st, phase);
duke@0 3201 if (st_off < 0) break; // return conservative answer
duke@0 3202
duke@0 3203 int st_size = st->as_Store()->memory_size();
duke@0 3204 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@0 3205 return st_off; // we found a complete word init
duke@0 3206 }
duke@0 3207
duke@0 3208 // update the map:
duke@0 3209
duke@0 3210 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@0 3211 if (this_int_off != int_map_off) {
duke@0 3212 // reset the map:
duke@0 3213 int_map = 0;
duke@0 3214 int_map_off = this_int_off;
duke@0 3215 }
duke@0 3216
duke@0 3217 int subword_off = st_off - this_int_off;
duke@0 3218 int_map |= right_n_bits(st_size) << subword_off;
duke@0 3219 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@0 3220 return this_int_off; // we found a complete word init
duke@0 3221 }
duke@0 3222
duke@0 3223 // Did this store hit or cross the word boundary?
duke@0 3224 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@0 3225 if (next_int_off == this_int_off + BytesPerInt) {
duke@0 3226 // We passed the current int, without fully initializing it.
duke@0 3227 int_map_off = next_int_off;
duke@0 3228 int_map >>= BytesPerInt;
duke@0 3229 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@0 3230 // We passed the current and next int.
duke@0 3231 return this_int_off + BytesPerInt;
duke@0 3232 }
duke@0 3233 }
duke@0 3234
duke@0 3235 return -1;
duke@0 3236 }
duke@0 3237
duke@0 3238
duke@0 3239 // Called when the associated AllocateNode is expanded into CFG.
duke@0 3240 // At this point, we may perform additional optimizations.
duke@0 3241 // Linearize the stores by ascending offset, to make memory
duke@0 3242 // activity as coherent as possible.
duke@0 3243 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@0 3244 intptr_t header_size,
duke@0 3245 Node* size_in_bytes,
duke@0 3246 PhaseGVN* phase) {
duke@0 3247 assert(!is_complete(), "not already complete");
duke@0 3248 assert(stores_are_sane(phase), "");
duke@0 3249 assert(allocation() != NULL, "must be present");
duke@0 3250
duke@0 3251 remove_extra_zeroes();
duke@0 3252
duke@0 3253 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@0 3254 // reduce instruction count for common initialization patterns
duke@0 3255 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@0 3256
duke@0 3257 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3258 Node* inits = zmem; // accumulating a linearized chain of inits
duke@0 3259 #ifdef ASSERT
coleenp@113 3260 intptr_t first_offset = allocation()->minimum_header_size();
coleenp@113 3261 intptr_t last_init_off = first_offset; // previous init offset
coleenp@113 3262 intptr_t last_init_end = first_offset; // previous init offset+size
coleenp@113 3263 intptr_t last_tile_end = first_offset; // previous tile offset+size
duke@0 3264 #endif
duke@0 3265 intptr_t zeroes_done = header_size;
duke@0 3266
duke@0 3267 bool do_zeroing = true; // we might give up if inits are very sparse
duke@0 3268 int big_init_gaps = 0; // how many large gaps have we seen?
duke@0 3269
duke@0 3270 if (ZeroTLAB) do_zeroing = false;
duke@0 3271 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@0 3272
duke@0 3273 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3274 Node* st = in(i);
duke@0 3275 intptr_t st_off = get_store_offset(st, phase);
duke@0 3276 if (st_off < 0)
duke@0 3277 break; // unknown junk in the inits
duke@0 3278 if (st->in(MemNode::Memory) != zmem)
duke@0 3279 break; // complicated store chains somehow in list
duke@0 3280
duke@0 3281 int st_size = st->as_Store()->memory_size();
duke@0 3282 intptr_t next_init_off = st_off + st_size;
duke@0 3283
duke@0 3284 if (do_zeroing && zeroes_done < next_init_off) {
duke@0 3285 // See if this store needs a zero before it or under it.
duke@0 3286 intptr_t zeroes_needed = st_off;
duke@0 3287
duke@0 3288 if (st_size < BytesPerInt) {
duke@0 3289 // Look for subword stores which only partially initialize words.
duke@0 3290 // If we find some, we must lay down some word-level zeroes first,
duke@0 3291 // underneath the subword stores.
duke@0 3292 //
duke@0 3293 // Examples:
duke@0 3294 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
duke@0 3295 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
duke@0 3296 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
duke@0 3297 //
duke@0 3298 // Note: coalesce_subword_stores may have already done this,
duke@0 3299 // if it was prompted by constant non-zero subword initializers.
duke@0 3300 // But this case can still arise with non-constant stores.
duke@0 3301
duke@0 3302 intptr_t next_full_store = find_next_fullword_store(i, phase);
duke@0 3303
duke@0 3304 // In the examples above:
duke@0 3305 // in(i) p q r s x y z
duke@0 3306 // st_off 12 13 14 15 12 13 14
duke@0 3307 // st_size 1 1 1 1 1 1 1
duke@0 3308 // next_full_s. 12 16 16 16 16 16 16
duke@0 3309 // z's_done 12 16 16 16 12 16 12
duke@0 3310 // z's_needed 12 16 16 16 16 16 16
duke@0 3311 // zsize 0 0 0 0 4 0 4
duke@0 3312 if (next_full_store < 0) {
duke@0 3313 // Conservative tack: Zero to end of current word.
duke@0 3314 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
duke@0 3315 } else {
duke@0 3316 // Zero to beginning of next fully initialized word.
duke@0 3317 // Or, don't zero at all, if we are already in that word.
duke@0 3318 assert(next_full_store >= zeroes_needed, "must go forward");
duke@0 3319 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
duke@0 3320 zeroes_needed = next_full_store;
duke@0 3321 }
duke@0 3322 }
duke@0 3323
duke@0 3324 if (zeroes_needed > zeroes_done) {
duke@0 3325 intptr_t zsize = zeroes_needed - zeroes_done;
duke@0 3326 // Do some incremental zeroing on rawmem, in parallel with inits.
duke@0 3327 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3328 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3329 zeroes_done, zeroes_needed,
duke@0 3330 phase);
duke@0 3331 zeroes_done = zeroes_needed;
duke@0 3332 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
duke@0 3333 do_zeroing = false; // leave the hole, next time
duke@0 3334 }
duke@0 3335 }
duke@0 3336
duke@0 3337 // Collect the store and move on:
duke@0 3338 st->set_req(MemNode::Memory, inits);
duke@0 3339 inits = st; // put it on the linearized chain
duke@0 3340 set_req(i, zmem); // unhook from previous position
duke@0 3341
duke@0 3342 if (zeroes_done == st_off)
duke@0 3343 zeroes_done = next_init_off;
duke@0 3344
duke@0 3345 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
duke@0 3346
duke@0 3347 #ifdef ASSERT
duke@0 3348 // Various order invariants. Weaker than stores_are_sane because
duke@0 3349 // a large constant tile can be filled in by smaller non-constant stores.
duke@0 3350 assert(st_off >= last_init_off, "inits do not reverse");
duke@0 3351 last_init_off = st_off;
duke@0 3352 const Type* val = NULL;
duke@0 3353 if (st_size >= BytesPerInt &&
duke@0 3354 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
duke@0 3355 (int)val->basic_type() < (int)T_OBJECT) {
duke@0 3356 assert(st_off >= last_tile_end, "tiles do not overlap");
duke@0 3357 assert(st_off >= last_init_end, "tiles do not overwrite inits");
duke@0 3358 last_tile_end = MAX2(last_tile_end, next_init_off);
duke@0 3359 } else {
duke@0 3360 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
duke@0 3361 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
duke@0 3362 assert(st_off >= last_init_end, "inits do not overlap");
duke@0 3363 last_init_end = next_init_off; // it's a non-tile
duke@0 3364 }
duke@0 3365 #endif //ASSERT
duke@0 3366 }
duke@0 3367
duke@0 3368 remove_extra_zeroes(); // clear out all the zmems left over
duke@0 3369 add_req(inits);
duke@0 3370
duke@0 3371 if (!ZeroTLAB) {
duke@0 3372 // If anything remains to be zeroed, zero it all now.
duke@0 3373 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3374 // if it is the last unused 4 bytes of an instance, forget about it
duke@0 3375 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
duke@0 3376 if (zeroes_done + BytesPerLong >= size_limit) {
duke@0 3377 assert(allocation() != NULL, "");
duke@0 3378 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
duke@0 3379 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
duke@0 3380 if (zeroes_done == k->layout_helper())
duke@0 3381 zeroes_done = size_limit;
duke@0 3382 }
duke@0 3383 if (zeroes_done < size_limit) {
duke@0 3384 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3385 zeroes_done, size_in_bytes, phase);
duke@0 3386 }
duke@0 3387 }
duke@0 3388
duke@0 3389 set_complete(phase);
duke@0 3390 return rawmem;
duke@0 3391 }
duke@0 3392
duke@0 3393
duke@0 3394 #ifdef ASSERT
duke@0 3395 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
duke@0 3396 if (is_complete())
duke@0 3397 return true; // stores could be anything at this point
coleenp@113 3398 assert(allocation() != NULL, "must be present");
coleenp@113 3399 intptr_t last_off = allocation()->minimum_header_size();
duke@0 3400 for (uint i = InitializeNode::RawStores; i < req(); i++) {
duke@0 3401 Node* st = in(i);
duke@0 3402 intptr_t st_off = get_store_offset(st, phase);
duke@0 3403 if (st_off < 0) continue; // ignore dead garbage
duke@0 3404 if (last_off > st_off) {
duke@0 3405 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
duke@0 3406 this->dump(2);
duke@0 3407 assert(false, "ascending store offsets");
duke@0 3408 return false;
duke@0 3409 }
duke@0 3410 last_off = st_off + st->as_Store()->memory_size();
duke@0 3411 }
duke@0 3412 return true;
duke@0 3413 }
duke@0 3414 #endif //ASSERT
duke@0 3415
duke@0 3416
duke@0 3417
duke@0 3418
duke@0 3419 //============================MergeMemNode=====================================
duke@0 3420 //
duke@0 3421 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
duke@0 3422 // contributing store or call operations. Each contributor provides the memory
duke@0 3423 // state for a particular "alias type" (see Compile::alias_type). For example,
duke@0 3424 // if a MergeMem has an input X for alias category #6, then any memory reference
duke@0 3425 // to alias category #6 may use X as its memory state input, as an exact equivalent
duke@0 3426 // to using the MergeMem as a whole.
duke@0 3427 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
duke@0 3428 //
duke@0 3429 // (Here, the <N> notation gives the index of the relevant adr_type.)
duke@0 3430 //
duke@0 3431 // In one special case (and more cases in the future), alias categories overlap.
duke@0 3432 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
duke@0 3433 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
duke@0 3434 // it is exactly equivalent to that state W:
duke@0 3435 // MergeMem(<Bot>: W) <==> W
duke@0 3436 //
duke@0 3437 // Usually, the merge has more than one input. In that case, where inputs
duke@0 3438 // overlap (i.e., one is Bot), the narrower alias type determines the memory
duke@0 3439 // state for that type, and the wider alias type (Bot) fills in everywhere else:
duke@0 3440 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
duke@0 3441 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
duke@0 3442 //
duke@0 3443 // A merge can take a "wide" memory state as one of its narrow inputs.
duke@0 3444 // This simply means that the merge observes out only the relevant parts of
duke@0 3445 // the wide input. That is, wide memory states arriving at narrow merge inputs
duke@0 3446 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
duke@0 3447 //
duke@0 3448 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
duke@0 3449 // and that memory slices "leak through":
duke@0 3450 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
duke@0 3451 //
duke@0 3452 // But, in such a cascade, repeated memory slices can "block the leak":
duke@0 3453 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
duke@0 3454 //
duke@0 3455 // In the last example, Y is not part of the combined memory state of the
duke@0 3456 // outermost MergeMem. The system must, of course, prevent unschedulable
duke@0 3457 // memory states from arising, so you can be sure that the state Y is somehow
duke@0 3458 // a precursor to state Y'.
duke@0 3459 //
duke@0 3460 //
duke@0 3461 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
duke@0 3462 // of each MergeMemNode array are exactly the numerical alias indexes, including
duke@0 3463 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
duke@0 3464 // Compile::alias_type (and kin) produce and manage these indexes.
duke@0 3465 //
duke@0 3466 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
duke@0 3467 // (Note that this provides quick access to the top node inside MergeMem methods,
duke@0 3468 // without the need to reach out via TLS to Compile::current.)
duke@0 3469 //
duke@0 3470 // As a consequence of what was just described, a MergeMem that represents a full
duke@0 3471 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
duke@0 3472 // containing all alias categories.
duke@0 3473 //
duke@0 3474 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
duke@0 3475 //
duke@0 3476 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
duke@0 3477 // a memory state for the alias type <N>, or else the top node, meaning that
duke@0 3478 // there is no particular input for that alias type. Note that the length of
duke@0 3479 // a MergeMem is variable, and may be extended at any time to accommodate new
duke@0 3480 // memory states at larger alias indexes. When merges grow, they are of course
duke@0 3481 // filled with "top" in the unused in() positions.
duke@0 3482 //
duke@0 3483 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
duke@0 3484 // (Top was chosen because it works smoothly with passes like GCM.)
duke@0 3485 //
duke@0 3486 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
duke@0 3487 // the type of random VM bits like TLS references.) Since it is always the
duke@0 3488 // first non-Bot memory slice, some low-level loops use it to initialize an
duke@0 3489 // index variable: for (i = AliasIdxRaw; i < req(); i++).
duke@0 3490 //
duke@0 3491 //
duke@0 3492 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
duke@0 3493 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
duke@0 3494 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
duke@0 3495 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
duke@0 3496 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
duke@0 3497 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
duke@0 3498 //
duke@0 3499 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
duke@0 3500 // really that different from the other memory inputs. An abbreviation called
duke@0 3501 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
duke@0 3502 //
duke@0 3503 //
duke@0 3504 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
duke@0 3505 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
duke@0 3506 // that "emerges though" the base memory will be marked as excluding the alias types
duke@0 3507 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
duke@0 3508 //
duke@0 3509 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
duke@0 3510 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
duke@0 3511 //
duke@0 3512 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
duke@0 3513 // (It is currently unimplemented.) As you can see, the resulting merge is
duke@0 3514 // actually a disjoint union of memory states, rather than an overlay.
duke@0 3515 //
duke@0 3516
duke@0 3517 //------------------------------MergeMemNode-----------------------------------
duke@0 3518 Node* MergeMemNode::make_empty_memory() {
duke@0 3519 Node* empty_memory = (Node*) Compile::current()->top();
duke@0 3520 assert(empty_memory->is_top(), "correct sentinel identity");
duke@0 3521 return empty_memory;
duke@0 3522 }
duke@0 3523
duke@0 3524 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
duke@0 3525 init_class_id(Class_MergeMem);
duke@0 3526 // all inputs are nullified in Node::Node(int)
duke@0 3527 // set_input(0, NULL); // no control input
duke@0 3528
duke@0 3529 // Initialize the edges uniformly to top, for starters.
duke@0 3530 Node* empty_mem = make_empty_memory();
duke@0 3531 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
duke@0 3532 init_req(i,empty_mem);
duke@0 3533 }
duke@0 3534 assert(empty_memory() == empty_mem, "");
duke@0 3535
duke@0 3536 if( new_base != NULL && new_base->is_MergeMem() ) {
duke@0 3537 MergeMemNode* mdef = new_base->as_MergeMem();
duke@0 3538 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
duke@0 3539 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
duke@0 3540 mms.set_memory(mms.memory2());
duke@0 3541 }
duke@0 3542 assert(base_memory() == mdef->base_memory(), "");
duke@0 3543 } else {
duke@0 3544 set_base_memory(new_base);
duke@0 3545 }
duke@0 3546 }
duke@0 3547
duke@0 3548 // Make a new, untransformed MergeMem with the same base as 'mem'.
duke@0 3549 // If mem is itself a MergeMem, populate the result with the same edges.
duke@0 3550 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
duke@0 3551 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
duke@0 3552 }
duke@0 3553
duke@0 3554 //------------------------------cmp--------------------------------------------
duke@0 3555 uint MergeMemNode::hash() const { return NO_HASH; }
duke@0 3556 uint MergeMemNode::cmp( const Node &n ) const {
duke@0 3557 return (&n == this); // Always fail except on self
duke@0 3558 }
duke@0 3559
duke@0 3560 //------------------------------Identity---------------------------------------
duke@0 3561 Node* MergeMemNode::Identity(PhaseTransform *phase) {
duke@0 3562 // Identity if this merge point does not record any interesting memory
duke@0 3563 // disambiguations.
duke@0 3564 Node* base_mem = base_memory();
duke@0 3565 Node* empty_mem = empty_memory(); <