annotate src/share/vm/opto/memnode.cpp @ 2722:a92cdbac8b9e

7081933: Use zeroing elimination optimization for large array Summary: Don't zero new typeArray during runtime call if the allocation is followed by arraycopy into it. Reviewed-by: twisti
author kvn
date Mon, 26 Sep 2011 10:24:05 -0700
parents 11a4af030e4b
children 670a74b863fc
rev   line source
duke@0 1 /*
kvn@2167 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. 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 *
trims@1472 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1472 20 * or visit www.oracle.com if you need additional information or have any
trims@1472 21 * questions.
duke@0 22 *
duke@0 23 */
duke@0 24
stefank@1879 25 #include "precompiled.hpp"
stefank@1879 26 #include "classfile/systemDictionary.hpp"
stefank@1879 27 #include "compiler/compileLog.hpp"
stefank@1879 28 #include "memory/allocation.inline.hpp"
stefank@1879 29 #include "oops/objArrayKlass.hpp"
stefank@1879 30 #include "opto/addnode.hpp"
stefank@1879 31 #include "opto/cfgnode.hpp"
stefank@1879 32 #include "opto/compile.hpp"
stefank@1879 33 #include "opto/connode.hpp"
stefank@1879 34 #include "opto/loopnode.hpp"
stefank@1879 35 #include "opto/machnode.hpp"
stefank@1879 36 #include "opto/matcher.hpp"
stefank@1879 37 #include "opto/memnode.hpp"
stefank@1879 38 #include "opto/mulnode.hpp"
stefank@1879 39 #include "opto/phaseX.hpp"
stefank@1879 40 #include "opto/regmask.hpp"
stefank@1879 41
duke@0 42 // Portions of code courtesy of Clifford Click
duke@0 43
duke@0 44 // Optimization - Graph Style
duke@0 45
kvn@74 46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
kvn@74 47
duke@0 48 //=============================================================================
duke@0 49 uint MemNode::size_of() const { return sizeof(*this); }
duke@0 50
duke@0 51 const TypePtr *MemNode::adr_type() const {
duke@0 52 Node* adr = in(Address);
duke@0 53 const TypePtr* cross_check = NULL;
duke@0 54 DEBUG_ONLY(cross_check = _adr_type);
duke@0 55 return calculate_adr_type(adr->bottom_type(), cross_check);
duke@0 56 }
duke@0 57
duke@0 58 #ifndef PRODUCT
duke@0 59 void MemNode::dump_spec(outputStream *st) const {
duke@0 60 if (in(Address) == NULL) return; // node is dead
duke@0 61 #ifndef ASSERT
duke@0 62 // fake the missing field
duke@0 63 const TypePtr* _adr_type = NULL;
duke@0 64 if (in(Address) != NULL)
duke@0 65 _adr_type = in(Address)->bottom_type()->isa_ptr();
duke@0 66 #endif
duke@0 67 dump_adr_type(this, _adr_type, st);
duke@0 68
duke@0 69 Compile* C = Compile::current();
duke@0 70 if( C->alias_type(_adr_type)->is_volatile() )
duke@0 71 st->print(" Volatile!");
duke@0 72 }
duke@0 73
duke@0 74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
duke@0 75 st->print(" @");
duke@0 76 if (adr_type == NULL) {
duke@0 77 st->print("NULL");
duke@0 78 } else {
duke@0 79 adr_type->dump_on(st);
duke@0 80 Compile* C = Compile::current();
duke@0 81 Compile::AliasType* atp = NULL;
duke@0 82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
duke@0 83 if (atp == NULL)
duke@0 84 st->print(", idx=?\?;");
duke@0 85 else if (atp->index() == Compile::AliasIdxBot)
duke@0 86 st->print(", idx=Bot;");
duke@0 87 else if (atp->index() == Compile::AliasIdxTop)
duke@0 88 st->print(", idx=Top;");
duke@0 89 else if (atp->index() == Compile::AliasIdxRaw)
duke@0 90 st->print(", idx=Raw;");
duke@0 91 else {
duke@0 92 ciField* field = atp->field();
duke@0 93 if (field) {
duke@0 94 st->print(", name=");
duke@0 95 field->print_name_on(st);
duke@0 96 }
duke@0 97 st->print(", idx=%d;", atp->index());
duke@0 98 }
duke@0 99 }
duke@0 100 }
duke@0 101
duke@0 102 extern void print_alias_types();
duke@0 103
duke@0 104 #endif
duke@0 105
kvn@74 106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 107 const TypeOopPtr *tinst = t_adr->isa_oopptr();
kvn@223 108 if (tinst == NULL || !tinst->is_known_instance_field())
kvn@74 109 return mchain; // don't try to optimize non-instance types
kvn@74 110 uint instance_id = tinst->instance_id();
kvn@253 111 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
kvn@74 112 Node *prev = NULL;
kvn@74 113 Node *result = mchain;
kvn@74 114 while (prev != result) {
kvn@74 115 prev = result;
kvn@253 116 if (result == start_mem)
twisti@605 117 break; // hit one of our sentinels
kvn@74 118 // skip over a call which does not affect this memory slice
kvn@74 119 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
kvn@74 120 Node *proj_in = result->in(0);
kvn@253 121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
twisti@605 122 break; // hit one of our sentinels
kvn@253 123 } else if (proj_in->is_Call()) {
kvn@74 124 CallNode *call = proj_in->as_Call();
kvn@74 125 if (!call->may_modify(t_adr, phase)) {
kvn@74 126 result = call->in(TypeFunc::Memory);
kvn@74 127 }
kvn@74 128 } else if (proj_in->is_Initialize()) {
kvn@74 129 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
kvn@74 130 // Stop if this is the initialization for the object instance which
kvn@74 131 // which contains this memory slice, otherwise skip over it.
kvn@74 132 if (alloc != NULL && alloc->_idx != instance_id) {
kvn@74 133 result = proj_in->in(TypeFunc::Memory);
kvn@74 134 }
kvn@74 135 } else if (proj_in->is_MemBar()) {
kvn@74 136 result = proj_in->in(TypeFunc::Memory);
kvn@253 137 } else {
kvn@253 138 assert(false, "unexpected projection");
kvn@74 139 }
kvn@1100 140 } else if (result->is_ClearArray()) {
kvn@1100 141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) {
kvn@1100 142 // Can not bypass initialization of the instance
kvn@1100 143 // we are looking for.
kvn@1100 144 break;
kvn@1100 145 }
kvn@1100 146 // Otherwise skip it (the call updated 'result' value).
kvn@74 147 } else if (result->is_MergeMem()) {
kvn@74 148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
kvn@74 149 }
kvn@74 150 }
kvn@74 151 return result;
kvn@74 152 }
kvn@74 153
kvn@74 154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 155 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
kvn@223 156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
kvn@74 157 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@74 158 Node *result = mchain;
kvn@74 159 result = optimize_simple_memory_chain(result, t_adr, phase);
kvn@74 160 if (is_instance && igvn != NULL && result->is_Phi()) {
kvn@74 161 PhiNode *mphi = result->as_Phi();
kvn@74 162 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
kvn@74 163 const TypePtr *t = mphi->adr_type();
kvn@163 164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
kvn@223 165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
kvn@247 166 t->is_oopptr()->cast_to_exactness(true)
kvn@247 167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
kvn@247 168 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
kvn@74 169 // clone the Phi with our address type
kvn@74 170 result = mphi->split_out_instance(t_adr, igvn);
kvn@74 171 } else {
kvn@74 172 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
kvn@74 173 }
kvn@74 174 }
kvn@74 175 return result;
kvn@74 176 }
kvn@74 177
kvn@64 178 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
kvn@64 179 uint alias_idx = phase->C->get_alias_index(tp);
kvn@64 180 Node *mem = mmem;
kvn@64 181 #ifdef ASSERT
kvn@64 182 {
kvn@64 183 // Check that current type is consistent with the alias index used during graph construction
kvn@64 184 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
kvn@64 185 bool consistent = adr_check == NULL || adr_check->empty() ||
kvn@64 186 phase->C->must_alias(adr_check, alias_idx );
kvn@64 187 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
kvn@64 188 if( !consistent && adr_check != NULL && !adr_check->empty() &&
rasbold@169 189 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
kvn@64 190 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
kvn@64 191 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
kvn@64 192 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
kvn@64 193 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
kvn@64 194 // don't assert if it is dead code.
kvn@64 195 consistent = true;
kvn@64 196 }
kvn@64 197 if( !consistent ) {
kvn@64 198 st->print("alias_idx==%d, adr_check==", alias_idx);
kvn@64 199 if( adr_check == NULL ) {
kvn@64 200 st->print("NULL");
kvn@64 201 } else {
kvn@64 202 adr_check->dump();
kvn@64 203 }
kvn@64 204 st->cr();
kvn@64 205 print_alias_types();
kvn@64 206 assert(consistent, "adr_check must match alias idx");
kvn@64 207 }
kvn@64 208 }
kvn@64 209 #endif
never@1735 210 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
kvn@64 211 // means an array I have not precisely typed yet. Do not do any
kvn@64 212 // alias stuff with it any time soon.
never@1735 213 const TypeOopPtr *toop = tp->isa_oopptr();
kvn@64 214 if( tp->base() != Type::AnyPtr &&
never@1735 215 !(toop &&
never@1735 216 toop->klass() != NULL &&
never@1735 217 toop->klass()->is_java_lang_Object() &&
never@1735 218 toop->offset() == Type::OffsetBot) ) {
kvn@64 219 // compress paths and change unreachable cycles to TOP
kvn@64 220 // If not, we can update the input infinitely along a MergeMem cycle
kvn@64 221 // Equivalent code in PhiNode::Ideal
kvn@64 222 Node* m = phase->transform(mmem);
twisti@605 223 // If transformed to a MergeMem, get the desired slice
kvn@64 224 // Otherwise the returned node represents memory for every slice
kvn@64 225 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
kvn@64 226 // Update input if it is progress over what we have now
kvn@64 227 }
kvn@64 228 return mem;
kvn@64 229 }
kvn@64 230
duke@0 231 //--------------------------Ideal_common---------------------------------------
duke@0 232 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
duke@0 233 // Unhook non-raw memories from complete (macro-expanded) initializations.
duke@0 234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
duke@0 235 // If our control input is a dead region, kill all below the region
duke@0 236 Node *ctl = in(MemNode::Control);
duke@0 237 if (ctl && remove_dead_region(phase, can_reshape))
duke@0 238 return this;
kvn@305 239 ctl = in(MemNode::Control);
kvn@305 240 // Don't bother trying to transform a dead node
kvn@305 241 if( ctl && ctl->is_top() ) return NodeSentinel;
duke@0 242
kvn@708 243 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@708 244 // Wait if control on the worklist.
kvn@708 245 if (ctl && can_reshape && igvn != NULL) {
kvn@708 246 Node* bol = NULL;
kvn@708 247 Node* cmp = NULL;
kvn@708 248 if (ctl->in(0)->is_If()) {
kvn@708 249 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
kvn@708 250 bol = ctl->in(0)->in(1);
kvn@708 251 if (bol->is_Bool())
kvn@708 252 cmp = ctl->in(0)->in(1)->in(1);
kvn@708 253 }
kvn@708 254 if (igvn->_worklist.member(ctl) ||
kvn@708 255 (bol != NULL && igvn->_worklist.member(bol)) ||
kvn@708 256 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
kvn@708 257 // This control path may be dead.
kvn@708 258 // Delay this memory node transformation until the control is processed.
kvn@708 259 phase->is_IterGVN()->_worklist.push(this);
kvn@708 260 return NodeSentinel; // caller will return NULL
kvn@708 261 }
kvn@708 262 }
duke@0 263 // Ignore if memory is dead, or self-loop
duke@0 264 Node *mem = in(MemNode::Memory);
duke@0 265 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 266 assert( mem != this, "dead loop in MemNode::Ideal" );
duke@0 267
duke@0 268 Node *address = in(MemNode::Address);
duke@0 269 const Type *t_adr = phase->type( address );
duke@0 270 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 271
kvn@708 272 if( can_reshape && igvn != NULL &&
kvn@1746 273 (igvn->_worklist.member(address) ||
kvn@1746 274 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) {
kvn@420 275 // The address's base and type may change when the address is processed.
kvn@420 276 // Delay this mem node transformation until the address is processed.
kvn@420 277 phase->is_IterGVN()->_worklist.push(this);
kvn@420 278 return NodeSentinel; // caller will return NULL
kvn@420 279 }
kvn@420 280
kvn@1062 281 // Do NOT remove or optimize the next lines: ensure a new alias index
kvn@1062 282 // is allocated for an oop pointer type before Escape Analysis.
kvn@1062 283 // Note: C++ will not remove it since the call has side effect.
kvn@1062 284 if ( t_adr->isa_oopptr() ) {
kvn@1062 285 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
kvn@1062 286 }
kvn@1062 287
kvn@708 288 #ifdef ASSERT
kvn@708 289 Node* base = NULL;
kvn@708 290 if (address->is_AddP())
kvn@708 291 base = address->in(AddPNode::Base);
kvn@708 292 assert(base == NULL || t_adr->isa_rawptr() ||
kvn@708 293 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
kvn@708 294 #endif
kvn@708 295
duke@0 296 // Avoid independent memory operations
duke@0 297 Node* old_mem = mem;
duke@0 298
kvn@36 299 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@36 300 // initializations was removed. After macro-expansion all stores catched
kvn@36 301 // by Initialize node became raw stores and there is no information
kvn@36 302 // which memory slices they modify. So it is unsafe to move any memory
kvn@36 303 // operation above these stores. Also in most cases hooked non-raw memories
kvn@36 304 // were already unhooked by using information from detect_ptr_independence()
kvn@36 305 // and find_previous_store().
duke@0 306
duke@0 307 if (mem->is_MergeMem()) {
duke@0 308 MergeMemNode* mmem = mem->as_MergeMem();
duke@0 309 const TypePtr *tp = t_adr->is_ptr();
kvn@64 310
kvn@64 311 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@0 312 }
duke@0 313
duke@0 314 if (mem != old_mem) {
duke@0 315 set_req(MemNode::Memory, mem);
kvn@305 316 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
duke@0 317 return this;
duke@0 318 }
duke@0 319
duke@0 320 // let the subclass continue analyzing...
duke@0 321 return NULL;
duke@0 322 }
duke@0 323
duke@0 324 // Helper function for proving some simple control dominations.
kvn@119 325 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
duke@0 326 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@0 327 // is not a constant (dominated by the method's StartNode).
duke@0 328 // Used by MemNode::find_previous_store to prove that the
duke@0 329 // control input of a memory operation predates (dominates)
duke@0 330 // an allocation it wants to look past.
kvn@119 331 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
kvn@119 332 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
kvn@119 333 return false; // Conservative answer for dead code
kvn@119 334
kvn@193 335 // Check 'dom'. Skip Proj and CatchProj nodes.
kvn@119 336 dom = dom->find_exact_control(dom);
kvn@119 337 if (dom == NULL || dom->is_top())
kvn@119 338 return false; // Conservative answer for dead code
kvn@119 339
kvn@193 340 if (dom == sub) {
kvn@193 341 // For the case when, for example, 'sub' is Initialize and the original
kvn@193 342 // 'dom' is Proj node of the 'sub'.
kvn@193 343 return false;
kvn@193 344 }
kvn@193 345
kvn@155 346 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
kvn@119 347 return true;
kvn@119 348
kvn@119 349 // 'dom' dominates 'sub' if its control edge and control edges
kvn@119 350 // of all its inputs dominate or equal to sub's control edge.
kvn@119 351
kvn@119 352 // Currently 'sub' is either Allocate, Initialize or Start nodes.
kvn@163 353 // Or Region for the check in LoadNode::Ideal();
kvn@163 354 // 'sub' should have sub->in(0) != NULL.
kvn@163 355 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
kvn@163 356 sub->is_Region(), "expecting only these nodes");
kvn@119 357
kvn@119 358 // Get control edge of 'sub'.
kvn@193 359 Node* orig_sub = sub;
kvn@119 360 sub = sub->find_exact_control(sub->in(0));
kvn@119 361 if (sub == NULL || sub->is_top())
kvn@119 362 return false; // Conservative answer for dead code
kvn@119 363
kvn@119 364 assert(sub->is_CFG(), "expecting control");
kvn@119 365
kvn@119 366 if (sub == dom)
kvn@119 367 return true;
kvn@119 368
kvn@119 369 if (sub->is_Start() || sub->is_Root())
kvn@119 370 return false;
kvn@119 371
kvn@119 372 {
kvn@119 373 // Check all control edges of 'dom'.
kvn@119 374
kvn@119 375 ResourceMark rm;
kvn@119 376 Arena* arena = Thread::current()->resource_area();
kvn@119 377 Node_List nlist(arena);
kvn@119 378 Unique_Node_List dom_list(arena);
kvn@119 379
kvn@119 380 dom_list.push(dom);
kvn@119 381 bool only_dominating_controls = false;
kvn@119 382
kvn@119 383 for (uint next = 0; next < dom_list.size(); next++) {
kvn@119 384 Node* n = dom_list.at(next);
kvn@193 385 if (n == orig_sub)
kvn@193 386 return false; // One of dom's inputs dominated by sub.
kvn@119 387 if (!n->is_CFG() && n->pinned()) {
kvn@119 388 // Check only own control edge for pinned non-control nodes.
kvn@119 389 n = n->find_exact_control(n->in(0));
kvn@119 390 if (n == NULL || n->is_top())
kvn@119 391 return false; // Conservative answer for dead code
kvn@119 392 assert(n->is_CFG(), "expecting control");
kvn@193 393 dom_list.push(n);
kvn@193 394 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
kvn@119 395 only_dominating_controls = true;
kvn@119 396 } else if (n->is_CFG()) {
kvn@119 397 if (n->dominates(sub, nlist))
kvn@119 398 only_dominating_controls = true;
kvn@119 399 else
kvn@119 400 return false;
kvn@119 401 } else {
kvn@119 402 // First, own control edge.
kvn@119 403 Node* m = n->find_exact_control(n->in(0));
kvn@155 404 if (m != NULL) {
kvn@155 405 if (m->is_top())
kvn@155 406 return false; // Conservative answer for dead code
kvn@155 407 dom_list.push(m);
kvn@155 408 }
kvn@119 409 // Now, the rest of edges.
kvn@119 410 uint cnt = n->req();
kvn@119 411 for (uint i = 1; i < cnt; i++) {
kvn@119 412 m = n->find_exact_control(n->in(i));
kvn@119 413 if (m == NULL || m->is_top())
kvn@119 414 continue;
kvn@119 415 dom_list.push(m);
duke@0 416 }
duke@0 417 }
duke@0 418 }
kvn@119 419 return only_dominating_controls;
duke@0 420 }
duke@0 421 }
duke@0 422
duke@0 423 //---------------------detect_ptr_independence---------------------------------
duke@0 424 // Used by MemNode::find_previous_store to prove that two base
duke@0 425 // pointers are never equal.
duke@0 426 // The pointers are accompanied by their associated allocations,
duke@0 427 // if any, which have been previously discovered by the caller.
duke@0 428 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@0 429 Node* p2, AllocateNode* a2,
duke@0 430 PhaseTransform* phase) {
duke@0 431 // Attempt to prove that these two pointers cannot be aliased.
duke@0 432 // They may both manifestly be allocations, and they should differ.
duke@0 433 // Or, if they are not both allocations, they can be distinct constants.
duke@0 434 // Otherwise, one is an allocation and the other a pre-existing value.
duke@0 435 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@0 436 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@0 437 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@0 438 return (a1 != a2);
duke@0 439 } else if (a1 != NULL) { // one allocation a1
duke@0 440 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
kvn@119 441 return all_controls_dominate(p2, a1);
duke@0 442 } else { //(a2 != NULL) // one allocation a2
kvn@119 443 return all_controls_dominate(p1, a2);
duke@0 444 }
duke@0 445 return false;
duke@0 446 }
duke@0 447
duke@0 448
duke@0 449 // The logic for reordering loads and stores uses four steps:
duke@0 450 // (a) Walk carefully past stores and initializations which we
duke@0 451 // can prove are independent of this load.
duke@0 452 // (b) Observe that the next memory state makes an exact match
duke@0 453 // with self (load or store), and locate the relevant store.
duke@0 454 // (c) Ensure that, if we were to wire self directly to the store,
duke@0 455 // the optimizer would fold it up somehow.
duke@0 456 // (d) Do the rewiring, and return, depending on some other part of
duke@0 457 // the optimizer to fold up the load.
duke@0 458 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@0 459 // specific to loads and stores, so they are handled by the callers.
duke@0 460 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@0 461 //
duke@0 462 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@0 463 Node* ctrl = in(MemNode::Control);
duke@0 464 Node* adr = in(MemNode::Address);
duke@0 465 intptr_t offset = 0;
duke@0 466 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 467 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@0 468
duke@0 469 if (offset == Type::OffsetBot)
duke@0 470 return NULL; // cannot unalias unless there are precise offsets
duke@0 471
kvn@74 472 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
kvn@74 473
duke@0 474 intptr_t size_in_bytes = memory_size();
duke@0 475
duke@0 476 Node* mem = in(MemNode::Memory); // start searching here...
duke@0 477
duke@0 478 int cnt = 50; // Cycle limiter
duke@0 479 for (;;) { // While we can dance past unrelated stores...
duke@0 480 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@0 481
duke@0 482 if (mem->is_Store()) {
duke@0 483 Node* st_adr = mem->in(MemNode::Address);
duke@0 484 intptr_t st_offset = 0;
duke@0 485 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@0 486 if (st_base == NULL)
duke@0 487 break; // inscrutable pointer
duke@0 488 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@0 489 const int MAX_STORE = BytesPerLong;
duke@0 490 if (st_offset >= offset + size_in_bytes ||
duke@0 491 st_offset <= offset - MAX_STORE ||
duke@0 492 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@0 493 // Success: The offsets are provably independent.
duke@0 494 // (You may ask, why not just test st_offset != offset and be done?
duke@0 495 // The answer is that stores of different sizes can co-exist
duke@0 496 // in the same sequence of RawMem effects. We sometimes initialize
duke@0 497 // a whole 'tile' of array elements with a single jint or jlong.)
duke@0 498 mem = mem->in(MemNode::Memory);
duke@0 499 continue; // (a) advance through independent store memory
duke@0 500 }
duke@0 501 }
duke@0 502 if (st_base != base &&
duke@0 503 detect_ptr_independence(base, alloc,
duke@0 504 st_base,
duke@0 505 AllocateNode::Ideal_allocation(st_base, phase),
duke@0 506 phase)) {
duke@0 507 // Success: The bases are provably independent.
duke@0 508 mem = mem->in(MemNode::Memory);
duke@0 509 continue; // (a) advance through independent store memory
duke@0 510 }
duke@0 511
duke@0 512 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@0 513 // since we have not managed to prove 'this' and 'mem' independent.
duke@0 514 if (st_base == base && st_offset == offset) {
duke@0 515 return mem; // let caller handle steps (c), (d)
duke@0 516 }
duke@0 517
duke@0 518 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 519 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@0 520 AllocateNode* st_alloc = st_init->allocation();
duke@0 521 if (st_alloc == NULL)
duke@0 522 break; // something degenerated
duke@0 523 bool known_identical = false;
duke@0 524 bool known_independent = false;
duke@0 525 if (alloc == st_alloc)
duke@0 526 known_identical = true;
duke@0 527 else if (alloc != NULL)
duke@0 528 known_independent = true;
kvn@119 529 else if (all_controls_dominate(this, st_alloc))
duke@0 530 known_independent = true;
duke@0 531
duke@0 532 if (known_independent) {
duke@0 533 // The bases are provably independent: Either they are
duke@0 534 // manifestly distinct allocations, or else the control
duke@0 535 // of this load dominates the store's allocation.
duke@0 536 int alias_idx = phase->C->get_alias_index(adr_type());
duke@0 537 if (alias_idx == Compile::AliasIdxRaw) {
duke@0 538 mem = st_alloc->in(TypeFunc::Memory);
duke@0 539 } else {
duke@0 540 mem = st_init->memory(alias_idx);
duke@0 541 }
duke@0 542 continue; // (a) advance through independent store memory
duke@0 543 }
duke@0 544
duke@0 545 // (b) at this point, if we are not looking at a store initializing
duke@0 546 // the same allocation we are loading from, we lose.
duke@0 547 if (known_identical) {
duke@0 548 // From caller, can_see_stored_value will consult find_captured_store.
duke@0 549 return mem; // let caller handle steps (c), (d)
duke@0 550 }
duke@0 551
kvn@223 552 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
kvn@74 553 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
kvn@74 554 if (mem->is_Proj() && mem->in(0)->is_Call()) {
kvn@74 555 CallNode *call = mem->in(0)->as_Call();
kvn@74 556 if (!call->may_modify(addr_t, phase)) {
kvn@74 557 mem = call->in(TypeFunc::Memory);
kvn@74 558 continue; // (a) advance through independent call memory
kvn@74 559 }
kvn@74 560 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
kvn@74 561 mem = mem->in(0)->in(TypeFunc::Memory);
kvn@74 562 continue; // (a) advance through independent MemBar memory
kvn@1100 563 } else if (mem->is_ClearArray()) {
kvn@1100 564 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
kvn@1100 565 // (the call updated 'mem' value)
kvn@1100 566 continue; // (a) advance through independent allocation memory
kvn@1100 567 } else {
kvn@1100 568 // Can not bypass initialization of the instance
kvn@1100 569 // we are looking for.
kvn@1100 570 return mem;
kvn@1100 571 }
kvn@74 572 } else if (mem->is_MergeMem()) {
kvn@74 573 int alias_idx = phase->C->get_alias_index(adr_type());
kvn@74 574 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@74 575 continue; // (a) advance through independent MergeMem memory
kvn@74 576 }
duke@0 577 }
duke@0 578
duke@0 579 // Unless there is an explicit 'continue', we must bail out here,
duke@0 580 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@0 581 break;
duke@0 582 }
duke@0 583
duke@0 584 return NULL; // bail out
duke@0 585 }
duke@0 586
duke@0 587 //----------------------calculate_adr_type-------------------------------------
duke@0 588 // Helper function. Notices when the given type of address hits top or bottom.
duke@0 589 // Also, asserts a cross-check of the type against the expected address type.
duke@0 590 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@0 591 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@0 592 #ifdef PRODUCT
duke@0 593 cross_check = NULL;
duke@0 594 #else
duke@0 595 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@0 596 #endif
duke@0 597 const TypePtr* tp = t->isa_ptr();
duke@0 598 if (tp == NULL) {
duke@0 599 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@0 600 return TypePtr::BOTTOM; // touches lots of memory
duke@0 601 } else {
duke@0 602 #ifdef ASSERT
duke@0 603 // %%%% [phh] We don't check the alias index if cross_check is
duke@0 604 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@0 605 if (cross_check != NULL &&
duke@0 606 cross_check != TypePtr::BOTTOM &&
duke@0 607 cross_check != TypeRawPtr::BOTTOM) {
duke@0 608 // Recheck the alias index, to see if it has changed (due to a bug).
duke@0 609 Compile* C = Compile::current();
duke@0 610 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@0 611 "must stay in the original alias category");
duke@0 612 // The type of the address must be contained in the adr_type,
duke@0 613 // disregarding "null"-ness.
duke@0 614 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@0 615 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
duke@0 616 assert(cross_check->meet(tp_notnull) == cross_check,
duke@0 617 "real address must not escape from expected memory type");
duke@0 618 }
duke@0 619 #endif
duke@0 620 return tp;
duke@0 621 }
duke@0 622 }
duke@0 623
duke@0 624 //------------------------adr_phi_is_loop_invariant----------------------------
duke@0 625 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@0 626 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@0 627 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@0 628 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@0 629 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@0 630 // with the same data. This implies that any path into the loop already
duke@0 631 // includes such a CastPP, and so the original cast, whatever its input,
duke@0 632 // must be covered by an equivalent cast, with an earlier control input.
duke@0 633 ResourceMark rm;
duke@0 634
duke@0 635 // The loop entry input of the phi should be the unique dominating
duke@0 636 // node for every Phi/CastPP in the loop.
duke@0 637 Unique_Node_List closure;
duke@0 638 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@0 639
duke@0 640 // Add the phi node and the cast to the worklist.
duke@0 641 Unique_Node_List worklist;
duke@0 642 worklist.push(adr_phi);
duke@0 643 if( cast != NULL ){
duke@0 644 if( !cast->is_ConstraintCast() ) return false;
duke@0 645 worklist.push(cast);
duke@0 646 }
duke@0 647
duke@0 648 // Begin recursive walk of phi nodes.
duke@0 649 while( worklist.size() ){
duke@0 650 // Take a node off the worklist
duke@0 651 Node *n = worklist.pop();
duke@0 652 if( !closure.member(n) ){
duke@0 653 // Add it to the closure.
duke@0 654 closure.push(n);
duke@0 655 // Make a sanity check to ensure we don't waste too much time here.
duke@0 656 if( closure.size() > 20) return false;
duke@0 657 // This node is OK if:
duke@0 658 // - it is a cast of an identical value
duke@0 659 // - or it is a phi node (then we add its inputs to the worklist)
duke@0 660 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@0 661 if( n->is_ConstraintCast() ){
duke@0 662 worklist.push(n->in(1));
duke@0 663 } else if( n->is_Phi() ) {
duke@0 664 for( uint i = 1; i < n->req(); i++ ) {
duke@0 665 worklist.push(n->in(i));
duke@0 666 }
duke@0 667 } else {
duke@0 668 return false;
duke@0 669 }
duke@0 670 }
duke@0 671 }
duke@0 672
duke@0 673 // Quit when the worklist is empty, and we've found no offending nodes.
duke@0 674 return true;
duke@0 675 }
duke@0 676
duke@0 677 //------------------------------Ideal_DU_postCCP-------------------------------
duke@0 678 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@0 679 // going away in this pass and we need to make this memory op depend on the
duke@0 680 // gating null check.
kvn@163 681 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
kvn@163 682 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
kvn@163 683 }
duke@0 684
duke@0 685 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@0 686 // some sense; we get to keep around the knowledge that an oop is not-null
duke@0 687 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@0 688 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@0 689 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@0 690 // some of the more trivial cases in the optimizer. Removing more useless
duke@0 691 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@0 692 // up on this approach. CNC 10/20/2000
kvn@163 693 // This static method may be called not from MemNode (EncodePNode calls it).
kvn@163 694 // Only the control edge of the node 'n' might be updated.
kvn@163 695 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
duke@0 696 Node *skipped_cast = NULL;
duke@0 697 // Need a null check? Regular static accesses do not because they are
duke@0 698 // from constant addresses. Array ops are gated by the range check (which
duke@0 699 // always includes a NULL check). Just check field ops.
kvn@163 700 if( n->in(MemNode::Control) == NULL ) {
duke@0 701 // Scan upwards for the highest location we can place this memory op.
duke@0 702 while( true ) {
duke@0 703 switch( adr->Opcode() ) {
duke@0 704
duke@0 705 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@0 706 adr = adr->in(AddPNode::Base);
duke@0 707 continue;
duke@0 708
coleenp@113 709 case Op_DecodeN: // No change to NULL-ness, so peek thru
coleenp@113 710 adr = adr->in(1);
coleenp@113 711 continue;
coleenp@113 712
duke@0 713 case Op_CastPP:
duke@0 714 // If the CastPP is useless, just peek on through it.
duke@0 715 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@0 716 // Remember the cast that we've peeked though. If we peek
duke@0 717 // through more than one, then we end up remembering the highest
duke@0 718 // one, that is, if in a loop, the one closest to the top.
duke@0 719 skipped_cast = adr;
duke@0 720 adr = adr->in(1);
duke@0 721 continue;
duke@0 722 }
duke@0 723 // CastPP is going away in this pass! We need this memory op to be
duke@0 724 // control-dependent on the test that is guarding the CastPP.
kvn@163 725 ccp->hash_delete(n);
kvn@163 726 n->set_req(MemNode::Control, adr->in(0));
kvn@163 727 ccp->hash_insert(n);
kvn@163 728 return n;
duke@0 729
duke@0 730 case Op_Phi:
duke@0 731 // Attempt to float above a Phi to some dominating point.
duke@0 732 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@0 733 // If we've already peeked through a Cast (which could have set the
duke@0 734 // control), we can't float above a Phi, because the skipped Cast
duke@0 735 // may not be loop invariant.
duke@0 736 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@0 737 adr = adr->in(1);
duke@0 738 continue;
duke@0 739 }
duke@0 740 }
duke@0 741
duke@0 742 // Intentional fallthrough!
duke@0 743
duke@0 744 // No obvious dominating point. The mem op is pinned below the Phi
duke@0 745 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@0 746 // then the mem op can float, but not indefinitely. It must be pinned
duke@0 747 // behind the controls leading to the Phi.
duke@0 748 case Op_CheckCastPP:
duke@0 749 // These usually stick around to change address type, however a
duke@0 750 // useless one can be elided and we still need to pick up a control edge
duke@0 751 if (adr->in(0) == NULL) {
duke@0 752 // This CheckCastPP node has NO control and is likely useless. But we
duke@0 753 // need check further up the ancestor chain for a control input to keep
duke@0 754 // the node in place. 4959717.
duke@0 755 skipped_cast = adr;
duke@0 756 adr = adr->in(1);
duke@0 757 continue;
duke@0 758 }
kvn@163 759 ccp->hash_delete(n);
kvn@163 760 n->set_req(MemNode::Control, adr->in(0));
kvn@163 761 ccp->hash_insert(n);
kvn@163 762 return n;
duke@0 763
duke@0 764 // List of "safe" opcodes; those that implicitly block the memory
duke@0 765 // op below any null check.
duke@0 766 case Op_CastX2P: // no null checks on native pointers
duke@0 767 case Op_Parm: // 'this' pointer is not null
duke@0 768 case Op_LoadP: // Loading from within a klass
coleenp@113 769 case Op_LoadN: // Loading from within a klass
duke@0 770 case Op_LoadKlass: // Loading from within a klass
kvn@164 771 case Op_LoadNKlass: // Loading from within a klass
duke@0 772 case Op_ConP: // Loading from a klass
kvn@163 773 case Op_ConN: // Loading from a klass
duke@0 774 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@0 775 case Op_Con: // Reading from TLS
duke@0 776 case Op_CMoveP: // CMoveP is pinned
kvn@164 777 case Op_CMoveN: // CMoveN is pinned
duke@0 778 break; // No progress
duke@0 779
duke@0 780 case Op_Proj: // Direct call to an allocation routine
duke@0 781 case Op_SCMemProj: // Memory state from store conditional ops
duke@0 782 #ifdef ASSERT
duke@0 783 {
duke@0 784 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@0 785 const Node* call = adr->in(0);
kvn@163 786 if (call->is_CallJava()) {
kvn@163 787 const CallJavaNode* call_java = call->as_CallJava();
kvn@64 788 const TypeTuple *r = call_java->tf()->range();
kvn@64 789 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@64 790 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@64 791 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@0 792 // We further presume that this is one of
duke@0 793 // new_instance_Java, new_array_Java, or
duke@0 794 // the like, but do not assert for this.
duke@0 795 } else if (call->is_Allocate()) {
duke@0 796 // similar case to new_instance_Java, etc.
duke@0 797 } else if (!call->is_CallLeaf()) {
duke@0 798 // Projections from fetch_oop (OSR) are allowed as well.
duke@0 799 ShouldNotReachHere();
duke@0 800 }
duke@0 801 }
duke@0 802 #endif
duke@0 803 break;
duke@0 804 default:
duke@0 805 ShouldNotReachHere();
duke@0 806 }
duke@0 807 break;
duke@0 808 }
duke@0 809 }
duke@0 810
duke@0 811 return NULL; // No progress
duke@0 812 }
duke@0 813
duke@0 814
duke@0 815 //=============================================================================
duke@0 816 uint LoadNode::size_of() const { return sizeof(*this); }
duke@0 817 uint LoadNode::cmp( const Node &n ) const
duke@0 818 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@0 819 const Type *LoadNode::bottom_type() const { return _type; }
duke@0 820 uint LoadNode::ideal_reg() const {
duke@0 821 return Matcher::base2reg[_type->base()];
duke@0 822 }
duke@0 823
duke@0 824 #ifndef PRODUCT
duke@0 825 void LoadNode::dump_spec(outputStream *st) const {
duke@0 826 MemNode::dump_spec(st);
duke@0 827 if( !Verbose && !WizardMode ) {
duke@0 828 // standard dump does this in Verbose and WizardMode
duke@0 829 st->print(" #"); _type->dump_on(st);
duke@0 830 }
duke@0 831 }
duke@0 832 #endif
duke@0 833
kvn@1529 834 #ifdef ASSERT
kvn@1529 835 //----------------------------is_immutable_value-------------------------------
kvn@1529 836 // Helper function to allow a raw load without control edge for some cases
kvn@1529 837 bool LoadNode::is_immutable_value(Node* adr) {
kvn@1529 838 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
kvn@1529 839 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
kvn@1529 840 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
kvn@1529 841 in_bytes(JavaThread::osthread_offset())));
kvn@1529 842 }
kvn@1529 843 #endif
duke@0 844
duke@0 845 //----------------------------LoadNode::make-----------------------------------
duke@0 846 // Polymorphic factory method:
coleenp@113 847 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
coleenp@113 848 Compile* C = gvn.C;
coleenp@113 849
duke@0 850 // sanity check the alias category against the created node type
duke@0 851 assert(!(adr_type->isa_oopptr() &&
duke@0 852 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@0 853 "use LoadKlassNode instead");
duke@0 854 assert(!(adr_type->isa_aryptr() &&
duke@0 855 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@0 856 "use LoadRangeNode instead");
kvn@1529 857 // Check control edge of raw loads
kvn@1529 858 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
kvn@1529 859 // oop will be recorded in oop map if load crosses safepoint
kvn@1529 860 rt->isa_oopptr() || is_immutable_value(adr),
kvn@1529 861 "raw memory operations should have control edge");
duke@0 862 switch (bt) {
twisti@624 863 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 864 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 865 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 866 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 867 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 868 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
twisti@558 869 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt );
twisti@558 870 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt );
twisti@558 871 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
coleenp@113 872 case T_OBJECT:
coleenp@113 873 #ifdef _LP64
kvn@163 874 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
kvn@221 875 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
kvn@221 876 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
coleenp@113 877 } else
coleenp@113 878 #endif
kvn@163 879 {
kvn@163 880 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@163 881 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
kvn@163 882 }
duke@0 883 }
duke@0 884 ShouldNotReachHere();
duke@0 885 return (LoadNode*)NULL;
duke@0 886 }
duke@0 887
duke@0 888 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
duke@0 889 bool require_atomic = true;
duke@0 890 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
duke@0 891 }
duke@0 892
duke@0 893
duke@0 894
duke@0 895
duke@0 896 //------------------------------hash-------------------------------------------
duke@0 897 uint LoadNode::hash() const {
duke@0 898 // unroll addition of interesting fields
duke@0 899 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@0 900 }
duke@0 901
duke@0 902 //---------------------------can_see_stored_value------------------------------
duke@0 903 // This routine exists to make sure this set of tests is done the same
duke@0 904 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@0 905 // will change the graph shape in a way which makes memory alive twice at the
duke@0 906 // same time (uses the Oracle model of aliasing), then some
duke@0 907 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@0 908 // of aliasing.
duke@0 909 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@0 910 Node* ld_adr = in(MemNode::Address);
duke@0 911
never@17 912 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
never@17 913 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
never@17 914 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
never@17 915 atp->field() != NULL && !atp->field()->is_volatile()) {
never@17 916 uint alias_idx = atp->index();
never@17 917 bool final = atp->field()->is_final();
never@17 918 Node* result = NULL;
never@17 919 Node* current = st;
never@17 920 // Skip through chains of MemBarNodes checking the MergeMems for
never@17 921 // new states for the slice of this load. Stop once any other
never@17 922 // kind of node is encountered. Loads from final memory can skip
never@17 923 // through any kind of MemBar but normal loads shouldn't skip
never@17 924 // through MemBarAcquire since the could allow them to move out of
never@17 925 // a synchronized region.
never@17 926 while (current->is_Proj()) {
never@17 927 int opc = current->in(0)->Opcode();
roland@2612 928 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) ||
roland@2612 929 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder ||
roland@2612 930 opc == Op_MemBarReleaseLock) {
never@17 931 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@17 932 if (mem->is_MergeMem()) {
never@17 933 MergeMemNode* merge = mem->as_MergeMem();
never@17 934 Node* new_st = merge->memory_at(alias_idx);
never@17 935 if (new_st == merge->base_memory()) {
never@17 936 // Keep searching
never@17 937 current = merge->base_memory();
never@17 938 continue;
never@17 939 }
never@17 940 // Save the new memory state for the slice and fall through
never@17 941 // to exit.
never@17 942 result = new_st;
never@17 943 }
never@17 944 }
never@17 945 break;
never@17 946 }
never@17 947 if (result != NULL) {
never@17 948 st = result;
never@17 949 }
never@17 950 }
never@17 951
never@17 952
duke@0 953 // Loop around twice in the case Load -> Initialize -> Store.
duke@0 954 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@0 955 for (int trip = 0; trip <= 1; trip++) {
duke@0 956
duke@0 957 if (st->is_Store()) {
duke@0 958 Node* st_adr = st->in(MemNode::Address);
duke@0 959 if (!phase->eqv(st_adr, ld_adr)) {
duke@0 960 // Try harder before giving up... Match raw and non-raw pointers.
duke@0 961 intptr_t st_off = 0;
duke@0 962 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@0 963 if (alloc == NULL) return NULL;
duke@0 964 intptr_t ld_off = 0;
duke@0 965 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
duke@0 966 if (alloc != allo2) return NULL;
duke@0 967 if (ld_off != st_off) return NULL;
duke@0 968 // At this point we have proven something like this setup:
duke@0 969 // A = Allocate(...)
duke@0 970 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@0 971 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@0 972 // (Actually, we haven't yet proven the Q's are the same.)
duke@0 973 // In other words, we are loading from a casted version of
duke@0 974 // the same pointer-and-offset that we stored to.
duke@0 975 // Thus, we are able to replace L by V.
duke@0 976 }
duke@0 977 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@0 978 if (store_Opcode() != st->Opcode())
duke@0 979 return NULL;
duke@0 980 return st->in(MemNode::ValueIn);
duke@0 981 }
duke@0 982
duke@0 983 intptr_t offset = 0; // scratch
duke@0 984
duke@0 985 // A load from a freshly-created object always returns zero.
duke@0 986 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@0 987 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@0 988 if (st->is_Proj() && st->in(0)->is_Allocate() &&
duke@0 989 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
duke@0 990 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
duke@0 991 // return a zero value for the load's basic type
duke@0 992 // (This is one of the few places where a generic PhaseTransform
duke@0 993 // can create new nodes. Think of it as lazily manifesting
duke@0 994 // virtually pre-existing constants.)
duke@0 995 return phase->zerocon(memory_type());
duke@0 996 }
duke@0 997
duke@0 998 // A load from an initialization barrier can match a captured store.
duke@0 999 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@0 1000 InitializeNode* init = st->in(0)->as_Initialize();
duke@0 1001 AllocateNode* alloc = init->allocation();
duke@0 1002 if (alloc != NULL &&
duke@0 1003 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
duke@0 1004 // examine a captured store value
duke@0 1005 st = init->find_captured_store(offset, memory_size(), phase);
duke@0 1006 if (st != NULL)
duke@0 1007 continue; // take one more trip around
duke@0 1008 }
duke@0 1009 }
duke@0 1010
duke@0 1011 break;
duke@0 1012 }
duke@0 1013
duke@0 1014 return NULL;
duke@0 1015 }
duke@0 1016
kvn@64 1017 //----------------------is_instance_field_load_with_local_phi------------------
kvn@64 1018 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@64 1019 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
kvn@64 1020 in(MemNode::Address)->is_AddP() ) {
kvn@64 1021 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@64 1022 // Only instances.
kvn@223 1023 if( t_oop != NULL && t_oop->is_known_instance_field() &&
kvn@64 1024 t_oop->offset() != Type::OffsetBot &&
kvn@64 1025 t_oop->offset() != Type::OffsetTop) {
kvn@64 1026 return true;
kvn@64 1027 }
kvn@64 1028 }
kvn@64 1029 return false;
kvn@64 1030 }
kvn@64 1031
duke@0 1032 //------------------------------Identity---------------------------------------
duke@0 1033 // Loads are identity if previous store is to same address
duke@0 1034 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@0 1035 // If the previous store-maker is the right kind of Store, and the store is
duke@0 1036 // to the same address, then we are equal to the value stored.
duke@0 1037 Node* mem = in(MemNode::Memory);
duke@0 1038 Node* value = can_see_stored_value(mem, phase);
duke@0 1039 if( value ) {
duke@0 1040 // byte, short & char stores truncate naturally.
duke@0 1041 // A load has to load the truncated value which requires
duke@0 1042 // some sort of masking operation and that requires an
duke@0 1043 // Ideal call instead of an Identity call.
duke@0 1044 if (memory_size() < BytesPerInt) {
duke@0 1045 // If the input to the store does not fit with the load's result type,
duke@0 1046 // it must be truncated via an Ideal call.
duke@0 1047 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@0 1048 return this;
duke@0 1049 }
duke@0 1050 // (This works even when value is a Con, but LoadNode::Value
duke@0 1051 // usually runs first, producing the singleton type of the Con.)
duke@0 1052 return value;
duke@0 1053 }
kvn@64 1054
kvn@64 1055 // Search for an existing data phi which was generated before for the same
twisti@605 1056 // instance's field to avoid infinite generation of phis in a loop.
kvn@64 1057 Node *region = mem->in(0);
kvn@64 1058 if (is_instance_field_load_with_local_phi(region)) {
kvn@64 1059 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
kvn@64 1060 int this_index = phase->C->get_alias_index(addr_t);
kvn@64 1061 int this_offset = addr_t->offset();
kvn@64 1062 int this_id = addr_t->is_oopptr()->instance_id();
kvn@64 1063 const Type* this_type = bottom_type();
kvn@64 1064 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@64 1065 Node* phi = region->fast_out(i);
kvn@64 1066 if (phi->is_Phi() && phi != mem &&
kvn@64 1067 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
kvn@64 1068 return phi;
kvn@64 1069 }
kvn@64 1070 }
kvn@64 1071 }
kvn@64 1072
duke@0 1073 return this;
duke@0 1074 }
duke@0 1075
never@17 1076
never@17 1077 // Returns true if the AliasType refers to the field that holds the
never@17 1078 // cached box array. Currently only handles the IntegerCache case.
never@17 1079 static bool is_autobox_cache(Compile::AliasType* atp) {
never@17 1080 if (atp != NULL && atp->field() != NULL) {
never@17 1081 ciField* field = atp->field();
never@17 1082 ciSymbol* klass = field->holder()->name();
never@17 1083 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1084 field->holder()->uses_default_loader() &&
never@17 1085 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1086 return true;
never@17 1087 }
never@17 1088 }
never@17 1089 return false;
never@17 1090 }
never@17 1091
never@17 1092 // Fetch the base value in the autobox array
never@17 1093 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
never@17 1094 if (atp != NULL && atp->field() != NULL) {
never@17 1095 ciField* field = atp->field();
never@17 1096 ciSymbol* klass = field->holder()->name();
never@17 1097 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1098 field->holder()->uses_default_loader() &&
never@17 1099 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1100 assert(field->is_constant(), "what?");
never@17 1101 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
never@17 1102 // Fetch the box object at the base of the array and get its value
never@17 1103 ciInstance* box = array->obj_at(0)->as_instance();
never@17 1104 ciInstanceKlass* ik = box->klass()->as_instance_klass();
never@17 1105 if (ik->nof_nonstatic_fields() == 1) {
never@17 1106 // This should be true nonstatic_field_at requires calling
never@17 1107 // nof_nonstatic_fields so check it anyway
never@17 1108 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
never@17 1109 cache_offset = c.as_int();
never@17 1110 }
never@17 1111 return true;
never@17 1112 }
never@17 1113 }
never@17 1114 return false;
never@17 1115 }
never@17 1116
never@17 1117 // Returns true if the AliasType refers to the value field of an
never@17 1118 // autobox object. Currently only handles Integer.
never@17 1119 static bool is_autobox_object(Compile::AliasType* atp) {
never@17 1120 if (atp != NULL && atp->field() != NULL) {
never@17 1121 ciField* field = atp->field();
never@17 1122 ciSymbol* klass = field->holder()->name();
never@17 1123 if (field->name() == ciSymbol::value_name() &&
never@17 1124 field->holder()->uses_default_loader() &&
never@17 1125 klass == ciSymbol::java_lang_Integer()) {
never@17 1126 return true;
never@17 1127 }
never@17 1128 }
never@17 1129 return false;
never@17 1130 }
never@17 1131
never@17 1132
never@17 1133 // We're loading from an object which has autobox behaviour.
never@17 1134 // If this object is result of a valueOf call we'll have a phi
never@17 1135 // merging a newly allocated object and a load from the cache.
never@17 1136 // We want to replace this load with the original incoming
never@17 1137 // argument to the valueOf call.
never@17 1138 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
never@17 1139 Node* base = in(Address)->in(AddPNode::Base);
never@17 1140 if (base->is_Phi() && base->req() == 3) {
never@17 1141 AllocateNode* allocation = NULL;
never@17 1142 int allocation_index = -1;
never@17 1143 int load_index = -1;
never@17 1144 for (uint i = 1; i < base->req(); i++) {
never@17 1145 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
never@17 1146 if (allocation != NULL) {
never@17 1147 allocation_index = i;
never@17 1148 load_index = 3 - allocation_index;
never@17 1149 break;
never@17 1150 }
never@17 1151 }
kvn@583 1152 bool has_load = ( allocation != NULL &&
kvn@583 1153 (base->in(load_index)->is_Load() ||
kvn@583 1154 base->in(load_index)->is_DecodeN() &&
kvn@583 1155 base->in(load_index)->in(1)->is_Load()) );
kvn@583 1156 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
never@17 1157 // Push the loads from the phi that comes from valueOf up
never@17 1158 // through it to allow elimination of the loads and the recovery
never@17 1159 // of the original value.
never@17 1160 Node* mem_phi = in(Memory);
never@17 1161 Node* offset = in(Address)->in(AddPNode::Offset);
never@553 1162 Node* region = base->in(0);
never@17 1163
never@17 1164 Node* in1 = clone();
never@17 1165 Node* in1_addr = in1->in(Address)->clone();
never@17 1166 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
never@17 1167 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
never@17 1168 in1_addr->set_req(AddPNode::Offset, offset);
never@553 1169 in1->set_req(0, region->in(allocation_index));
never@17 1170 in1->set_req(Address, in1_addr);
never@17 1171 in1->set_req(Memory, mem_phi->in(allocation_index));
never@17 1172
never@17 1173 Node* in2 = clone();
never@17 1174 Node* in2_addr = in2->in(Address)->clone();
never@17 1175 in2_addr->set_req(AddPNode::Base, base->in(load_index));
never@17 1176 in2_addr->set_req(AddPNode::Address, base->in(load_index));
never@17 1177 in2_addr->set_req(AddPNode::Offset, offset);
never@553 1178 in2->set_req(0, region->in(load_index));
never@17 1179 in2->set_req(Address, in2_addr);
never@17 1180 in2->set_req(Memory, mem_phi->in(load_index));
never@17 1181
never@17 1182 in1_addr = phase->transform(in1_addr);
never@17 1183 in1 = phase->transform(in1);
never@17 1184 in2_addr = phase->transform(in2_addr);
never@17 1185 in2 = phase->transform(in2);
never@17 1186
never@553 1187 PhiNode* result = PhiNode::make_blank(region, this);
never@17 1188 result->set_req(allocation_index, in1);
never@17 1189 result->set_req(load_index, in2);
never@17 1190 return result;
never@17 1191 }
kvn@583 1192 } else if (base->is_Load() ||
kvn@583 1193 base->is_DecodeN() && base->in(1)->is_Load()) {
kvn@583 1194 if (base->is_DecodeN()) {
kvn@583 1195 // Get LoadN node which loads cached Integer object
kvn@583 1196 base = base->in(1);
kvn@583 1197 }
never@17 1198 // Eliminate the load of Integer.value for integers from the cache
never@17 1199 // array by deriving the value from the index into the array.
never@17 1200 // Capture the offset of the load and then reverse the computation.
never@17 1201 Node* load_base = base->in(Address)->in(AddPNode::Base);
kvn@583 1202 if (load_base->is_DecodeN()) {
kvn@583 1203 // Get LoadN node which loads IntegerCache.cache field
kvn@583 1204 load_base = load_base->in(1);
kvn@583 1205 }
never@17 1206 if (load_base != NULL) {
never@17 1207 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
never@17 1208 intptr_t cache_offset;
never@17 1209 int shift = -1;
never@17 1210 Node* cache = NULL;
never@17 1211 if (is_autobox_cache(atp)) {
kvn@29 1212 shift = exact_log2(type2aelembytes(T_OBJECT));
never@17 1213 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
never@17 1214 }
never@17 1215 if (cache != NULL && base->in(Address)->is_AddP()) {
never@17 1216 Node* elements[4];
never@17 1217 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@17 1218 int cache_low;
never@17 1219 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
never@17 1220 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
never@17 1221 // Add up all the offsets making of the address of the load
never@17 1222 Node* result = elements[0];
never@17 1223 for (int i = 1; i < count; i++) {
never@17 1224 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
never@17 1225 }
never@17 1226 // Remove the constant offset from the address and then
never@17 1227 // remove the scaling of the offset to recover the original index.
never@17 1228 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
never@17 1229 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
never@17 1230 // Peel the shift off directly but wrap it in a dummy node
never@17 1231 // since Ideal can't return existing nodes
never@17 1232 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
never@17 1233 } else {
never@17 1234 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
never@17 1235 }
never@17 1236 #ifdef _LP64
never@17 1237 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
never@17 1238 #endif
never@17 1239 return result;
never@17 1240 }
never@17 1241 }
never@17 1242 }
never@17 1243 }
never@17 1244 return NULL;
never@17 1245 }
never@17 1246
kvn@163 1247 //------------------------------split_through_phi------------------------------
kvn@163 1248 // Split instance field load through Phi.
kvn@163 1249 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
kvn@163 1250 Node* mem = in(MemNode::Memory);
kvn@163 1251 Node* address = in(MemNode::Address);
kvn@163 1252 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@163 1253 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@163 1254
kvn@163 1255 assert(mem->is_Phi() && (t_oop != NULL) &&
kvn@223 1256 t_oop->is_known_instance_field(), "invalide conditions");
kvn@163 1257
kvn@163 1258 Node *region = mem->in(0);
kvn@163 1259 if (region == NULL) {
kvn@163 1260 return NULL; // Wait stable graph
kvn@163 1261 }
kvn@163 1262 uint cnt = mem->req();
kvn@2375 1263 for (uint i = 1; i < cnt; i++) {
kvn@2375 1264 Node* rc = region->in(i);
kvn@2375 1265 if (rc == NULL || phase->type(rc) == Type::TOP)
kvn@2375 1266 return NULL; // Wait stable graph
kvn@163 1267 Node *in = mem->in(i);
kvn@2375 1268 if (in == NULL) {
kvn@163 1269 return NULL; // Wait stable graph
kvn@163 1270 }
kvn@163 1271 }
kvn@163 1272 // Check for loop invariant.
kvn@163 1273 if (cnt == 3) {
kvn@2375 1274 for (uint i = 1; i < cnt; i++) {
kvn@163 1275 Node *in = mem->in(i);
kvn@163 1276 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
kvn@163 1277 if (m == mem) {
kvn@163 1278 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
kvn@163 1279 return this;
kvn@163 1280 }
kvn@163 1281 }
kvn@163 1282 }
kvn@163 1283 // Split through Phi (see original code in loopopts.cpp).
kvn@163 1284 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
kvn@163 1285
kvn@163 1286 // Do nothing here if Identity will find a value
kvn@163 1287 // (to avoid infinite chain of value phis generation).
kvn@2375 1288 if (!phase->eqv(this, this->Identity(phase)))
kvn@163 1289 return NULL;
kvn@163 1290
kvn@163 1291 // Skip the split if the region dominates some control edge of the address.
kvn@2375 1292 if (!MemNode::all_controls_dominate(address, region))
kvn@163 1293 return NULL;
kvn@163 1294
kvn@163 1295 const Type* this_type = this->bottom_type();
kvn@163 1296 int this_index = phase->C->get_alias_index(addr_t);
kvn@163 1297 int this_offset = addr_t->offset();
kvn@163 1298 int this_iid = addr_t->is_oopptr()->instance_id();
kvn@163 1299 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@163 1300 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
kvn@2375 1301 for (uint i = 1; i < region->req(); i++) {
kvn@163 1302 Node *x;
kvn@163 1303 Node* the_clone = NULL;
kvn@2375 1304 if (region->in(i) == phase->C->top()) {
kvn@163 1305 x = phase->C->top(); // Dead path? Use a dead data op
kvn@163 1306 } else {
kvn@163 1307 x = this->clone(); // Else clone up the data op
kvn@163 1308 the_clone = x; // Remember for possible deletion.
kvn@163 1309 // Alter data node to use pre-phi inputs
kvn@2375 1310 if (this->in(0) == region) {
kvn@2375 1311 x->set_req(0, region->in(i));
kvn@163 1312 } else {
kvn@2375 1313 x->set_req(0, NULL);
kvn@163 1314 }
kvn@2375 1315 for (uint j = 1; j < this->req(); j++) {
kvn@163 1316 Node *in = this->in(j);
kvn@2375 1317 if (in->is_Phi() && in->in(0) == region)
kvn@2375 1318 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
kvn@163 1319 }
kvn@163 1320 }
kvn@163 1321 // Check for a 'win' on some paths
kvn@163 1322 const Type *t = x->Value(igvn);
kvn@163 1323
kvn@163 1324 bool singleton = t->singleton();
kvn@163 1325
kvn@163 1326 // See comments in PhaseIdealLoop::split_thru_phi().
kvn@2375 1327 if (singleton && t == Type::TOP) {
kvn@163 1328 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
kvn@163 1329 }
kvn@163 1330
kvn@2375 1331 if (singleton) {
kvn@163 1332 x = igvn->makecon(t);
kvn@163 1333 } else {
kvn@163 1334 // We now call Identity to try to simplify the cloned node.
kvn@163 1335 // Note that some Identity methods call phase->type(this).
kvn@163 1336 // Make sure that the type array is big enough for
kvn@163 1337 // our new node, even though we may throw the node away.
kvn@163 1338 // (This tweaking with igvn only works because x is a new node.)
kvn@163 1339 igvn->set_type(x, t);
kvn@293 1340 // If x is a TypeNode, capture any more-precise type permanently into Node
twisti@605 1341 // otherwise it will be not updated during igvn->transform since
kvn@293 1342 // igvn->type(x) is set to x->Value() already.
kvn@293 1343 x->raise_bottom_type(t);
kvn@163 1344 Node *y = x->Identity(igvn);
kvn@2375 1345 if (y != x) {
kvn@163 1346 x = y;
kvn@163 1347 } else {
kvn@163 1348 y = igvn->hash_find(x);
kvn@2375 1349 if (y) {
kvn@163 1350 x = y;
kvn@163 1351 } else {
kvn@163 1352 // Else x is a new node we are keeping
kvn@163 1353 // We do not need register_new_node_with_optimizer
kvn@163 1354 // because set_type has already been called.
kvn@163 1355 igvn->_worklist.push(x);
kvn@163 1356 }
kvn@163 1357 }
kvn@163 1358 }
kvn@163 1359 if (x != the_clone && the_clone != NULL)
kvn@163 1360 igvn->remove_dead_node(the_clone);
kvn@163 1361 phi->set_req(i, x);
kvn@163 1362 }
kvn@2375 1363 // Record Phi
kvn@2375 1364 igvn->register_new_node_with_optimizer(phi);
kvn@2375 1365 return phi;
kvn@163 1366 }
never@17 1367
duke@0 1368 //------------------------------Ideal------------------------------------------
duke@0 1369 // If the load is from Field memory and the pointer is non-null, we can
duke@0 1370 // zero out the control input.
duke@0 1371 // If the offset is constant and the base is an object allocation,
duke@0 1372 // try to hook me up to the exact initializing store.
duke@0 1373 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1374 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1375 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1376
duke@0 1377 Node* ctrl = in(MemNode::Control);
duke@0 1378 Node* address = in(MemNode::Address);
duke@0 1379
duke@0 1380 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@0 1381 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@0 1382 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@0 1383 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@0 1384 ctrl = ctrl->in(0);
duke@0 1385 set_req(MemNode::Control,ctrl);
duke@0 1386 }
duke@0 1387
kvn@708 1388 intptr_t ignore = 0;
kvn@708 1389 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
kvn@708 1390 if (base != NULL
kvn@708 1391 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
kvn@708 1392 // Check for useless control edge in some common special cases
kvn@708 1393 if (in(MemNode::Control) != NULL
duke@0 1394 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
kvn@119 1395 && all_controls_dominate(base, phase->C->start())) {
duke@0 1396 // A method-invariant, non-null address (constant or 'this' argument).
duke@0 1397 set_req(MemNode::Control, NULL);
duke@0 1398 }
kvn@708 1399
kvn@708 1400 if (EliminateAutoBox && can_reshape) {
kvn@708 1401 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
never@17 1402 Compile::AliasType* atp = phase->C->alias_type(adr_type());
never@17 1403 if (is_autobox_object(atp)) {
never@17 1404 Node* result = eliminate_autobox(phase);
never@17 1405 if (result != NULL) return result;
never@17 1406 }
never@17 1407 }
never@17 1408 }
never@17 1409
kvn@74 1410 Node* mem = in(MemNode::Memory);
kvn@74 1411 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@74 1412
kvn@74 1413 if (addr_t != NULL) {
kvn@74 1414 // try to optimize our memory input
kvn@74 1415 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
kvn@74 1416 if (opt_mem != mem) {
kvn@74 1417 set_req(MemNode::Memory, opt_mem);
kvn@305 1418 if (phase->type( opt_mem ) == Type::TOP) return NULL;
kvn@74 1419 return this;
kvn@74 1420 }
kvn@74 1421 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@74 1422 if (can_reshape && opt_mem->is_Phi() &&
kvn@223 1423 (t_oop != NULL) && t_oop->is_known_instance_field()) {
kvn@163 1424 // Split instance field load through Phi.
kvn@163 1425 Node* result = split_through_phi(phase);
kvn@163 1426 if (result != NULL) return result;
kvn@74 1427 }
kvn@74 1428 }
kvn@74 1429
duke@0 1430 // Check for prior store with a different base or offset; make Load
duke@0 1431 // independent. Skip through any number of them. Bail out if the stores
duke@0 1432 // are in an endless dead cycle and report no progress. This is a key
duke@0 1433 // transform for Reflection. However, if after skipping through the Stores
duke@0 1434 // we can't then fold up against a prior store do NOT do the transform as
duke@0 1435 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@0 1436 // array memory alive twice: once for the hoisted Load and again after the
duke@0 1437 // bypassed Store. This situation only works if EVERYBODY who does
duke@0 1438 // anti-dependence work knows how to bypass. I.e. we need all
duke@0 1439 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@0 1440 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@0 1441 // fold up, do so.
duke@0 1442 Node* prev_mem = find_previous_store(phase);
duke@0 1443 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1444 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@0 1445 // (c) See if we can fold up on the spot, but don't fold up here.
twisti@558 1446 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
duke@0 1447 // just return a prior value, which is done by Identity calls.
duke@0 1448 if (can_see_stored_value(prev_mem, phase)) {
duke@0 1449 // Make ready for step (d):
duke@0 1450 set_req(MemNode::Memory, prev_mem);
duke@0 1451 return this;
duke@0 1452 }
duke@0 1453 }
duke@0 1454
duke@0 1455 return NULL; // No further progress
duke@0 1456 }
duke@0 1457
duke@0 1458 // Helper to recognize certain Klass fields which are invariant across
duke@0 1459 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@0 1460 const Type*
duke@0 1461 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@0 1462 ciKlass* klass) const {
duke@0 1463 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1464 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@0 1465 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@0 1466 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@0 1467 return TypeInt::make(klass->modifier_flags());
duke@0 1468 }
duke@0 1469 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1470 // The field is Klass::_access_flags. Return its (constant) value.
duke@0 1471 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@0 1472 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@0 1473 return TypeInt::make(klass->access_flags());
duke@0 1474 }
duke@0 1475 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1476 // The field is Klass::_layout_helper. Return its constant value if known.
duke@0 1477 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1478 return TypeInt::make(klass->layout_helper());
duke@0 1479 }
duke@0 1480
duke@0 1481 // No match.
duke@0 1482 return NULL;
duke@0 1483 }
duke@0 1484
duke@0 1485 //------------------------------Value-----------------------------------------
duke@0 1486 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@0 1487 // Either input is TOP ==> the result is TOP
duke@0 1488 Node* mem = in(MemNode::Memory);
duke@0 1489 const Type *t1 = phase->type(mem);
duke@0 1490 if (t1 == Type::TOP) return Type::TOP;
duke@0 1491 Node* adr = in(MemNode::Address);
duke@0 1492 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@0 1493 if (tp == NULL || tp->empty()) return Type::TOP;
duke@0 1494 int off = tp->offset();
duke@0 1495 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
twisti@2667 1496 Compile* C = phase->C;
duke@0 1497
duke@0 1498 // Try to guess loaded type from pointer type
duke@0 1499 if (tp->base() == Type::AryPtr) {
duke@0 1500 const Type *t = tp->is_aryptr()->elem();
duke@0 1501 // Don't do this for integer types. There is only potential profit if
duke@0 1502 // the element type t is lower than _type; that is, for int types, if _type is
duke@0 1503 // more restrictive than t. This only happens here if one is short and the other
duke@0 1504 // char (both 16 bits), and in those cases we've made an intentional decision
duke@0 1505 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@0 1506 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@0 1507 //
duke@0 1508 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@0 1509 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@0 1510 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@0 1511 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@0 1512 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@0 1513 // In fact, that could have been the original type of p1, and p1 could have
duke@0 1514 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@0 1515 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@0 1516 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
kvn@293 1517 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
duke@0 1518 // t might actually be lower than _type, if _type is a unique
duke@0 1519 // concrete subclass of abstract class t.
duke@0 1520 // Make sure the reference is not into the header, by comparing
duke@0 1521 // the offset against the offset of the start of the array's data.
duke@0 1522 // Different array types begin at slightly different offsets (12 vs. 16).
duke@0 1523 // We choose T_BYTE as an example base type that is least restrictive
duke@0 1524 // as to alignment, which will therefore produce the smallest
duke@0 1525 // possible base offset.
duke@0 1526 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1527 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
duke@0 1528 const Type* jt = t->join(_type);
duke@0 1529 // In any case, do not allow the join, per se, to empty out the type.
duke@0 1530 if (jt->empty() && !t->empty()) {
duke@0 1531 // This can happen if a interface-typed array narrows to a class type.
duke@0 1532 jt = _type;
duke@0 1533 }
never@17 1534
kvn@708 1535 if (EliminateAutoBox && adr->is_AddP()) {
never@17 1536 // The pointers in the autobox arrays are always non-null
kvn@708 1537 Node* base = adr->in(AddPNode::Base);
kvn@708 1538 if (base != NULL &&
kvn@708 1539 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
twisti@2667 1540 Compile::AliasType* atp = C->alias_type(base->adr_type());
never@17 1541 if (is_autobox_cache(atp)) {
never@17 1542 return jt->join(TypePtr::NOTNULL)->is_ptr();
never@17 1543 }
never@17 1544 }
never@17 1545 }
duke@0 1546 return jt;
duke@0 1547 }
duke@0 1548 }
duke@0 1549 } else if (tp->base() == Type::InstPtr) {
twisti@2667 1550 ciEnv* env = C->env();
never@1080 1551 const TypeInstPtr* tinst = tp->is_instptr();
never@1080 1552 ciKlass* klass = tinst->klass();
duke@0 1553 assert( off != Type::OffsetBot ||
duke@0 1554 // arrays can be cast to Objects
duke@0 1555 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@0 1556 // unsafe field access may not have a constant offset
twisti@2667 1557 C->has_unsafe_access(),
duke@0 1558 "Field accesses must be precise" );
duke@0 1559 // For oop loads, we expect the _type to be precise
twisti@2667 1560 if (klass == env->String_klass() &&
never@1080 1561 adr->is_AddP() && off != Type::OffsetBot) {
kvn@2167 1562 // For constant Strings treat the final fields as compile time constants.
never@1080 1563 Node* base = adr->in(AddPNode::Base);
never@1686 1564 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
never@1686 1565 if (t != NULL && t->singleton()) {
twisti@2667 1566 ciField* field = env->String_klass()->get_field_by_offset(off, false);
kvn@2167 1567 if (field != NULL && field->is_final()) {
kvn@2167 1568 ciObject* string = t->const_oop();
kvn@2167 1569 ciConstant constant = string->as_instance()->field_value(field);
kvn@2167 1570 if (constant.basic_type() == T_INT) {
kvn@2167 1571 return TypeInt::make(constant.as_int());
kvn@2167 1572 } else if (constant.basic_type() == T_ARRAY) {
kvn@2167 1573 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
jcoomes@2226 1574 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
kvn@2167 1575 } else {
jcoomes@2226 1576 return TypeOopPtr::make_from_constant(constant.as_object(), true);
kvn@2167 1577 }
never@1080 1578 }
never@1080 1579 }
never@1080 1580 }
never@1080 1581 }
twisti@2667 1582 // Optimizations for constant objects
twisti@2667 1583 ciObject* const_oop = tinst->const_oop();
twisti@2667 1584 if (const_oop != NULL) {
twisti@2667 1585 // For constant CallSites treat the target field as a compile time constant.
twisti@2667 1586 if (const_oop->is_call_site()) {
twisti@2667 1587 ciCallSite* call_site = const_oop->as_call_site();
twisti@2667 1588 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
twisti@2667 1589 if (field != NULL && field->is_call_site_target()) {
twisti@2667 1590 ciMethodHandle* target = call_site->get_target();
twisti@2667 1591 if (target != NULL) { // just in case
twisti@2667 1592 ciConstant constant(T_OBJECT, target);
twisti@2667 1593 const Type* t;
twisti@2667 1594 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
twisti@2667 1595 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
twisti@2667 1596 } else {
twisti@2667 1597 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
twisti@2667 1598 }
twisti@2667 1599 // Add a dependence for invalidation of the optimization.
twisti@2667 1600 if (!call_site->is_constant_call_site()) {
twisti@2667 1601 C->dependencies()->assert_call_site_target_value(call_site, target);
twisti@2667 1602 }
twisti@2667 1603 return t;
twisti@2667 1604 }
twisti@2667 1605 }
twisti@2667 1606 }
twisti@2667 1607 }
duke@0 1608 } else if (tp->base() == Type::KlassPtr) {
duke@0 1609 assert( off != Type::OffsetBot ||
duke@0 1610 // arrays can be cast to Objects
duke@0 1611 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@0 1612 // also allow array-loading from the primary supertype
duke@0 1613 // array during subtype checks
duke@0 1614 Opcode() == Op_LoadKlass,
duke@0 1615 "Field accesses must be precise" );
duke@0 1616 // For klass/static loads, we expect the _type to be precise
duke@0 1617 }
duke@0 1618
duke@0 1619 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1620 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1621 ciKlass* klass = tkls->klass();
duke@0 1622 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@0 1623 // We are loading a field from a Klass metaobject whose identity
duke@0 1624 // is known at compile time (the type is "exact" or "precise").
duke@0 1625 // Check for fields we know are maintained as constants by the VM.
duke@0 1626 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1627 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@0 1628 // (Folds up type checking code.)
duke@0 1629 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@0 1630 return TypeInt::make(klass->super_check_offset());
duke@0 1631 }
duke@0 1632 // Compute index into primary_supers array
duke@0 1633 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1634 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1635 if( depth < ciKlass::primary_super_limit() ) {
duke@0 1636 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1637 // (Folds up type checking code.)
duke@0 1638 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1639 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1640 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1641 }
duke@0 1642 const Type* aift = load_array_final_field(tkls, klass);
duke@0 1643 if (aift != NULL) return aift;
duke@0 1644 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
duke@0 1645 && klass->is_array_klass()) {
duke@0 1646 // The field is arrayKlass::_component_mirror. Return its (constant) value.
duke@0 1647 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@0 1648 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@0 1649 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@0 1650 }
duke@0 1651 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1652 // The field is Klass::_java_mirror. Return its (constant) value.
duke@0 1653 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@0 1654 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@0 1655 return TypeInstPtr::make(klass->java_mirror());
duke@0 1656 }
duke@0 1657 }
duke@0 1658
duke@0 1659 // We can still check if we are loading from the primary_supers array at a
duke@0 1660 // shallow enough depth. Even though the klass is not exact, entries less
duke@0 1661 // than or equal to its super depth are correct.
duke@0 1662 if (klass->is_loaded() ) {
duke@0 1663 ciType *inner = klass->klass();
duke@0 1664 while( inner->is_obj_array_klass() )
duke@0 1665 inner = inner->as_obj_array_klass()->base_element_type();
duke@0 1666 if( inner->is_instance_klass() &&
duke@0 1667 !inner->as_instance_klass()->flags().is_interface() ) {
duke@0 1668 // Compute index into primary_supers array
duke@0 1669 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1670 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1671 if( depth < ciKlass::primary_super_limit() &&
duke@0 1672 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@0 1673 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1674 // (Folds up type checking code.)
duke@0 1675 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1676 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1677 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1678 }
duke@0 1679 }
duke@0 1680 }
duke@0 1681
duke@0 1682 // If the type is enough to determine that the thing is not an array,
duke@0 1683 // we can give the layout_helper a positive interval type.
duke@0 1684 // This will help short-circuit some reflective code.
duke@0 1685 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
duke@0 1686 && !klass->is_array_klass() // not directly typed as an array
duke@0 1687 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@0 1688 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@0 1689 ) {
duke@0 1690 // Note: When interfaces are reliable, we can narrow the interface
duke@0 1691 // test to (klass != Serializable && klass != Cloneable).
duke@0 1692 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1693 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@0 1694 // The key property of this type is that it folds up tests
duke@0 1695 // for array-ness, since it proves that the layout_helper is positive.
duke@0 1696 // Thus, a generic value like the basic object layout helper works fine.
duke@0 1697 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@0 1698 }
duke@0 1699 }
duke@0 1700
duke@0 1701 // If we are loading from a freshly-allocated object, produce a zero,
duke@0 1702 // if the load is provably beyond the header of the object.
duke@0 1703 // (Also allow a variable load from a fresh array to produce zero.)
kvn@2375 1704 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@2375 1705 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
kvn@2375 1706 if (ReduceFieldZeroing || is_instance) {
duke@0 1707 Node* value = can_see_stored_value(mem,phase);
duke@0 1708 if (value != NULL && value->is_Con())
duke@0 1709 return value->bottom_type();
duke@0 1710 }
duke@0 1711
kvn@2375 1712 if (is_instance) {
kvn@64 1713 // If we have an instance type and our memory input is the
kvn@64 1714 // programs's initial memory state, there is no matching store,
kvn@64 1715 // so just return a zero of the appropriate type
kvn@64 1716 Node *mem = in(MemNode::Memory);
kvn@64 1717 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@64 1718 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@64 1719 return Type::get_zero_type(_type->basic_type());
kvn@64 1720 }
kvn@64 1721 }
duke@0 1722 return _type;
duke@0 1723 }
duke@0 1724
duke@0 1725 //------------------------------match_edge-------------------------------------
duke@0 1726 // Do we Match on this edge index or not? Match only the address.
duke@0 1727 uint LoadNode::match_edge(uint idx) const {
duke@0 1728 return idx == MemNode::Address;
duke@0 1729 }
duke@0 1730
duke@0 1731 //--------------------------LoadBNode::Ideal--------------------------------------
duke@0 1732 //
duke@0 1733 // If the previous store is to the same address as this load,
duke@0 1734 // and the value stored was larger than a byte, replace this load
duke@0 1735 // with the value stored truncated to a byte. If no truncation is
duke@0 1736 // needed, the replacement is done in LoadNode::Identity().
duke@0 1737 //
duke@0 1738 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1739 Node* mem = in(MemNode::Memory);
duke@0 1740 Node* value = can_see_stored_value(mem,phase);
duke@0 1741 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1742 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
duke@0 1743 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
duke@0 1744 }
duke@0 1745 // Identity call will handle the case where truncation is not needed.
duke@0 1746 return LoadNode::Ideal(phase, can_reshape);
duke@0 1747 }
duke@0 1748
twisti@624 1749 //--------------------------LoadUBNode::Ideal-------------------------------------
twisti@624 1750 //
twisti@624 1751 // If the previous store is to the same address as this load,
twisti@624 1752 // and the value stored was larger than a byte, replace this load
twisti@624 1753 // with the value stored truncated to a byte. If no truncation is
twisti@624 1754 // needed, the replacement is done in LoadNode::Identity().
twisti@624 1755 //
twisti@624 1756 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
twisti@624 1757 Node* mem = in(MemNode::Memory);
twisti@624 1758 Node* value = can_see_stored_value(mem, phase);
twisti@624 1759 if (value && !phase->type(value)->higher_equal(_type))
twisti@624 1760 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF));
twisti@624 1761 // Identity call will handle the case where truncation is not needed.
twisti@624 1762 return LoadNode::Ideal(phase, can_reshape);
twisti@624 1763 }
twisti@624 1764
twisti@558 1765 //--------------------------LoadUSNode::Ideal-------------------------------------
duke@0 1766 //
duke@0 1767 // If the previous store is to the same address as this load,
duke@0 1768 // and the value stored was larger than a char, replace this load
duke@0 1769 // with the value stored truncated to a char. If no truncation is
duke@0 1770 // needed, the replacement is done in LoadNode::Identity().
duke@0 1771 //
twisti@558 1772 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1773 Node* mem = in(MemNode::Memory);
duke@0 1774 Node* value = can_see_stored_value(mem,phase);
duke@0 1775 if( value && !phase->type(value)->higher_equal( _type ) )
duke@0 1776 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
duke@0 1777 // Identity call will handle the case where truncation is not needed.
duke@0 1778 return LoadNode::Ideal(phase, can_reshape);
duke@0 1779 }
duke@0 1780
duke@0 1781 //--------------------------LoadSNode::Ideal--------------------------------------
duke@0 1782 //
duke@0 1783 // If the previous store is to the same address as this load,
duke@0 1784 // and the value stored was larger than a short, replace this load
duke@0 1785 // with the value stored truncated to a short. If no truncation is
duke@0 1786 // needed, the replacement is done in LoadNode::Identity().
duke@0 1787 //
duke@0 1788 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1789 Node* mem = in(MemNode::Memory);
duke@0 1790 Node* value = can_see_stored_value(mem,phase);
duke@0 1791 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1792 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
duke@0 1793 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
duke@0 1794 }
duke@0 1795 // Identity call will handle the case where truncation is not needed.
duke@0 1796 return LoadNode::Ideal(phase, can_reshape);
duke@0 1797 }
duke@0 1798
duke@0 1799 //=============================================================================
kvn@164 1800 //----------------------------LoadKlassNode::make------------------------------
kvn@164 1801 // Polymorphic factory method:
kvn@164 1802 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
kvn@164 1803 Compile* C = gvn.C;
kvn@164 1804 Node *ctl = NULL;
kvn@164 1805 // sanity check the alias category against the created node type
kvn@164 1806 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
kvn@164 1807 assert(adr_type != NULL, "expecting TypeOopPtr");
kvn@164 1808 #ifdef _LP64
kvn@164 1809 if (adr_type->is_ptr_to_narrowoop()) {
kvn@221 1810 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
kvn@221 1811 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
kvn@168 1812 }
kvn@164 1813 #endif
kvn@168 1814 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@168 1815 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
kvn@164 1816 }
kvn@164 1817
duke@0 1818 //------------------------------Value------------------------------------------
duke@0 1819 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1820 return klass_value_common(phase);
kvn@164 1821 }
kvn@164 1822
kvn@164 1823 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
duke@0 1824 // Either input is TOP ==> the result is TOP
duke@0 1825 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1826 if (t1 == Type::TOP) return Type::TOP;
duke@0 1827 Node *adr = in(MemNode::Address);
duke@0 1828 const Type *t2 = phase->type( adr );
duke@0 1829 if (t2 == Type::TOP) return Type::TOP;
duke@0 1830 const TypePtr *tp = t2->is_ptr();
duke@0 1831 if (TypePtr::above_centerline(tp->ptr()) ||
duke@0 1832 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@0 1833
duke@0 1834 // Return a more precise klass, if possible
duke@0 1835 const TypeInstPtr *tinst = tp->isa_instptr();
duke@0 1836 if (tinst != NULL) {
duke@0 1837 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@0 1838 int offset = tinst->offset();
duke@0 1839 if (ik == phase->C->env()->Class_klass()
duke@0 1840 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1841 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1842 // We are loading a special hidden field from a Class mirror object,
duke@0 1843 // the field which points to the VM's Klass metaobject.
duke@0 1844 ciType* t = tinst->java_mirror_type();
duke@0 1845 // java_mirror_type returns non-null for compile-time Class constants.
duke@0 1846 if (t != NULL) {
duke@0 1847 // constant oop => constant klass
duke@0 1848 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1849 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@0 1850 }
duke@0 1851 if (!t->is_klass()) {
duke@0 1852 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@0 1853 return TypePtr::NULL_PTR;
duke@0 1854 }
duke@0 1855 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@0 1856 return TypeKlassPtr::make(t->as_klass());
duke@0 1857 }
duke@0 1858 // non-constant mirror, so we can't tell what's going on
duke@0 1859 }
duke@0 1860 if( !ik->is_loaded() )
duke@0 1861 return _type; // Bail out if not loaded
duke@0 1862 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1863 if (tinst->klass_is_exact()) {
duke@0 1864 return TypeKlassPtr::make(ik);
duke@0 1865 }
duke@0 1866 // See if we can become precise: no subklasses and no interface
duke@0 1867 // (Note: We need to support verified interfaces.)
duke@0 1868 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1869 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1870 // Add a dependence; if any subclass added we need to recompile
duke@0 1871 if (!ik->is_final()) {
duke@0 1872 // %%% should use stronger assert_unique_concrete_subtype instead
duke@0 1873 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1874 }
duke@0 1875 // Return precise klass
duke@0 1876 return TypeKlassPtr::make(ik);
duke@0 1877 }
duke@0 1878
duke@0 1879 // Return root of possible klass
duke@0 1880 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@0 1881 }
duke@0 1882 }
duke@0 1883
duke@0 1884 // Check for loading klass from an array
duke@0 1885 const TypeAryPtr *tary = tp->isa_aryptr();
duke@0 1886 if( tary != NULL ) {
duke@0 1887 ciKlass *tary_klass = tary->klass();
duke@0 1888 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@0 1889 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@0 1890 if (tary->klass_is_exact()) {
duke@0 1891 return TypeKlassPtr::make(tary_klass);
duke@0 1892 }
duke@0 1893 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@0 1894 // If the klass is an object array, we defer the question to the
duke@0 1895 // array component klass.
duke@0 1896 if( ak->is_obj_array_klass() ) {
duke@0 1897 assert( ak->is_loaded(), "" );
duke@0 1898 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@0 1899 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@0 1900 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@0 1901 // See if we can become precise: no subklasses and no interface
duke@0 1902 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1903 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1904 // Add a dependence; if any subclass added we need to recompile
duke@0 1905 if (!ik->is_final()) {
duke@0 1906 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1907 }
duke@0 1908 // Return precise array klass
duke@0 1909 return TypeKlassPtr::make(ak);
duke@0 1910 }
duke@0 1911 }
duke@0 1912 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 1913 } else { // Found a type-array?
duke@0 1914 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1915 assert( ak->is_type_array_klass(), "" );
duke@0 1916 return TypeKlassPtr::make(ak); // These are always precise
duke@0 1917 }
duke@0 1918 }
duke@0 1919 }
duke@0 1920
duke@0 1921 // Check for loading klass from an array klass
duke@0 1922 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1923 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1924 ciKlass* klass = tkls->klass();
duke@0 1925 if( !klass->is_loaded() )
duke@0 1926 return _type; // Bail out if not loaded
duke@0 1927 if( klass->is_obj_array_klass() &&
duke@0 1928 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1929 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@0 1930 // // Always returning precise element type is incorrect,
duke@0 1931 // // e.g., element type could be object and array may contain strings
duke@0 1932 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@0 1933
duke@0 1934 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@0 1935 // according to the element type's subclassing.
duke@0 1936 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@0 1937 }
duke@0 1938 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
duke@0 1939 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1940 ciKlass* sup = klass->as_instance_klass()->super();
duke@0 1941 // The field is Klass::_super. Return its (constant) value.
duke@0 1942 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@0 1943 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@0 1944 }
duke@0 1945 }
duke@0 1946
duke@0 1947 // Bailout case
duke@0 1948 return LoadNode::Value(phase);
duke@0 1949 }
duke@0 1950
duke@0 1951 //------------------------------Identity---------------------------------------
duke@0 1952 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@0 1953 // Also feed through the klass in Allocate(...klass...)._klass.
duke@0 1954 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1955 return klass_identity_common(phase);
kvn@164 1956 }
kvn@164 1957
kvn@164 1958 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
duke@0 1959 Node* x = LoadNode::Identity(phase);
duke@0 1960 if (x != this) return x;
duke@0 1961
duke@0 1962 // Take apart the address into an oop and and offset.
duke@0 1963 // Return 'this' if we cannot.
duke@0 1964 Node* adr = in(MemNode::Address);
duke@0 1965 intptr_t offset = 0;
duke@0 1966 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1967 if (base == NULL) return this;
duke@0 1968 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@0 1969 if (toop == NULL) return this;
duke@0 1970
duke@0 1971 // We can fetch the klass directly through an AllocateNode.
duke@0 1972 // This works even if the klass is not constant (clone or newArray).
duke@0 1973 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1974 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@0 1975 if (allocated_klass != NULL) {
duke@0 1976 return allocated_klass;
duke@0 1977 }
duke@0 1978 }
duke@0 1979
duke@0 1980 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
duke@0 1981 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
duke@0 1982 // See inline_native_Class_query for occurrences of these patterns.
duke@0 1983 // Java Example: x.getClass().isAssignableFrom(y)
duke@0 1984 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@0 1985 //
duke@0 1986 // This improves reflective code, often making the Class
duke@0 1987 // mirror go completely dead. (Current exception: Class
duke@0 1988 // mirrors may appear in debug info, but we could clean them out by
duke@0 1989 // introducing a new debug info operator for klassOop.java_mirror).
duke@0 1990 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@0 1991 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1992 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1993 // We are loading a special hidden field from a Class mirror,
duke@0 1994 // the field which points to its Klass or arrayKlass metaobject.
duke@0 1995 if (base->is_Load()) {
duke@0 1996 Node* adr2 = base->in(MemNode::Address);
duke@0 1997 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@0 1998 if (tkls != NULL && !tkls->empty()
duke@0 1999 && (tkls->klass()->is_instance_klass() ||
duke@0 2000 tkls->klass()->is_array_klass())
duke@0 2001 && adr2->is_AddP()
duke@0 2002 ) {
duke@0 2003 int mirror_field = Klass::java_mirror_offset_in_bytes();
duke@0 2004 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 2005 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
duke@0 2006 }
duke@0 2007 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
duke@0 2008 return adr2->in(AddPNode::Base);
duke@0 2009 }
duke@0 2010 }
duke@0 2011 }
duke@0 2012 }
duke@0 2013
duke@0 2014 return this;
duke@0 2015 }
duke@0 2016
kvn@164 2017
kvn@164 2018 //------------------------------Value------------------------------------------
kvn@164 2019 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 2020 const Type *t = klass_value_common(phase);
kvn@221 2021 if (t == Type::TOP)
kvn@221 2022 return t;
kvn@221 2023
kvn@221 2024 return t->make_narrowoop();
kvn@164 2025 }
kvn@164 2026
kvn@164 2027 //------------------------------Identity---------------------------------------
kvn@164 2028 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
kvn@164 2029 // Also feed through the klass in Allocate(...klass...)._klass.
kvn@164 2030 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 2031 Node *x = klass_identity_common(phase);
kvn@164 2032
kvn@164 2033 const Type *t = phase->type( x );
kvn@164 2034 if( t == Type::TOP ) return x;
kvn@164 2035 if( t->isa_narrowoop()) return x;
kvn@164 2036
kvn@221 2037 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
kvn@164 2038 }
kvn@164 2039
duke@0 2040 //------------------------------Value-----------------------------------------
duke@0 2041 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@0 2042 // Either input is TOP ==> the result is TOP
duke@0 2043 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 2044 if( t1 == Type::TOP ) return Type::TOP;
duke@0 2045 Node *adr = in(MemNode::Address);
duke@0 2046 const Type *t2 = phase->type( adr );
duke@0 2047 if( t2 == Type::TOP ) return Type::TOP;
duke@0 2048 const TypePtr *tp = t2->is_ptr();
duke@0 2049 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@0 2050 const TypeAryPtr *tap = tp->isa_aryptr();
duke@0 2051 if( !tap ) return _type;
duke@0 2052 return tap->size();
duke@0 2053 }
duke@0 2054
rasbold@366 2055 //-------------------------------Ideal---------------------------------------
rasbold@366 2056 // Feed through the length in AllocateArray(...length...)._length.
rasbold@366 2057 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
rasbold@366 2058 Node* p = MemNode::Ideal_common(phase, can_reshape);
rasbold@366 2059 if (p) return (p == NodeSentinel) ? NULL : p;
rasbold@366 2060
rasbold@366 2061 // Take apart the address into an oop and and offset.
rasbold@366 2062 // Return 'this' if we cannot.
rasbold@366 2063 Node* adr = in(MemNode::Address);
rasbold@366 2064 intptr_t offset = 0;
rasbold@366 2065 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
rasbold@366 2066 if (base == NULL) return NULL;
rasbold@366 2067 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
rasbold@366 2068 if (tary == NULL) return NULL;
rasbold@366 2069
rasbold@366 2070 // We can fetch the length directly through an AllocateArrayNode.
rasbold@366 2071 // This works even if the length is not constant (clone or newArray).
rasbold@366 2072 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 2073 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 2074 if (alloc != NULL) {
rasbold@366 2075 Node* allocated_length = alloc->Ideal_length();
rasbold@366 2076 Node* len = alloc->make_ideal_length(tary, phase);
rasbold@366 2077 if (allocated_length != len) {
rasbold@366 2078 // New CastII improves on this.
rasbold@366 2079 return len;
rasbold@366 2080 }
rasbold@366 2081 }
rasbold@366 2082 }
rasbold@366 2083
rasbold@366 2084 return NULL;
rasbold@366 2085 }
rasbold@366 2086
duke@0 2087 //------------------------------Identity---------------------------------------
duke@0 2088 // Feed through the length in AllocateArray(...length...)._length.
duke@0 2089 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@0 2090 Node* x = LoadINode::Identity(phase);
duke@0 2091 if (x != this) return x;
duke@0 2092
duke@0 2093 // Take apart the address into an oop and and offset.
duke@0 2094 // Return 'this' if we cannot.
duke@0 2095 Node* adr = in(MemNode::Address);
duke@0 2096 intptr_t offset = 0;
duke@0 2097 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 2098 if (base == NULL) return this;
duke@0 2099 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@0 2100 if (tary == NULL) return this;
duke@0 2101
duke@0 2102 // We can fetch the length directly through an AllocateArrayNode.
duke@0 2103 // This works even if the length is not constant (clone or newArray).
duke@0 2104 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 2105 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 2106 if (alloc != NULL) {
rasbold@366 2107 Node* allocated_length = alloc->Ideal_length();
rasbold@366 2108 // Do not allow make_ideal_length to allocate a CastII node.
rasbold@366 2109 Node* len = alloc->make_ideal_length(tary, phase, false);
rasbold@366 2110 if (allocated_length == len) {
rasbold@366 2111 // Return allocated_length only if it would not be improved by a CastII.
rasbold@366 2112 return allocated_length;
rasbold@366 2113 }
duke@0 2114 }
duke@0 2115 }
duke@0 2116
duke@0 2117 return this;
duke@0 2118
duke@0 2119 }
rasbold@366 2120
duke@0 2121 //=============================================================================
duke@0 2122 //---------------------------StoreNode::make-----------------------------------
duke@0 2123 // Polymorphic factory method:
coleenp@113 2124 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
coleenp@113 2125 Compile* C = gvn.C;
kvn@1529 2126 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
kvn@1529 2127 ctl != NULL, "raw memory operations should have control edge");
coleenp@113 2128
duke@0 2129 switch (bt) {
duke@0 2130 case T_BOOLEAN:
duke@0 2131 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
duke@0 2132 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
duke@0 2133 case T_CHAR:
duke@0 2134 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
duke@0 2135 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
duke@0 2136 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
duke@0 2137 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
duke@0 2138 case T_ADDRESS:
coleenp@113 2139 case T_OBJECT:
coleenp@113 2140 #ifdef _LP64
kvn@163 2141 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
coleenp@113 2142 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
coleenp@113 2143 adr->bottom_type()->isa_rawptr())) {
kvn@221 2144 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
kvn@221 2145 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
coleenp@113 2146 } else
coleenp@113 2147 #endif
kvn@221 2148 {
kvn@221 2149 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
kvn@221 2150 }
duke@0 2151 }
duke@0 2152 ShouldNotReachHere();
duke@0 2153 return (StoreNode*)NULL;
duke@0 2154 }
duke@0 2155
duke@0 2156 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
duke@0 2157 bool require_atomic = true;
duke@0 2158 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
duke@0 2159 }
duke@0 2160
duke@0 2161
duke@0 2162 //--------------------------bottom_type----------------------------------------
duke@0 2163 const Type *StoreNode::bottom_type() const {
duke@0 2164 return Type::MEMORY;
duke@0 2165 }
duke@0 2166
duke@0 2167 //------------------------------hash-------------------------------------------
duke@0 2168 uint StoreNode::hash() const {
duke@0 2169 // unroll addition of interesting fields
duke@0 2170 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@0 2171
duke@0 2172 // Since they are not commoned, do not hash them:
duke@0 2173 return NO_HASH;
duke@0 2174 }
duke@0 2175
duke@0 2176 //------------------------------Ideal------------------------------------------
duke@0 2177 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@0 2178 // When a store immediately follows a relevant allocation/initialization,
duke@0 2179 // try to capture it into the initialization, or hoist it above.
duke@0 2180 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 2181 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 2182 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 2183
duke@0 2184 Node* mem = in(MemNode::Memory);
duke@0 2185 Node* address = in(MemNode::Address);
duke@0 2186
never@2345 2187 // Back-to-back stores to same address? Fold em up. Generally
never@2345 2188 // unsafe if I have intervening uses... Also disallowed for StoreCM
never@2345 2189 // since they must follow each StoreP operation. Redundant StoreCMs
never@2345 2190 // are eliminated just before matching in final_graph_reshape.
never@2345 2191 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address) &&
never@2345 2192 mem->Opcode() != Op_StoreCM) {
duke@0 2193 // Looking at a dead closed cycle of memory?
duke@0 2194 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@0 2195
duke@0 2196 assert(Opcode() == mem->Opcode() ||
duke@0 2197 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@0 2198 "no mismatched stores, except on raw memory");
duke@0 2199
duke@0 2200 if (mem->outcnt() == 1 && // check for intervening uses
duke@0 2201 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@0 2202 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@0 2203 // For example, 'mem' might be the final state at a conditional return.
duke@0 2204 // Or, 'mem' might be used by some node which is live at the same time
duke@0 2205 // 'this' is live, which might be unschedulable. So, require exactly
duke@0 2206 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@0 2207 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@0 2208 if (can_reshape) { // (%%% is this an anachronism?)
duke@0 2209 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@0 2210 phase->is_IterGVN());
duke@0 2211 } else {
duke@0 2212 // It's OK to do this in the parser, since DU info is always accurate,
duke@0 2213 // and the parser always refers to nodes via SafePointNode maps.
duke@0 2214 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@0 2215 }
duke@0 2216 return this;
duke@0 2217 }
duke@0 2218 }
duke@0 2219
duke@0 2220 // Capture an unaliased, unconditional, simple store into an initializer.
duke@0 2221 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@0 2222 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@0 2223 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 2224 InitializeNode* init = mem->in(0)->as_Initialize();
duke@0 2225 intptr_t offset = init->can_capture_store(this, phase);
duke@0 2226 if (offset > 0) {
duke@0 2227 Node* moved = init->capture_store(this, offset, phase);
duke@0 2228 // If the InitializeNode captured me, it made a raw copy of me,
duke@0 2229 // and I need to disappear.
duke@0 2230 if (moved != NULL) {
duke@0 2231 // %%% hack to ensure that Ideal returns a new node:
duke@0 2232 mem = MergeMemNode::make(phase->C, mem);
duke@0 2233 return mem; // fold me away
duke@0 2234 }
duke@0 2235 }
duke@0 2236 }
duke@0 2237
duke@0 2238 return NULL; // No further progress
duke@0 2239 }
duke@0 2240
duke@0 2241 //------------------------------Value-----------------------------------------
duke@0 2242 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@0 2243 // Either input is TOP ==> the result is TOP
duke@0 2244 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 2245 if( t1 == Type::TOP ) return Type::TOP;
duke@0 2246 const Type *t2 = phase->type( in(MemNode::Address) );
duke@0 2247 if( t2 == Type::TOP ) return Type::TOP;
duke@0 2248 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@0 2249 if( t3 == Type::TOP ) return Type::TOP;
duke@0 2250 return Type::MEMORY;
duke@0 2251 }
duke@0 2252
duke@0 2253 //------------------------------Identity---------------------------------------
duke@0 2254 // Remove redundant stores:
duke@0 2255 // Store(m, p, Load(m, p)) changes to m.
duke@0 2256 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@0 2257 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@0 2258 Node* mem = in(MemNode::Memory);
duke@0 2259 Node* adr = in(MemNode::Address);
duke@0 2260 Node* val = in(MemNode::ValueIn);
duke@0 2261
duke@0 2262 // Load then Store? Then the Store is useless
duke@0 2263 if (val->is_Load() &&
duke@0 2264 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
duke@0 2265 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
duke@0 2266 val->as_Load()->store_Opcode() == Opcode()) {
duke@0 2267 return mem;
duke@0 2268 }
duke@0 2269
duke@0 2270 // Two stores in a row of the same value?
duke@0 2271 if (mem->is_Store() &&
duke@0 2272 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
duke@0 2273 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
duke@0 2274 mem->Opcode() == Opcode()) {
duke@0 2275 return mem;
duke@0 2276 }
duke@0 2277
duke@0 2278 // Store of zero anywhere into a freshly-allocated object?
duke@0 2279 // Then the store is useless.
duke@0 2280 // (It must already have been captured by the InitializeNode.)
duke@0 2281 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@0 2282 // a newly allocated object is already all-zeroes everywhere
duke@0 2283 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@0 2284 return mem;
duke@0 2285 }
duke@0 2286
duke@0 2287 // the store may also apply to zero-bits in an earlier object
duke@0 2288 Node* prev_mem = find_previous_store(phase);
duke@0 2289 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 2290 if (prev_mem != NULL) {
duke@0 2291 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@0 2292 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@0 2293 // prev_val and val might differ by a cast; it would be good
duke@0 2294 // to keep the more informative of the two.
duke@0 2295 return mem;
duke@0 2296 }
duke@0 2297 }
duke@0 2298 }
duke@0 2299
duke@0 2300 return this;
duke@0 2301 }
duke@0 2302
duke@0 2303 //------------------------------match_edge-------------------------------------
duke@0 2304 // Do we Match on this edge index or not? Match only memory & value
duke@0 2305 uint StoreNode::match_edge(uint idx) const {
duke@0 2306 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@0 2307 }
duke@0 2308
duke@0 2309 //------------------------------cmp--------------------------------------------
duke@0 2310 // Do not common stores up together. They generally have to be split
duke@0 2311 // back up anyways, so do not bother.
duke@0 2312 uint StoreNode::cmp( const Node &n ) const {
duke@0 2313 return (&n == this); // Always fail except on self
duke@0 2314 }
duke@0 2315
duke@0 2316 //------------------------------Ideal_masked_input-----------------------------
duke@0 2317 // Check for a useless mask before a partial-word store
duke@0 2318 // (StoreB ... (AndI valIn conIa) )
duke@0 2319 // If (conIa & mask == mask) this simplifies to
duke@0 2320 // (StoreB ... (valIn) )
duke@0 2321 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@0 2322 Node *val = in(MemNode::ValueIn);
duke@0 2323 if( val->Opcode() == Op_AndI ) {
duke@0 2324 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2325 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@0 2326 set_req(MemNode::ValueIn, val->in(1));
duke@0 2327 return this;
duke@0 2328 }
duke@0 2329 }
duke@0 2330 return NULL;
duke@0 2331 }
duke@0 2332
duke@0 2333
duke@0 2334 //------------------------------Ideal_sign_extended_input----------------------
duke@0 2335 // Check for useless sign-extension before a partial-word store
duke@0 2336 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@0 2337 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@0 2338 // (StoreB ... (valIn) )
duke@0 2339 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@0 2340 Node *val = in(MemNode::ValueIn);
duke@0 2341 if( val->Opcode() == Op_RShiftI ) {
duke@0 2342 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2343 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@0 2344 Node *shl = val->in(1);
duke@0 2345 if( shl->Opcode() == Op_LShiftI ) {
duke@0 2346 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@0 2347 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@0 2348 set_req(MemNode::ValueIn, shl->in(1));
duke@0 2349 return this;
duke@0 2350 }
duke@0 2351 }
duke@0 2352 }
duke@0 2353 }
duke@0 2354 return NULL;
duke@0 2355 }
duke@0 2356
duke@0 2357 //------------------------------value_never_loaded-----------------------------------
duke@0 2358 // Determine whether there are any possible loads of the value stored.
duke@0 2359 // For simplicity, we actually check if there are any loads from the
duke@0 2360 // address stored to, not just for loads of the value stored by this node.
duke@0 2361 //
duke@0 2362 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@0 2363 Node *adr = in(Address);
duke@0 2364 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@0 2365 if (adr_oop == NULL)
duke@0 2366 return false;
kvn@223 2367 if (!adr_oop->is_known_instance_field())
duke@0 2368 return false; // if not a distinct instance, there may be aliases of the address
duke@0 2369 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@0 2370 Node *use = adr->fast_out(i);
duke@0 2371 int opc = use->Opcode();
duke@0 2372 if (use->is_Load() || use->is_LoadStore()) {
duke@0 2373 return false;
duke@0 2374 }
duke@0 2375 }
duke@0 2376 return true;
duke@0 2377 }
duke@0 2378
duke@0 2379 //=============================================================================
duke@0 2380 //------------------------------Ideal------------------------------------------
duke@0 2381 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2382 // we can skip the AND operation. If the store is from a sign-extension
duke@0 2383 // (a left shift, then right shift) we can skip both.
duke@0 2384 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2385 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@0 2386 if( progress != NULL ) return progress;
duke@0 2387
duke@0 2388 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@0 2389 if( progress != NULL ) return progress;
duke@0 2390
duke@0 2391 // Finally check the default case
duke@0 2392 return StoreNode::Ideal(phase, can_reshape);
duke@0 2393 }
duke@0 2394
duke@0 2395 //=============================================================================
duke@0 2396 //------------------------------Ideal------------------------------------------
duke@0 2397 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2398 // we can skip the AND operation
duke@0 2399 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2400 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@0 2401 if( progress != NULL ) return progress;
duke@0 2402
duke@0 2403 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@0 2404 if( progress != NULL ) return progress;
duke@0 2405
duke@0 2406 // Finally check the default case
duke@0 2407 return StoreNode::Ideal(phase, can_reshape);
duke@0 2408 }
duke@0 2409
duke@0 2410 //=============================================================================
duke@0 2411 //------------------------------Identity---------------------------------------
duke@0 2412 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@0 2413 // No need to card mark when storing a null ptr
duke@0 2414 Node* my_store = in(MemNode::OopStore);
duke@0 2415 if (my_store->is_Store()) {
duke@0 2416 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@0 2417 if( t1 == TypePtr::NULL_PTR ) {
duke@0 2418 return in(MemNode::Memory);
duke@0 2419 }
duke@0 2420 }
duke@0 2421 return this;
duke@0 2422 }
duke@0 2423
cfang@985 2424 //=============================================================================
cfang@985 2425 //------------------------------Ideal---------------------------------------
cfang@985 2426 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
cfang@985 2427 Node* progress = StoreNode::Ideal(phase, can_reshape);
cfang@985 2428 if (progress != NULL) return progress;
cfang@985 2429
cfang@985 2430 Node* my_store = in(MemNode::OopStore);
cfang@985 2431 if (my_store->is_MergeMem()) {
cfang@985 2432 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
cfang@985 2433 set_req(MemNode::OopStore, mem);
cfang@985 2434 return this;
cfang@985 2435 }
cfang@985 2436
cfang@985 2437 return NULL;
cfang@985 2438 }
cfang@985 2439
duke@0 2440 //------------------------------Value-----------------------------------------
duke@0 2441 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@43 2442 // Either input is TOP ==> the result is TOP
kvn@43 2443 const Type *t = phase->type( in(MemNode::Memory) );
kvn@43 2444 if( t == Type::TOP ) return Type::TOP;
kvn@43 2445 t = phase->type( in(MemNode::Address) );
kvn@43 2446 if( t == Type::TOP ) return Type::TOP;
kvn@43 2447 t = phase->type( in(MemNode::ValueIn) );
kvn@43 2448 if( t == Type::TOP ) return Type::TOP;
duke@0 2449 // If extra input is TOP ==> the result is TOP
kvn@43 2450 t = phase->type( in(MemNode::OopStore) );
kvn@43 2451 if( t == Type::TOP ) return Type::TOP;
duke@0 2452
duke@0 2453 return StoreNode::Value( phase );
duke@0 2454 }
duke@0 2455
duke@0 2456
duke@0 2457 //=============================================================================
duke@0 2458 //----------------------------------SCMemProjNode------------------------------
duke@0 2459 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@0 2460 {
duke@0 2461 return bottom_type();
duke@0 2462 }
duke@0 2463
duke@0 2464 //=============================================================================
duke@0 2465 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
duke@0 2466 init_req(MemNode::Control, c );
duke@0 2467 init_req(MemNode::Memory , mem);
duke@0 2468 init_req(MemNode::Address, adr);
duke@0 2469 init_req(MemNode::ValueIn, val);
duke@0 2470 init_req( ExpectedIn, ex );
duke@0 2471 init_class_id(Class_LoadStore);
duke@0 2472
duke@0 2473 }
duke@0 2474
duke@0 2475 //=============================================================================
duke@0 2476 //-------------------------------adr_type--------------------------------------
duke@0 2477 // Do we Match on this edge index or not? Do not match memory
duke@0 2478 const TypePtr* ClearArrayNode::adr_type() const {
duke@0 2479 Node *adr = in(3);
duke@0 2480 return MemNode::calculate_adr_type(adr->bottom_type());
duke@0 2481 }
duke@0 2482
duke@0 2483 //------------------------------match_edge-------------------------------------
duke@0 2484 // Do we Match on this edge index or not? Do not match memory
duke@0 2485 uint ClearArrayNode::match_edge(uint idx) const {
duke@0 2486 return idx > 1;
duke@0 2487 }
duke@0 2488
duke@0 2489 //------------------------------Identity---------------------------------------
duke@0 2490 // Clearing a zero length array does nothing
duke@0 2491 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@68 2492 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@0 2493 }
duke@0 2494
duke@0 2495 //------------------------------Idealize---------------------------------------
duke@0 2496 // Clearing a short array is faster with stores
duke@0 2497 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2498 const int unit = BytesPerLong;
duke@0 2499 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@0 2500 if (!t) return NULL;
duke@0 2501 if (!t->is_con()) return NULL;
duke@0 2502 intptr_t raw_count = t->get_con();
duke@0 2503 intptr_t size = raw_count;
duke@0 2504 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@0 2505 // Clearing nothing uses the Identity call.
duke@0 2506 // Negative clears are possible on dead ClearArrays
duke@0 2507 // (see jck test stmt114.stmt11402.val).
duke@0 2508 if (size <= 0 || size % unit != 0) return NULL;
duke@0 2509 intptr_t count = size / unit;
duke@0 2510 // Length too long; use fast hardware clear
duke@0 2511 if (size > Matcher::init_array_short_size) return NULL;
duke@0 2512 Node *mem = in(1);
duke@0 2513 if( phase->type(mem)==Type::TOP ) return NULL;
duke@0 2514 Node *adr = in(3);
duke@0 2515 const Type* at = phase->type(adr);
duke@0 2516 if( at==Type::TOP ) return NULL;
duke@0 2517 const TypePtr* atp = at->isa_ptr();
duke@0 2518 // adjust atp to be the correct array element address type
duke@0 2519 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@0 2520 else atp = atp->add_offset(Type::OffsetBot);
duke@0 2521 // Get base for derived pointer purposes
duke@0 2522 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@0 2523 Node *base = adr->in(1);
duke@0 2524
duke@0 2525 Node *zero = phase->makecon(TypeLong::ZERO);
duke@0 2526 Node *off = phase->MakeConX(BytesPerLong);
duke@0 2527 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2528 count--;
duke@0 2529 while( count-- ) {
duke@0 2530 mem = phase->transform(mem);
duke@0 2531 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
duke@0 2532 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2533 }
duke@0 2534 return mem;
duke@0 2535 }
duke@0 2536
kvn@1100 2537 //----------------------------step_through----------------------------------
kvn@1100 2538 // Return allocation input memory edge if it is different instance
kvn@1100 2539 // or itself if it is the one we are looking for.
kvn@1100 2540 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
kvn@1100 2541 Node* n = *np;
kvn@1100 2542 assert(n->is_ClearArray(), "sanity");
kvn@1100 2543 intptr_t offset;
kvn@1100 2544 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
kvn@1100 2545 // This method is called only before Allocate nodes are expanded during
kvn@1100 2546 // macro nodes expansion. Before that ClearArray nodes are only generated
kvn@1100 2547 // in LibraryCallKit::generate_arraycopy() which follows allocations.
kvn@1100 2548 assert(alloc != NULL, "should have allocation");
kvn@1100 2549 if (alloc->_idx == instance_id) {
kvn@1100 2550 // Can not bypass initialization of the instance we are looking for.
kvn@1100 2551 return false;
kvn@1100 2552 }
kvn@1100 2553 // Otherwise skip it.
kvn@1100 2554 InitializeNode* init = alloc->initialization();
kvn@1100 2555 if (init != NULL)
kvn@1100 2556 *np = init->in(TypeFunc::Memory);
kvn@1100 2557 else
kvn@1100 2558 *np = alloc->in(TypeFunc::Memory);
kvn@1100 2559 return true;
kvn@1100 2560 }
kvn@1100 2561
duke@0 2562 //----------------------------clear_memory-------------------------------------
duke@0 2563 // Generate code to initialize object storage to zero.
duke@0 2564 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2565 intptr_t start_offset,
duke@0 2566 Node* end_offset,
duke@0 2567 PhaseGVN* phase) {
duke@0 2568 Compile* C = phase->C;
duke@0 2569 intptr_t offset = start_offset;
duke@0 2570
duke@0 2571 int unit = BytesPerLong;
duke@0 2572 if ((offset % unit) != 0) {
duke@0 2573 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
duke@0 2574 adr = phase->transform(adr);
duke@0 2575 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2576 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2577 mem = phase->transform(mem);
duke@0 2578 offset += BytesPerInt;
duke@0 2579 }
duke@0 2580 assert((offset % unit) == 0, "");
duke@0 2581
duke@0 2582 // Initialize the remaining stuff, if any, with a ClearArray.
duke@0 2583 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@0 2584 }
duke@0 2585
duke@0 2586 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2587 Node* start_offset,
duke@0 2588 Node* end_offset,
duke@0 2589 PhaseGVN* phase) {
never@68 2590 if (start_offset == end_offset) {
never@68 2591 // nothing to do
never@68 2592 return mem;
never@68 2593 }
never@68 2594
duke@0 2595 Compile* C = phase->C;
duke@0 2596 int unit = BytesPerLong;
duke@0 2597 Node* zbase = start_offset;
duke@0 2598 Node* zend = end_offset;
duke@0 2599
duke@0 2600 // Scale to the unit required by the CPU:
duke@0 2601 if (!Matcher::init_array_count_is_in_bytes) {
duke@0 2602 Node* shift = phase->intcon(exact_log2(unit));
duke@0 2603 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
duke@0 2604 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
duke@0 2605 }
duke@0 2606
duke@0 2607 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
duke@0 2608 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
duke@0 2609
duke@0 2610 // Bulk clear double-words
duke@0 2611 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
duke@0 2612 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
duke@0 2613 return phase->transform(mem);
duke@0 2614 }
duke@0 2615
duke@0 2616 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2617 intptr_t start_offset,
duke@0 2618 intptr_t end_offset,
duke@0 2619 PhaseGVN* phase) {
never@68 2620 if (start_offset == end_offset) {
never@68 2621 // nothing to do
never@68 2622 return mem;
never@68 2623 }
never@68 2624
duke@0 2625 Compile* C = phase->C;
duke@0 2626 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@0 2627 intptr_t done_offset = end_offset;
duke@0 2628 if ((done_offset % BytesPerLong) != 0) {
duke@0 2629 done_offset -= BytesPerInt;
duke@0 2630 }
duke@0 2631 if (done_offset > start_offset) {
duke@0 2632 mem = clear_memory(ctl, mem, dest,
duke@0 2633 start_offset, phase->MakeConX(done_offset), phase);
duke@0 2634 }
duke@0 2635 if (done_offset < end_offset) { // emit the final 32-bit store
duke@0 2636 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@0 2637 adr = phase->transform(adr);
duke@0 2638 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2639 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2640 mem = phase->transform(mem);
duke@0 2641 done_offset += BytesPerInt;
duke@0 2642 }
duke@0 2643 assert(done_offset == end_offset, "");
duke@0 2644 return mem;
duke@0 2645 }
duke@0 2646
duke@0 2647 //=============================================================================
kvn@2259 2648 // Do not match memory edge.
kvn@2259 2649 uint StrIntrinsicNode::match_edge(uint idx) const {
kvn@2259 2650 return idx == 2 || idx == 3;
duke@0 2651 }
duke@0 2652
duke@0 2653 //------------------------------Ideal------------------------------------------
duke@0 2654 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2655 // control copies
kvn@2259 2656 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@2259 2657 if (remove_dead_region(phase, can_reshape)) return this;
kvn@2259 2658
kvn@2264 2659 if (can_reshape) {
kvn@2264 2660 Node* mem = phase->transform(in(MemNode::Memory));
kvn@2264 2661 // If transformed to a MergeMem, get the desired slice
kvn@2264 2662 uint alias_idx = phase->C->get_alias_index(adr_type());
kvn@2264 2663 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
kvn@2264 2664 if (mem != in(MemNode::Memory)) {
kvn@2264 2665 set_req(MemNode::Memory, mem);
kvn@2264 2666 return this;
kvn@2264 2667 }
kvn@2264 2668 }
kvn@2259 2669 return NULL;
rasbold@169 2670 }
rasbold@169 2671
duke@0 2672 //=============================================================================
duke@0 2673 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@0 2674 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@0 2675 _adr_type(C->get_adr_type(alias_idx))
duke@0 2676 {
duke@0 2677 init_class_id(Class_MemBar);
duke@0 2678 Node* top = C->top();
duke@0 2679 init_req(TypeFunc::I_O,top);
duke@0 2680 init_req(TypeFunc::FramePtr,top);
duke@0 2681 init_req(TypeFunc::ReturnAdr,top);
duke@0 2682 if (precedent != NULL)
duke@0 2683 init_req(TypeFunc::Parms, precedent);
duke@0 2684 }
duke@0 2685
duke@0 2686 //------------------------------cmp--------------------------------------------
duke@0 2687 uint MemBarNode::hash() const { return NO_HASH; }
duke@0 2688 uint MemBarNode::cmp( const Node &n ) const {
duke@0 2689 return (&n == this); // Always fail except on self
duke@0 2690 }
duke@0 2691
duke@0 2692 //------------------------------make-------------------------------------------
duke@0 2693 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@0 2694 int len = Precedent + (pn == NULL? 0: 1);
duke@0 2695 switch (opcode) {
duke@0 2696 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
duke@0 2697 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
roland@2612 2698 case Op_MemBarAcquireLock: return new(C, len) MemBarAcquireLockNode(C, atp, pn);
roland@2612 2699 case Op_MemBarReleaseLock: return new(C, len) MemBarReleaseLockNode(C, atp, pn);
duke@0 2700 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
duke@0 2701 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
duke@0 2702 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
duke@0 2703 default: ShouldNotReachHere(); return NULL;
duke@0 2704 }
duke@0 2705 }
duke@0 2706
duke@0 2707 //------------------------------Ideal------------------------------------------
duke@0 2708 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2709 // control copies
duke@0 2710 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@1100 2711 if (remove_dead_region(phase, can_reshape)) return this;
kvn@1100 2712
kvn@1100 2713 // Eliminate volatile MemBars for scalar replaced objects.
kvn@1100 2714 if (can_reshape && req() == (Precedent+1) &&
kvn@1100 2715 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) {
kvn@1100 2716 // Volatile field loads and stores.
kvn@1100 2717 Node* my_mem = in(MemBarNode::Precedent);
kvn@1100 2718 if (my_mem != NULL && my_mem->is_Mem()) {
kvn@1100 2719 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@1100 2720 // Check for scalar replaced object reference.
kvn@1100 2721 if( t_oop != NULL && t_oop->is_known_instance_field() &&
kvn@1100 2722 t_oop->offset() != Type::OffsetBot &&
kvn@1100 2723 t_oop->offset() != Type::OffsetTop) {
kvn@1100 2724 // Replace MemBar projections by its inputs.
kvn@1100 2725 PhaseIterGVN* igvn = phase->is_IterGVN();
kvn@1100 2726 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
kvn@1100 2727 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
kvn@1100 2728 // Must return either the original node (now dead) or a new node
kvn@1100 2729 // (Do not return a top here, since that would break the uniqueness of top.)
kvn@1100 2730 return new (phase->C, 1) ConINode(TypeInt::ZERO);
kvn@1100 2731 }
kvn@1100 2732 }
kvn@1100 2733 }
kvn@1100 2734 return NULL;
duke@0 2735 }
duke@0 2736
duke@0 2737 //------------------------------Value------------------------------------------
duke@0 2738 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@0 2739 if( !in(0) ) return Type::TOP;
duke@0 2740 if( phase->type(in(0)) == Type::TOP )
duke@0 2741 return Type::TOP;
duke@0 2742 return TypeTuple::MEMBAR;
duke@0 2743 }
duke@0 2744
duke@0 2745 //------------------------------match------------------------------------------
duke@0 2746 // Construct projections for memory.
duke@0 2747 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@0 2748 switch (proj->_con) {
duke@0 2749 case TypeFunc::Control:
duke@0 2750 case TypeFunc::Memory:
duke@0 2751 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@0 2752 }
duke@0 2753 ShouldNotReachHere();
duke@0 2754 return NULL;
duke@0 2755 }
duke@0 2756
duke@0 2757 //===========================InitializeNode====================================
duke@0 2758 // SUMMARY:
duke@0 2759 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@0 2760 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@0 2761 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@0 2762 // It can coalesce related raw stores into larger units (called 'tiles').
duke@0 2763 // It can avoid zeroing new storage for memory units which have raw inits.
duke@0 2764 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@0 2765 //
duke@0 2766 // EXAMPLE:
duke@0 2767 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@0 2768 // ctl = incoming control; mem* = incoming memory
duke@0 2769 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@0 2770 // First allocate uninitialized memory and fill in the header:
duke@0 2771 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@0 2772 // ctl := alloc.Control; mem* := alloc.Memory*
duke@0 2773 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@0 2774 // Then initialize to zero the non-header parts of the raw memory block:
duke@0 2775 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@0 2776 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@0 2777 // After the initialize node executes, the object is ready for service:
duke@0 2778 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@0 2779 // Suppose its body is immediately initialized as {1,2}:
duke@0 2780 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2781 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2782 // mem.SLICE(#short[*]) := store2
duke@0 2783 //
duke@0 2784 // DETAILS:
duke@0 2785 // An InitializeNode collects and isolates object initialization after
duke@0 2786 // an AllocateNode and before the next possible safepoint. As a
duke@0 2787 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@0 2788 // down past any safepoint or any publication of the allocation.
duke@0 2789 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@0 2790 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@0 2791 //
duke@0 2792 // The semantics of the InitializeNode include an implicit zeroing of
duke@0 2793 // the new object from object header to the end of the object.
duke@0 2794 // (The object header and end are determined by the AllocateNode.)
duke@0 2795 //
duke@0 2796 // Certain stores may be added as direct inputs to the InitializeNode.
duke@0 2797 // These stores must update raw memory, and they must be to addresses
duke@0 2798 // derived from the raw address produced by AllocateNode, and with
duke@0 2799 // a constant offset. They must be ordered by increasing offset.
duke@0 2800 // The first one is at in(RawStores), the last at in(req()-1).
duke@0 2801 // Unlike most memory operations, they are not linked in a chain,
duke@0 2802 // but are displayed in parallel as users of the rawmem output of
duke@0 2803 // the allocation.
duke@0 2804 //
duke@0 2805 // (See comments in InitializeNode::capture_store, which continue
duke@0 2806 // the example given above.)
duke@0 2807 //
duke@0 2808 // When the associated Allocate is macro-expanded, the InitializeNode
duke@0 2809 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@0 2810 // may also be created at that point to represent any required zeroing.
duke@0 2811 // The InitializeNode is then marked 'complete', prohibiting further
duke@0 2812 // capturing of nearby memory operations.
duke@0 2813 //
duke@0 2814 // During macro-expansion, all captured initializations which store
twisti@605 2815 // constant values of 32 bits or smaller are coalesced (if advantageous)
duke@0 2816 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@0 2817 // initialized in fewer memory operations. Memory words which are
duke@0 2818 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@0 2819 // by explicit stores of zero. (The code shape happens to do all
duke@0 2820 // zeroing first, then all other stores, with both sequences occurring
duke@0 2821 // in order of ascending offsets.)
duke@0 2822 //
duke@0 2823 // Alternatively, code may be inserted between an AllocateNode and its
duke@0 2824 // InitializeNode, to perform arbitrary initialization of the new object.
duke@0 2825 // E.g., the object copying intrinsics insert complex data transfers here.
duke@0 2826 // The initialization must then be marked as 'complete' disable the
duke@0 2827 // built-in zeroing semantics and the collection of initializing stores.
duke@0 2828 //
duke@0 2829 // While an InitializeNode is incomplete, reads from the memory state
duke@0 2830 // produced by it are optimizable if they match the control edge and
duke@0 2831 // new oop address associated with the allocation/initialization.
duke@0 2832 // They return a stored value (if the offset matches) or else zero.
duke@0 2833 // A write to the memory state, if it matches control and address,
duke@0 2834 // and if it is to a constant offset, may be 'captured' by the
duke@0 2835 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@0 2836 // inside the initialization, to the raw oop produced by the allocation.
duke@0 2837 // Operations on addresses which are provably distinct (e.g., to
duke@0 2838 // other AllocateNodes) are allowed to bypass the initialization.
duke@0 2839 //
duke@0 2840 // The effect of all this is to consolidate object initialization
duke@0 2841 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@0 2842 // single location, where it can be optimized as a unit.
duke@0 2843 //
duke@0 2844 // Only stores with an offset less than TrackedInitializationLimit words
duke@0 2845 // will be considered for capture by an InitializeNode. This puts a
duke@0 2846 // reasonable limit on the complexity of optimized initializations.
duke@0 2847
duke@0 2848 //---------------------------InitializeNode------------------------------------
duke@0 2849 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
kvn@2722 2850 : _is_complete(Incomplete),
duke@0 2851 MemBarNode(C, adr_type, rawoop)
duke@0 2852 {
duke@0 2853 init_class_id(Class_Initialize);
duke@0 2854
duke@0 2855 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@0 2856 assert(in(RawAddress) == rawoop, "proper init");
duke@0 2857 // Note: allocation() can be NULL, for secondary initialization barriers
duke@0 2858 }
duke@0 2859
duke@0 2860 // Since this node is not matched, it will be processed by the
duke@0 2861 // register allocator. Declare that there are no constraints
duke@0 2862 // on the allocation of the RawAddress edge.
duke@0 2863 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@0 2864 // This edge should be set to top, by the set_complete. But be conservative.
duke@0 2865 if (idx == InitializeNode::RawAddress)
duke@0 2866 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@0 2867 return RegMask::Empty;
duke@0 2868 }
duke@0 2869
duke@0 2870 Node* InitializeNode::memory(uint alias_idx) {
duke@0 2871 Node* mem = in(Memory);
duke@0 2872 if (mem->is_MergeMem()) {
duke@0 2873 return mem->as_MergeMem()->memory_at(alias_idx);
duke@0 2874 } else {
duke@0 2875 // incoming raw memory is not split
duke@0 2876 return mem;
duke@0 2877 }
duke@0 2878 }
duke@0 2879
duke@0 2880 bool InitializeNode::is_non_zero() {
duke@0 2881 if (is_complete()) return false;
duke@0 2882 remove_extra_zeroes();
duke@0 2883 return (req() > RawStores);
duke@0 2884 }
duke@0 2885
duke@0 2886 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@0 2887 assert(!is_complete(), "caller responsibility");
kvn@2722 2888 _is_complete = Complete;
duke@0 2889
duke@0 2890 // After this node is complete, it contains a bunch of
duke@0 2891 // raw-memory initializations. There is no need for
duke@0 2892 // it to have anything to do with non-raw memory effects.
duke@0 2893 // Therefore, tell all non-raw users to re-optimize themselves,
duke@0 2894 // after skipping the memory effects of this initialization.
duke@0 2895 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@0 2896 if (igvn) igvn->add_users_to_worklist(this);
duke@0 2897 }
duke@0 2898
duke@0 2899 // convenience function
duke@0 2900 // return false if the init contains any stores already
duke@0 2901 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@0 2902 InitializeNode* init = initialization();
duke@0 2903 if (init == NULL || init->is_complete()) return false;
duke@0 2904 init->remove_extra_zeroes();
duke@0 2905 // for now, if this allocation has already collected any inits, bail:
duke@0 2906 if (init->is_non_zero()) return false;
duke@0 2907 init->set_complete(phase);
duke@0 2908 return true;
duke@0 2909 }
duke@0 2910
duke@0 2911 void InitializeNode::remove_extra_zeroes() {
duke@0 2912 if (req() == RawStores) return;
duke@0 2913 Node* zmem = zero_memory();
duke@0 2914 uint fill = RawStores;
duke@0 2915 for (uint i = fill; i < req(); i++) {
duke@0 2916 Node* n = in(i);
duke@0 2917 if (n->is_top() || n == zmem) continue; // skip
duke@0 2918 if (fill < i) set_req(fill, n); // compact
duke@0 2919 ++fill;
duke@0 2920 }
duke@0 2921 // delete any empty spaces created:
duke@0 2922 while (fill < req()) {
duke@0 2923 del_req(fill);
duke@0 2924 }
duke@0 2925 }
duke@0 2926
duke@0 2927 // Helper for remembering which stores go with which offsets.
duke@0 2928 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@0 2929 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@0 2930 intptr_t offset = -1;
duke@0 2931 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@0 2932 phase, offset);
duke@0 2933 if (base == NULL) return -1; // something is dead,
duke@0 2934 if (offset < 0) return -1; // dead, dead
duke@0 2935 return offset;
duke@0 2936 }
duke@0 2937
duke@0 2938 // Helper for proving that an initialization expression is
duke@0 2939 // "simple enough" to be folded into an object initialization.
duke@0 2940 // Attempts to prove that a store's initial value 'n' can be captured
duke@0 2941 // within the initialization without creating a vicious cycle, such as:
duke@0 2942 // { Foo p = new Foo(); p.next = p; }
duke@0 2943 // True for constants and parameters and small combinations thereof.
duke@0 2944 bool InitializeNode::detect_init_independence(Node* n,
duke@0 2945 bool st_is_pinned,
duke@0 2946 int& count) {
duke@0 2947 if (n == NULL) return true; // (can this really happen?)
duke@0 2948 if (n->is_Proj()) n = n->in(0);
duke@0 2949 if (n == this) return false; // found a cycle
duke@0 2950 if (n->is_Con()) return true;
duke@0 2951 if (n->is_Start()) return true; // params, etc., are OK
duke@0 2952 if (n->is_Root()) return true; // even better
duke@0 2953
duke@0 2954 Node* ctl = n->in(0);
duke@0 2955 if (ctl != NULL && !ctl->is_top()) {
duke@0 2956 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@0 2957 if (ctl == this) return false;
duke@0 2958
duke@0 2959 // If we already know that the enclosing memory op is pinned right after
duke@0 2960 // the init, then any control flow that the store has picked up
duke@0 2961 // must have preceded the init, or else be equal to the init.
duke@0 2962 // Even after loop optimizations (which might change control edges)
duke@0 2963 // a store is never pinned *before* the availability of its inputs.
kvn@119 2964 if (!MemNode::all_controls_dominate(n, this))
duke@0 2965 return false; // failed to prove a good control
duke@0 2966
duke@0 2967 }
duke@0 2968
duke@0 2969 // Check data edges for possible dependencies on 'this'.
duke@0 2970 if ((count += 1) > 20) return false; // complexity limit
duke@0 2971 for (uint i = 1; i < n->req(); i++) {
duke@0 2972 Node* m = n->in(i);
duke@0 2973 if (m == NULL || m == n || m->is_top()) continue;
duke@0 2974 uint first_i = n->find_edge(m);
duke@0 2975 if (i != first_i) continue; // process duplicate edge just once
duke@0 2976 if (!detect_init_independence(m, st_is_pinned, count)) {
duke@0 2977 return false;
duke@0 2978 }
duke@0 2979 }
duke@0 2980
duke@0 2981 return true;
duke@0 2982 }
duke@0 2983
duke@0 2984 // Here are all the checks a Store must pass before it can be moved into
duke@0 2985 // an initialization. Returns zero if a check fails.
duke@0 2986 // On success, returns the (constant) offset to which the store applies,
duke@0 2987 // within the initialized memory.
duke@0 2988 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
duke@0 2989 const int FAIL = 0;
duke@0 2990 if (st->req() != MemNode::ValueIn + 1)
duke@0 2991 return FAIL; // an inscrutable StoreNode (card mark?)
duke@0 2992 Node* ctl = st->in(MemNode::Control);
duke@0 2993 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@0 2994 return FAIL; // must be unconditional after the initialization
duke@0 2995 Node* mem = st->in(MemNode::Memory);
duke@0 2996 if (!(mem->is_Proj() && mem->in(0) == this))
duke@0 2997 return FAIL; // must not be preceded by other stores
duke@0 2998 Node* adr = st->in(MemNode::Address);
duke@0 2999 intptr_t offset;
duke@0 3000 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@0 3001 if (alloc == NULL)
duke@0 3002 return FAIL; // inscrutable address
duke@0 3003 if (alloc != allocation())
duke@0 3004 return FAIL; // wrong allocation! (store needs to float up)
duke@0 3005 Node* val = st->in(MemNode::ValueIn);
duke@0 3006 int complexity_count = 0;
duke@0 3007 if (!detect_init_independence(val, true, complexity_count))
duke@0 3008 return FAIL; // stored value must be 'simple enough'
duke@0 3009
duke@0 3010 return offset; // success
duke@0 3011 }
duke@0 3012
duke@0 3013 // Find the captured store in(i) which corresponds to the range
duke@0 3014 // [start..start+size) in the initialized object.
duke@0 3015 // If there is one, return its index i. If there isn't, return the
duke@0 3016 // negative of the index where it should be inserted.
duke@0 3017 // Return 0 if the queried range overlaps an initialization boundary
duke@0 3018 // or if dead code is encountered.
duke@0 3019 // If size_in_bytes is zero, do not bother with overlap checks.
duke@0 3020 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@0 3021 int size_in_bytes,
duke@0 3022 PhaseTransform* phase) {
duke@0 3023 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@0 3024
duke@0 3025 if (is_complete())
duke@0 3026 return FAIL; // arraycopy got here first; punt
duke@0 3027
duke@0 3028 assert(allocation() != NULL, "must be present");
duke@0 3029
duke@0 3030 // no negatives, no header fields:
coleenp@113 3031 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@0 3032
duke@0 3033 // after a certain size, we bail out on tracking all the stores:
duke@0 3034 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 3035 if (start >= ti_limit) return FAIL;
duke@0 3036
duke@0 3037 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@0 3038 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@0 3039
duke@0 3040 Node* st = in(i);
duke@0 3041 intptr_t st_off = get_store_offset(st, phase);
duke@0 3042 if (st_off < 0) {
duke@0 3043 if (st != zero_memory()) {
duke@0 3044 return FAIL; // bail out if there is dead garbage
duke@0 3045 }
duke@0 3046 } else if (st_off > start) {
duke@0 3047 // ...we are done, since stores are ordered
duke@0 3048 if (st_off < start + size_in_bytes) {
duke@0 3049 return FAIL; // the next store overlaps
duke@0 3050 }
duke@0 3051 return -(int)i; // not found; here is where to put it
duke@0 3052 } else if (st_off < start) {
duke@0 3053 if (size_in_bytes != 0 &&
duke@0 3054 start < st_off + MAX_STORE &&
duke@0 3055 start < st_off + st->as_Store()->memory_size()) {
duke@0 3056 return FAIL; // the previous store overlaps
duke@0 3057 }
duke@0 3058 } else {
duke@0 3059 if (size_in_bytes != 0 &&
duke@0 3060 st->as_Store()->memory_size() != size_in_bytes) {
duke@0 3061 return FAIL; // mismatched store size
duke@0 3062 }
duke@0 3063 return i;
duke@0 3064 }
duke@0 3065
duke@0 3066 ++i;
duke@0 3067 }
duke@0 3068 }
duke@0 3069
duke@0 3070 // Look for a captured store which initializes at the offset 'start'
duke@0 3071 // with the given size. If there is no such store, and no other
duke@0 3072 // initialization interferes, then return zero_memory (the memory
duke@0 3073 // projection of the AllocateNode).
duke@0 3074 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@0 3075 PhaseTransform* phase) {
duke@0 3076 assert(stores_are_sane(phase), "");
duke@0 3077 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 3078 if (i == 0) {
duke@0 3079 return NULL; // something is dead
duke@0 3080 } else if (i < 0) {
duke@0 3081 return zero_memory(); // just primordial zero bits here
duke@0 3082 } else {
duke@0 3083 Node* st = in(i); // here is the store at this position
duke@0 3084 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@0 3085 return st;
duke@0 3086 }
duke@0 3087 }
duke@0 3088
duke@0 3089 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@0 3090 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@0 3091 PhaseTransform* phase) {
duke@0 3092 Node* addr = in(RawAddress);
duke@0 3093 if (offset != 0) {
duke@0 3094 Compile* C = phase->C;
duke@0 3095 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
duke@0 3096 phase->MakeConX(offset)) );
duke@0 3097 }
duke@0 3098 return addr;
duke@0 3099 }
duke@0 3100
duke@0 3101 // Clone the given store, converting it into a raw store
duke@0 3102 // initializing a field or element of my new object.
duke@0 3103 // Caller is responsible for retiring the original store,
duke@0 3104 // with subsume_node or the like.
duke@0 3105 //
duke@0 3106 // From the example above InitializeNode::InitializeNode,
duke@0 3107 // here are the old stores to be captured:
duke@0 3108 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 3109 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 3110 //
duke@0 3111 // Here is the changed code; note the extra edges on init:
duke@0 3112 // alloc = (Allocate ...)
duke@0 3113 // rawoop = alloc.RawAddress
duke@0 3114 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@0 3115 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@0 3116 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@0 3117 // rawstore1 rawstore2)
duke@0 3118 //
duke@0 3119 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
duke@0 3120 PhaseTransform* phase) {
duke@0 3121 assert(stores_are_sane(phase), "");
duke@0 3122
duke@0 3123 if (start < 0) return NULL;
duke@0 3124 assert(can_capture_store(st, phase) == start, "sanity");
duke@0 3125
duke@0 3126 Compile* C = phase->C;
duke@0 3127 int size_in_bytes = st->memory_size();
duke@0 3128 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 3129 if (i == 0) return NULL; // bail out
duke@0 3130 Node* prev_mem = NULL; // raw memory for the captured store
duke@0 3131 if (i > 0) {
duke@0 3132 prev_mem = in(i); // there is a pre-existing store under this one
duke@0 3133 set_req(i, C->top()); // temporarily disconnect it
duke@0 3134 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@0 3135 } else {
duke@0 3136 i = -i; // no pre-existing store
duke@0 3137 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@0 3138 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@0 3139 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@0 3140 else
duke@0 3141 ins_req(i, C->top()); // build a new edge
duke@0 3142 }
duke@0 3143 Node* new_st = st->clone();
duke@0 3144 new_st->set_req(MemNode::Control, in(Control));
duke@0 3145 new_st->set_req(MemNode::Memory, prev_mem);
duke@0 3146 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@0 3147 new_st = phase->transform(new_st);
duke@0 3148
duke@0 3149 // At this point, new_st might have swallowed a pre-existing store
duke@0 3150 // at the same offset, or perhaps new_st might have disappeared,
duke@0 3151 // if it redundantly stored the same value (or zero to fresh memory).
duke@0 3152
duke@0 3153 // In any case, wire it in:
duke@0 3154 set_req(i, new_st);
duke@0 3155
duke@0 3156 // The caller may now kill the old guy.
duke@0 3157 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@0 3158 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@0 3159 assert(!is_complete(), "");
duke@0 3160 return new_st;
duke@0 3161 }
duke@0 3162
duke@0 3163 static bool store_constant(jlong* tiles, int num_tiles,
duke@0 3164 intptr_t st_off, int st_size,
duke@0 3165 jlong con) {
duke@0 3166 if ((st_off & (st_size-1)) != 0)
duke@0 3167 return false; // strange store offset (assume size==2**N)
duke@0 3168 address addr = (address)tiles + st_off;
duke@0 3169 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@0 3170 switch (st_size) {
duke@0 3171 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@0 3172 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@0 3173 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@0 3174 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@0 3175 default: return false; // strange store size (detect size!=2**N here)
duke@0 3176 }
duke@0 3177 return true; // return success to caller
duke@0 3178 }
duke@0 3179
duke@0 3180 // Coalesce subword constants into int constants and possibly
duke@0 3181 // into long constants. The goal, if the CPU permits,
duke@0 3182 // is to initialize the object with a small number of 64-bit tiles.
duke@0 3183 // Also, convert floating-point constants to bit patterns.
duke@0 3184 // Non-constants are not relevant to this pass.
duke@0 3185 //
duke@0 3186 // In terms of the running example on InitializeNode::InitializeNode
duke@0 3187 // and InitializeNode::capture_store, here is the transformation
duke@0 3188 // of rawstore1 and rawstore2 into rawstore12:
duke@0 3189 // alloc = (Allocate ...)
duke@0 3190 // rawoop = alloc.RawAddress
duke@0 3191 // tile12 = 0x00010002
duke@0 3192 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@0 3193 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@0 3194 //
duke@0 3195 void
duke@0 3196 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@0 3197 Node* size_in_bytes,
duke@0 3198 PhaseGVN* phase) {
duke@0 3199 Compile* C = phase->C;
duke@0 3200
duke@0 3201 assert(stores_are_sane(phase), "");
duke@0 3202 // Note: After this pass, they are not completely sane,
duke@0 3203 // since there may be some overlaps.
duke@0 3204
duke@0 3205 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@0 3206
duke@0 3207 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 3208 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@0 3209 size_limit = MIN2(size_limit, ti_limit);
duke@0 3210 size_limit = align_size_up(size_limit, BytesPerLong);
duke@0 3211 int num_tiles = size_limit / BytesPerLong;
duke@0 3212
duke@0 3213 // allocate space for the tile map:
duke@0 3214 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@0 3215 jlong tiles_buf[small_len];
duke@0 3216 Node* nodes_buf[small_len];
duke@0 3217 jlong inits_buf[small_len];
duke@0 3218 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@0 3219 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 3220 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@0 3221 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@0 3222 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@0 3223 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 3224 // tiles: exact bitwise model of all primitive constants
duke@0 3225 // nodes: last constant-storing node subsumed into the tiles model
duke@0 3226 // inits: which bytes (in each tile) are touched by any initializations
duke@0 3227
duke@0 3228 //// Pass A: Fill in the tile model with any relevant stores.
duke@0 3229
duke@0 3230 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@0 3231 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@0 3232 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@0 3233 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3234 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3235 Node* st = in(i);
duke@0 3236 intptr_t st_off = get_store_offset(st, phase);
duke@0 3237
duke@0 3238 // Figure out the store's offset and constant value:
duke@0 3239 if (st_off < header_size) continue; //skip (ignore header)
duke@0 3240 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@0 3241 int st_size = st->as_Store()->memory_size();
duke@0 3242 if (st_off + st_size > size_limit) break;
duke@0 3243
duke@0 3244 // Record which bytes are touched, whether by constant or not.
duke@0 3245 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@0 3246 continue; // skip (strange store size)
duke@0 3247
duke@0 3248 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@0 3249 if (!val->singleton()) continue; //skip (non-con store)
duke@0 3250 BasicType type = val->basic_type();
duke@0 3251
duke@0 3252 jlong con = 0;
duke@0 3253 switch (type) {
duke@0 3254 case T_INT: con = val->is_int()->get_con(); break;
duke@0 3255 case T_LONG: con = val->is_long()->get_con(); break;
duke@0 3256 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@0 3257 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@0 3258 default: continue; //skip (odd store type)
duke@0 3259 }
duke@0 3260
duke@0 3261 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@0 3262 st->Opcode() == Op_StoreL) {
duke@0 3263 continue; // This StoreL is already optimal.
duke@0 3264 }
duke@0 3265
duke@0 3266 // Store down the constant.
duke@0 3267 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@0 3268
duke@0 3269 intptr_t j = st_off >> LogBytesPerLong;
duke@0 3270
duke@0 3271 if (type == T_INT && st_size == BytesPerInt
duke@0 3272 && (st_off & BytesPerInt) == BytesPerInt) {
duke@0 3273 jlong lcon = tiles[j];
duke@0 3274 if (!Matcher::isSimpleConstant64(lcon) &&
duke@0 3275 st->Opcode() == Op_StoreI) {
duke@0 3276 // This StoreI is already optimal by itself.
duke@0 3277 jint* intcon = (jint*) &tiles[j];
duke@0 3278 intcon[1] = 0; // undo the store_constant()
duke@0 3279
duke@0 3280 // If the previous store is also optimal by itself, back up and
duke@0 3281 // undo the action of the previous loop iteration... if we can.
duke@0 3282 // But if we can't, just let the previous half take care of itself.
duke@0 3283 st = nodes[j];
duke@0 3284 st_off -= BytesPerInt;
duke@0 3285 con = intcon[0];
duke@0 3286 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@0 3287 assert(st_off >= header_size, "still ignoring header");
duke@0 3288 assert(get_store_offset(st, phase) == st_off, "must be");
duke@0 3289 assert(in(i-1) == zmem, "must be");
duke@0 3290 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@0 3291 assert(con == tcon->is_int()->get_con(), "must be");
duke@0 3292 // Undo the effects of the previous loop trip, which swallowed st:
duke@0 3293 intcon[0] = 0; // undo store_constant()
duke@0 3294 set_req(i-1, st); // undo set_req(i, zmem)
duke@0 3295 nodes[j] = NULL; // undo nodes[j] = st
duke@0 3296 --old_subword; // undo ++old_subword
duke@0 3297 }
duke@0 3298 continue; // This StoreI is already optimal.
duke@0 3299 }
duke@0 3300 }
duke@0 3301
duke@0 3302 // This store is not needed.
duke@0 3303 set_req(i, zmem);
duke@0 3304 nodes[j] = st; // record for the moment
duke@0 3305 if (st_size < BytesPerLong) // something has changed
duke@0 3306 ++old_subword; // includes int/float, but who's counting...
duke@0 3307 else ++old_long;
duke@0 3308 }
duke@0 3309
duke@0 3310 if ((old_subword + old_long) == 0)
duke@0 3311 return; // nothing more to do
duke@0 3312
duke@0 3313 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@0 3314 // Be sure to insert them before overlapping non-constant stores.
duke@0 3315 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@0 3316 for (int j = 0; j < num_tiles; j++) {
duke@0 3317 jlong con = tiles[j];
duke@0 3318 jlong init = inits[j];
duke@0 3319 if (con == 0) continue;
duke@0 3320 jint con0, con1; // split the constant, address-wise
duke@0 3321 jint init0, init1; // split the init map, address-wise
duke@0 3322 { union { jlong con; jint intcon[2]; } u;
duke@0 3323 u.con = con;
duke@0 3324 con0 = u.intcon[0];
duke@0 3325 con1 = u.intcon[1];
duke@0 3326 u.con = init;
duke@0 3327 init0 = u.intcon[0];
duke@0 3328 init1 = u.intcon[1];
duke@0 3329 }
duke@0 3330
duke@0 3331 Node* old = nodes[j];
duke@0 3332 assert(old != NULL, "need the prior store");
duke@0 3333 intptr_t offset = (j * BytesPerLong);
duke@0 3334
duke@0 3335 bool split = !Matcher::isSimpleConstant64(con);
duke@0 3336
duke@0 3337 if (offset < header_size) {
duke@0 3338 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@0 3339 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@0 3340 split = true; // only the second word counts
duke@0 3341 // Example: int a[] = { 42 ... }
duke@0 3342 } else if (con0 == 0 && init0 == -1) {
duke@0 3343 split = true; // first word is covered by full inits
duke@0 3344 // Example: int a[] = { ... foo(), 42 ... }
duke@0 3345 } else if (con1 == 0 && init1 == -1) {
duke@0 3346 split = true; // second word is covered by full inits
duke@0 3347 // Example: int a[] = { ... 42, foo() ... }
duke@0 3348 }
duke@0 3349
duke@0 3350 // Here's a case where init0 is neither 0 nor -1:
duke@0 3351 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@0 3352 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@0 3353 // In this case the tile is not split; it is (jlong)42.
duke@0 3354 // The big tile is stored down, and then the foo() value is inserted.
duke@0 3355 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@0 3356
duke@0 3357 Node* ctl = old->in(MemNode::Control);
duke@0 3358 Node* adr = make_raw_address(offset, phase);
duke@0 3359 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 3360
duke@0 3361 // One or two coalesced stores to plop down.
duke@0 3362 Node* st[2];
duke@0 3363 intptr_t off[2];
duke@0 3364 int nst = 0;
duke@0 3365 if (!split) {
duke@0 3366 ++new_long;
duke@0 3367 off[nst] = offset;
coleenp@113 3368 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3369 phase->longcon(con), T_LONG);
duke@0 3370 } else {
duke@0 3371 // Omit either if it is a zero.
duke@0 3372 if (con0 != 0) {
duke@0 3373 ++new_int;
duke@0 3374 off[nst] = offset;
coleenp@113 3375 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3376 phase->intcon(con0), T_INT);
duke@0 3377 }
duke@0 3378 if (con1 != 0) {
duke@0 3379 ++new_int;
duke@0 3380 offset += BytesPerInt;
duke@0 3381 adr = make_raw_address(offset, phase);
duke@0 3382 off[nst] = offset;
coleenp@113 3383 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3384 phase->intcon(con1), T_INT);
duke@0 3385 }
duke@0 3386 }
duke@0 3387
duke@0 3388 // Insert second store first, then the first before the second.
duke@0 3389 // Insert each one just before any overlapping non-constant stores.
duke@0 3390 while (nst > 0) {
duke@0 3391 Node* st1 = st[--nst];
duke@0 3392 C->copy_node_notes_to(st1, old);
duke@0 3393 st1 = phase->transform(st1);
duke@0 3394 offset = off[nst];
duke@0 3395 assert(offset >= header_size, "do not smash header");
duke@0 3396 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@0 3397 guarantee(ins_idx != 0, "must re-insert constant store");
duke@0 3398 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@0 3399 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@0 3400 set_req(--ins_idx, st1);
duke@0 3401 else
duke@0 3402 ins_req(ins_idx, st1);
duke@0 3403 }
duke@0 3404 }
duke@0 3405
duke@0 3406 if (PrintCompilation && WizardMode)
duke@0 3407 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@0 3408 old_subword, old_long, new_int, new_long);
duke@0 3409 if (C->log() != NULL)
duke@0 3410 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@0 3411 old_subword, old_long, new_int, new_long);
duke@0 3412
duke@0 3413 // Clean up any remaining occurrences of zmem:
duke@0 3414 remove_extra_zeroes();
duke@0 3415 }
duke@0 3416
duke@0 3417 // Explore forward from in(start) to find the first fully initialized
duke@0 3418 // word, and return its offset. Skip groups of subword stores which
duke@0 3419 // together initialize full words. If in(start) is itself part of a
duke@0 3420 // fully initialized word, return the offset of in(start). If there
duke@0 3421 // are no following full-word stores, or if something is fishy, return
duke@0 3422 // a negative value.
duke@0 3423 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@0 3424 int int_map = 0;
duke@0 3425 intptr_t int_map_off = 0;
duke@0 3426 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@0 3427
duke@0 3428 for (uint i = start, limit = req(); i < limit; i++) {
duke@0 3429 Node* st = in(i);
duke@0 3430
duke@0 3431 intptr_t st_off = get_store_offset(st, phase);
duke@0 3432 if (st_off < 0) break; // return conservative answer
duke@0 3433
duke@0 3434 int st_size = st->as_Store()->memory_size();
duke@0 3435 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@0 3436 return st_off; // we found a complete word init
duke@0 3437 }
duke@0 3438
duke@0 3439 // update the map:
duke@0 3440
duke@0 3441 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@0 3442 if (this_int_off != int_map_off) {
duke@0 3443 // reset the map:
duke@0 3444 int_map = 0;
duke@0 3445 int_map_off = this_int_off;
duke@0 3446 }
duke@0 3447
duke@0 3448 int subword_off = st_off - this_int_off;
duke@0 3449 int_map |= right_n_bits(st_size) << subword_off;
duke@0 3450 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@0 3451 return this_int_off; // we found a complete word init
duke@0 3452 }
duke@0 3453
duke@0 3454 // Did this store hit or cross the word boundary?
duke@0 3455 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@0 3456 if (next_int_off == this_int_off + BytesPerInt) {
duke@0 3457 // We passed the current int, without fully initializing it.
duke@0 3458 int_map_off = next_int_off;
duke@0 3459 int_map >>= BytesPerInt;
duke@0 3460 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@0 3461 // We passed the current and next int.
duke@0 3462 return this_int_off + BytesPerInt;
duke@0 3463 }
duke@0 3464 }
duke@0 3465
duke@0 3466 return -1;
duke@0 3467 }
duke@0 3468
duke@0 3469
duke@0 3470 // Called when the associated AllocateNode is expanded into CFG.
duke@0 3471 // At this point, we may perform additional optimizations.
duke@0 3472 // Linearize the stores by ascending offset, to make memory
duke@0 3473 // activity as coherent as possible.
duke@0 3474 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@0 3475 intptr_t header_size,
duke@0 3476 Node* size_in_bytes,
duke@0 3477 PhaseGVN* phase) {
duke@0 3478 assert(!is_complete(), "not already complete");
duke@0 3479 assert(stores_are_sane(phase), "");
duke@0 3480 assert(allocation() != NULL, "must be present");
duke@0 3481
duke@0 3482 remove_extra_zeroes();
duke@0 3483
duke@0 3484 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@0 3485 // reduce instruction count for common initialization patterns
duke@0 3486 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@0 3487
duke@0 3488 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3489 Node* inits = zmem; // accumulating a linearized chain of inits
duke@0 3490 #ifdef ASSERT
coleenp@113 3491 intptr_t first_offset = allocation()->minimum_header_size();
coleenp@113 3492 intptr_t last_init_off = first_offset; // previous init offset
coleenp@113 3493 intptr_t last_init_end = first_offset; // previous init offset+size
coleenp@113 3494 intptr_t last_tile_end = first_offset; // previous tile offset+size
duke@0 3495 #endif
duke@0 3496 intptr_t zeroes_done = header_size;
duke@0 3497
duke@0 3498 bool do_zeroing = true; // we might give up if inits are very sparse
duke@0 3499 int big_init_gaps = 0; // how many large gaps have we seen?
duke@0 3500
duke@0 3501 if (ZeroTLAB) do_zeroing = false;
duke@0 3502 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@0 3503
duke@0 3504 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3505 Node* st = in(i);