annotate src/share/vm/opto/memnode.cpp @ 1080:7c57aead6d3e

6892658: C2 should optimize some stringbuilder patterns Reviewed-by: kvn, twisti
author never
date Thu, 12 Nov 2009 09:24:21 -0800
parents dcdcc8c16e20
children f96a1a986f7b
rev   line source
duke@0 1 /*
twisti@624 2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
duke@0 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
duke@0 20 * CA 95054 USA or visit www.sun.com if you need additional information or
duke@0 21 * have any questions.
duke@0 22 *
duke@0 23 */
duke@0 24
duke@0 25 // Portions of code courtesy of Clifford Click
duke@0 26
duke@0 27 // Optimization - Graph Style
duke@0 28
duke@0 29 #include "incls/_precompiled.incl"
duke@0 30 #include "incls/_memnode.cpp.incl"
duke@0 31
kvn@74 32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
kvn@74 33
duke@0 34 //=============================================================================
duke@0 35 uint MemNode::size_of() const { return sizeof(*this); }
duke@0 36
duke@0 37 const TypePtr *MemNode::adr_type() const {
duke@0 38 Node* adr = in(Address);
duke@0 39 const TypePtr* cross_check = NULL;
duke@0 40 DEBUG_ONLY(cross_check = _adr_type);
duke@0 41 return calculate_adr_type(adr->bottom_type(), cross_check);
duke@0 42 }
duke@0 43
duke@0 44 #ifndef PRODUCT
duke@0 45 void MemNode::dump_spec(outputStream *st) const {
duke@0 46 if (in(Address) == NULL) return; // node is dead
duke@0 47 #ifndef ASSERT
duke@0 48 // fake the missing field
duke@0 49 const TypePtr* _adr_type = NULL;
duke@0 50 if (in(Address) != NULL)
duke@0 51 _adr_type = in(Address)->bottom_type()->isa_ptr();
duke@0 52 #endif
duke@0 53 dump_adr_type(this, _adr_type, st);
duke@0 54
duke@0 55 Compile* C = Compile::current();
duke@0 56 if( C->alias_type(_adr_type)->is_volatile() )
duke@0 57 st->print(" Volatile!");
duke@0 58 }
duke@0 59
duke@0 60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
duke@0 61 st->print(" @");
duke@0 62 if (adr_type == NULL) {
duke@0 63 st->print("NULL");
duke@0 64 } else {
duke@0 65 adr_type->dump_on(st);
duke@0 66 Compile* C = Compile::current();
duke@0 67 Compile::AliasType* atp = NULL;
duke@0 68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
duke@0 69 if (atp == NULL)
duke@0 70 st->print(", idx=?\?;");
duke@0 71 else if (atp->index() == Compile::AliasIdxBot)
duke@0 72 st->print(", idx=Bot;");
duke@0 73 else if (atp->index() == Compile::AliasIdxTop)
duke@0 74 st->print(", idx=Top;");
duke@0 75 else if (atp->index() == Compile::AliasIdxRaw)
duke@0 76 st->print(", idx=Raw;");
duke@0 77 else {
duke@0 78 ciField* field = atp->field();
duke@0 79 if (field) {
duke@0 80 st->print(", name=");
duke@0 81 field->print_name_on(st);
duke@0 82 }
duke@0 83 st->print(", idx=%d;", atp->index());
duke@0 84 }
duke@0 85 }
duke@0 86 }
duke@0 87
duke@0 88 extern void print_alias_types();
duke@0 89
duke@0 90 #endif
duke@0 91
kvn@74 92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 93 const TypeOopPtr *tinst = t_adr->isa_oopptr();
kvn@223 94 if (tinst == NULL || !tinst->is_known_instance_field())
kvn@74 95 return mchain; // don't try to optimize non-instance types
kvn@74 96 uint instance_id = tinst->instance_id();
kvn@253 97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
kvn@74 98 Node *prev = NULL;
kvn@74 99 Node *result = mchain;
kvn@74 100 while (prev != result) {
kvn@74 101 prev = result;
kvn@253 102 if (result == start_mem)
twisti@605 103 break; // hit one of our sentinels
kvn@74 104 // skip over a call which does not affect this memory slice
kvn@74 105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
kvn@74 106 Node *proj_in = result->in(0);
kvn@253 107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
twisti@605 108 break; // hit one of our sentinels
kvn@253 109 } else if (proj_in->is_Call()) {
kvn@74 110 CallNode *call = proj_in->as_Call();
kvn@74 111 if (!call->may_modify(t_adr, phase)) {
kvn@74 112 result = call->in(TypeFunc::Memory);
kvn@74 113 }
kvn@74 114 } else if (proj_in->is_Initialize()) {
kvn@74 115 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
kvn@74 116 // Stop if this is the initialization for the object instance which
kvn@74 117 // which contains this memory slice, otherwise skip over it.
kvn@74 118 if (alloc != NULL && alloc->_idx != instance_id) {
kvn@74 119 result = proj_in->in(TypeFunc::Memory);
kvn@74 120 }
kvn@74 121 } else if (proj_in->is_MemBar()) {
kvn@74 122 result = proj_in->in(TypeFunc::Memory);
kvn@253 123 } else {
kvn@253 124 assert(false, "unexpected projection");
kvn@74 125 }
kvn@74 126 } else if (result->is_MergeMem()) {
kvn@74 127 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
kvn@74 128 }
kvn@74 129 }
kvn@74 130 return result;
kvn@74 131 }
kvn@74 132
kvn@74 133 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 134 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
kvn@223 135 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
kvn@74 136 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@74 137 Node *result = mchain;
kvn@74 138 result = optimize_simple_memory_chain(result, t_adr, phase);
kvn@74 139 if (is_instance && igvn != NULL && result->is_Phi()) {
kvn@74 140 PhiNode *mphi = result->as_Phi();
kvn@74 141 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
kvn@74 142 const TypePtr *t = mphi->adr_type();
kvn@163 143 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
kvn@223 144 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
kvn@247 145 t->is_oopptr()->cast_to_exactness(true)
kvn@247 146 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
kvn@247 147 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
kvn@74 148 // clone the Phi with our address type
kvn@74 149 result = mphi->split_out_instance(t_adr, igvn);
kvn@74 150 } else {
kvn@74 151 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
kvn@74 152 }
kvn@74 153 }
kvn@74 154 return result;
kvn@74 155 }
kvn@74 156
kvn@64 157 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
kvn@64 158 uint alias_idx = phase->C->get_alias_index(tp);
kvn@64 159 Node *mem = mmem;
kvn@64 160 #ifdef ASSERT
kvn@64 161 {
kvn@64 162 // Check that current type is consistent with the alias index used during graph construction
kvn@64 163 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
kvn@64 164 bool consistent = adr_check == NULL || adr_check->empty() ||
kvn@64 165 phase->C->must_alias(adr_check, alias_idx );
kvn@64 166 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
kvn@64 167 if( !consistent && adr_check != NULL && !adr_check->empty() &&
rasbold@169 168 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
kvn@64 169 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
kvn@64 170 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
kvn@64 171 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
kvn@64 172 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
kvn@64 173 // don't assert if it is dead code.
kvn@64 174 consistent = true;
kvn@64 175 }
kvn@64 176 if( !consistent ) {
kvn@64 177 st->print("alias_idx==%d, adr_check==", alias_idx);
kvn@64 178 if( adr_check == NULL ) {
kvn@64 179 st->print("NULL");
kvn@64 180 } else {
kvn@64 181 adr_check->dump();
kvn@64 182 }
kvn@64 183 st->cr();
kvn@64 184 print_alias_types();
kvn@64 185 assert(consistent, "adr_check must match alias idx");
kvn@64 186 }
kvn@64 187 }
kvn@64 188 #endif
kvn@64 189 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
kvn@64 190 // means an array I have not precisely typed yet. Do not do any
kvn@64 191 // alias stuff with it any time soon.
kvn@64 192 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 193 if( tp->base() != Type::AnyPtr &&
kvn@64 194 !(tinst &&
kvn@64 195 tinst->klass()->is_java_lang_Object() &&
kvn@64 196 tinst->offset() == Type::OffsetBot) ) {
kvn@64 197 // compress paths and change unreachable cycles to TOP
kvn@64 198 // If not, we can update the input infinitely along a MergeMem cycle
kvn@64 199 // Equivalent code in PhiNode::Ideal
kvn@64 200 Node* m = phase->transform(mmem);
twisti@605 201 // If transformed to a MergeMem, get the desired slice
kvn@64 202 // Otherwise the returned node represents memory for every slice
kvn@64 203 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
kvn@64 204 // Update input if it is progress over what we have now
kvn@64 205 }
kvn@64 206 return mem;
kvn@64 207 }
kvn@64 208
duke@0 209 //--------------------------Ideal_common---------------------------------------
duke@0 210 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
duke@0 211 // Unhook non-raw memories from complete (macro-expanded) initializations.
duke@0 212 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
duke@0 213 // If our control input is a dead region, kill all below the region
duke@0 214 Node *ctl = in(MemNode::Control);
duke@0 215 if (ctl && remove_dead_region(phase, can_reshape))
duke@0 216 return this;
kvn@305 217 ctl = in(MemNode::Control);
kvn@305 218 // Don't bother trying to transform a dead node
kvn@305 219 if( ctl && ctl->is_top() ) return NodeSentinel;
duke@0 220
kvn@708 221 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@708 222 // Wait if control on the worklist.
kvn@708 223 if (ctl && can_reshape && igvn != NULL) {
kvn@708 224 Node* bol = NULL;
kvn@708 225 Node* cmp = NULL;
kvn@708 226 if (ctl->in(0)->is_If()) {
kvn@708 227 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
kvn@708 228 bol = ctl->in(0)->in(1);
kvn@708 229 if (bol->is_Bool())
kvn@708 230 cmp = ctl->in(0)->in(1)->in(1);
kvn@708 231 }
kvn@708 232 if (igvn->_worklist.member(ctl) ||
kvn@708 233 (bol != NULL && igvn->_worklist.member(bol)) ||
kvn@708 234 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
kvn@708 235 // This control path may be dead.
kvn@708 236 // Delay this memory node transformation until the control is processed.
kvn@708 237 phase->is_IterGVN()->_worklist.push(this);
kvn@708 238 return NodeSentinel; // caller will return NULL
kvn@708 239 }
kvn@708 240 }
duke@0 241 // Ignore if memory is dead, or self-loop
duke@0 242 Node *mem = in(MemNode::Memory);
duke@0 243 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 244 assert( mem != this, "dead loop in MemNode::Ideal" );
duke@0 245
duke@0 246 Node *address = in(MemNode::Address);
duke@0 247 const Type *t_adr = phase->type( address );
duke@0 248 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 249
kvn@708 250 if( can_reshape && igvn != NULL &&
kvn@708 251 (igvn->_worklist.member(address) || phase->type(address) != adr_type()) ) {
kvn@420 252 // The address's base and type may change when the address is processed.
kvn@420 253 // Delay this mem node transformation until the address is processed.
kvn@420 254 phase->is_IterGVN()->_worklist.push(this);
kvn@420 255 return NodeSentinel; // caller will return NULL
kvn@420 256 }
kvn@420 257
kvn@1062 258 // Do NOT remove or optimize the next lines: ensure a new alias index
kvn@1062 259 // is allocated for an oop pointer type before Escape Analysis.
kvn@1062 260 // Note: C++ will not remove it since the call has side effect.
kvn@1062 261 if ( t_adr->isa_oopptr() ) {
kvn@1062 262 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
kvn@1062 263 }
kvn@1062 264
kvn@708 265 #ifdef ASSERT
kvn@708 266 Node* base = NULL;
kvn@708 267 if (address->is_AddP())
kvn@708 268 base = address->in(AddPNode::Base);
kvn@708 269 assert(base == NULL || t_adr->isa_rawptr() ||
kvn@708 270 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
kvn@708 271 #endif
kvn@708 272
duke@0 273 // Avoid independent memory operations
duke@0 274 Node* old_mem = mem;
duke@0 275
kvn@36 276 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@36 277 // initializations was removed. After macro-expansion all stores catched
kvn@36 278 // by Initialize node became raw stores and there is no information
kvn@36 279 // which memory slices they modify. So it is unsafe to move any memory
kvn@36 280 // operation above these stores. Also in most cases hooked non-raw memories
kvn@36 281 // were already unhooked by using information from detect_ptr_independence()
kvn@36 282 // and find_previous_store().
duke@0 283
duke@0 284 if (mem->is_MergeMem()) {
duke@0 285 MergeMemNode* mmem = mem->as_MergeMem();
duke@0 286 const TypePtr *tp = t_adr->is_ptr();
kvn@64 287
kvn@64 288 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@0 289 }
duke@0 290
duke@0 291 if (mem != old_mem) {
duke@0 292 set_req(MemNode::Memory, mem);
kvn@305 293 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
duke@0 294 return this;
duke@0 295 }
duke@0 296
duke@0 297 // let the subclass continue analyzing...
duke@0 298 return NULL;
duke@0 299 }
duke@0 300
duke@0 301 // Helper function for proving some simple control dominations.
kvn@119 302 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
duke@0 303 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@0 304 // is not a constant (dominated by the method's StartNode).
duke@0 305 // Used by MemNode::find_previous_store to prove that the
duke@0 306 // control input of a memory operation predates (dominates)
duke@0 307 // an allocation it wants to look past.
kvn@119 308 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
kvn@119 309 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
kvn@119 310 return false; // Conservative answer for dead code
kvn@119 311
kvn@193 312 // Check 'dom'. Skip Proj and CatchProj nodes.
kvn@119 313 dom = dom->find_exact_control(dom);
kvn@119 314 if (dom == NULL || dom->is_top())
kvn@119 315 return false; // Conservative answer for dead code
kvn@119 316
kvn@193 317 if (dom == sub) {
kvn@193 318 // For the case when, for example, 'sub' is Initialize and the original
kvn@193 319 // 'dom' is Proj node of the 'sub'.
kvn@193 320 return false;
kvn@193 321 }
kvn@193 322
kvn@155 323 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
kvn@119 324 return true;
kvn@119 325
kvn@119 326 // 'dom' dominates 'sub' if its control edge and control edges
kvn@119 327 // of all its inputs dominate or equal to sub's control edge.
kvn@119 328
kvn@119 329 // Currently 'sub' is either Allocate, Initialize or Start nodes.
kvn@163 330 // Or Region for the check in LoadNode::Ideal();
kvn@163 331 // 'sub' should have sub->in(0) != NULL.
kvn@163 332 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
kvn@163 333 sub->is_Region(), "expecting only these nodes");
kvn@119 334
kvn@119 335 // Get control edge of 'sub'.
kvn@193 336 Node* orig_sub = sub;
kvn@119 337 sub = sub->find_exact_control(sub->in(0));
kvn@119 338 if (sub == NULL || sub->is_top())
kvn@119 339 return false; // Conservative answer for dead code
kvn@119 340
kvn@119 341 assert(sub->is_CFG(), "expecting control");
kvn@119 342
kvn@119 343 if (sub == dom)
kvn@119 344 return true;
kvn@119 345
kvn@119 346 if (sub->is_Start() || sub->is_Root())
kvn@119 347 return false;
kvn@119 348
kvn@119 349 {
kvn@119 350 // Check all control edges of 'dom'.
kvn@119 351
kvn@119 352 ResourceMark rm;
kvn@119 353 Arena* arena = Thread::current()->resource_area();
kvn@119 354 Node_List nlist(arena);
kvn@119 355 Unique_Node_List dom_list(arena);
kvn@119 356
kvn@119 357 dom_list.push(dom);
kvn@119 358 bool only_dominating_controls = false;
kvn@119 359
kvn@119 360 for (uint next = 0; next < dom_list.size(); next++) {
kvn@119 361 Node* n = dom_list.at(next);
kvn@193 362 if (n == orig_sub)
kvn@193 363 return false; // One of dom's inputs dominated by sub.
kvn@119 364 if (!n->is_CFG() && n->pinned()) {
kvn@119 365 // Check only own control edge for pinned non-control nodes.
kvn@119 366 n = n->find_exact_control(n->in(0));
kvn@119 367 if (n == NULL || n->is_top())
kvn@119 368 return false; // Conservative answer for dead code
kvn@119 369 assert(n->is_CFG(), "expecting control");
kvn@193 370 dom_list.push(n);
kvn@193 371 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
kvn@119 372 only_dominating_controls = true;
kvn@119 373 } else if (n->is_CFG()) {
kvn@119 374 if (n->dominates(sub, nlist))
kvn@119 375 only_dominating_controls = true;
kvn@119 376 else
kvn@119 377 return false;
kvn@119 378 } else {
kvn@119 379 // First, own control edge.
kvn@119 380 Node* m = n->find_exact_control(n->in(0));
kvn@155 381 if (m != NULL) {
kvn@155 382 if (m->is_top())
kvn@155 383 return false; // Conservative answer for dead code
kvn@155 384 dom_list.push(m);
kvn@155 385 }
kvn@119 386 // Now, the rest of edges.
kvn@119 387 uint cnt = n->req();
kvn@119 388 for (uint i = 1; i < cnt; i++) {
kvn@119 389 m = n->find_exact_control(n->in(i));
kvn@119 390 if (m == NULL || m->is_top())
kvn@119 391 continue;
kvn@119 392 dom_list.push(m);
duke@0 393 }
duke@0 394 }
duke@0 395 }
kvn@119 396 return only_dominating_controls;
duke@0 397 }
duke@0 398 }
duke@0 399
duke@0 400 //---------------------detect_ptr_independence---------------------------------
duke@0 401 // Used by MemNode::find_previous_store to prove that two base
duke@0 402 // pointers are never equal.
duke@0 403 // The pointers are accompanied by their associated allocations,
duke@0 404 // if any, which have been previously discovered by the caller.
duke@0 405 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@0 406 Node* p2, AllocateNode* a2,
duke@0 407 PhaseTransform* phase) {
duke@0 408 // Attempt to prove that these two pointers cannot be aliased.
duke@0 409 // They may both manifestly be allocations, and they should differ.
duke@0 410 // Or, if they are not both allocations, they can be distinct constants.
duke@0 411 // Otherwise, one is an allocation and the other a pre-existing value.
duke@0 412 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@0 413 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@0 414 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@0 415 return (a1 != a2);
duke@0 416 } else if (a1 != NULL) { // one allocation a1
duke@0 417 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
kvn@119 418 return all_controls_dominate(p2, a1);
duke@0 419 } else { //(a2 != NULL) // one allocation a2
kvn@119 420 return all_controls_dominate(p1, a2);
duke@0 421 }
duke@0 422 return false;
duke@0 423 }
duke@0 424
duke@0 425
duke@0 426 // The logic for reordering loads and stores uses four steps:
duke@0 427 // (a) Walk carefully past stores and initializations which we
duke@0 428 // can prove are independent of this load.
duke@0 429 // (b) Observe that the next memory state makes an exact match
duke@0 430 // with self (load or store), and locate the relevant store.
duke@0 431 // (c) Ensure that, if we were to wire self directly to the store,
duke@0 432 // the optimizer would fold it up somehow.
duke@0 433 // (d) Do the rewiring, and return, depending on some other part of
duke@0 434 // the optimizer to fold up the load.
duke@0 435 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@0 436 // specific to loads and stores, so they are handled by the callers.
duke@0 437 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@0 438 //
duke@0 439 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@0 440 Node* ctrl = in(MemNode::Control);
duke@0 441 Node* adr = in(MemNode::Address);
duke@0 442 intptr_t offset = 0;
duke@0 443 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 444 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@0 445
duke@0 446 if (offset == Type::OffsetBot)
duke@0 447 return NULL; // cannot unalias unless there are precise offsets
duke@0 448
kvn@74 449 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
kvn@74 450
duke@0 451 intptr_t size_in_bytes = memory_size();
duke@0 452
duke@0 453 Node* mem = in(MemNode::Memory); // start searching here...
duke@0 454
duke@0 455 int cnt = 50; // Cycle limiter
duke@0 456 for (;;) { // While we can dance past unrelated stores...
duke@0 457 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@0 458
duke@0 459 if (mem->is_Store()) {
duke@0 460 Node* st_adr = mem->in(MemNode::Address);
duke@0 461 intptr_t st_offset = 0;
duke@0 462 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@0 463 if (st_base == NULL)
duke@0 464 break; // inscrutable pointer
duke@0 465 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@0 466 const int MAX_STORE = BytesPerLong;
duke@0 467 if (st_offset >= offset + size_in_bytes ||
duke@0 468 st_offset <= offset - MAX_STORE ||
duke@0 469 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@0 470 // Success: The offsets are provably independent.
duke@0 471 // (You may ask, why not just test st_offset != offset and be done?
duke@0 472 // The answer is that stores of different sizes can co-exist
duke@0 473 // in the same sequence of RawMem effects. We sometimes initialize
duke@0 474 // a whole 'tile' of array elements with a single jint or jlong.)
duke@0 475 mem = mem->in(MemNode::Memory);
duke@0 476 continue; // (a) advance through independent store memory
duke@0 477 }
duke@0 478 }
duke@0 479 if (st_base != base &&
duke@0 480 detect_ptr_independence(base, alloc,
duke@0 481 st_base,
duke@0 482 AllocateNode::Ideal_allocation(st_base, phase),
duke@0 483 phase)) {
duke@0 484 // Success: The bases are provably independent.
duke@0 485 mem = mem->in(MemNode::Memory);
duke@0 486 continue; // (a) advance through independent store memory
duke@0 487 }
duke@0 488
duke@0 489 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@0 490 // since we have not managed to prove 'this' and 'mem' independent.
duke@0 491 if (st_base == base && st_offset == offset) {
duke@0 492 return mem; // let caller handle steps (c), (d)
duke@0 493 }
duke@0 494
duke@0 495 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 496 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@0 497 AllocateNode* st_alloc = st_init->allocation();
duke@0 498 if (st_alloc == NULL)
duke@0 499 break; // something degenerated
duke@0 500 bool known_identical = false;
duke@0 501 bool known_independent = false;
duke@0 502 if (alloc == st_alloc)
duke@0 503 known_identical = true;
duke@0 504 else if (alloc != NULL)
duke@0 505 known_independent = true;
kvn@119 506 else if (all_controls_dominate(this, st_alloc))
duke@0 507 known_independent = true;
duke@0 508
duke@0 509 if (known_independent) {
duke@0 510 // The bases are provably independent: Either they are
duke@0 511 // manifestly distinct allocations, or else the control
duke@0 512 // of this load dominates the store's allocation.
duke@0 513 int alias_idx = phase->C->get_alias_index(adr_type());
duke@0 514 if (alias_idx == Compile::AliasIdxRaw) {
duke@0 515 mem = st_alloc->in(TypeFunc::Memory);
duke@0 516 } else {
duke@0 517 mem = st_init->memory(alias_idx);
duke@0 518 }
duke@0 519 continue; // (a) advance through independent store memory
duke@0 520 }
duke@0 521
duke@0 522 // (b) at this point, if we are not looking at a store initializing
duke@0 523 // the same allocation we are loading from, we lose.
duke@0 524 if (known_identical) {
duke@0 525 // From caller, can_see_stored_value will consult find_captured_store.
duke@0 526 return mem; // let caller handle steps (c), (d)
duke@0 527 }
duke@0 528
kvn@223 529 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
kvn@74 530 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
kvn@74 531 if (mem->is_Proj() && mem->in(0)->is_Call()) {
kvn@74 532 CallNode *call = mem->in(0)->as_Call();
kvn@74 533 if (!call->may_modify(addr_t, phase)) {
kvn@74 534 mem = call->in(TypeFunc::Memory);
kvn@74 535 continue; // (a) advance through independent call memory
kvn@74 536 }
kvn@74 537 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
kvn@74 538 mem = mem->in(0)->in(TypeFunc::Memory);
kvn@74 539 continue; // (a) advance through independent MemBar memory
kvn@74 540 } else if (mem->is_MergeMem()) {
kvn@74 541 int alias_idx = phase->C->get_alias_index(adr_type());
kvn@74 542 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@74 543 continue; // (a) advance through independent MergeMem memory
kvn@74 544 }
duke@0 545 }
duke@0 546
duke@0 547 // Unless there is an explicit 'continue', we must bail out here,
duke@0 548 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@0 549 break;
duke@0 550 }
duke@0 551
duke@0 552 return NULL; // bail out
duke@0 553 }
duke@0 554
duke@0 555 //----------------------calculate_adr_type-------------------------------------
duke@0 556 // Helper function. Notices when the given type of address hits top or bottom.
duke@0 557 // Also, asserts a cross-check of the type against the expected address type.
duke@0 558 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@0 559 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@0 560 #ifdef PRODUCT
duke@0 561 cross_check = NULL;
duke@0 562 #else
duke@0 563 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@0 564 #endif
duke@0 565 const TypePtr* tp = t->isa_ptr();
duke@0 566 if (tp == NULL) {
duke@0 567 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@0 568 return TypePtr::BOTTOM; // touches lots of memory
duke@0 569 } else {
duke@0 570 #ifdef ASSERT
duke@0 571 // %%%% [phh] We don't check the alias index if cross_check is
duke@0 572 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@0 573 if (cross_check != NULL &&
duke@0 574 cross_check != TypePtr::BOTTOM &&
duke@0 575 cross_check != TypeRawPtr::BOTTOM) {
duke@0 576 // Recheck the alias index, to see if it has changed (due to a bug).
duke@0 577 Compile* C = Compile::current();
duke@0 578 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@0 579 "must stay in the original alias category");
duke@0 580 // The type of the address must be contained in the adr_type,
duke@0 581 // disregarding "null"-ness.
duke@0 582 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@0 583 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
duke@0 584 assert(cross_check->meet(tp_notnull) == cross_check,
duke@0 585 "real address must not escape from expected memory type");
duke@0 586 }
duke@0 587 #endif
duke@0 588 return tp;
duke@0 589 }
duke@0 590 }
duke@0 591
duke@0 592 //------------------------adr_phi_is_loop_invariant----------------------------
duke@0 593 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@0 594 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@0 595 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@0 596 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@0 597 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@0 598 // with the same data. This implies that any path into the loop already
duke@0 599 // includes such a CastPP, and so the original cast, whatever its input,
duke@0 600 // must be covered by an equivalent cast, with an earlier control input.
duke@0 601 ResourceMark rm;
duke@0 602
duke@0 603 // The loop entry input of the phi should be the unique dominating
duke@0 604 // node for every Phi/CastPP in the loop.
duke@0 605 Unique_Node_List closure;
duke@0 606 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@0 607
duke@0 608 // Add the phi node and the cast to the worklist.
duke@0 609 Unique_Node_List worklist;
duke@0 610 worklist.push(adr_phi);
duke@0 611 if( cast != NULL ){
duke@0 612 if( !cast->is_ConstraintCast() ) return false;
duke@0 613 worklist.push(cast);
duke@0 614 }
duke@0 615
duke@0 616 // Begin recursive walk of phi nodes.
duke@0 617 while( worklist.size() ){
duke@0 618 // Take a node off the worklist
duke@0 619 Node *n = worklist.pop();
duke@0 620 if( !closure.member(n) ){
duke@0 621 // Add it to the closure.
duke@0 622 closure.push(n);
duke@0 623 // Make a sanity check to ensure we don't waste too much time here.
duke@0 624 if( closure.size() > 20) return false;
duke@0 625 // This node is OK if:
duke@0 626 // - it is a cast of an identical value
duke@0 627 // - or it is a phi node (then we add its inputs to the worklist)
duke@0 628 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@0 629 if( n->is_ConstraintCast() ){
duke@0 630 worklist.push(n->in(1));
duke@0 631 } else if( n->is_Phi() ) {
duke@0 632 for( uint i = 1; i < n->req(); i++ ) {
duke@0 633 worklist.push(n->in(i));
duke@0 634 }
duke@0 635 } else {
duke@0 636 return false;
duke@0 637 }
duke@0 638 }
duke@0 639 }
duke@0 640
duke@0 641 // Quit when the worklist is empty, and we've found no offending nodes.
duke@0 642 return true;
duke@0 643 }
duke@0 644
duke@0 645 //------------------------------Ideal_DU_postCCP-------------------------------
duke@0 646 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@0 647 // going away in this pass and we need to make this memory op depend on the
duke@0 648 // gating null check.
kvn@163 649 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
kvn@163 650 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
kvn@163 651 }
duke@0 652
duke@0 653 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@0 654 // some sense; we get to keep around the knowledge that an oop is not-null
duke@0 655 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@0 656 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@0 657 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@0 658 // some of the more trivial cases in the optimizer. Removing more useless
duke@0 659 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@0 660 // up on this approach. CNC 10/20/2000
kvn@163 661 // This static method may be called not from MemNode (EncodePNode calls it).
kvn@163 662 // Only the control edge of the node 'n' might be updated.
kvn@163 663 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
duke@0 664 Node *skipped_cast = NULL;
duke@0 665 // Need a null check? Regular static accesses do not because they are
duke@0 666 // from constant addresses. Array ops are gated by the range check (which
duke@0 667 // always includes a NULL check). Just check field ops.
kvn@163 668 if( n->in(MemNode::Control) == NULL ) {
duke@0 669 // Scan upwards for the highest location we can place this memory op.
duke@0 670 while( true ) {
duke@0 671 switch( adr->Opcode() ) {
duke@0 672
duke@0 673 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@0 674 adr = adr->in(AddPNode::Base);
duke@0 675 continue;
duke@0 676
coleenp@113 677 case Op_DecodeN: // No change to NULL-ness, so peek thru
coleenp@113 678 adr = adr->in(1);
coleenp@113 679 continue;
coleenp@113 680
duke@0 681 case Op_CastPP:
duke@0 682 // If the CastPP is useless, just peek on through it.
duke@0 683 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@0 684 // Remember the cast that we've peeked though. If we peek
duke@0 685 // through more than one, then we end up remembering the highest
duke@0 686 // one, that is, if in a loop, the one closest to the top.
duke@0 687 skipped_cast = adr;
duke@0 688 adr = adr->in(1);
duke@0 689 continue;
duke@0 690 }
duke@0 691 // CastPP is going away in this pass! We need this memory op to be
duke@0 692 // control-dependent on the test that is guarding the CastPP.
kvn@163 693 ccp->hash_delete(n);
kvn@163 694 n->set_req(MemNode::Control, adr->in(0));
kvn@163 695 ccp->hash_insert(n);
kvn@163 696 return n;
duke@0 697
duke@0 698 case Op_Phi:
duke@0 699 // Attempt to float above a Phi to some dominating point.
duke@0 700 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@0 701 // If we've already peeked through a Cast (which could have set the
duke@0 702 // control), we can't float above a Phi, because the skipped Cast
duke@0 703 // may not be loop invariant.
duke@0 704 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@0 705 adr = adr->in(1);
duke@0 706 continue;
duke@0 707 }
duke@0 708 }
duke@0 709
duke@0 710 // Intentional fallthrough!
duke@0 711
duke@0 712 // No obvious dominating point. The mem op is pinned below the Phi
duke@0 713 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@0 714 // then the mem op can float, but not indefinitely. It must be pinned
duke@0 715 // behind the controls leading to the Phi.
duke@0 716 case Op_CheckCastPP:
duke@0 717 // These usually stick around to change address type, however a
duke@0 718 // useless one can be elided and we still need to pick up a control edge
duke@0 719 if (adr->in(0) == NULL) {
duke@0 720 // This CheckCastPP node has NO control and is likely useless. But we
duke@0 721 // need check further up the ancestor chain for a control input to keep
duke@0 722 // the node in place. 4959717.
duke@0 723 skipped_cast = adr;
duke@0 724 adr = adr->in(1);
duke@0 725 continue;
duke@0 726 }
kvn@163 727 ccp->hash_delete(n);
kvn@163 728 n->set_req(MemNode::Control, adr->in(0));
kvn@163 729 ccp->hash_insert(n);
kvn@163 730 return n;
duke@0 731
duke@0 732 // List of "safe" opcodes; those that implicitly block the memory
duke@0 733 // op below any null check.
duke@0 734 case Op_CastX2P: // no null checks on native pointers
duke@0 735 case Op_Parm: // 'this' pointer is not null
duke@0 736 case Op_LoadP: // Loading from within a klass
coleenp@113 737 case Op_LoadN: // Loading from within a klass
duke@0 738 case Op_LoadKlass: // Loading from within a klass
kvn@164 739 case Op_LoadNKlass: // Loading from within a klass
duke@0 740 case Op_ConP: // Loading from a klass
kvn@163 741 case Op_ConN: // Loading from a klass
duke@0 742 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@0 743 case Op_Con: // Reading from TLS
duke@0 744 case Op_CMoveP: // CMoveP is pinned
kvn@164 745 case Op_CMoveN: // CMoveN is pinned
duke@0 746 break; // No progress
duke@0 747
duke@0 748 case Op_Proj: // Direct call to an allocation routine
duke@0 749 case Op_SCMemProj: // Memory state from store conditional ops
duke@0 750 #ifdef ASSERT
duke@0 751 {
duke@0 752 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@0 753 const Node* call = adr->in(0);
kvn@163 754 if (call->is_CallJava()) {
kvn@163 755 const CallJavaNode* call_java = call->as_CallJava();
kvn@64 756 const TypeTuple *r = call_java->tf()->range();
kvn@64 757 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@64 758 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@64 759 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@0 760 // We further presume that this is one of
duke@0 761 // new_instance_Java, new_array_Java, or
duke@0 762 // the like, but do not assert for this.
duke@0 763 } else if (call->is_Allocate()) {
duke@0 764 // similar case to new_instance_Java, etc.
duke@0 765 } else if (!call->is_CallLeaf()) {
duke@0 766 // Projections from fetch_oop (OSR) are allowed as well.
duke@0 767 ShouldNotReachHere();
duke@0 768 }
duke@0 769 }
duke@0 770 #endif
duke@0 771 break;
duke@0 772 default:
duke@0 773 ShouldNotReachHere();
duke@0 774 }
duke@0 775 break;
duke@0 776 }
duke@0 777 }
duke@0 778
duke@0 779 return NULL; // No progress
duke@0 780 }
duke@0 781
duke@0 782
duke@0 783 //=============================================================================
duke@0 784 uint LoadNode::size_of() const { return sizeof(*this); }
duke@0 785 uint LoadNode::cmp( const Node &n ) const
duke@0 786 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@0 787 const Type *LoadNode::bottom_type() const { return _type; }
duke@0 788 uint LoadNode::ideal_reg() const {
duke@0 789 return Matcher::base2reg[_type->base()];
duke@0 790 }
duke@0 791
duke@0 792 #ifndef PRODUCT
duke@0 793 void LoadNode::dump_spec(outputStream *st) const {
duke@0 794 MemNode::dump_spec(st);
duke@0 795 if( !Verbose && !WizardMode ) {
duke@0 796 // standard dump does this in Verbose and WizardMode
duke@0 797 st->print(" #"); _type->dump_on(st);
duke@0 798 }
duke@0 799 }
duke@0 800 #endif
duke@0 801
duke@0 802
duke@0 803 //----------------------------LoadNode::make-----------------------------------
duke@0 804 // Polymorphic factory method:
coleenp@113 805 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
coleenp@113 806 Compile* C = gvn.C;
coleenp@113 807
duke@0 808 // sanity check the alias category against the created node type
duke@0 809 assert(!(adr_type->isa_oopptr() &&
duke@0 810 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@0 811 "use LoadKlassNode instead");
duke@0 812 assert(!(adr_type->isa_aryptr() &&
duke@0 813 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@0 814 "use LoadRangeNode instead");
duke@0 815 switch (bt) {
twisti@624 816 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 817 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 818 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 819 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 820 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 821 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
twisti@558 822 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt );
twisti@558 823 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt );
twisti@558 824 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
coleenp@113 825 case T_OBJECT:
coleenp@113 826 #ifdef _LP64
kvn@163 827 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
kvn@221 828 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
kvn@221 829 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
coleenp@113 830 } else
coleenp@113 831 #endif
kvn@163 832 {
kvn@163 833 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@163 834 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
kvn@163 835 }
duke@0 836 }
duke@0 837 ShouldNotReachHere();
duke@0 838 return (LoadNode*)NULL;
duke@0 839 }
duke@0 840
duke@0 841 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
duke@0 842 bool require_atomic = true;
duke@0 843 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
duke@0 844 }
duke@0 845
duke@0 846
duke@0 847
duke@0 848
duke@0 849 //------------------------------hash-------------------------------------------
duke@0 850 uint LoadNode::hash() const {
duke@0 851 // unroll addition of interesting fields
duke@0 852 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@0 853 }
duke@0 854
duke@0 855 //---------------------------can_see_stored_value------------------------------
duke@0 856 // This routine exists to make sure this set of tests is done the same
duke@0 857 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@0 858 // will change the graph shape in a way which makes memory alive twice at the
duke@0 859 // same time (uses the Oracle model of aliasing), then some
duke@0 860 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@0 861 // of aliasing.
duke@0 862 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@0 863 Node* ld_adr = in(MemNode::Address);
duke@0 864
never@17 865 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
never@17 866 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
never@17 867 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
never@17 868 atp->field() != NULL && !atp->field()->is_volatile()) {
never@17 869 uint alias_idx = atp->index();
never@17 870 bool final = atp->field()->is_final();
never@17 871 Node* result = NULL;
never@17 872 Node* current = st;
never@17 873 // Skip through chains of MemBarNodes checking the MergeMems for
never@17 874 // new states for the slice of this load. Stop once any other
never@17 875 // kind of node is encountered. Loads from final memory can skip
never@17 876 // through any kind of MemBar but normal loads shouldn't skip
never@17 877 // through MemBarAcquire since the could allow them to move out of
never@17 878 // a synchronized region.
never@17 879 while (current->is_Proj()) {
never@17 880 int opc = current->in(0)->Opcode();
never@17 881 if ((final && opc == Op_MemBarAcquire) ||
never@17 882 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
never@17 883 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@17 884 if (mem->is_MergeMem()) {
never@17 885 MergeMemNode* merge = mem->as_MergeMem();
never@17 886 Node* new_st = merge->memory_at(alias_idx);
never@17 887 if (new_st == merge->base_memory()) {
never@17 888 // Keep searching
never@17 889 current = merge->base_memory();
never@17 890 continue;
never@17 891 }
never@17 892 // Save the new memory state for the slice and fall through
never@17 893 // to exit.
never@17 894 result = new_st;
never@17 895 }
never@17 896 }
never@17 897 break;
never@17 898 }
never@17 899 if (result != NULL) {
never@17 900 st = result;
never@17 901 }
never@17 902 }
never@17 903
never@17 904
duke@0 905 // Loop around twice in the case Load -> Initialize -> Store.
duke@0 906 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@0 907 for (int trip = 0; trip <= 1; trip++) {
duke@0 908
duke@0 909 if (st->is_Store()) {
duke@0 910 Node* st_adr = st->in(MemNode::Address);
duke@0 911 if (!phase->eqv(st_adr, ld_adr)) {
duke@0 912 // Try harder before giving up... Match raw and non-raw pointers.
duke@0 913 intptr_t st_off = 0;
duke@0 914 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@0 915 if (alloc == NULL) return NULL;
duke@0 916 intptr_t ld_off = 0;
duke@0 917 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
duke@0 918 if (alloc != allo2) return NULL;
duke@0 919 if (ld_off != st_off) return NULL;
duke@0 920 // At this point we have proven something like this setup:
duke@0 921 // A = Allocate(...)
duke@0 922 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@0 923 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@0 924 // (Actually, we haven't yet proven the Q's are the same.)
duke@0 925 // In other words, we are loading from a casted version of
duke@0 926 // the same pointer-and-offset that we stored to.
duke@0 927 // Thus, we are able to replace L by V.
duke@0 928 }
duke@0 929 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@0 930 if (store_Opcode() != st->Opcode())
duke@0 931 return NULL;
duke@0 932 return st->in(MemNode::ValueIn);
duke@0 933 }
duke@0 934
duke@0 935 intptr_t offset = 0; // scratch
duke@0 936
duke@0 937 // A load from a freshly-created object always returns zero.
duke@0 938 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@0 939 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@0 940 if (st->is_Proj() && st->in(0)->is_Allocate() &&
duke@0 941 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
duke@0 942 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
duke@0 943 // return a zero value for the load's basic type
duke@0 944 // (This is one of the few places where a generic PhaseTransform
duke@0 945 // can create new nodes. Think of it as lazily manifesting
duke@0 946 // virtually pre-existing constants.)
duke@0 947 return phase->zerocon(memory_type());
duke@0 948 }
duke@0 949
duke@0 950 // A load from an initialization barrier can match a captured store.
duke@0 951 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@0 952 InitializeNode* init = st->in(0)->as_Initialize();
duke@0 953 AllocateNode* alloc = init->allocation();
duke@0 954 if (alloc != NULL &&
duke@0 955 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
duke@0 956 // examine a captured store value
duke@0 957 st = init->find_captured_store(offset, memory_size(), phase);
duke@0 958 if (st != NULL)
duke@0 959 continue; // take one more trip around
duke@0 960 }
duke@0 961 }
duke@0 962
duke@0 963 break;
duke@0 964 }
duke@0 965
duke@0 966 return NULL;
duke@0 967 }
duke@0 968
kvn@64 969 //----------------------is_instance_field_load_with_local_phi------------------
kvn@64 970 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@64 971 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
kvn@64 972 in(MemNode::Address)->is_AddP() ) {
kvn@64 973 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@64 974 // Only instances.
kvn@223 975 if( t_oop != NULL && t_oop->is_known_instance_field() &&
kvn@64 976 t_oop->offset() != Type::OffsetBot &&
kvn@64 977 t_oop->offset() != Type::OffsetTop) {
kvn@64 978 return true;
kvn@64 979 }
kvn@64 980 }
kvn@64 981 return false;
kvn@64 982 }
kvn@64 983
duke@0 984 //------------------------------Identity---------------------------------------
duke@0 985 // Loads are identity if previous store is to same address
duke@0 986 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@0 987 // If the previous store-maker is the right kind of Store, and the store is
duke@0 988 // to the same address, then we are equal to the value stored.
duke@0 989 Node* mem = in(MemNode::Memory);
duke@0 990 Node* value = can_see_stored_value(mem, phase);
duke@0 991 if( value ) {
duke@0 992 // byte, short & char stores truncate naturally.
duke@0 993 // A load has to load the truncated value which requires
duke@0 994 // some sort of masking operation and that requires an
duke@0 995 // Ideal call instead of an Identity call.
duke@0 996 if (memory_size() < BytesPerInt) {
duke@0 997 // If the input to the store does not fit with the load's result type,
duke@0 998 // it must be truncated via an Ideal call.
duke@0 999 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@0 1000 return this;
duke@0 1001 }
duke@0 1002 // (This works even when value is a Con, but LoadNode::Value
duke@0 1003 // usually runs first, producing the singleton type of the Con.)
duke@0 1004 return value;
duke@0 1005 }
kvn@64 1006
kvn@64 1007 // Search for an existing data phi which was generated before for the same
twisti@605 1008 // instance's field to avoid infinite generation of phis in a loop.
kvn@64 1009 Node *region = mem->in(0);
kvn@64 1010 if (is_instance_field_load_with_local_phi(region)) {
kvn@64 1011 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
kvn@64 1012 int this_index = phase->C->get_alias_index(addr_t);
kvn@64 1013 int this_offset = addr_t->offset();
kvn@64 1014 int this_id = addr_t->is_oopptr()->instance_id();
kvn@64 1015 const Type* this_type = bottom_type();
kvn@64 1016 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@64 1017 Node* phi = region->fast_out(i);
kvn@64 1018 if (phi->is_Phi() && phi != mem &&
kvn@64 1019 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
kvn@64 1020 return phi;
kvn@64 1021 }
kvn@64 1022 }
kvn@64 1023 }
kvn@64 1024
duke@0 1025 return this;
duke@0 1026 }
duke@0 1027
never@17 1028
never@17 1029 // Returns true if the AliasType refers to the field that holds the
never@17 1030 // cached box array. Currently only handles the IntegerCache case.
never@17 1031 static bool is_autobox_cache(Compile::AliasType* atp) {
never@17 1032 if (atp != NULL && atp->field() != NULL) {
never@17 1033 ciField* field = atp->field();
never@17 1034 ciSymbol* klass = field->holder()->name();
never@17 1035 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1036 field->holder()->uses_default_loader() &&
never@17 1037 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1038 return true;
never@17 1039 }
never@17 1040 }
never@17 1041 return false;
never@17 1042 }
never@17 1043
never@17 1044 // Fetch the base value in the autobox array
never@17 1045 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
never@17 1046 if (atp != NULL && atp->field() != NULL) {
never@17 1047 ciField* field = atp->field();
never@17 1048 ciSymbol* klass = field->holder()->name();
never@17 1049 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1050 field->holder()->uses_default_loader() &&
never@17 1051 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1052 assert(field->is_constant(), "what?");
never@17 1053 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
never@17 1054 // Fetch the box object at the base of the array and get its value
never@17 1055 ciInstance* box = array->obj_at(0)->as_instance();
never@17 1056 ciInstanceKlass* ik = box->klass()->as_instance_klass();
never@17 1057 if (ik->nof_nonstatic_fields() == 1) {
never@17 1058 // This should be true nonstatic_field_at requires calling
never@17 1059 // nof_nonstatic_fields so check it anyway
never@17 1060 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
never@17 1061 cache_offset = c.as_int();
never@17 1062 }
never@17 1063 return true;
never@17 1064 }
never@17 1065 }
never@17 1066 return false;
never@17 1067 }
never@17 1068
never@17 1069 // Returns true if the AliasType refers to the value field of an
never@17 1070 // autobox object. Currently only handles Integer.
never@17 1071 static bool is_autobox_object(Compile::AliasType* atp) {
never@17 1072 if (atp != NULL && atp->field() != NULL) {
never@17 1073 ciField* field = atp->field();
never@17 1074 ciSymbol* klass = field->holder()->name();
never@17 1075 if (field->name() == ciSymbol::value_name() &&
never@17 1076 field->holder()->uses_default_loader() &&
never@17 1077 klass == ciSymbol::java_lang_Integer()) {
never@17 1078 return true;
never@17 1079 }
never@17 1080 }
never@17 1081 return false;
never@17 1082 }
never@17 1083
never@17 1084
never@17 1085 // We're loading from an object which has autobox behaviour.
never@17 1086 // If this object is result of a valueOf call we'll have a phi
never@17 1087 // merging a newly allocated object and a load from the cache.
never@17 1088 // We want to replace this load with the original incoming
never@17 1089 // argument to the valueOf call.
never@17 1090 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
never@17 1091 Node* base = in(Address)->in(AddPNode::Base);
never@17 1092 if (base->is_Phi() && base->req() == 3) {
never@17 1093 AllocateNode* allocation = NULL;
never@17 1094 int allocation_index = -1;
never@17 1095 int load_index = -1;
never@17 1096 for (uint i = 1; i < base->req(); i++) {
never@17 1097 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
never@17 1098 if (allocation != NULL) {
never@17 1099 allocation_index = i;
never@17 1100 load_index = 3 - allocation_index;
never@17 1101 break;
never@17 1102 }
never@17 1103 }
kvn@583 1104 bool has_load = ( allocation != NULL &&
kvn@583 1105 (base->in(load_index)->is_Load() ||
kvn@583 1106 base->in(load_index)->is_DecodeN() &&
kvn@583 1107 base->in(load_index)->in(1)->is_Load()) );
kvn@583 1108 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
never@17 1109 // Push the loads from the phi that comes from valueOf up
never@17 1110 // through it to allow elimination of the loads and the recovery
never@17 1111 // of the original value.
never@17 1112 Node* mem_phi = in(Memory);
never@17 1113 Node* offset = in(Address)->in(AddPNode::Offset);
never@553 1114 Node* region = base->in(0);
never@17 1115
never@17 1116 Node* in1 = clone();
never@17 1117 Node* in1_addr = in1->in(Address)->clone();
never@17 1118 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
never@17 1119 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
never@17 1120 in1_addr->set_req(AddPNode::Offset, offset);
never@553 1121 in1->set_req(0, region->in(allocation_index));
never@17 1122 in1->set_req(Address, in1_addr);
never@17 1123 in1->set_req(Memory, mem_phi->in(allocation_index));
never@17 1124
never@17 1125 Node* in2 = clone();
never@17 1126 Node* in2_addr = in2->in(Address)->clone();
never@17 1127 in2_addr->set_req(AddPNode::Base, base->in(load_index));
never@17 1128 in2_addr->set_req(AddPNode::Address, base->in(load_index));
never@17 1129 in2_addr->set_req(AddPNode::Offset, offset);
never@553 1130 in2->set_req(0, region->in(load_index));
never@17 1131 in2->set_req(Address, in2_addr);
never@17 1132 in2->set_req(Memory, mem_phi->in(load_index));
never@17 1133
never@17 1134 in1_addr = phase->transform(in1_addr);
never@17 1135 in1 = phase->transform(in1);
never@17 1136 in2_addr = phase->transform(in2_addr);
never@17 1137 in2 = phase->transform(in2);
never@17 1138
never@553 1139 PhiNode* result = PhiNode::make_blank(region, this);
never@17 1140 result->set_req(allocation_index, in1);
never@17 1141 result->set_req(load_index, in2);
never@17 1142 return result;
never@17 1143 }
kvn@583 1144 } else if (base->is_Load() ||
kvn@583 1145 base->is_DecodeN() && base->in(1)->is_Load()) {
kvn@583 1146 if (base->is_DecodeN()) {
kvn@583 1147 // Get LoadN node which loads cached Integer object
kvn@583 1148 base = base->in(1);
kvn@583 1149 }
never@17 1150 // Eliminate the load of Integer.value for integers from the cache
never@17 1151 // array by deriving the value from the index into the array.
never@17 1152 // Capture the offset of the load and then reverse the computation.
never@17 1153 Node* load_base = base->in(Address)->in(AddPNode::Base);
kvn@583 1154 if (load_base->is_DecodeN()) {
kvn@583 1155 // Get LoadN node which loads IntegerCache.cache field
kvn@583 1156 load_base = load_base->in(1);
kvn@583 1157 }
never@17 1158 if (load_base != NULL) {
never@17 1159 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
never@17 1160 intptr_t cache_offset;
never@17 1161 int shift = -1;
never@17 1162 Node* cache = NULL;
never@17 1163 if (is_autobox_cache(atp)) {
kvn@29 1164 shift = exact_log2(type2aelembytes(T_OBJECT));
never@17 1165 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
never@17 1166 }
never@17 1167 if (cache != NULL && base->in(Address)->is_AddP()) {
never@17 1168 Node* elements[4];
never@17 1169 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@17 1170 int cache_low;
never@17 1171 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
never@17 1172 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
never@17 1173 // Add up all the offsets making of the address of the load
never@17 1174 Node* result = elements[0];
never@17 1175 for (int i = 1; i < count; i++) {
never@17 1176 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
never@17 1177 }
never@17 1178 // Remove the constant offset from the address and then
never@17 1179 // remove the scaling of the offset to recover the original index.
never@17 1180 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
never@17 1181 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
never@17 1182 // Peel the shift off directly but wrap it in a dummy node
never@17 1183 // since Ideal can't return existing nodes
never@17 1184 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
never@17 1185 } else {
never@17 1186 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
never@17 1187 }
never@17 1188 #ifdef _LP64
never@17 1189 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
never@17 1190 #endif
never@17 1191 return result;
never@17 1192 }
never@17 1193 }
never@17 1194 }
never@17 1195 }
never@17 1196 return NULL;
never@17 1197 }
never@17 1198
kvn@163 1199 //------------------------------split_through_phi------------------------------
kvn@163 1200 // Split instance field load through Phi.
kvn@163 1201 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
kvn@163 1202 Node* mem = in(MemNode::Memory);
kvn@163 1203 Node* address = in(MemNode::Address);
kvn@163 1204 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@163 1205 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@163 1206
kvn@163 1207 assert(mem->is_Phi() && (t_oop != NULL) &&
kvn@223 1208 t_oop->is_known_instance_field(), "invalide conditions");
kvn@163 1209
kvn@163 1210 Node *region = mem->in(0);
kvn@163 1211 if (region == NULL) {
kvn@163 1212 return NULL; // Wait stable graph
kvn@163 1213 }
kvn@163 1214 uint cnt = mem->req();
kvn@163 1215 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1216 Node *in = mem->in(i);
kvn@163 1217 if( in == NULL ) {
kvn@163 1218 return NULL; // Wait stable graph
kvn@163 1219 }
kvn@163 1220 }
kvn@163 1221 // Check for loop invariant.
kvn@163 1222 if (cnt == 3) {
kvn@163 1223 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1224 Node *in = mem->in(i);
kvn@163 1225 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
kvn@163 1226 if (m == mem) {
kvn@163 1227 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
kvn@163 1228 return this;
kvn@163 1229 }
kvn@163 1230 }
kvn@163 1231 }
kvn@163 1232 // Split through Phi (see original code in loopopts.cpp).
kvn@163 1233 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
kvn@163 1234
kvn@163 1235 // Do nothing here if Identity will find a value
kvn@163 1236 // (to avoid infinite chain of value phis generation).
kvn@163 1237 if ( !phase->eqv(this, this->Identity(phase)) )
kvn@163 1238 return NULL;
kvn@163 1239
kvn@163 1240 // Skip the split if the region dominates some control edge of the address.
kvn@163 1241 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
kvn@163 1242 return NULL;
kvn@163 1243
kvn@163 1244 const Type* this_type = this->bottom_type();
kvn@163 1245 int this_index = phase->C->get_alias_index(addr_t);
kvn@163 1246 int this_offset = addr_t->offset();
kvn@163 1247 int this_iid = addr_t->is_oopptr()->instance_id();
kvn@163 1248 int wins = 0;
kvn@163 1249 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@163 1250 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
kvn@163 1251 for( uint i = 1; i < region->req(); i++ ) {
kvn@163 1252 Node *x;
kvn@163 1253 Node* the_clone = NULL;
kvn@163 1254 if( region->in(i) == phase->C->top() ) {
kvn@163 1255 x = phase->C->top(); // Dead path? Use a dead data op
kvn@163 1256 } else {
kvn@163 1257 x = this->clone(); // Else clone up the data op
kvn@163 1258 the_clone = x; // Remember for possible deletion.
kvn@163 1259 // Alter data node to use pre-phi inputs
kvn@163 1260 if( this->in(0) == region ) {
kvn@163 1261 x->set_req( 0, region->in(i) );
kvn@163 1262 } else {
kvn@163 1263 x->set_req( 0, NULL );
kvn@163 1264 }
kvn@163 1265 for( uint j = 1; j < this->req(); j++ ) {
kvn@163 1266 Node *in = this->in(j);
kvn@163 1267 if( in->is_Phi() && in->in(0) == region )
kvn@163 1268 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
kvn@163 1269 }
kvn@163 1270 }
kvn@163 1271 // Check for a 'win' on some paths
kvn@163 1272 const Type *t = x->Value(igvn);
kvn@163 1273
kvn@163 1274 bool singleton = t->singleton();
kvn@163 1275
kvn@163 1276 // See comments in PhaseIdealLoop::split_thru_phi().
kvn@163 1277 if( singleton && t == Type::TOP ) {
kvn@163 1278 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
kvn@163 1279 }
kvn@163 1280
kvn@163 1281 if( singleton ) {
kvn@163 1282 wins++;
kvn@163 1283 x = igvn->makecon(t);
kvn@163 1284 } else {
kvn@163 1285 // We now call Identity to try to simplify the cloned node.
kvn@163 1286 // Note that some Identity methods call phase->type(this).
kvn@163 1287 // Make sure that the type array is big enough for
kvn@163 1288 // our new node, even though we may throw the node away.
kvn@163 1289 // (This tweaking with igvn only works because x is a new node.)
kvn@163 1290 igvn->set_type(x, t);
kvn@293 1291 // If x is a TypeNode, capture any more-precise type permanently into Node
twisti@605 1292 // otherwise it will be not updated during igvn->transform since
kvn@293 1293 // igvn->type(x) is set to x->Value() already.
kvn@293 1294 x->raise_bottom_type(t);
kvn@163 1295 Node *y = x->Identity(igvn);
kvn@163 1296 if( y != x ) {
kvn@163 1297 wins++;
kvn@163 1298 x = y;
kvn@163 1299 } else {
kvn@163 1300 y = igvn->hash_find(x);
kvn@163 1301 if( y ) {
kvn@163 1302 wins++;
kvn@163 1303 x = y;
kvn@163 1304 } else {
kvn@163 1305 // Else x is a new node we are keeping
kvn@163 1306 // We do not need register_new_node_with_optimizer
kvn@163 1307 // because set_type has already been called.
kvn@163 1308 igvn->_worklist.push(x);
kvn@163 1309 }
kvn@163 1310 }
kvn@163 1311 }
kvn@163 1312 if (x != the_clone && the_clone != NULL)
kvn@163 1313 igvn->remove_dead_node(the_clone);
kvn@163 1314 phi->set_req(i, x);
kvn@163 1315 }
kvn@163 1316 if( wins > 0 ) {
kvn@163 1317 // Record Phi
kvn@163 1318 igvn->register_new_node_with_optimizer(phi);
kvn@163 1319 return phi;
kvn@163 1320 }
kvn@163 1321 igvn->remove_dead_node(phi);
kvn@163 1322 return NULL;
kvn@163 1323 }
never@17 1324
duke@0 1325 //------------------------------Ideal------------------------------------------
duke@0 1326 // If the load is from Field memory and the pointer is non-null, we can
duke@0 1327 // zero out the control input.
duke@0 1328 // If the offset is constant and the base is an object allocation,
duke@0 1329 // try to hook me up to the exact initializing store.
duke@0 1330 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1331 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1332 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1333
duke@0 1334 Node* ctrl = in(MemNode::Control);
duke@0 1335 Node* address = in(MemNode::Address);
duke@0 1336
duke@0 1337 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@0 1338 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@0 1339 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@0 1340 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@0 1341 ctrl = ctrl->in(0);
duke@0 1342 set_req(MemNode::Control,ctrl);
duke@0 1343 }
duke@0 1344
kvn@708 1345 intptr_t ignore = 0;
kvn@708 1346 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
kvn@708 1347 if (base != NULL
kvn@708 1348 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
kvn@708 1349 // Check for useless control edge in some common special cases
kvn@708 1350 if (in(MemNode::Control) != NULL
duke@0 1351 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
kvn@119 1352 && all_controls_dominate(base, phase->C->start())) {
duke@0 1353 // A method-invariant, non-null address (constant or 'this' argument).
duke@0 1354 set_req(MemNode::Control, NULL);
duke@0 1355 }
kvn@708 1356
kvn@708 1357 if (EliminateAutoBox && can_reshape) {
kvn@708 1358 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
never@17 1359 Compile::AliasType* atp = phase->C->alias_type(adr_type());
never@17 1360 if (is_autobox_object(atp)) {
never@17 1361 Node* result = eliminate_autobox(phase);
never@17 1362 if (result != NULL) return result;
never@17 1363 }
never@17 1364 }
never@17 1365 }
never@17 1366
kvn@74 1367 Node* mem = in(MemNode::Memory);
kvn@74 1368 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@74 1369
kvn@74 1370 if (addr_t != NULL) {
kvn@74 1371 // try to optimize our memory input
kvn@74 1372 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
kvn@74 1373 if (opt_mem != mem) {
kvn@74 1374 set_req(MemNode::Memory, opt_mem);
kvn@305 1375 if (phase->type( opt_mem ) == Type::TOP) return NULL;
kvn@74 1376 return this;
kvn@74 1377 }
kvn@74 1378 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@74 1379 if (can_reshape && opt_mem->is_Phi() &&
kvn@223 1380 (t_oop != NULL) && t_oop->is_known_instance_field()) {
kvn@163 1381 // Split instance field load through Phi.
kvn@163 1382 Node* result = split_through_phi(phase);
kvn@163 1383 if (result != NULL) return result;
kvn@74 1384 }
kvn@74 1385 }
kvn@74 1386
duke@0 1387 // Check for prior store with a different base or offset; make Load
duke@0 1388 // independent. Skip through any number of them. Bail out if the stores
duke@0 1389 // are in an endless dead cycle and report no progress. This is a key
duke@0 1390 // transform for Reflection. However, if after skipping through the Stores
duke@0 1391 // we can't then fold up against a prior store do NOT do the transform as
duke@0 1392 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@0 1393 // array memory alive twice: once for the hoisted Load and again after the
duke@0 1394 // bypassed Store. This situation only works if EVERYBODY who does
duke@0 1395 // anti-dependence work knows how to bypass. I.e. we need all
duke@0 1396 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@0 1397 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@0 1398 // fold up, do so.
duke@0 1399 Node* prev_mem = find_previous_store(phase);
duke@0 1400 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1401 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@0 1402 // (c) See if we can fold up on the spot, but don't fold up here.
twisti@558 1403 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
duke@0 1404 // just return a prior value, which is done by Identity calls.
duke@0 1405 if (can_see_stored_value(prev_mem, phase)) {
duke@0 1406 // Make ready for step (d):
duke@0 1407 set_req(MemNode::Memory, prev_mem);
duke@0 1408 return this;
duke@0 1409 }
duke@0 1410 }
duke@0 1411
duke@0 1412 return NULL; // No further progress
duke@0 1413 }
duke@0 1414
duke@0 1415 // Helper to recognize certain Klass fields which are invariant across
duke@0 1416 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@0 1417 const Type*
duke@0 1418 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@0 1419 ciKlass* klass) const {
duke@0 1420 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1421 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@0 1422 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@0 1423 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@0 1424 return TypeInt::make(klass->modifier_flags());
duke@0 1425 }
duke@0 1426 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1427 // The field is Klass::_access_flags. Return its (constant) value.
duke@0 1428 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@0 1429 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@0 1430 return TypeInt::make(klass->access_flags());
duke@0 1431 }
duke@0 1432 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1433 // The field is Klass::_layout_helper. Return its constant value if known.
duke@0 1434 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1435 return TypeInt::make(klass->layout_helper());
duke@0 1436 }
duke@0 1437
duke@0 1438 // No match.
duke@0 1439 return NULL;
duke@0 1440 }
duke@0 1441
duke@0 1442 //------------------------------Value-----------------------------------------
duke@0 1443 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@0 1444 // Either input is TOP ==> the result is TOP
duke@0 1445 Node* mem = in(MemNode::Memory);
duke@0 1446 const Type *t1 = phase->type(mem);
duke@0 1447 if (t1 == Type::TOP) return Type::TOP;
duke@0 1448 Node* adr = in(MemNode::Address);
duke@0 1449 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@0 1450 if (tp == NULL || tp->empty()) return Type::TOP;
duke@0 1451 int off = tp->offset();
duke@0 1452 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
duke@0 1453
duke@0 1454 // Try to guess loaded type from pointer type
duke@0 1455 if (tp->base() == Type::AryPtr) {
duke@0 1456 const Type *t = tp->is_aryptr()->elem();
duke@0 1457 // Don't do this for integer types. There is only potential profit if
duke@0 1458 // the element type t is lower than _type; that is, for int types, if _type is
duke@0 1459 // more restrictive than t. This only happens here if one is short and the other
duke@0 1460 // char (both 16 bits), and in those cases we've made an intentional decision
duke@0 1461 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@0 1462 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@0 1463 //
duke@0 1464 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@0 1465 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@0 1466 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@0 1467 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@0 1468 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@0 1469 // In fact, that could have been the original type of p1, and p1 could have
duke@0 1470 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@0 1471 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@0 1472 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
kvn@293 1473 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
duke@0 1474 // t might actually be lower than _type, if _type is a unique
duke@0 1475 // concrete subclass of abstract class t.
duke@0 1476 // Make sure the reference is not into the header, by comparing
duke@0 1477 // the offset against the offset of the start of the array's data.
duke@0 1478 // Different array types begin at slightly different offsets (12 vs. 16).
duke@0 1479 // We choose T_BYTE as an example base type that is least restrictive
duke@0 1480 // as to alignment, which will therefore produce the smallest
duke@0 1481 // possible base offset.
duke@0 1482 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1483 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
duke@0 1484 const Type* jt = t->join(_type);
duke@0 1485 // In any case, do not allow the join, per se, to empty out the type.
duke@0 1486 if (jt->empty() && !t->empty()) {
duke@0 1487 // This can happen if a interface-typed array narrows to a class type.
duke@0 1488 jt = _type;
duke@0 1489 }
never@17 1490
kvn@708 1491 if (EliminateAutoBox && adr->is_AddP()) {
never@17 1492 // The pointers in the autobox arrays are always non-null
kvn@708 1493 Node* base = adr->in(AddPNode::Base);
kvn@708 1494 if (base != NULL &&
kvn@708 1495 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
never@17 1496 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
never@17 1497 if (is_autobox_cache(atp)) {
never@17 1498 return jt->join(TypePtr::NOTNULL)->is_ptr();
never@17 1499 }
never@17 1500 }
never@17 1501 }
duke@0 1502 return jt;
duke@0 1503 }
duke@0 1504 }
duke@0 1505 } else if (tp->base() == Type::InstPtr) {
never@1080 1506 const TypeInstPtr* tinst = tp->is_instptr();
never@1080 1507 ciKlass* klass = tinst->klass();
duke@0 1508 assert( off != Type::OffsetBot ||
duke@0 1509 // arrays can be cast to Objects
duke@0 1510 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@0 1511 // unsafe field access may not have a constant offset
duke@0 1512 phase->C->has_unsafe_access(),
duke@0 1513 "Field accesses must be precise" );
duke@0 1514 // For oop loads, we expect the _type to be precise
never@1080 1515 if (OptimizeStringConcat && klass == phase->C->env()->String_klass() &&
never@1080 1516 adr->is_AddP() && off != Type::OffsetBot) {
never@1080 1517 // For constant Strings treat the fields as compile time constants.
never@1080 1518 Node* base = adr->in(AddPNode::Base);
never@1080 1519 if (base->Opcode() == Op_ConP) {
never@1080 1520 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
never@1080 1521 ciObject* string = t->const_oop();
never@1080 1522 ciConstant constant = string->as_instance()->field_value_by_offset(off);
never@1080 1523 if (constant.basic_type() == T_INT) {
never@1080 1524 return TypeInt::make(constant.as_int());
never@1080 1525 } else if (constant.basic_type() == T_ARRAY) {
never@1080 1526 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
never@1080 1527 return TypeNarrowOop::make_from_constant(constant.as_object());
never@1080 1528 } else {
never@1080 1529 return TypeOopPtr::make_from_constant(constant.as_object());
never@1080 1530 }
never@1080 1531 }
never@1080 1532 }
never@1080 1533 }
duke@0 1534 } else if (tp->base() == Type::KlassPtr) {
duke@0 1535 assert( off != Type::OffsetBot ||
duke@0 1536 // arrays can be cast to Objects
duke@0 1537 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@0 1538 // also allow array-loading from the primary supertype
duke@0 1539 // array during subtype checks
duke@0 1540 Opcode() == Op_LoadKlass,
duke@0 1541 "Field accesses must be precise" );
duke@0 1542 // For klass/static loads, we expect the _type to be precise
duke@0 1543 }
duke@0 1544
duke@0 1545 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1546 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1547 ciKlass* klass = tkls->klass();
duke@0 1548 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@0 1549 // We are loading a field from a Klass metaobject whose identity
duke@0 1550 // is known at compile time (the type is "exact" or "precise").
duke@0 1551 // Check for fields we know are maintained as constants by the VM.
duke@0 1552 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1553 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@0 1554 // (Folds up type checking code.)
duke@0 1555 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@0 1556 return TypeInt::make(klass->super_check_offset());
duke@0 1557 }
duke@0 1558 // Compute index into primary_supers array
duke@0 1559 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1560 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1561 if( depth < ciKlass::primary_super_limit() ) {
duke@0 1562 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1563 // (Folds up type checking code.)
duke@0 1564 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1565 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1566 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1567 }
duke@0 1568 const Type* aift = load_array_final_field(tkls, klass);
duke@0 1569 if (aift != NULL) return aift;
duke@0 1570 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
duke@0 1571 && klass->is_array_klass()) {
duke@0 1572 // The field is arrayKlass::_component_mirror. Return its (constant) value.
duke@0 1573 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@0 1574 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@0 1575 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@0 1576 }
duke@0 1577 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1578 // The field is Klass::_java_mirror. Return its (constant) value.
duke@0 1579 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@0 1580 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@0 1581 return TypeInstPtr::make(klass->java_mirror());
duke@0 1582 }
duke@0 1583 }
duke@0 1584
duke@0 1585 // We can still check if we are loading from the primary_supers array at a
duke@0 1586 // shallow enough depth. Even though the klass is not exact, entries less
duke@0 1587 // than or equal to its super depth are correct.
duke@0 1588 if (klass->is_loaded() ) {
duke@0 1589 ciType *inner = klass->klass();
duke@0 1590 while( inner->is_obj_array_klass() )
duke@0 1591 inner = inner->as_obj_array_klass()->base_element_type();
duke@0 1592 if( inner->is_instance_klass() &&
duke@0 1593 !inner->as_instance_klass()->flags().is_interface() ) {
duke@0 1594 // Compute index into primary_supers array
duke@0 1595 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1596 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1597 if( depth < ciKlass::primary_super_limit() &&
duke@0 1598 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@0 1599 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1600 // (Folds up type checking code.)
duke@0 1601 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1602 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1603 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1604 }
duke@0 1605 }
duke@0 1606 }
duke@0 1607
duke@0 1608 // If the type is enough to determine that the thing is not an array,
duke@0 1609 // we can give the layout_helper a positive interval type.
duke@0 1610 // This will help short-circuit some reflective code.
duke@0 1611 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
duke@0 1612 && !klass->is_array_klass() // not directly typed as an array
duke@0 1613 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@0 1614 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@0 1615 ) {
duke@0 1616 // Note: When interfaces are reliable, we can narrow the interface
duke@0 1617 // test to (klass != Serializable && klass != Cloneable).
duke@0 1618 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1619 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@0 1620 // The key property of this type is that it folds up tests
duke@0 1621 // for array-ness, since it proves that the layout_helper is positive.
duke@0 1622 // Thus, a generic value like the basic object layout helper works fine.
duke@0 1623 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@0 1624 }
duke@0 1625 }
duke@0 1626
duke@0 1627 // If we are loading from a freshly-allocated object, produce a zero,
duke@0 1628 // if the load is provably beyond the header of the object.
duke@0 1629 // (Also allow a variable load from a fresh array to produce zero.)
duke@0 1630 if (ReduceFieldZeroing) {
duke@0 1631 Node* value = can_see_stored_value(mem,phase);
duke@0 1632 if (value != NULL && value->is_Con())
duke@0 1633 return value->bottom_type();
duke@0 1634 }
duke@0 1635
kvn@64 1636 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@223 1637 if (tinst != NULL && tinst->is_known_instance_field()) {
kvn@64 1638 // If we have an instance type and our memory input is the
kvn@64 1639 // programs's initial memory state, there is no matching store,
kvn@64 1640 // so just return a zero of the appropriate type
kvn@64 1641 Node *mem = in(MemNode::Memory);
kvn@64 1642 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@64 1643 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@64 1644 return Type::get_zero_type(_type->basic_type());
kvn@64 1645 }
kvn@64 1646 }
duke@0 1647 return _type;
duke@0 1648 }
duke@0 1649
duke@0 1650 //------------------------------match_edge-------------------------------------
duke@0 1651 // Do we Match on this edge index or not? Match only the address.
duke@0 1652 uint LoadNode::match_edge(uint idx) const {
duke@0 1653 return idx == MemNode::Address;
duke@0 1654 }
duke@0 1655
duke@0 1656 //--------------------------LoadBNode::Ideal--------------------------------------
duke@0 1657 //
duke@0 1658 // If the previous store is to the same address as this load,
duke@0 1659 // and the value stored was larger than a byte, replace this load
duke@0 1660 // with the value stored truncated to a byte. If no truncation is
duke@0 1661 // needed, the replacement is done in LoadNode::Identity().
duke@0 1662 //
duke@0 1663 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1664 Node* mem = in(MemNode::Memory);
duke@0 1665 Node* value = can_see_stored_value(mem,phase);
duke@0 1666 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1667 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
duke@0 1668 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
duke@0 1669 }
duke@0 1670 // Identity call will handle the case where truncation is not needed.
duke@0 1671 return LoadNode::Ideal(phase, can_reshape);
duke@0 1672 }
duke@0 1673
twisti@624 1674 //--------------------------LoadUBNode::Ideal-------------------------------------
twisti@624 1675 //
twisti@624 1676 // If the previous store is to the same address as this load,
twisti@624 1677 // and the value stored was larger than a byte, replace this load
twisti@624 1678 // with the value stored truncated to a byte. If no truncation is
twisti@624 1679 // needed, the replacement is done in LoadNode::Identity().
twisti@624 1680 //
twisti@624 1681 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
twisti@624 1682 Node* mem = in(MemNode::Memory);
twisti@624 1683 Node* value = can_see_stored_value(mem, phase);
twisti@624 1684 if (value && !phase->type(value)->higher_equal(_type))
twisti@624 1685 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF));
twisti@624 1686 // Identity call will handle the case where truncation is not needed.
twisti@624 1687 return LoadNode::Ideal(phase, can_reshape);
twisti@624 1688 }
twisti@624 1689
twisti@558 1690 //--------------------------LoadUSNode::Ideal-------------------------------------
duke@0 1691 //
duke@0 1692 // If the previous store is to the same address as this load,
duke@0 1693 // and the value stored was larger than a char, replace this load
duke@0 1694 // with the value stored truncated to a char. If no truncation is
duke@0 1695 // needed, the replacement is done in LoadNode::Identity().
duke@0 1696 //
twisti@558 1697 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1698 Node* mem = in(MemNode::Memory);
duke@0 1699 Node* value = can_see_stored_value(mem,phase);
duke@0 1700 if( value && !phase->type(value)->higher_equal( _type ) )
duke@0 1701 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
duke@0 1702 // Identity call will handle the case where truncation is not needed.
duke@0 1703 return LoadNode::Ideal(phase, can_reshape);
duke@0 1704 }
duke@0 1705
duke@0 1706 //--------------------------LoadSNode::Ideal--------------------------------------
duke@0 1707 //
duke@0 1708 // If the previous store is to the same address as this load,
duke@0 1709 // and the value stored was larger than a short, replace this load
duke@0 1710 // with the value stored truncated to a short. If no truncation is
duke@0 1711 // needed, the replacement is done in LoadNode::Identity().
duke@0 1712 //
duke@0 1713 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1714 Node* mem = in(MemNode::Memory);
duke@0 1715 Node* value = can_see_stored_value(mem,phase);
duke@0 1716 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1717 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
duke@0 1718 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
duke@0 1719 }
duke@0 1720 // Identity call will handle the case where truncation is not needed.
duke@0 1721 return LoadNode::Ideal(phase, can_reshape);
duke@0 1722 }
duke@0 1723
duke@0 1724 //=============================================================================
kvn@164 1725 //----------------------------LoadKlassNode::make------------------------------
kvn@164 1726 // Polymorphic factory method:
kvn@164 1727 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
kvn@164 1728 Compile* C = gvn.C;
kvn@164 1729 Node *ctl = NULL;
kvn@164 1730 // sanity check the alias category against the created node type
kvn@164 1731 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
kvn@164 1732 assert(adr_type != NULL, "expecting TypeOopPtr");
kvn@164 1733 #ifdef _LP64
kvn@164 1734 if (adr_type->is_ptr_to_narrowoop()) {
kvn@221 1735 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
kvn@221 1736 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
kvn@168 1737 }
kvn@164 1738 #endif
kvn@168 1739 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@168 1740 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
kvn@164 1741 }
kvn@164 1742
duke@0 1743 //------------------------------Value------------------------------------------
duke@0 1744 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1745 return klass_value_common(phase);
kvn@164 1746 }
kvn@164 1747
kvn@164 1748 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
duke@0 1749 // Either input is TOP ==> the result is TOP
duke@0 1750 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1751 if (t1 == Type::TOP) return Type::TOP;
duke@0 1752 Node *adr = in(MemNode::Address);
duke@0 1753 const Type *t2 = phase->type( adr );
duke@0 1754 if (t2 == Type::TOP) return Type::TOP;
duke@0 1755 const TypePtr *tp = t2->is_ptr();
duke@0 1756 if (TypePtr::above_centerline(tp->ptr()) ||
duke@0 1757 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@0 1758
duke@0 1759 // Return a more precise klass, if possible
duke@0 1760 const TypeInstPtr *tinst = tp->isa_instptr();
duke@0 1761 if (tinst != NULL) {
duke@0 1762 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@0 1763 int offset = tinst->offset();
duke@0 1764 if (ik == phase->C->env()->Class_klass()
duke@0 1765 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1766 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1767 // We are loading a special hidden field from a Class mirror object,
duke@0 1768 // the field which points to the VM's Klass metaobject.
duke@0 1769 ciType* t = tinst->java_mirror_type();
duke@0 1770 // java_mirror_type returns non-null for compile-time Class constants.
duke@0 1771 if (t != NULL) {
duke@0 1772 // constant oop => constant klass
duke@0 1773 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1774 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@0 1775 }
duke@0 1776 if (!t->is_klass()) {
duke@0 1777 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@0 1778 return TypePtr::NULL_PTR;
duke@0 1779 }
duke@0 1780 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@0 1781 return TypeKlassPtr::make(t->as_klass());
duke@0 1782 }
duke@0 1783 // non-constant mirror, so we can't tell what's going on
duke@0 1784 }
duke@0 1785 if( !ik->is_loaded() )
duke@0 1786 return _type; // Bail out if not loaded
duke@0 1787 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1788 if (tinst->klass_is_exact()) {
duke@0 1789 return TypeKlassPtr::make(ik);
duke@0 1790 }
duke@0 1791 // See if we can become precise: no subklasses and no interface
duke@0 1792 // (Note: We need to support verified interfaces.)
duke@0 1793 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1794 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1795 // Add a dependence; if any subclass added we need to recompile
duke@0 1796 if (!ik->is_final()) {
duke@0 1797 // %%% should use stronger assert_unique_concrete_subtype instead
duke@0 1798 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1799 }
duke@0 1800 // Return precise klass
duke@0 1801 return TypeKlassPtr::make(ik);
duke@0 1802 }
duke@0 1803
duke@0 1804 // Return root of possible klass
duke@0 1805 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@0 1806 }
duke@0 1807 }
duke@0 1808
duke@0 1809 // Check for loading klass from an array
duke@0 1810 const TypeAryPtr *tary = tp->isa_aryptr();
duke@0 1811 if( tary != NULL ) {
duke@0 1812 ciKlass *tary_klass = tary->klass();
duke@0 1813 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@0 1814 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@0 1815 if (tary->klass_is_exact()) {
duke@0 1816 return TypeKlassPtr::make(tary_klass);
duke@0 1817 }
duke@0 1818 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@0 1819 // If the klass is an object array, we defer the question to the
duke@0 1820 // array component klass.
duke@0 1821 if( ak->is_obj_array_klass() ) {
duke@0 1822 assert( ak->is_loaded(), "" );
duke@0 1823 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@0 1824 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@0 1825 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@0 1826 // See if we can become precise: no subklasses and no interface
duke@0 1827 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1828 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1829 // Add a dependence; if any subclass added we need to recompile
duke@0 1830 if (!ik->is_final()) {
duke@0 1831 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1832 }
duke@0 1833 // Return precise array klass
duke@0 1834 return TypeKlassPtr::make(ak);
duke@0 1835 }
duke@0 1836 }
duke@0 1837 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 1838 } else { // Found a type-array?
duke@0 1839 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1840 assert( ak->is_type_array_klass(), "" );
duke@0 1841 return TypeKlassPtr::make(ak); // These are always precise
duke@0 1842 }
duke@0 1843 }
duke@0 1844 }
duke@0 1845
duke@0 1846 // Check for loading klass from an array klass
duke@0 1847 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1848 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1849 ciKlass* klass = tkls->klass();
duke@0 1850 if( !klass->is_loaded() )
duke@0 1851 return _type; // Bail out if not loaded
duke@0 1852 if( klass->is_obj_array_klass() &&
duke@0 1853 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1854 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@0 1855 // // Always returning precise element type is incorrect,
duke@0 1856 // // e.g., element type could be object and array may contain strings
duke@0 1857 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@0 1858
duke@0 1859 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@0 1860 // according to the element type's subclassing.
duke@0 1861 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@0 1862 }
duke@0 1863 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
duke@0 1864 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1865 ciKlass* sup = klass->as_instance_klass()->super();
duke@0 1866 // The field is Klass::_super. Return its (constant) value.
duke@0 1867 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@0 1868 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@0 1869 }
duke@0 1870 }
duke@0 1871
duke@0 1872 // Bailout case
duke@0 1873 return LoadNode::Value(phase);
duke@0 1874 }
duke@0 1875
duke@0 1876 //------------------------------Identity---------------------------------------
duke@0 1877 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@0 1878 // Also feed through the klass in Allocate(...klass...)._klass.
duke@0 1879 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1880 return klass_identity_common(phase);
kvn@164 1881 }
kvn@164 1882
kvn@164 1883 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
duke@0 1884 Node* x = LoadNode::Identity(phase);
duke@0 1885 if (x != this) return x;
duke@0 1886
duke@0 1887 // Take apart the address into an oop and and offset.
duke@0 1888 // Return 'this' if we cannot.
duke@0 1889 Node* adr = in(MemNode::Address);
duke@0 1890 intptr_t offset = 0;
duke@0 1891 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1892 if (base == NULL) return this;
duke@0 1893 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@0 1894 if (toop == NULL) return this;
duke@0 1895
duke@0 1896 // We can fetch the klass directly through an AllocateNode.
duke@0 1897 // This works even if the klass is not constant (clone or newArray).
duke@0 1898 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1899 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@0 1900 if (allocated_klass != NULL) {
duke@0 1901 return allocated_klass;
duke@0 1902 }
duke@0 1903 }
duke@0 1904
duke@0 1905 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
duke@0 1906 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
duke@0 1907 // See inline_native_Class_query for occurrences of these patterns.
duke@0 1908 // Java Example: x.getClass().isAssignableFrom(y)
duke@0 1909 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@0 1910 //
duke@0 1911 // This improves reflective code, often making the Class
duke@0 1912 // mirror go completely dead. (Current exception: Class
duke@0 1913 // mirrors may appear in debug info, but we could clean them out by
duke@0 1914 // introducing a new debug info operator for klassOop.java_mirror).
duke@0 1915 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@0 1916 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1917 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1918 // We are loading a special hidden field from a Class mirror,
duke@0 1919 // the field which points to its Klass or arrayKlass metaobject.
duke@0 1920 if (base->is_Load()) {
duke@0 1921 Node* adr2 = base->in(MemNode::Address);
duke@0 1922 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@0 1923 if (tkls != NULL && !tkls->empty()
duke@0 1924 && (tkls->klass()->is_instance_klass() ||
duke@0 1925 tkls->klass()->is_array_klass())
duke@0 1926 && adr2->is_AddP()
duke@0 1927 ) {
duke@0 1928 int mirror_field = Klass::java_mirror_offset_in_bytes();
duke@0 1929 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1930 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
duke@0 1931 }
duke@0 1932 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
duke@0 1933 return adr2->in(AddPNode::Base);
duke@0 1934 }
duke@0 1935 }
duke@0 1936 }
duke@0 1937 }
duke@0 1938
duke@0 1939 return this;
duke@0 1940 }
duke@0 1941
kvn@164 1942
kvn@164 1943 //------------------------------Value------------------------------------------
kvn@164 1944 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1945 const Type *t = klass_value_common(phase);
kvn@221 1946 if (t == Type::TOP)
kvn@221 1947 return t;
kvn@221 1948
kvn@221 1949 return t->make_narrowoop();
kvn@164 1950 }
kvn@164 1951
kvn@164 1952 //------------------------------Identity---------------------------------------
kvn@164 1953 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
kvn@164 1954 // Also feed through the klass in Allocate(...klass...)._klass.
kvn@164 1955 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1956 Node *x = klass_identity_common(phase);
kvn@164 1957
kvn@164 1958 const Type *t = phase->type( x );
kvn@164 1959 if( t == Type::TOP ) return x;
kvn@164 1960 if( t->isa_narrowoop()) return x;
kvn@164 1961
kvn@221 1962 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
kvn@164 1963 }
kvn@164 1964
duke@0 1965 //------------------------------Value-----------------------------------------
duke@0 1966 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@0 1967 // Either input is TOP ==> the result is TOP
duke@0 1968 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1969 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1970 Node *adr = in(MemNode::Address);
duke@0 1971 const Type *t2 = phase->type( adr );
duke@0 1972 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1973 const TypePtr *tp = t2->is_ptr();
duke@0 1974 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@0 1975 const TypeAryPtr *tap = tp->isa_aryptr();
duke@0 1976 if( !tap ) return _type;
duke@0 1977 return tap->size();
duke@0 1978 }
duke@0 1979
rasbold@366 1980 //-------------------------------Ideal---------------------------------------
rasbold@366 1981 // Feed through the length in AllocateArray(...length...)._length.
rasbold@366 1982 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
rasbold@366 1983 Node* p = MemNode::Ideal_common(phase, can_reshape);
rasbold@366 1984 if (p) return (p == NodeSentinel) ? NULL : p;
rasbold@366 1985
rasbold@366 1986 // Take apart the address into an oop and and offset.
rasbold@366 1987 // Return 'this' if we cannot.
rasbold@366 1988 Node* adr = in(MemNode::Address);
rasbold@366 1989 intptr_t offset = 0;
rasbold@366 1990 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
rasbold@366 1991 if (base == NULL) return NULL;
rasbold@366 1992 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
rasbold@366 1993 if (tary == NULL) return NULL;
rasbold@366 1994
rasbold@366 1995 // We can fetch the length directly through an AllocateArrayNode.
rasbold@366 1996 // This works even if the length is not constant (clone or newArray).
rasbold@366 1997 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 1998 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 1999 if (alloc != NULL) {
rasbold@366 2000 Node* allocated_length = alloc->Ideal_length();
rasbold@366 2001 Node* len = alloc->make_ideal_length(tary, phase);
rasbold@366 2002 if (allocated_length != len) {
rasbold@366 2003 // New CastII improves on this.
rasbold@366 2004 return len;
rasbold@366 2005 }
rasbold@366 2006 }
rasbold@366 2007 }
rasbold@366 2008
rasbold@366 2009 return NULL;
rasbold@366 2010 }
rasbold@366 2011
duke@0 2012 //------------------------------Identity---------------------------------------
duke@0 2013 // Feed through the length in AllocateArray(...length...)._length.
duke@0 2014 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@0 2015 Node* x = LoadINode::Identity(phase);
duke@0 2016 if (x != this) return x;
duke@0 2017
duke@0 2018 // Take apart the address into an oop and and offset.
duke@0 2019 // Return 'this' if we cannot.
duke@0 2020 Node* adr = in(MemNode::Address);
duke@0 2021 intptr_t offset = 0;
duke@0 2022 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 2023 if (base == NULL) return this;
duke@0 2024 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@0 2025 if (tary == NULL) return this;
duke@0 2026
duke@0 2027 // We can fetch the length directly through an AllocateArrayNode.
duke@0 2028 // This works even if the length is not constant (clone or newArray).
duke@0 2029 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 2030 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 2031 if (alloc != NULL) {
rasbold@366 2032 Node* allocated_length = alloc->Ideal_length();
rasbold@366 2033 // Do not allow make_ideal_length to allocate a CastII node.
rasbold@366 2034 Node* len = alloc->make_ideal_length(tary, phase, false);
rasbold@366 2035 if (allocated_length == len) {
rasbold@366 2036 // Return allocated_length only if it would not be improved by a CastII.
rasbold@366 2037 return allocated_length;
rasbold@366 2038 }
duke@0 2039 }
duke@0 2040 }
duke@0 2041
duke@0 2042 return this;
duke@0 2043
duke@0 2044 }
rasbold@366 2045
duke@0 2046 //=============================================================================
duke@0 2047 //---------------------------StoreNode::make-----------------------------------
duke@0 2048 // Polymorphic factory method:
coleenp@113 2049 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
coleenp@113 2050 Compile* C = gvn.C;
coleenp@113 2051
duke@0 2052 switch (bt) {
duke@0 2053 case T_BOOLEAN:
duke@0 2054 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
duke@0 2055 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
duke@0 2056 case T_CHAR:
duke@0 2057 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
duke@0 2058 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
duke@0 2059 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
duke@0 2060 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
duke@0 2061 case T_ADDRESS:
coleenp@113 2062 case T_OBJECT:
coleenp@113 2063 #ifdef _LP64
kvn@163 2064 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
coleenp@113 2065 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
coleenp@113 2066 adr->bottom_type()->isa_rawptr())) {
kvn@221 2067 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
kvn@221 2068 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
coleenp@113 2069 } else
coleenp@113 2070 #endif
kvn@221 2071 {
kvn@221 2072 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
kvn@221 2073 }
duke@0 2074 }
duke@0 2075 ShouldNotReachHere();
duke@0 2076 return (StoreNode*)NULL;
duke@0 2077 }
duke@0 2078
duke@0 2079 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
duke@0 2080 bool require_atomic = true;
duke@0 2081 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
duke@0 2082 }
duke@0 2083
duke@0 2084
duke@0 2085 //--------------------------bottom_type----------------------------------------
duke@0 2086 const Type *StoreNode::bottom_type() const {
duke@0 2087 return Type::MEMORY;
duke@0 2088 }
duke@0 2089
duke@0 2090 //------------------------------hash-------------------------------------------
duke@0 2091 uint StoreNode::hash() const {
duke@0 2092 // unroll addition of interesting fields
duke@0 2093 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@0 2094
duke@0 2095 // Since they are not commoned, do not hash them:
duke@0 2096 return NO_HASH;
duke@0 2097 }
duke@0 2098
duke@0 2099 //------------------------------Ideal------------------------------------------
duke@0 2100 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@0 2101 // When a store immediately follows a relevant allocation/initialization,
duke@0 2102 // try to capture it into the initialization, or hoist it above.
duke@0 2103 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 2104 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 2105 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 2106
duke@0 2107 Node* mem = in(MemNode::Memory);
duke@0 2108 Node* address = in(MemNode::Address);
duke@0 2109
duke@0 2110 // Back-to-back stores to same address? Fold em up.
duke@0 2111 // Generally unsafe if I have intervening uses...
duke@0 2112 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
duke@0 2113 // Looking at a dead closed cycle of memory?
duke@0 2114 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@0 2115
duke@0 2116 assert(Opcode() == mem->Opcode() ||
duke@0 2117 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@0 2118 "no mismatched stores, except on raw memory");
duke@0 2119
duke@0 2120 if (mem->outcnt() == 1 && // check for intervening uses
duke@0 2121 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@0 2122 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@0 2123 // For example, 'mem' might be the final state at a conditional return.
duke@0 2124 // Or, 'mem' might be used by some node which is live at the same time
duke@0 2125 // 'this' is live, which might be unschedulable. So, require exactly
duke@0 2126 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@0 2127 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@0 2128 if (can_reshape) { // (%%% is this an anachronism?)
duke@0 2129 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@0 2130 phase->is_IterGVN());
duke@0 2131 } else {
duke@0 2132 // It's OK to do this in the parser, since DU info is always accurate,
duke@0 2133 // and the parser always refers to nodes via SafePointNode maps.
duke@0 2134 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@0 2135 }
duke@0 2136 return this;
duke@0 2137 }
duke@0 2138 }
duke@0 2139
duke@0 2140 // Capture an unaliased, unconditional, simple store into an initializer.
duke@0 2141 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@0 2142 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@0 2143 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 2144 InitializeNode* init = mem->in(0)->as_Initialize();
duke@0 2145 intptr_t offset = init->can_capture_store(this, phase);
duke@0 2146 if (offset > 0) {
duke@0 2147 Node* moved = init->capture_store(this, offset, phase);
duke@0 2148 // If the InitializeNode captured me, it made a raw copy of me,
duke@0 2149 // and I need to disappear.
duke@0 2150 if (moved != NULL) {
duke@0 2151 // %%% hack to ensure that Ideal returns a new node:
duke@0 2152 mem = MergeMemNode::make(phase->C, mem);
duke@0 2153 return mem; // fold me away
duke@0 2154 }
duke@0 2155 }
duke@0 2156 }
duke@0 2157
duke@0 2158 return NULL; // No further progress
duke@0 2159 }
duke@0 2160
duke@0 2161 //------------------------------Value-----------------------------------------
duke@0 2162 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@0 2163 // Either input is TOP ==> the result is TOP
duke@0 2164 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 2165 if( t1 == Type::TOP ) return Type::TOP;
duke@0 2166 const Type *t2 = phase->type( in(MemNode::Address) );
duke@0 2167 if( t2 == Type::TOP ) return Type::TOP;
duke@0 2168 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@0 2169 if( t3 == Type::TOP ) return Type::TOP;
duke@0 2170 return Type::MEMORY;
duke@0 2171 }
duke@0 2172
duke@0 2173 //------------------------------Identity---------------------------------------
duke@0 2174 // Remove redundant stores:
duke@0 2175 // Store(m, p, Load(m, p)) changes to m.
duke@0 2176 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@0 2177 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@0 2178 Node* mem = in(MemNode::Memory);
duke@0 2179 Node* adr = in(MemNode::Address);
duke@0 2180 Node* val = in(MemNode::ValueIn);
duke@0 2181
duke@0 2182 // Load then Store? Then the Store is useless
duke@0 2183 if (val->is_Load() &&
duke@0 2184 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
duke@0 2185 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
duke@0 2186 val->as_Load()->store_Opcode() == Opcode()) {
duke@0 2187 return mem;
duke@0 2188 }
duke@0 2189
duke@0 2190 // Two stores in a row of the same value?
duke@0 2191 if (mem->is_Store() &&
duke@0 2192 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
duke@0 2193 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
duke@0 2194 mem->Opcode() == Opcode()) {
duke@0 2195 return mem;
duke@0 2196 }
duke@0 2197
duke@0 2198 // Store of zero anywhere into a freshly-allocated object?
duke@0 2199 // Then the store is useless.
duke@0 2200 // (It must already have been captured by the InitializeNode.)
duke@0 2201 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@0 2202 // a newly allocated object is already all-zeroes everywhere
duke@0 2203 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@0 2204 return mem;
duke@0 2205 }
duke@0 2206
duke@0 2207 // the store may also apply to zero-bits in an earlier object
duke@0 2208 Node* prev_mem = find_previous_store(phase);
duke@0 2209 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 2210 if (prev_mem != NULL) {
duke@0 2211 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@0 2212 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@0 2213 // prev_val and val might differ by a cast; it would be good
duke@0 2214 // to keep the more informative of the two.
duke@0 2215 return mem;
duke@0 2216 }
duke@0 2217 }
duke@0 2218 }
duke@0 2219
duke@0 2220 return this;
duke@0 2221 }
duke@0 2222
duke@0 2223 //------------------------------match_edge-------------------------------------
duke@0 2224 // Do we Match on this edge index or not? Match only memory & value
duke@0 2225 uint StoreNode::match_edge(uint idx) const {
duke@0 2226 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@0 2227 }
duke@0 2228
duke@0 2229 //------------------------------cmp--------------------------------------------
duke@0 2230 // Do not common stores up together. They generally have to be split
duke@0 2231 // back up anyways, so do not bother.
duke@0 2232 uint StoreNode::cmp( const Node &n ) const {
duke@0 2233 return (&n == this); // Always fail except on self
duke@0 2234 }
duke@0 2235
duke@0 2236 //------------------------------Ideal_masked_input-----------------------------
duke@0 2237 // Check for a useless mask before a partial-word store
duke@0 2238 // (StoreB ... (AndI valIn conIa) )
duke@0 2239 // If (conIa & mask == mask) this simplifies to
duke@0 2240 // (StoreB ... (valIn) )
duke@0 2241 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@0 2242 Node *val = in(MemNode::ValueIn);
duke@0 2243 if( val->Opcode() == Op_AndI ) {
duke@0 2244 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2245 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@0 2246 set_req(MemNode::ValueIn, val->in(1));
duke@0 2247 return this;
duke@0 2248 }
duke@0 2249 }
duke@0 2250 return NULL;
duke@0 2251 }
duke@0 2252
duke@0 2253
duke@0 2254 //------------------------------Ideal_sign_extended_input----------------------
duke@0 2255 // Check for useless sign-extension before a partial-word store
duke@0 2256 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@0 2257 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@0 2258 // (StoreB ... (valIn) )
duke@0 2259 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@0 2260 Node *val = in(MemNode::ValueIn);
duke@0 2261 if( val->Opcode() == Op_RShiftI ) {
duke@0 2262 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2263 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@0 2264 Node *shl = val->in(1);
duke@0 2265 if( shl->Opcode() == Op_LShiftI ) {
duke@0 2266 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@0 2267 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@0 2268 set_req(MemNode::ValueIn, shl->in(1));
duke@0 2269 return this;
duke@0 2270 }
duke@0 2271 }
duke@0 2272 }
duke@0 2273 }
duke@0 2274 return NULL;
duke@0 2275 }
duke@0 2276
duke@0 2277 //------------------------------value_never_loaded-----------------------------------
duke@0 2278 // Determine whether there are any possible loads of the value stored.
duke@0 2279 // For simplicity, we actually check if there are any loads from the
duke@0 2280 // address stored to, not just for loads of the value stored by this node.
duke@0 2281 //
duke@0 2282 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@0 2283 Node *adr = in(Address);
duke@0 2284 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@0 2285 if (adr_oop == NULL)
duke@0 2286 return false;
kvn@223 2287 if (!adr_oop->is_known_instance_field())
duke@0 2288 return false; // if not a distinct instance, there may be aliases of the address
duke@0 2289 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@0 2290 Node *use = adr->fast_out(i);
duke@0 2291 int opc = use->Opcode();
duke@0 2292 if (use->is_Load() || use->is_LoadStore()) {
duke@0 2293 return false;
duke@0 2294 }
duke@0 2295 }
duke@0 2296 return true;
duke@0 2297 }
duke@0 2298
duke@0 2299 //=============================================================================
duke@0 2300 //------------------------------Ideal------------------------------------------
duke@0 2301 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2302 // we can skip the AND operation. If the store is from a sign-extension
duke@0 2303 // (a left shift, then right shift) we can skip both.
duke@0 2304 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2305 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@0 2306 if( progress != NULL ) return progress;
duke@0 2307
duke@0 2308 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@0 2309 if( progress != NULL ) return progress;
duke@0 2310
duke@0 2311 // Finally check the default case
duke@0 2312 return StoreNode::Ideal(phase, can_reshape);
duke@0 2313 }
duke@0 2314
duke@0 2315 //=============================================================================
duke@0 2316 //------------------------------Ideal------------------------------------------
duke@0 2317 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2318 // we can skip the AND operation
duke@0 2319 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2320 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@0 2321 if( progress != NULL ) return progress;
duke@0 2322
duke@0 2323 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@0 2324 if( progress != NULL ) return progress;
duke@0 2325
duke@0 2326 // Finally check the default case
duke@0 2327 return StoreNode::Ideal(phase, can_reshape);
duke@0 2328 }
duke@0 2329
duke@0 2330 //=============================================================================
duke@0 2331 //------------------------------Identity---------------------------------------
duke@0 2332 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@0 2333 // No need to card mark when storing a null ptr
duke@0 2334 Node* my_store = in(MemNode::OopStore);
duke@0 2335 if (my_store->is_Store()) {
duke@0 2336 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@0 2337 if( t1 == TypePtr::NULL_PTR ) {
duke@0 2338 return in(MemNode::Memory);
duke@0 2339 }
duke@0 2340 }
duke@0 2341 return this;
duke@0 2342 }
duke@0 2343
cfang@985 2344 //=============================================================================
cfang@985 2345 //------------------------------Ideal---------------------------------------
cfang@985 2346 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
cfang@985 2347 Node* progress = StoreNode::Ideal(phase, can_reshape);
cfang@985 2348 if (progress != NULL) return progress;
cfang@985 2349
cfang@985 2350 Node* my_store = in(MemNode::OopStore);
cfang@985 2351 if (my_store->is_MergeMem()) {
cfang@985 2352 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
cfang@985 2353 set_req(MemNode::OopStore, mem);
cfang@985 2354 return this;
cfang@985 2355 }
cfang@985 2356
cfang@985 2357 return NULL;
cfang@985 2358 }
cfang@985 2359
duke@0 2360 //------------------------------Value-----------------------------------------
duke@0 2361 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@43 2362 // Either input is TOP ==> the result is TOP
kvn@43 2363 const Type *t = phase->type( in(MemNode::Memory) );
kvn@43 2364 if( t == Type::TOP ) return Type::TOP;
kvn@43 2365 t = phase->type( in(MemNode::Address) );
kvn@43 2366 if( t == Type::TOP ) return Type::TOP;
kvn@43 2367 t = phase->type( in(MemNode::ValueIn) );
kvn@43 2368 if( t == Type::TOP ) return Type::TOP;
duke@0 2369 // If extra input is TOP ==> the result is TOP
kvn@43 2370 t = phase->type( in(MemNode::OopStore) );
kvn@43 2371 if( t == Type::TOP ) return Type::TOP;
duke@0 2372
duke@0 2373 return StoreNode::Value( phase );
duke@0 2374 }
duke@0 2375
duke@0 2376
duke@0 2377 //=============================================================================
duke@0 2378 //----------------------------------SCMemProjNode------------------------------
duke@0 2379 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@0 2380 {
duke@0 2381 return bottom_type();
duke@0 2382 }
duke@0 2383
duke@0 2384 //=============================================================================
duke@0 2385 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
duke@0 2386 init_req(MemNode::Control, c );
duke@0 2387 init_req(MemNode::Memory , mem);
duke@0 2388 init_req(MemNode::Address, adr);
duke@0 2389 init_req(MemNode::ValueIn, val);
duke@0 2390 init_req( ExpectedIn, ex );
duke@0 2391 init_class_id(Class_LoadStore);
duke@0 2392
duke@0 2393 }
duke@0 2394
duke@0 2395 //=============================================================================
duke@0 2396 //-------------------------------adr_type--------------------------------------
duke@0 2397 // Do we Match on this edge index or not? Do not match memory
duke@0 2398 const TypePtr* ClearArrayNode::adr_type() const {
duke@0 2399 Node *adr = in(3);
duke@0 2400 return MemNode::calculate_adr_type(adr->bottom_type());
duke@0 2401 }
duke@0 2402
duke@0 2403 //------------------------------match_edge-------------------------------------
duke@0 2404 // Do we Match on this edge index or not? Do not match memory
duke@0 2405 uint ClearArrayNode::match_edge(uint idx) const {
duke@0 2406 return idx > 1;
duke@0 2407 }
duke@0 2408
duke@0 2409 //------------------------------Identity---------------------------------------
duke@0 2410 // Clearing a zero length array does nothing
duke@0 2411 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@68 2412 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@0 2413 }
duke@0 2414
duke@0 2415 //------------------------------Idealize---------------------------------------
duke@0 2416 // Clearing a short array is faster with stores
duke@0 2417 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2418 const int unit = BytesPerLong;
duke@0 2419 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@0 2420 if (!t) return NULL;
duke@0 2421 if (!t->is_con()) return NULL;
duke@0 2422 intptr_t raw_count = t->get_con();
duke@0 2423 intptr_t size = raw_count;
duke@0 2424 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@0 2425 // Clearing nothing uses the Identity call.
duke@0 2426 // Negative clears are possible on dead ClearArrays
duke@0 2427 // (see jck test stmt114.stmt11402.val).
duke@0 2428 if (size <= 0 || size % unit != 0) return NULL;
duke@0 2429 intptr_t count = size / unit;
duke@0 2430 // Length too long; use fast hardware clear
duke@0 2431 if (size > Matcher::init_array_short_size) return NULL;
duke@0 2432 Node *mem = in(1);
duke@0 2433 if( phase->type(mem)==Type::TOP ) return NULL;
duke@0 2434 Node *adr = in(3);
duke@0 2435 const Type* at = phase->type(adr);
duke@0 2436 if( at==Type::TOP ) return NULL;
duke@0 2437 const TypePtr* atp = at->isa_ptr();
duke@0 2438 // adjust atp to be the correct array element address type
duke@0 2439 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@0 2440 else atp = atp->add_offset(Type::OffsetBot);
duke@0 2441 // Get base for derived pointer purposes
duke@0 2442 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@0 2443 Node *base = adr->in(1);
duke@0 2444
duke@0 2445 Node *zero = phase->makecon(TypeLong::ZERO);
duke@0 2446 Node *off = phase->MakeConX(BytesPerLong);
duke@0 2447 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2448 count--;
duke@0 2449 while( count-- ) {
duke@0 2450 mem = phase->transform(mem);
duke@0 2451 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
duke@0 2452 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2453 }
duke@0 2454 return mem;
duke@0 2455 }
duke@0 2456
duke@0 2457 //----------------------------clear_memory-------------------------------------
duke@0 2458 // Generate code to initialize object storage to zero.
duke@0 2459 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2460 intptr_t start_offset,
duke@0 2461 Node* end_offset,
duke@0 2462 PhaseGVN* phase) {
duke@0 2463 Compile* C = phase->C;
duke@0 2464 intptr_t offset = start_offset;
duke@0 2465
duke@0 2466 int unit = BytesPerLong;
duke@0 2467 if ((offset % unit) != 0) {
duke@0 2468 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
duke@0 2469 adr = phase->transform(adr);
duke@0 2470 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2471 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2472 mem = phase->transform(mem);
duke@0 2473 offset += BytesPerInt;
duke@0 2474 }
duke@0 2475 assert((offset % unit) == 0, "");
duke@0 2476
duke@0 2477 // Initialize the remaining stuff, if any, with a ClearArray.
duke@0 2478 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@0 2479 }
duke@0 2480
duke@0 2481 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2482 Node* start_offset,
duke@0 2483 Node* end_offset,
duke@0 2484 PhaseGVN* phase) {
never@68 2485 if (start_offset == end_offset) {
never@68 2486 // nothing to do
never@68 2487 return mem;
never@68 2488 }
never@68 2489
duke@0 2490 Compile* C = phase->C;
duke@0 2491 int unit = BytesPerLong;
duke@0 2492 Node* zbase = start_offset;
duke@0 2493 Node* zend = end_offset;
duke@0 2494
duke@0 2495 // Scale to the unit required by the CPU:
duke@0 2496 if (!Matcher::init_array_count_is_in_bytes) {
duke@0 2497 Node* shift = phase->intcon(exact_log2(unit));
duke@0 2498 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
duke@0 2499 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
duke@0 2500 }
duke@0 2501
duke@0 2502 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
duke@0 2503 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
duke@0 2504
duke@0 2505 // Bulk clear double-words
duke@0 2506 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
duke@0 2507 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
duke@0 2508 return phase->transform(mem);
duke@0 2509 }
duke@0 2510
duke@0 2511 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2512 intptr_t start_offset,
duke@0 2513 intptr_t end_offset,
duke@0 2514 PhaseGVN* phase) {
never@68 2515 if (start_offset == end_offset) {
never@68 2516 // nothing to do
never@68 2517 return mem;
never@68 2518 }
never@68 2519
duke@0 2520 Compile* C = phase->C;
duke@0 2521 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@0 2522 intptr_t done_offset = end_offset;
duke@0 2523 if ((done_offset % BytesPerLong) != 0) {
duke@0 2524 done_offset -= BytesPerInt;
duke@0 2525 }
duke@0 2526 if (done_offset > start_offset) {
duke@0 2527 mem = clear_memory(ctl, mem, dest,
duke@0 2528 start_offset, phase->MakeConX(done_offset), phase);
duke@0 2529 }
duke@0 2530 if (done_offset < end_offset) { // emit the final 32-bit store
duke@0 2531 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@0 2532 adr = phase->transform(adr);
duke@0 2533 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2534 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2535 mem = phase->transform(mem);
duke@0 2536 done_offset += BytesPerInt;
duke@0 2537 }
duke@0 2538 assert(done_offset == end_offset, "");
duke@0 2539 return mem;
duke@0 2540 }
duke@0 2541
duke@0 2542 //=============================================================================
duke@0 2543 // Do we match on this edge? No memory edges
duke@0 2544 uint StrCompNode::match_edge(uint idx) const {
kvn@986 2545 return idx == 2 || idx == 3; // StrComp (Binary str1 cnt1) (Binary str2 cnt2)
duke@0 2546 }
duke@0 2547
duke@0 2548 //------------------------------Ideal------------------------------------------
duke@0 2549 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2550 // control copies
duke@0 2551 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2552 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2553 }
duke@0 2554
kvn@986 2555 //=============================================================================
cfang@681 2556 // Do we match on this edge? No memory edges
cfang@681 2557 uint StrEqualsNode::match_edge(uint idx) const {
kvn@986 2558 return idx == 2 || idx == 3; // StrEquals (Binary str1 str2) cnt
cfang@681 2559 }
cfang@681 2560
cfang@681 2561 //------------------------------Ideal------------------------------------------
cfang@681 2562 // Return a node which is more "ideal" than the current node. Strip out
cfang@681 2563 // control copies
cfang@681 2564 Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){
cfang@681 2565 return remove_dead_region(phase, can_reshape) ? this : NULL;
cfang@681 2566 }
cfang@681 2567
cfang@681 2568 //=============================================================================
cfang@681 2569 // Do we match on this edge? No memory edges
cfang@681 2570 uint StrIndexOfNode::match_edge(uint idx) const {
kvn@986 2571 return idx == 2 || idx == 3; // StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2)
cfang@681 2572 }
cfang@681 2573
cfang@681 2574 //------------------------------Ideal------------------------------------------
cfang@681 2575 // Return a node which is more "ideal" than the current node. Strip out
cfang@681 2576 // control copies
cfang@681 2577 Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){
cfang@681 2578 return remove_dead_region(phase, can_reshape) ? this : NULL;
cfang@681 2579 }
cfang@681 2580
kvn@986 2581 //=============================================================================
kvn@986 2582 // Do we match on this edge? No memory edges
kvn@986 2583 uint AryEqNode::match_edge(uint idx) const {
kvn@986 2584 return idx == 2 || idx == 3; // StrEquals ary1 ary2
kvn@986 2585 }
rasbold@169 2586 //------------------------------Ideal------------------------------------------
rasbold@169 2587 // Return a node which is more "ideal" than the current node. Strip out
rasbold@169 2588 // control copies
rasbold@169 2589 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
rasbold@169 2590 return remove_dead_region(phase, can_reshape) ? this : NULL;
rasbold@169 2591 }
rasbold@169 2592
duke@0 2593 //=============================================================================
duke@0 2594 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@0 2595 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@0 2596 _adr_type(C->get_adr_type(alias_idx))
duke@0 2597 {
duke@0 2598 init_class_id(Class_MemBar);
duke@0 2599 Node* top = C->top();
duke@0 2600 init_req(TypeFunc::I_O,top);
duke@0 2601 init_req(TypeFunc::FramePtr,top);
duke@0 2602 init_req(TypeFunc::ReturnAdr,top);
duke@0 2603 if (precedent != NULL)
duke@0 2604 init_req(TypeFunc::Parms, precedent);
duke@0 2605 }
duke@0 2606
duke@0 2607 //------------------------------cmp--------------------------------------------
duke@0 2608 uint MemBarNode::hash() const { return NO_HASH; }
duke@0 2609 uint MemBarNode::cmp( const Node &n ) const {
duke@0 2610 return (&n == this); // Always fail except on self
duke@0 2611 }
duke@0 2612
duke@0 2613 //------------------------------make-------------------------------------------
duke@0 2614 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@0 2615 int len = Precedent + (pn == NULL? 0: 1);
duke@0 2616 switch (opcode) {
duke@0 2617 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
duke@0 2618 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
duke@0 2619 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
duke@0 2620 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
duke@0 2621 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
duke@0 2622 default: ShouldNotReachHere(); return NULL;
duke@0 2623 }
duke@0 2624 }
duke@0 2625
duke@0 2626 //------------------------------Ideal------------------------------------------
duke@0 2627 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2628 // control copies
duke@0 2629 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@305 2630 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2631 }
duke@0 2632
duke@0 2633 //------------------------------Value------------------------------------------
duke@0 2634 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@0 2635 if( !in(0) ) return Type::TOP;
duke@0 2636 if( phase->type(in(0)) == Type::TOP )
duke@0 2637 return Type::TOP;
duke@0 2638 return TypeTuple::MEMBAR;
duke@0 2639 }
duke@0 2640
duke@0 2641 //------------------------------match------------------------------------------
duke@0 2642 // Construct projections for memory.
duke@0 2643 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@0 2644 switch (proj->_con) {
duke@0 2645 case TypeFunc::Control:
duke@0 2646 case TypeFunc::Memory:
duke@0 2647 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@0 2648 }
duke@0 2649 ShouldNotReachHere();
duke@0 2650 return NULL;
duke@0 2651 }
duke@0 2652
duke@0 2653 //===========================InitializeNode====================================
duke@0 2654 // SUMMARY:
duke@0 2655 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@0 2656 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@0 2657 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@0 2658 // It can coalesce related raw stores into larger units (called 'tiles').
duke@0 2659 // It can avoid zeroing new storage for memory units which have raw inits.
duke@0 2660 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@0 2661 //
duke@0 2662 // EXAMPLE:
duke@0 2663 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@0 2664 // ctl = incoming control; mem* = incoming memory
duke@0 2665 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@0 2666 // First allocate uninitialized memory and fill in the header:
duke@0 2667 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@0 2668 // ctl := alloc.Control; mem* := alloc.Memory*
duke@0 2669 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@0 2670 // Then initialize to zero the non-header parts of the raw memory block:
duke@0 2671 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@0 2672 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@0 2673 // After the initialize node executes, the object is ready for service:
duke@0 2674 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@0 2675 // Suppose its body is immediately initialized as {1,2}:
duke@0 2676 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2677 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2678 // mem.SLICE(#short[*]) := store2
duke@0 2679 //
duke@0 2680 // DETAILS:
duke@0 2681 // An InitializeNode collects and isolates object initialization after
duke@0 2682 // an AllocateNode and before the next possible safepoint. As a
duke@0 2683 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@0 2684 // down past any safepoint or any publication of the allocation.
duke@0 2685 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@0 2686 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@0 2687 //
duke@0 2688 // The semantics of the InitializeNode include an implicit zeroing of
duke@0 2689 // the new object from object header to the end of the object.
duke@0 2690 // (The object header and end are determined by the AllocateNode.)
duke@0 2691 //
duke@0 2692 // Certain stores may be added as direct inputs to the InitializeNode.
duke@0 2693 // These stores must update raw memory, and they must be to addresses
duke@0 2694 // derived from the raw address produced by AllocateNode, and with
duke@0 2695 // a constant offset. They must be ordered by increasing offset.
duke@0 2696 // The first one is at in(RawStores), the last at in(req()-1).
duke@0 2697 // Unlike most memory operations, they are not linked in a chain,
duke@0 2698 // but are displayed in parallel as users of the rawmem output of
duke@0 2699 // the allocation.
duke@0 2700 //
duke@0 2701 // (See comments in InitializeNode::capture_store, which continue
duke@0 2702 // the example given above.)
duke@0 2703 //
duke@0 2704 // When the associated Allocate is macro-expanded, the InitializeNode
duke@0 2705 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@0 2706 // may also be created at that point to represent any required zeroing.
duke@0 2707 // The InitializeNode is then marked 'complete', prohibiting further
duke@0 2708 // capturing of nearby memory operations.
duke@0 2709 //
duke@0 2710 // During macro-expansion, all captured initializations which store
twisti@605 2711 // constant values of 32 bits or smaller are coalesced (if advantageous)
duke@0 2712 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@0 2713 // initialized in fewer memory operations. Memory words which are
duke@0 2714 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@0 2715 // by explicit stores of zero. (The code shape happens to do all
duke@0 2716 // zeroing first, then all other stores, with both sequences occurring
duke@0 2717 // in order of ascending offsets.)
duke@0 2718 //
duke@0 2719 // Alternatively, code may be inserted between an AllocateNode and its
duke@0 2720 // InitializeNode, to perform arbitrary initialization of the new object.
duke@0 2721 // E.g., the object copying intrinsics insert complex data transfers here.
duke@0 2722 // The initialization must then be marked as 'complete' disable the
duke@0 2723 // built-in zeroing semantics and the collection of initializing stores.
duke@0 2724 //
duke@0 2725 // While an InitializeNode is incomplete, reads from the memory state
duke@0 2726 // produced by it are optimizable if they match the control edge and
duke@0 2727 // new oop address associated with the allocation/initialization.
duke@0 2728 // They return a stored value (if the offset matches) or else zero.
duke@0 2729 // A write to the memory state, if it matches control and address,
duke@0 2730 // and if it is to a constant offset, may be 'captured' by the
duke@0 2731 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@0 2732 // inside the initialization, to the raw oop produced by the allocation.
duke@0 2733 // Operations on addresses which are provably distinct (e.g., to
duke@0 2734 // other AllocateNodes) are allowed to bypass the initialization.
duke@0 2735 //
duke@0 2736 // The effect of all this is to consolidate object initialization
duke@0 2737 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@0 2738 // single location, where it can be optimized as a unit.
duke@0 2739 //
duke@0 2740 // Only stores with an offset less than TrackedInitializationLimit words
duke@0 2741 // will be considered for capture by an InitializeNode. This puts a
duke@0 2742 // reasonable limit on the complexity of optimized initializations.
duke@0 2743
duke@0 2744 //---------------------------InitializeNode------------------------------------
duke@0 2745 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
duke@0 2746 : _is_complete(false),
duke@0 2747 MemBarNode(C, adr_type, rawoop)
duke@0 2748 {
duke@0 2749 init_class_id(Class_Initialize);
duke@0 2750
duke@0 2751 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@0 2752 assert(in(RawAddress) == rawoop, "proper init");
duke@0 2753 // Note: allocation() can be NULL, for secondary initialization barriers
duke@0 2754 }
duke@0 2755
duke@0 2756 // Since this node is not matched, it will be processed by the
duke@0 2757 // register allocator. Declare that there are no constraints
duke@0 2758 // on the allocation of the RawAddress edge.
duke@0 2759 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@0 2760 // This edge should be set to top, by the set_complete. But be conservative.
duke@0 2761 if (idx == InitializeNode::RawAddress)
duke@0 2762 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@0 2763 return RegMask::Empty;
duke@0 2764 }
duke@0 2765
duke@0 2766 Node* InitializeNode::memory(uint alias_idx) {
duke@0 2767 Node* mem = in(Memory);
duke@0 2768 if (mem->is_MergeMem()) {
duke@0 2769 return mem->as_MergeMem()->memory_at(alias_idx);
duke@0 2770 } else {
duke@0 2771 // incoming raw memory is not split
duke@0 2772 return mem;
duke@0 2773 }
duke@0 2774 }
duke@0 2775
duke@0 2776 bool InitializeNode::is_non_zero() {
duke@0 2777 if (is_complete()) return false;
duke@0 2778 remove_extra_zeroes();
duke@0 2779 return (req() > RawStores);
duke@0 2780 }
duke@0 2781
duke@0 2782 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@0 2783 assert(!is_complete(), "caller responsibility");
duke@0 2784 _is_complete = true;
duke@0 2785
duke@0 2786 // After this node is complete, it contains a bunch of
duke@0 2787 // raw-memory initializations. There is no need for
duke@0 2788 // it to have anything to do with non-raw memory effects.
duke@0 2789 // Therefore, tell all non-raw users to re-optimize themselves,
duke@0 2790 // after skipping the memory effects of this initialization.
duke@0 2791 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@0 2792 if (igvn) igvn->add_users_to_worklist(this);
duke@0 2793 }
duke@0 2794
duke@0 2795 // convenience function
duke@0 2796 // return false if the init contains any stores already
duke@0 2797 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@0 2798 InitializeNode* init = initialization();
duke@0 2799 if (init == NULL || init->is_complete()) return false;
duke@0 2800 init->remove_extra_zeroes();
duke@0 2801 // for now, if this allocation has already collected any inits, bail:
duke@0 2802 if (init->is_non_zero()) return false;
duke@0 2803 init->set_complete(phase);
duke@0 2804 return true;
duke@0 2805 }
duke@0 2806
duke@0 2807 void InitializeNode::remove_extra_zeroes() {
duke@0 2808 if (req() == RawStores) return;
duke@0 2809 Node* zmem = zero_memory();
duke@0 2810 uint fill = RawStores;
duke@0 2811 for (uint i = fill; i < req(); i++) {
duke@0 2812 Node* n = in(i);
duke@0 2813 if (n->is_top() || n == zmem) continue; // skip
duke@0 2814 if (fill < i) set_req(fill, n); // compact
duke@0 2815 ++fill;
duke@0 2816 }
duke@0 2817 // delete any empty spaces created:
duke@0 2818 while (fill < req()) {
duke@0 2819 del_req(fill);
duke@0 2820 }
duke@0 2821 }
duke@0 2822
duke@0 2823 // Helper for remembering which stores go with which offsets.
duke@0 2824 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@0 2825 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@0 2826 intptr_t offset = -1;
duke@0 2827 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@0 2828 phase, offset);
duke@0 2829 if (base == NULL) return -1; // something is dead,
duke@0 2830 if (offset < 0) return -1; // dead, dead
duke@0 2831 return offset;
duke@0 2832 }
duke@0 2833
duke@0 2834 // Helper for proving that an initialization expression is
duke@0 2835 // "simple enough" to be folded into an object initialization.
duke@0 2836 // Attempts to prove that a store's initial value 'n' can be captured
duke@0 2837 // within the initialization without creating a vicious cycle, such as:
duke@0 2838 // { Foo p = new Foo(); p.next = p; }
duke@0 2839 // True for constants and parameters and small combinations thereof.
duke@0 2840 bool InitializeNode::detect_init_independence(Node* n,
duke@0 2841 bool st_is_pinned,
duke@0 2842 int& count) {
duke@0 2843 if (n == NULL) return true; // (can this really happen?)
duke@0 2844 if (n->is_Proj()) n = n->in(0);
duke@0 2845 if (n == this) return false; // found a cycle
duke@0 2846 if (n->is_Con()) return true;
duke@0 2847 if (n->is_Start()) return true; // params, etc., are OK
duke@0 2848 if (n->is_Root()) return true; // even better
duke@0 2849
duke@0 2850 Node* ctl = n->in(0);
duke@0 2851 if (ctl != NULL && !ctl->is_top()) {
duke@0 2852 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@0 2853 if (ctl == this) return false;
duke@0 2854
duke@0 2855 // If we already know that the enclosing memory op is pinned right after
duke@0 2856 // the init, then any control flow that the store has picked up
duke@0 2857 // must have preceded the init, or else be equal to the init.
duke@0 2858 // Even after loop optimizations (which might change control edges)
duke@0 2859 // a store is never pinned *before* the availability of its inputs.
kvn@119 2860 if (!MemNode::all_controls_dominate(n, this))
duke@0 2861 return false; // failed to prove a good control
duke@0 2862
duke@0 2863 }
duke@0 2864
duke@0 2865 // Check data edges for possible dependencies on 'this'.
duke@0 2866 if ((count += 1) > 20) return false; // complexity limit
duke@0 2867 for (uint i = 1; i < n->req(); i++) {
duke@0 2868 Node* m = n->in(i);
duke@0 2869 if (m == NULL || m == n || m->is_top()) continue;
duke@0 2870 uint first_i = n->find_edge(m);
duke@0 2871 if (i != first_i) continue; // process duplicate edge just once
duke@0 2872 if (!detect_init_independence(m, st_is_pinned, count)) {
duke@0 2873 return false;
duke@0 2874 }
duke@0 2875 }
duke@0 2876
duke@0 2877 return true;
duke@0 2878 }
duke@0 2879
duke@0 2880 // Here are all the checks a Store must pass before it can be moved into
duke@0 2881 // an initialization. Returns zero if a check fails.
duke@0 2882 // On success, returns the (constant) offset to which the store applies,
duke@0 2883 // within the initialized memory.
duke@0 2884 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
duke@0 2885 const int FAIL = 0;
duke@0 2886 if (st->req() != MemNode::ValueIn + 1)
duke@0 2887 return FAIL; // an inscrutable StoreNode (card mark?)
duke@0 2888 Node* ctl = st->in(MemNode::Control);
duke@0 2889 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@0 2890 return FAIL; // must be unconditional after the initialization
duke@0 2891 Node* mem = st->in(MemNode::Memory);
duke@0 2892 if (!(mem->is_Proj() && mem->in(0) == this))
duke@0 2893 return FAIL; // must not be preceded by other stores
duke@0 2894 Node* adr = st->in(MemNode::Address);
duke@0 2895 intptr_t offset;
duke@0 2896 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@0 2897 if (alloc == NULL)
duke@0 2898 return FAIL; // inscrutable address
duke@0 2899 if (alloc != allocation())
duke@0 2900 return FAIL; // wrong allocation! (store needs to float up)
duke@0 2901 Node* val = st->in(MemNode::ValueIn);
duke@0 2902 int complexity_count = 0;
duke@0 2903 if (!detect_init_independence(val, true, complexity_count))
duke@0 2904 return FAIL; // stored value must be 'simple enough'
duke@0 2905
duke@0 2906 return offset; // success
duke@0 2907 }
duke@0 2908
duke@0 2909 // Find the captured store in(i) which corresponds to the range
duke@0 2910 // [start..start+size) in the initialized object.
duke@0 2911 // If there is one, return its index i. If there isn't, return the
duke@0 2912 // negative of the index where it should be inserted.
duke@0 2913 // Return 0 if the queried range overlaps an initialization boundary
duke@0 2914 // or if dead code is encountered.
duke@0 2915 // If size_in_bytes is zero, do not bother with overlap checks.
duke@0 2916 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@0 2917 int size_in_bytes,
duke@0 2918 PhaseTransform* phase) {
duke@0 2919 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@0 2920
duke@0 2921 if (is_complete())
duke@0 2922 return FAIL; // arraycopy got here first; punt
duke@0 2923
duke@0 2924 assert(allocation() != NULL, "must be present");
duke@0 2925
duke@0 2926 // no negatives, no header fields:
coleenp@113 2927 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@0 2928
duke@0 2929 // after a certain size, we bail out on tracking all the stores:
duke@0 2930 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2931 if (start >= ti_limit) return FAIL;
duke@0 2932
duke@0 2933 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@0 2934 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@0 2935
duke@0 2936 Node* st = in(i);
duke@0 2937 intptr_t st_off = get_store_offset(st, phase);
duke@0 2938 if (st_off < 0) {
duke@0 2939 if (st != zero_memory()) {
duke@0 2940 return FAIL; // bail out if there is dead garbage
duke@0 2941 }
duke@0 2942 } else if (st_off > start) {
duke@0 2943 // ...we are done, since stores are ordered
duke@0 2944 if (st_off < start + size_in_bytes) {
duke@0 2945 return FAIL; // the next store overlaps
duke@0 2946 }
duke@0 2947 return -(int)i; // not found; here is where to put it
duke@0 2948 } else if (st_off < start) {
duke@0 2949 if (size_in_bytes != 0 &&
duke@0 2950 start < st_off + MAX_STORE &&
duke@0 2951 start < st_off + st->as_Store()->memory_size()) {
duke@0 2952 return FAIL; // the previous store overlaps
duke@0 2953 }
duke@0 2954 } else {
duke@0 2955 if (size_in_bytes != 0 &&
duke@0 2956 st->as_Store()->memory_size() != size_in_bytes) {
duke@0 2957 return FAIL; // mismatched store size
duke@0 2958 }
duke@0 2959 return i;
duke@0 2960 }
duke@0 2961
duke@0 2962 ++i;
duke@0 2963 }
duke@0 2964 }
duke@0 2965
duke@0 2966 // Look for a captured store which initializes at the offset 'start'
duke@0 2967 // with the given size. If there is no such store, and no other
duke@0 2968 // initialization interferes, then return zero_memory (the memory
duke@0 2969 // projection of the AllocateNode).
duke@0 2970 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@0 2971 PhaseTransform* phase) {
duke@0 2972 assert(stores_are_sane(phase), "");
duke@0 2973 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2974 if (i == 0) {
duke@0 2975 return NULL; // something is dead
duke@0 2976 } else if (i < 0) {
duke@0 2977 return zero_memory(); // just primordial zero bits here
duke@0 2978 } else {
duke@0 2979 Node* st = in(i); // here is the store at this position
duke@0 2980 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@0 2981 return st;
duke@0 2982 }
duke@0 2983 }
duke@0 2984
duke@0 2985 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@0 2986 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@0 2987 PhaseTransform* phase) {
duke@0 2988 Node* addr = in(RawAddress);
duke@0 2989 if (offset != 0) {
duke@0 2990 Compile* C = phase->C;
duke@0 2991 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
duke@0 2992 phase->MakeConX(offset)) );
duke@0 2993 }
duke@0 2994 return addr;
duke@0 2995 }
duke@0 2996
duke@0 2997 // Clone the given store, converting it into a raw store
duke@0 2998 // initializing a field or element of my new object.
duke@0 2999 // Caller is responsible for retiring the original store,
duke@0 3000 // with subsume_node or the like.
duke@0 3001 //
duke@0 3002 // From the example above InitializeNode::InitializeNode,
duke@0 3003 // here are the old stores to be captured:
duke@0 3004 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 3005 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 3006 //
duke@0 3007 // Here is the changed code; note the extra edges on init:
duke@0 3008 // alloc = (Allocate ...)
duke@0 3009 // rawoop = alloc.RawAddress
duke@0 3010 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@0 3011 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@0 3012 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@0 3013 // rawstore1 rawstore2)
duke@0 3014 //
duke@0 3015 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
duke@0 3016 PhaseTransform* phase) {
duke@0 3017 assert(stores_are_sane(phase), "");
duke@0 3018
duke@0 3019 if (start < 0) return NULL;
duke@0 3020 assert(can_capture_store(st, phase) == start, "sanity");
duke@0 3021
duke@0 3022 Compile* C = phase->C;
duke@0 3023 int size_in_bytes = st->memory_size();
duke@0 3024 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 3025 if (i == 0) return NULL; // bail out
duke@0 3026 Node* prev_mem = NULL; // raw memory for the captured store
duke@0 3027 if (i > 0) {
duke@0 3028 prev_mem = in(i); // there is a pre-existing store under this one
duke@0 3029 set_req(i, C->top()); // temporarily disconnect it
duke@0 3030 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@0 3031 } else {
duke@0 3032 i = -i; // no pre-existing store
duke@0 3033 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@0 3034 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@0 3035 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@0 3036 else
duke@0 3037 ins_req(i, C->top()); // build a new edge
duke@0 3038 }
duke@0 3039 Node* new_st = st->clone();
duke@0 3040 new_st->set_req(MemNode::Control, in(Control));
duke@0 3041 new_st->set_req(MemNode::Memory, prev_mem);
duke@0 3042 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@0 3043 new_st = phase->transform(new_st);
duke@0 3044
duke@0 3045 // At this point, new_st might have swallowed a pre-existing store
duke@0 3046 // at the same offset, or perhaps new_st might have disappeared,
duke@0 3047 // if it redundantly stored the same value (or zero to fresh memory).
duke@0 3048
duke@0 3049 // In any case, wire it in:
duke@0 3050 set_req(i, new_st);
duke@0 3051
duke@0 3052 // The caller may now kill the old guy.
duke@0 3053 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@0 3054 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@0 3055 assert(!is_complete(), "");
duke@0 3056 return new_st;
duke@0 3057 }
duke@0 3058
duke@0 3059 static bool store_constant(jlong* tiles, int num_tiles,
duke@0 3060 intptr_t st_off, int st_size,
duke@0 3061 jlong con) {
duke@0 3062 if ((st_off & (st_size-1)) != 0)
duke@0 3063 return false; // strange store offset (assume size==2**N)
duke@0 3064 address addr = (address)tiles + st_off;
duke@0 3065 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@0 3066 switch (st_size) {
duke@0 3067 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@0 3068 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@0 3069 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@0 3070 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@0 3071 default: return false; // strange store size (detect size!=2**N here)
duke@0 3072 }
duke@0 3073 return true; // return success to caller
duke@0 3074 }
duke@0 3075
duke@0 3076 // Coalesce subword constants into int constants and possibly
duke@0 3077 // into long constants. The goal, if the CPU permits,
duke@0 3078 // is to initialize the object with a small number of 64-bit tiles.
duke@0 3079 // Also, convert floating-point constants to bit patterns.
duke@0 3080 // Non-constants are not relevant to this pass.
duke@0 3081 //
duke@0 3082 // In terms of the running example on InitializeNode::InitializeNode
duke@0 3083 // and InitializeNode::capture_store, here is the transformation
duke@0 3084 // of rawstore1 and rawstore2 into rawstore12:
duke@0 3085 // alloc = (Allocate ...)
duke@0 3086 // rawoop = alloc.RawAddress
duke@0 3087 // tile12 = 0x00010002
duke@0 3088 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@0 3089 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@0 3090 //
duke@0 3091 void
duke@0 3092 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@0 3093 Node* size_in_bytes,
duke@0 3094 PhaseGVN* phase) {
duke@0 3095 Compile* C = phase->C;
duke@0 3096
duke@0 3097 assert(stores_are_sane(phase), "");
duke@0 3098 // Note: After this pass, they are not completely sane,
duke@0 3099 // since there may be some overlaps.
duke@0 3100
duke@0 3101 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@0 3102
duke@0 3103 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 3104 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@0 3105 size_limit = MIN2(size_limit, ti_limit);
duke@0 3106 size_limit = align_size_up(size_limit, BytesPerLong);
duke@0 3107 int num_tiles = size_limit / BytesPerLong;
duke@0 3108
duke@0 3109 // allocate space for the tile map:
duke@0 3110 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@0 3111 jlong tiles_buf[small_len];
duke@0 3112 Node* nodes_buf[small_len];
duke@0 3113 jlong inits_buf[small_len];
duke@0 3114 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@0 3115 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 3116 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@0 3117 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@0 3118 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@0 3119 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 3120 // tiles: exact bitwise model of all primitive constants
duke@0 3121 // nodes: last constant-storing node subsumed into the tiles model
duke@0 3122 // inits: which bytes (in each tile) are touched by any initializations
duke@0 3123
duke@0 3124 //// Pass A: Fill in the tile model with any relevant stores.
duke@0 3125
duke@0 3126 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@0 3127 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@0 3128 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@0 3129 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3130 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3131 Node* st = in(i);
duke@0 3132 intptr_t st_off = get_store_offset(st, phase);
duke@0 3133
duke@0 3134 // Figure out the store's offset and constant value:
duke@0 3135 if (st_off < header_size) continue; //skip (ignore header)
duke@0 3136 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@0 3137 int st_size = st->as_Store()->memory_size();
duke@0 3138 if (st_off + st_size > size_limit) break;
duke@0 3139
duke@0 3140 // Record which bytes are touched, whether by constant or not.
duke@0 3141 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@0 3142 continue; // skip (strange store size)
duke@0 3143
duke@0 3144 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@0 3145 if (!val->singleton()) continue; //skip (non-con store)
duke@0 3146 BasicType type = val->basic_type();
duke@0 3147
duke@0 3148 jlong con = 0;
duke@0 3149 switch (type) {
duke@0 3150 case T_INT: con = val->is_int()->get_con(); break;
duke@0 3151 case T_LONG: con = val->is_long()->get_con(); break;
duke@0 3152 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@0 3153 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@0 3154 default: continue; //skip (odd store type)
duke@0 3155 }
duke@0 3156
duke@0 3157 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@0 3158 st->Opcode() == Op_StoreL) {
duke@0 3159 continue; // This StoreL is already optimal.
duke@0 3160 }
duke@0 3161
duke@0 3162 // Store down the constant.
duke@0 3163 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@0 3164
duke@0 3165 intptr_t j = st_off >> LogBytesPerLong;
duke@0 3166
duke@0 3167 if (type == T_INT && st_size == BytesPerInt
duke@0 3168 && (st_off & BytesPerInt) == BytesPerInt) {
duke@0 3169 jlong lcon = tiles[j];
duke@0 3170 if (!Matcher::isSimpleConstant64(lcon) &&
duke@0 3171 st->Opcode() == Op_StoreI) {
duke@0 3172 // This StoreI is already optimal by itself.
duke@0 3173 jint* intcon = (jint*) &tiles[j];
duke@0 3174 intcon[1] = 0; // undo the store_constant()
duke@0 3175
duke@0 3176 // If the previous store is also optimal by itself, back up and
duke@0 3177 // undo the action of the previous loop iteration... if we can.
duke@0 3178 // But if we can't, just let the previous half take care of itself.
duke@0 3179 st = nodes[j];
duke@0 3180 st_off -= BytesPerInt;
duke@0 3181 con = intcon[0];
duke@0 3182 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@0 3183 assert(st_off >= header_size, "still ignoring header");
duke@0 3184 assert(get_store_offset(st, phase) == st_off, "must be");
duke@0 3185 assert(in(i-1) == zmem, "must be");
duke@0 3186 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@0 3187 assert(con == tcon->is_int()->get_con(), "must be");
duke@0 3188 // Undo the effects of the previous loop trip, which swallowed st:
duke@0 3189 intcon[0] = 0; // undo store_constant()
duke@0 3190 set_req(i-1, st); // undo set_req(i, zmem)
duke@0 3191 nodes[j] = NULL; // undo nodes[j] = st
duke@0 3192 --old_subword; // undo ++old_subword
duke@0 3193 }
duke@0 3194 continue; // This StoreI is already optimal.
duke@0 3195 }
duke@0 3196 }
duke@0 3197
duke@0 3198 // This store is not needed.
duke@0 3199 set_req(i, zmem);
duke@0 3200 nodes[j] = st; // record for the moment
duke@0 3201 if (st_size < BytesPerLong) // something has changed
duke@0 3202 ++old_subword; // includes int/float, but who's counting...
duke@0 3203 else ++old_long;
duke@0 3204 }
duke@0 3205
duke@0 3206 if ((old_subword + old_long) == 0)
duke@0 3207 return; // nothing more to do
duke@0 3208
duke@0 3209 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@0 3210 // Be sure to insert them before overlapping non-constant stores.
duke@0 3211 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@0 3212 for (int j = 0; j < num_tiles; j++) {
duke@0 3213 jlong con = tiles[j];
duke@0 3214 jlong init = inits[j];
duke@0 3215 if (con == 0) continue;
duke@0 3216 jint con0, con1; // split the constant, address-wise
duke@0 3217 jint init0, init1; // split the init map, address-wise
duke@0 3218 { union { jlong con; jint intcon[2]; } u;
duke@0 3219 u.con = con;
duke@0 3220 con0 = u.intcon[0];
duke@0 3221 con1 = u.intcon[1];
duke@0 3222 u.con = init;
duke@0 3223 init0 = u.intcon[0];
duke@0 3224 init1 = u.intcon[1];
duke@0 3225 }
duke@0 3226
duke@0 3227 Node* old = nodes[j];
duke@0 3228 assert(old != NULL, "need the prior store");
duke@0 3229 intptr_t offset = (j * BytesPerLong);
duke@0 3230
duke@0 3231 bool split = !Matcher::isSimpleConstant64(con);
duke@0 3232
duke@0 3233 if (offset < header_size) {
duke@0 3234 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@0 3235 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@0 3236 split = true; // only the second word counts
duke@0 3237 // Example: int a[] = { 42 ... }
duke@0 3238 } else if (con0 == 0 && init0 == -1) {
duke@0 3239 split = true; // first word is covered by full inits
duke@0 3240 // Example: int a[] = { ... foo(), 42 ... }
duke@0 3241 } else if (con1 == 0 && init1 == -1) {
duke@0 3242 split = true; // second word is covered by full inits
duke@0 3243 // Example: int a[] = { ... 42, foo() ... }
duke@0 3244 }
duke@0 3245
duke@0 3246 // Here's a case where init0 is neither 0 nor -1:
duke@0 3247 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@0 3248 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@0 3249 // In this case the tile is not split; it is (jlong)42.
duke@0 3250 // The big tile is stored down, and then the foo() value is inserted.
duke@0 3251 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@0 3252
duke@0 3253 Node* ctl = old->in(MemNode::Control);
duke@0 3254 Node* adr = make_raw_address(offset, phase);
duke@0 3255 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 3256
duke@0 3257 // One or two coalesced stores to plop down.
duke@0 3258 Node* st[2];
duke@0 3259 intptr_t off[2];
duke@0 3260 int nst = 0;
duke@0 3261 if (!split) {
duke@0 3262 ++new_long;
duke@0 3263 off[nst] = offset;
coleenp@113 3264 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3265 phase->longcon(con), T_LONG);
duke@0 3266 } else {
duke@0 3267 // Omit either if it is a zero.
duke@0 3268 if (con0 != 0) {
duke@0 3269 ++new_int;
duke@0 3270 off[nst] = offset;
coleenp@113 3271 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3272 phase->intcon(con0), T_INT);
duke@0 3273 }
duke@0 3274 if (con1 != 0) {
duke@0 3275 ++new_int;
duke@0 3276 offset += BytesPerInt;
duke@0 3277 adr = make_raw_address(offset, phase);
duke@0 3278 off[nst] = offset;
coleenp@113 3279 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3280 phase->intcon(con1), T_INT);
duke@0 3281 }
duke@0 3282 }
duke@0 3283
duke@0 3284 // Insert second store first, then the first before the second.
duke@0 3285 // Insert each one just before any overlapping non-constant stores.
duke@0 3286 while (nst > 0) {
duke@0 3287 Node* st1 = st[--nst];
duke@0 3288 C->copy_node_notes_to(st1, old);
duke@0 3289 st1 = phase->transform(st1);
duke@0 3290 offset = off[nst];
duke@0 3291 assert(offset >= header_size, "do not smash header");
duke@0 3292 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@0 3293 guarantee(ins_idx != 0, "must re-insert constant store");
duke@0 3294 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@0 3295 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@0 3296 set_req(--ins_idx, st1);
duke@0 3297 else
duke@0 3298 ins_req(ins_idx, st1);
duke@0 3299 }
duke@0 3300 }
duke@0 3301
duke@0 3302 if (PrintCompilation && WizardMode)
duke@0 3303 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@0 3304 old_subword, old_long, new_int, new_long);
duke@0 3305 if (C->log() != NULL)
duke@0 3306 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@0 3307 old_subword, old_long, new_int, new_long);
duke@0 3308
duke@0 3309 // Clean up any remaining occurrences of zmem:
duke@0 3310 remove_extra_zeroes();
duke@0 3311 }
duke@0 3312
duke@0 3313 // Explore forward from in(start) to find the first fully initialized
duke@0 3314 // word, and return its offset. Skip groups of subword stores which
duke@0 3315 // together initialize full words. If in(start) is itself part of a
duke@0 3316 // fully initialized word, return the offset of in(start). If there
duke@0 3317 // are no following full-word stores, or if something is fishy, return
duke@0 3318 // a negative value.
duke@0 3319 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@0 3320 int int_map = 0;
duke@0 3321 intptr_t int_map_off = 0;
duke@0 3322 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@0 3323
duke@0 3324 for (uint i = start, limit = req(); i < limit; i++) {
duke@0 3325 Node* st = in(i);
duke@0 3326
duke@0 3327 intptr_t st_off = get_store_offset(st, phase);
duke@0 3328 if (st_off < 0) break; // return conservative answer
duke@0 3329
duke@0 3330 int st_size = st->as_Store()->memory_size();
duke@0 3331 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@0 3332 return st_off; // we found a complete word init
duke@0 3333 }
duke@0 3334
duke@0 3335 // update the map:
duke@0 3336
duke@0 3337 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@0 3338 if (this_int_off != int_map_off) {
duke@0 3339 // reset the map:
duke@0 3340 int_map = 0;
duke@0 3341 int_map_off = this_int_off;
duke@0 3342 }
duke@0 3343
duke@0 3344 int subword_off = st_off - this_int_off;
duke@0 3345 int_map |= right_n_bits(st_size) << subword_off;
duke@0 3346 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@0 3347 return this_int_off; // we found a complete word init
duke@0 3348 }
duke@0 3349
duke@0 3350 // Did this store hit or cross the word boundary?
duke@0 3351 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@0 3352 if (next_int_off == this_int_off + BytesPerInt) {
duke@0 3353 // We passed the current int, without fully initializing it.
duke@0 3354 int_map_off = next_int_off;
duke@0 3355 int_map >>= BytesPerInt;
duke@0 3356 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@0 3357 // We passed the current and next int.
duke@0 3358 return this_int_off + BytesPerInt;
duke@0 3359 }
duke@0 3360 }
duke@0 3361
duke@0 3362 return -1;
duke@0 3363 }
duke@0 3364
duke@0 3365
duke@0 3366 // Called when the associated AllocateNode is expanded into CFG.
duke@0 3367 // At this point, we may perform additional optimizations.
duke@0 3368 // Linearize the stores by ascending offset, to make memory
duke@0 3369 // activity as coherent as possible.
duke@0 3370 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@0 3371 intptr_t header_size,
duke@0 3372 Node* size_in_bytes,
duke@0 3373 PhaseGVN* phase) {
duke@0 3374 assert(!is_complete(), "not already complete");
duke@0 3375 assert(stores_are_sane(phase), "");
duke@0 3376 assert(allocation() != NULL, "must be present");
duke@0 3377
duke@0 3378 remove_extra_zeroes();
duke@0 3379
duke@0 3380 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@0 3381 // reduce instruction count for common initialization patterns
duke@0 3382 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@0 3383
duke@0 3384 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3385 Node* inits = zmem; // accumulating a linearized chain of inits
duke@0 3386 #ifdef ASSERT
coleenp@113 3387 intptr_t first_offset = allocation()->minimum_header_size();
coleenp@113 3388 intptr_t last_init_off = first_offset; // previous init offset
coleenp@113 3389 intptr_t last_init_end = first_offset; // previous init offset+size
coleenp@113 3390 intptr_t last_tile_end = first_offset; // previous tile offset+size
duke@0 3391 #endif
duke@0 3392 intptr_t zeroes_done = header_size;
duke@0 3393
duke@0 3394 bool do_zeroing = true; // we might give up if inits are very sparse
duke@0 3395 int big_init_gaps = 0; // how many large gaps have we seen?
duke@0 3396
duke@0 3397 if (ZeroTLAB) do_zeroing = false;
duke@0 3398 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@0 3399
duke@0 3400 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; 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)
duke@0 3404 break; // unknown junk in the inits
duke@0 3405 if (st->in(MemNode::Memory) != zmem)
duke@0 3406 break; // complicated store chains somehow in list
duke@0 3407
duke@0 3408 int st_size = st->as_Store()->memory_size();
duke@0 3409 intptr_t next_init_off = st_off + st_size;
duke@0 3410
duke@0 3411 if (do_zeroing && zeroes_done < next_init_off) {
duke@0 3412 // See if this store needs a zero before it or under it.
duke@0 3413 intptr_t zeroes_needed = st_off;
duke@0 3414
duke@0 3415 if (st_size < BytesPerInt) {
duke@0 3416 // Look for subword stores which only partially initialize words.
duke@0 3417 // If we find some, we must lay down some word-level zeroes first,
duke@0 3418 // underneath the subword stores.
duke@0 3419 //
duke@0 3420 // Examples:
duke@0 3421 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
duke@0 3422 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
duke@0 3423 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
duke@0 3424 //
duke@0 3425 // Note: coalesce_subword_stores may have already done this,
duke@0 3426 // if it was prompted by constant non-zero subword initializers.
duke@0 3427 // But this case can still arise with non-constant stores.
duke@0 3428
duke@0 3429 intptr_t next_full_store = find_next_fullword_store(i, phase);
duke@0 3430
duke@0 3431 // In the examples above:
duke@0 3432 // in(i) p q r s x y z
duke@0 3433 // st_off 12 13 14 15 12 13 14
duke@0 3434 // st_size 1 1 1 1 1 1 1
duke@0 3435 // next_full_s. 12 16 16 16 16 16 16
duke@0 3436 // z's_done 12 16 16 16 12 16 12
duke@0 3437 // z's_needed 12 16 16 16 16 16 16
duke@0 3438 // zsize 0 0 0 0 4 0 4
duke@0 3439 if (next_full_store < 0) {
duke@0 3440 // Conservative tack: Zero to end of current word.
duke@0 3441 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
duke@0 3442 } else {
duke@0 3443 // Zero to beginning of next fully initialized word.
duke@0 3444 // Or, don't zero at all, if we are already in that word.
duke@0 3445 assert(next_full_store >= zeroes_needed, "must go forward");
duke@0 3446 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
duke@0 3447 zeroes_needed = next_full_store;
duke@0 3448 }
duke@0 3449 }
duke@0 3450
duke@0 3451 if (zeroes_needed > zeroes_done) {
duke@0 3452 intptr_t zsize = zeroes_needed - zeroes_done;
duke@0 3453 // Do some incremental zeroing on rawmem, in parallel with inits.
duke@0 3454 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3455 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3456 zeroes_done, zeroes_needed,
duke@0 3457 phase);
duke@0 3458 zeroes_done = zeroes_needed;
duke@0 3459 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
duke@0 3460 do_zeroing = false; // leave the hole, next time
duke@0 3461 }
duke@0 3462 }
duke@0 3463
duke@0 3464 // Collect the store and move on:
duke@0 3465 st->set_req(MemNode::Memory, inits);
duke@0 3466 inits = st; // put it on the linearized chain
duke@0 3467 set_req(i, zmem); // unhook from previous position
duke@0 3468
duke@0 3469 if (zeroes_done == st_off)
duke@0 3470 zeroes_done = next_init_off;
duke@0 3471
duke@0 3472 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
duke@0 3473
duke@0 3474 #ifdef ASSERT
duke@0 3475 // Various order invariants. Weaker than stores_are_sane because
duke@0 3476 // a large constant tile can be filled in by smaller non-constant stores.
duke@0 3477 assert(st_off >= last_init_off, "inits do not reverse");
duke@0 3478 last_init_off = st_off;
duke@0 3479 const Type* val = NULL;
duke@0 3480 if (st_size >= BytesPerInt &&
duke@0 3481 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
duke@0 3482 (int)val->basic_type() < (int)T_OBJECT) {
duke@0 3483 assert(st_off >= last_tile_end, "tiles do not overlap");
duke@0 3484 assert(st_off >= last_init_end, "tiles do not overwrite inits");
duke@0 3485 last_tile_end = MAX2(last_tile_end, next_init_off);
duke@0 3486 } else {
duke@0 3487 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
duke@0 3488 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
duke@0 3489 assert(st_off >= last_init_end, "inits do not overlap");
duke@0 3490 last_init_end = next_init_off; // it's a non-tile
duke@0 3491 }
duke@0 3492 #endif //ASSERT
duke@0 3493 }
duke@0 3494
duke@0 3495 remove_extra_zeroes(); // clear out all the zmems left over
duke@0 3496 add_req(inits);
duke@0 3497
duke@0 3498 if (!ZeroTLAB) {
duke@0 3499 // If anything remains to be zeroed, zero it all now.
duke@0 3500 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3501 // if it is the last unused 4 bytes of an instance, forget about it
duke@0 3502 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
duke@0 3503 if (zeroes_done + BytesPerLong >= size_limit) {
duke@0 3504 assert(allocation() != NULL, "");
duke@0 3505 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
duke@0 3506 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
duke@0 3507 if (zeroes_done == k->layout_helper())
duke@0 3508 zeroes_done = size_limit;
duke@0 3509 }
duke@0 3510 if (zeroes_done < size_limit) {
duke@0 3511 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3512 zeroes_done, size_in_bytes, phase);
duke@0 3513 }
duke@0 3514 }
duke@0 3515
duke@0 3516 set_complete(phase);
duke@0 3517 return rawmem;
duke@0 3518 }
duke@0 3519
duke@0 3520
duke@0 3521 #ifdef ASSERT
duke@0 3522 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
duke@0 3523 if (is_complete())
duke@0 3524 return true; // stores could be anything at this point
coleenp@113 3525 assert(allocation() != NULL, "must be present");
coleenp@113 3526 intptr_t last_off = allocation()->minimum_header_size();
duke@0 3527 for (uint i = InitializeNode::RawStores; i < req(); i++) {
duke@0 3528 Node* st = in(i);
duke@0 3529 intptr_t st_off = get_store_offset(st, phase);
duke@0 3530 if (st_off < 0) continue; // ignore dead garbage
duke@0 3531 if (last_off > st_off) {
duke@0 3532 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
duke@0 3533 this->dump(2);
duke@0 3534 assert(false, "ascending store offsets");
duke@0 3535 return false;
duke@0 3536 }
duke@0 3537 last_off = st_off + st->as_Store()->memory_size();
duke@0