src/share/vm/opto/memnode.cpp

Wed, 03 Jun 2015 14:22:57 +0200

author
roland
date
Wed, 03 Jun 2015 14:22:57 +0200
changeset 7859
c1c199dde5c9
parent 7858
55d07ec5bde4
child 7994
04ff2f6cd0eb
child 8368
32b682649973
permissions
-rw-r--r--

8077504: Unsafe load can loose control dependency and cause crash
Summary: Node::depends_only_on_test() should return false for Unsafe loads
Reviewed-by: kvn, adinn

duke@435 1 /*
drchase@6680 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
duke@435 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435 4 *
duke@435 5 * This code is free software; you can redistribute it and/or modify it
duke@435 6 * under the terms of the GNU General Public License version 2 only, as
duke@435 7 * published by the Free Software Foundation.
duke@435 8 *
duke@435 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@435 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@435 13 * accompanied this code).
duke@435 14 *
duke@435 15 * You should have received a copy of the GNU General Public License version
duke@435 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@435 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435 18 *
trims@1907 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907 20 * or visit www.oracle.com if you need additional information or have any
trims@1907 21 * questions.
duke@435 22 *
duke@435 23 */
duke@435 24
stefank@2314 25 #include "precompiled.hpp"
stefank@2314 26 #include "classfile/systemDictionary.hpp"
stefank@2314 27 #include "compiler/compileLog.hpp"
stefank@2314 28 #include "memory/allocation.inline.hpp"
stefank@2314 29 #include "oops/objArrayKlass.hpp"
stefank@2314 30 #include "opto/addnode.hpp"
stefank@2314 31 #include "opto/cfgnode.hpp"
stefank@2314 32 #include "opto/compile.hpp"
stefank@2314 33 #include "opto/connode.hpp"
stefank@2314 34 #include "opto/loopnode.hpp"
stefank@2314 35 #include "opto/machnode.hpp"
stefank@2314 36 #include "opto/matcher.hpp"
stefank@2314 37 #include "opto/memnode.hpp"
stefank@2314 38 #include "opto/mulnode.hpp"
stefank@2314 39 #include "opto/phaseX.hpp"
stefank@2314 40 #include "opto/regmask.hpp"
stefank@2314 41
duke@435 42 // Portions of code courtesy of Clifford Click
duke@435 43
duke@435 44 // Optimization - Graph Style
duke@435 45
kvn@509 46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
kvn@509 47
duke@435 48 //=============================================================================
duke@435 49 uint MemNode::size_of() const { return sizeof(*this); }
duke@435 50
duke@435 51 const TypePtr *MemNode::adr_type() const {
duke@435 52 Node* adr = in(Address);
duke@435 53 const TypePtr* cross_check = NULL;
duke@435 54 DEBUG_ONLY(cross_check = _adr_type);
duke@435 55 return calculate_adr_type(adr->bottom_type(), cross_check);
duke@435 56 }
duke@435 57
duke@435 58 #ifndef PRODUCT
duke@435 59 void MemNode::dump_spec(outputStream *st) const {
duke@435 60 if (in(Address) == NULL) return; // node is dead
duke@435 61 #ifndef ASSERT
duke@435 62 // fake the missing field
duke@435 63 const TypePtr* _adr_type = NULL;
duke@435 64 if (in(Address) != NULL)
duke@435 65 _adr_type = in(Address)->bottom_type()->isa_ptr();
duke@435 66 #endif
duke@435 67 dump_adr_type(this, _adr_type, st);
duke@435 68
duke@435 69 Compile* C = Compile::current();
duke@435 70 if( C->alias_type(_adr_type)->is_volatile() )
duke@435 71 st->print(" Volatile!");
duke@435 72 }
duke@435 73
duke@435 74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
duke@435 75 st->print(" @");
duke@435 76 if (adr_type == NULL) {
duke@435 77 st->print("NULL");
duke@435 78 } else {
duke@435 79 adr_type->dump_on(st);
duke@435 80 Compile* C = Compile::current();
duke@435 81 Compile::AliasType* atp = NULL;
duke@435 82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
duke@435 83 if (atp == NULL)
duke@435 84 st->print(", idx=?\?;");
duke@435 85 else if (atp->index() == Compile::AliasIdxBot)
duke@435 86 st->print(", idx=Bot;");
duke@435 87 else if (atp->index() == Compile::AliasIdxTop)
duke@435 88 st->print(", idx=Top;");
duke@435 89 else if (atp->index() == Compile::AliasIdxRaw)
duke@435 90 st->print(", idx=Raw;");
duke@435 91 else {
duke@435 92 ciField* field = atp->field();
duke@435 93 if (field) {
duke@435 94 st->print(", name=");
duke@435 95 field->print_name_on(st);
duke@435 96 }
duke@435 97 st->print(", idx=%d;", atp->index());
duke@435 98 }
duke@435 99 }
duke@435 100 }
duke@435 101
duke@435 102 extern void print_alias_types();
duke@435 103
duke@435 104 #endif
duke@435 105
kvn@5110 106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
kvn@5110 107 assert((t_oop != NULL), "sanity");
kvn@5110 108 bool is_instance = t_oop->is_known_instance_field();
kvn@5110 109 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
kvn@5110 110 (load != NULL) && load->is_Load() &&
kvn@5110 111 (phase->is_IterGVN() != NULL);
kvn@5110 112 if (!(is_instance || is_boxed_value_load))
kvn@509 113 return mchain; // don't try to optimize non-instance types
kvn@5110 114 uint instance_id = t_oop->instance_id();
kvn@688 115 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
kvn@509 116 Node *prev = NULL;
kvn@509 117 Node *result = mchain;
kvn@509 118 while (prev != result) {
kvn@509 119 prev = result;
kvn@688 120 if (result == start_mem)
twisti@1040 121 break; // hit one of our sentinels
kvn@509 122 // skip over a call which does not affect this memory slice
kvn@509 123 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
kvn@509 124 Node *proj_in = result->in(0);
kvn@688 125 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
twisti@1040 126 break; // hit one of our sentinels
kvn@688 127 } else if (proj_in->is_Call()) {
kvn@509 128 CallNode *call = proj_in->as_Call();
kvn@5110 129 if (!call->may_modify(t_oop, phase)) { // returns false for instances
kvn@509 130 result = call->in(TypeFunc::Memory);
kvn@509 131 }
kvn@509 132 } else if (proj_in->is_Initialize()) {
kvn@509 133 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
kvn@509 134 // Stop if this is the initialization for the object instance which
kvn@509 135 // which contains this memory slice, otherwise skip over it.
kvn@5110 136 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
kvn@5110 137 break;
kvn@5110 138 }
kvn@5110 139 if (is_instance) {
kvn@509 140 result = proj_in->in(TypeFunc::Memory);
kvn@5110 141 } else if (is_boxed_value_load) {
kvn@5110 142 Node* klass = alloc->in(AllocateNode::KlassNode);
kvn@5110 143 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
kvn@5110 144 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
kvn@5110 145 result = proj_in->in(TypeFunc::Memory); // not related allocation
kvn@5110 146 }
kvn@509 147 }
kvn@509 148 } else if (proj_in->is_MemBar()) {
kvn@509 149 result = proj_in->in(TypeFunc::Memory);
kvn@688 150 } else {
kvn@688 151 assert(false, "unexpected projection");
kvn@509 152 }
kvn@1535 153 } else if (result->is_ClearArray()) {
kvn@5110 154 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
kvn@1535 155 // Can not bypass initialization of the instance
kvn@1535 156 // we are looking for.
kvn@1535 157 break;
kvn@1535 158 }
kvn@1535 159 // Otherwise skip it (the call updated 'result' value).
kvn@509 160 } else if (result->is_MergeMem()) {
kvn@5110 161 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
kvn@509 162 }
kvn@509 163 }
kvn@509 164 return result;
kvn@509 165 }
kvn@509 166
kvn@5110 167 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
kvn@5110 168 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
kvn@5110 169 if (t_oop == NULL)
kvn@5110 170 return mchain; // don't try to optimize non-oop types
kvn@5110 171 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
kvn@5110 172 bool is_instance = t_oop->is_known_instance_field();
kvn@509 173 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@509 174 if (is_instance && igvn != NULL && result->is_Phi()) {
kvn@509 175 PhiNode *mphi = result->as_Phi();
kvn@509 176 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
kvn@509 177 const TypePtr *t = mphi->adr_type();
kvn@598 178 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
kvn@658 179 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
kvn@682 180 t->is_oopptr()->cast_to_exactness(true)
kvn@682 181 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
kvn@682 182 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
kvn@509 183 // clone the Phi with our address type
kvn@509 184 result = mphi->split_out_instance(t_adr, igvn);
kvn@509 185 } else {
kvn@509 186 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
kvn@509 187 }
kvn@509 188 }
kvn@509 189 return result;
kvn@509 190 }
kvn@509 191
kvn@499 192 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
kvn@499 193 uint alias_idx = phase->C->get_alias_index(tp);
kvn@499 194 Node *mem = mmem;
kvn@499 195 #ifdef ASSERT
kvn@499 196 {
kvn@499 197 // Check that current type is consistent with the alias index used during graph construction
kvn@499 198 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
kvn@499 199 bool consistent = adr_check == NULL || adr_check->empty() ||
kvn@499 200 phase->C->must_alias(adr_check, alias_idx );
kvn@499 201 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
kvn@499 202 if( !consistent && adr_check != NULL && !adr_check->empty() &&
rasbold@604 203 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
kvn@499 204 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
kvn@499 205 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
kvn@499 206 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
kvn@499 207 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
kvn@499 208 // don't assert if it is dead code.
kvn@499 209 consistent = true;
kvn@499 210 }
kvn@499 211 if( !consistent ) {
kvn@499 212 st->print("alias_idx==%d, adr_check==", alias_idx);
kvn@499 213 if( adr_check == NULL ) {
kvn@499 214 st->print("NULL");
kvn@499 215 } else {
kvn@499 216 adr_check->dump();
kvn@499 217 }
kvn@499 218 st->cr();
kvn@499 219 print_alias_types();
kvn@499 220 assert(consistent, "adr_check must match alias idx");
kvn@499 221 }
kvn@499 222 }
kvn@499 223 #endif
never@2170 224 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
kvn@499 225 // means an array I have not precisely typed yet. Do not do any
kvn@499 226 // alias stuff with it any time soon.
never@2170 227 const TypeOopPtr *toop = tp->isa_oopptr();
kvn@499 228 if( tp->base() != Type::AnyPtr &&
never@2170 229 !(toop &&
never@2170 230 toop->klass() != NULL &&
never@2170 231 toop->klass()->is_java_lang_Object() &&
never@2170 232 toop->offset() == Type::OffsetBot) ) {
kvn@499 233 // compress paths and change unreachable cycles to TOP
kvn@499 234 // If not, we can update the input infinitely along a MergeMem cycle
kvn@499 235 // Equivalent code in PhiNode::Ideal
kvn@499 236 Node* m = phase->transform(mmem);
twisti@1040 237 // If transformed to a MergeMem, get the desired slice
kvn@499 238 // Otherwise the returned node represents memory for every slice
kvn@499 239 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
kvn@499 240 // Update input if it is progress over what we have now
kvn@499 241 }
kvn@499 242 return mem;
kvn@499 243 }
kvn@499 244
duke@435 245 //--------------------------Ideal_common---------------------------------------
duke@435 246 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
duke@435 247 // Unhook non-raw memories from complete (macro-expanded) initializations.
duke@435 248 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
duke@435 249 // If our control input is a dead region, kill all below the region
duke@435 250 Node *ctl = in(MemNode::Control);
duke@435 251 if (ctl && remove_dead_region(phase, can_reshape))
duke@435 252 return this;
kvn@740 253 ctl = in(MemNode::Control);
kvn@740 254 // Don't bother trying to transform a dead node
kvn@4695 255 if (ctl && ctl->is_top()) return NodeSentinel;
duke@435 256
kvn@1143 257 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@1143 258 // Wait if control on the worklist.
kvn@1143 259 if (ctl && can_reshape && igvn != NULL) {
kvn@1143 260 Node* bol = NULL;
kvn@1143 261 Node* cmp = NULL;
kvn@1143 262 if (ctl->in(0)->is_If()) {
kvn@1143 263 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
kvn@1143 264 bol = ctl->in(0)->in(1);
kvn@1143 265 if (bol->is_Bool())
kvn@1143 266 cmp = ctl->in(0)->in(1)->in(1);
kvn@1143 267 }
kvn@1143 268 if (igvn->_worklist.member(ctl) ||
kvn@1143 269 (bol != NULL && igvn->_worklist.member(bol)) ||
kvn@1143 270 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
kvn@1143 271 // This control path may be dead.
kvn@1143 272 // Delay this memory node transformation until the control is processed.
kvn@1143 273 phase->is_IterGVN()->_worklist.push(this);
kvn@1143 274 return NodeSentinel; // caller will return NULL
kvn@1143 275 }
kvn@1143 276 }
duke@435 277 // Ignore if memory is dead, or self-loop
duke@435 278 Node *mem = in(MemNode::Memory);
kvn@4695 279 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
kvn@4695 280 assert(mem != this, "dead loop in MemNode::Ideal");
duke@435 281
kvn@3320 282 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
kvn@3320 283 // This memory slice may be dead.
kvn@3320 284 // Delay this mem node transformation until the memory is processed.
kvn@3320 285 phase->is_IterGVN()->_worklist.push(this);
kvn@3320 286 return NodeSentinel; // caller will return NULL
kvn@3320 287 }
kvn@3320 288
duke@435 289 Node *address = in(MemNode::Address);
kvn@4695 290 const Type *t_adr = phase->type(address);
kvn@4695 291 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
kvn@4695 292
kvn@4695 293 if (can_reshape && igvn != NULL &&
kvn@2181 294 (igvn->_worklist.member(address) ||
kvn@4695 295 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
kvn@855 296 // The address's base and type may change when the address is processed.
kvn@855 297 // Delay this mem node transformation until the address is processed.
kvn@855 298 phase->is_IterGVN()->_worklist.push(this);
kvn@855 299 return NodeSentinel; // caller will return NULL
kvn@855 300 }
kvn@855 301
kvn@1497 302 // Do NOT remove or optimize the next lines: ensure a new alias index
kvn@1497 303 // is allocated for an oop pointer type before Escape Analysis.
kvn@1497 304 // Note: C++ will not remove it since the call has side effect.
kvn@4695 305 if (t_adr->isa_oopptr()) {
kvn@1497 306 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
kvn@1497 307 }
kvn@1497 308
kvn@1143 309 Node* base = NULL;
kvn@6657 310 if (address->is_AddP()) {
kvn@1143 311 base = address->in(AddPNode::Base);
kvn@6657 312 }
kvn@4695 313 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
kvn@4695 314 !t_adr->isa_rawptr()) {
kvn@4695 315 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
kvn@6657 316 // Skip this node optimization if its address has TOP base.
kvn@6657 317 return NodeSentinel; // caller will return NULL
kvn@4695 318 }
kvn@1143 319
duke@435 320 // Avoid independent memory operations
duke@435 321 Node* old_mem = mem;
duke@435 322
kvn@471 323 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@471 324 // initializations was removed. After macro-expansion all stores catched
kvn@471 325 // by Initialize node became raw stores and there is no information
kvn@471 326 // which memory slices they modify. So it is unsafe to move any memory
kvn@471 327 // operation above these stores. Also in most cases hooked non-raw memories
kvn@471 328 // were already unhooked by using information from detect_ptr_independence()
kvn@471 329 // and find_previous_store().
duke@435 330
duke@435 331 if (mem->is_MergeMem()) {
duke@435 332 MergeMemNode* mmem = mem->as_MergeMem();
duke@435 333 const TypePtr *tp = t_adr->is_ptr();
kvn@499 334
kvn@499 335 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@435 336 }
duke@435 337
duke@435 338 if (mem != old_mem) {
duke@435 339 set_req(MemNode::Memory, mem);
roland@4657 340 if (can_reshape && old_mem->outcnt() == 0) {
roland@4657 341 igvn->_worklist.push(old_mem);
roland@4657 342 }
kvn@740 343 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
duke@435 344 return this;
duke@435 345 }
duke@435 346
duke@435 347 // let the subclass continue analyzing...
duke@435 348 return NULL;
duke@435 349 }
duke@435 350
duke@435 351 // Helper function for proving some simple control dominations.
kvn@554 352 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
duke@435 353 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@435 354 // is not a constant (dominated by the method's StartNode).
duke@435 355 // Used by MemNode::find_previous_store to prove that the
duke@435 356 // control input of a memory operation predates (dominates)
duke@435 357 // an allocation it wants to look past.
kvn@554 358 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
kvn@554 359 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
kvn@554 360 return false; // Conservative answer for dead code
kvn@554 361
kvn@628 362 // Check 'dom'. Skip Proj and CatchProj nodes.
kvn@554 363 dom = dom->find_exact_control(dom);
kvn@554 364 if (dom == NULL || dom->is_top())
kvn@554 365 return false; // Conservative answer for dead code
kvn@554 366
kvn@628 367 if (dom == sub) {
kvn@628 368 // For the case when, for example, 'sub' is Initialize and the original
kvn@628 369 // 'dom' is Proj node of the 'sub'.
kvn@628 370 return false;
kvn@628 371 }
kvn@628 372
kvn@590 373 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
kvn@554 374 return true;
kvn@554 375
kvn@554 376 // 'dom' dominates 'sub' if its control edge and control edges
kvn@554 377 // of all its inputs dominate or equal to sub's control edge.
kvn@554 378
kvn@554 379 // Currently 'sub' is either Allocate, Initialize or Start nodes.
kvn@598 380 // Or Region for the check in LoadNode::Ideal();
kvn@598 381 // 'sub' should have sub->in(0) != NULL.
kvn@598 382 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
kvn@5110 383 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
kvn@554 384
kvn@554 385 // Get control edge of 'sub'.
kvn@628 386 Node* orig_sub = sub;
kvn@554 387 sub = sub->find_exact_control(sub->in(0));
kvn@554 388 if (sub == NULL || sub->is_top())
kvn@554 389 return false; // Conservative answer for dead code
kvn@554 390
kvn@554 391 assert(sub->is_CFG(), "expecting control");
kvn@554 392
kvn@554 393 if (sub == dom)
kvn@554 394 return true;
kvn@554 395
kvn@554 396 if (sub->is_Start() || sub->is_Root())
kvn@554 397 return false;
kvn@554 398
kvn@554 399 {
kvn@554 400 // Check all control edges of 'dom'.
kvn@554 401
kvn@554 402 ResourceMark rm;
kvn@554 403 Arena* arena = Thread::current()->resource_area();
kvn@554 404 Node_List nlist(arena);
kvn@554 405 Unique_Node_List dom_list(arena);
kvn@554 406
kvn@554 407 dom_list.push(dom);
kvn@554 408 bool only_dominating_controls = false;
kvn@554 409
kvn@554 410 for (uint next = 0; next < dom_list.size(); next++) {
kvn@554 411 Node* n = dom_list.at(next);
kvn@628 412 if (n == orig_sub)
kvn@628 413 return false; // One of dom's inputs dominated by sub.
kvn@554 414 if (!n->is_CFG() && n->pinned()) {
kvn@554 415 // Check only own control edge for pinned non-control nodes.
kvn@554 416 n = n->find_exact_control(n->in(0));
kvn@554 417 if (n == NULL || n->is_top())
kvn@554 418 return false; // Conservative answer for dead code
kvn@554 419 assert(n->is_CFG(), "expecting control");
kvn@628 420 dom_list.push(n);
kvn@628 421 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
kvn@554 422 only_dominating_controls = true;
kvn@554 423 } else if (n->is_CFG()) {
kvn@554 424 if (n->dominates(sub, nlist))
kvn@554 425 only_dominating_controls = true;
kvn@554 426 else
kvn@554 427 return false;
kvn@554 428 } else {
kvn@554 429 // First, own control edge.
kvn@554 430 Node* m = n->find_exact_control(n->in(0));
kvn@590 431 if (m != NULL) {
kvn@590 432 if (m->is_top())
kvn@590 433 return false; // Conservative answer for dead code
kvn@590 434 dom_list.push(m);
kvn@590 435 }
kvn@554 436 // Now, the rest of edges.
kvn@554 437 uint cnt = n->req();
kvn@554 438 for (uint i = 1; i < cnt; i++) {
kvn@554 439 m = n->find_exact_control(n->in(i));
kvn@554 440 if (m == NULL || m->is_top())
kvn@554 441 continue;
kvn@554 442 dom_list.push(m);
duke@435 443 }
duke@435 444 }
duke@435 445 }
kvn@554 446 return only_dominating_controls;
duke@435 447 }
duke@435 448 }
duke@435 449
duke@435 450 //---------------------detect_ptr_independence---------------------------------
duke@435 451 // Used by MemNode::find_previous_store to prove that two base
duke@435 452 // pointers are never equal.
duke@435 453 // The pointers are accompanied by their associated allocations,
duke@435 454 // if any, which have been previously discovered by the caller.
duke@435 455 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@435 456 Node* p2, AllocateNode* a2,
duke@435 457 PhaseTransform* phase) {
duke@435 458 // Attempt to prove that these two pointers cannot be aliased.
duke@435 459 // They may both manifestly be allocations, and they should differ.
duke@435 460 // Or, if they are not both allocations, they can be distinct constants.
duke@435 461 // Otherwise, one is an allocation and the other a pre-existing value.
duke@435 462 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@435 463 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@435 464 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@435 465 return (a1 != a2);
duke@435 466 } else if (a1 != NULL) { // one allocation a1
duke@435 467 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
kvn@554 468 return all_controls_dominate(p2, a1);
duke@435 469 } else { //(a2 != NULL) // one allocation a2
kvn@554 470 return all_controls_dominate(p1, a2);
duke@435 471 }
duke@435 472 return false;
duke@435 473 }
duke@435 474
duke@435 475
duke@435 476 // The logic for reordering loads and stores uses four steps:
duke@435 477 // (a) Walk carefully past stores and initializations which we
duke@435 478 // can prove are independent of this load.
duke@435 479 // (b) Observe that the next memory state makes an exact match
duke@435 480 // with self (load or store), and locate the relevant store.
duke@435 481 // (c) Ensure that, if we were to wire self directly to the store,
duke@435 482 // the optimizer would fold it up somehow.
duke@435 483 // (d) Do the rewiring, and return, depending on some other part of
duke@435 484 // the optimizer to fold up the load.
duke@435 485 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@435 486 // specific to loads and stores, so they are handled by the callers.
duke@435 487 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@435 488 //
duke@435 489 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@435 490 Node* ctrl = in(MemNode::Control);
duke@435 491 Node* adr = in(MemNode::Address);
duke@435 492 intptr_t offset = 0;
duke@435 493 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@435 494 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@435 495
duke@435 496 if (offset == Type::OffsetBot)
duke@435 497 return NULL; // cannot unalias unless there are precise offsets
duke@435 498
kvn@509 499 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
kvn@509 500
duke@435 501 intptr_t size_in_bytes = memory_size();
duke@435 502
duke@435 503 Node* mem = in(MemNode::Memory); // start searching here...
duke@435 504
duke@435 505 int cnt = 50; // Cycle limiter
duke@435 506 for (;;) { // While we can dance past unrelated stores...
duke@435 507 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@435 508
duke@435 509 if (mem->is_Store()) {
duke@435 510 Node* st_adr = mem->in(MemNode::Address);
duke@435 511 intptr_t st_offset = 0;
duke@435 512 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@435 513 if (st_base == NULL)
duke@435 514 break; // inscrutable pointer
duke@435 515 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@435 516 const int MAX_STORE = BytesPerLong;
duke@435 517 if (st_offset >= offset + size_in_bytes ||
duke@435 518 st_offset <= offset - MAX_STORE ||
duke@435 519 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@435 520 // Success: The offsets are provably independent.
duke@435 521 // (You may ask, why not just test st_offset != offset and be done?
duke@435 522 // The answer is that stores of different sizes can co-exist
duke@435 523 // in the same sequence of RawMem effects. We sometimes initialize
duke@435 524 // a whole 'tile' of array elements with a single jint or jlong.)
duke@435 525 mem = mem->in(MemNode::Memory);
duke@435 526 continue; // (a) advance through independent store memory
duke@435 527 }
duke@435 528 }
duke@435 529 if (st_base != base &&
duke@435 530 detect_ptr_independence(base, alloc,
duke@435 531 st_base,
duke@435 532 AllocateNode::Ideal_allocation(st_base, phase),
duke@435 533 phase)) {
duke@435 534 // Success: The bases are provably independent.
duke@435 535 mem = mem->in(MemNode::Memory);
duke@435 536 continue; // (a) advance through independent store memory
duke@435 537 }
duke@435 538
duke@435 539 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@435 540 // since we have not managed to prove 'this' and 'mem' independent.
duke@435 541 if (st_base == base && st_offset == offset) {
duke@435 542 return mem; // let caller handle steps (c), (d)
duke@435 543 }
duke@435 544
duke@435 545 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@435 546 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@435 547 AllocateNode* st_alloc = st_init->allocation();
duke@435 548 if (st_alloc == NULL)
duke@435 549 break; // something degenerated
duke@435 550 bool known_identical = false;
duke@435 551 bool known_independent = false;
duke@435 552 if (alloc == st_alloc)
duke@435 553 known_identical = true;
duke@435 554 else if (alloc != NULL)
duke@435 555 known_independent = true;
kvn@554 556 else if (all_controls_dominate(this, st_alloc))
duke@435 557 known_independent = true;
duke@435 558
duke@435 559 if (known_independent) {
duke@435 560 // The bases are provably independent: Either they are
duke@435 561 // manifestly distinct allocations, or else the control
duke@435 562 // of this load dominates the store's allocation.
duke@435 563 int alias_idx = phase->C->get_alias_index(adr_type());
duke@435 564 if (alias_idx == Compile::AliasIdxRaw) {
duke@435 565 mem = st_alloc->in(TypeFunc::Memory);
duke@435 566 } else {
duke@435 567 mem = st_init->memory(alias_idx);
duke@435 568 }
duke@435 569 continue; // (a) advance through independent store memory
duke@435 570 }
duke@435 571
duke@435 572 // (b) at this point, if we are not looking at a store initializing
duke@435 573 // the same allocation we are loading from, we lose.
duke@435 574 if (known_identical) {
duke@435 575 // From caller, can_see_stored_value will consult find_captured_store.
duke@435 576 return mem; // let caller handle steps (c), (d)
duke@435 577 }
duke@435 578
kvn@658 579 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
kvn@509 580 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
kvn@509 581 if (mem->is_Proj() && mem->in(0)->is_Call()) {
kvn@509 582 CallNode *call = mem->in(0)->as_Call();
kvn@509 583 if (!call->may_modify(addr_t, phase)) {
kvn@509 584 mem = call->in(TypeFunc::Memory);
kvn@509 585 continue; // (a) advance through independent call memory
kvn@509 586 }
kvn@509 587 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
kvn@509 588 mem = mem->in(0)->in(TypeFunc::Memory);
kvn@509 589 continue; // (a) advance through independent MemBar memory
kvn@1535 590 } else if (mem->is_ClearArray()) {
kvn@1535 591 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
kvn@1535 592 // (the call updated 'mem' value)
kvn@1535 593 continue; // (a) advance through independent allocation memory
kvn@1535 594 } else {
kvn@1535 595 // Can not bypass initialization of the instance
kvn@1535 596 // we are looking for.
kvn@1535 597 return mem;
kvn@1535 598 }
kvn@509 599 } else if (mem->is_MergeMem()) {
kvn@509 600 int alias_idx = phase->C->get_alias_index(adr_type());
kvn@509 601 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@509 602 continue; // (a) advance through independent MergeMem memory
kvn@509 603 }
duke@435 604 }
duke@435 605
duke@435 606 // Unless there is an explicit 'continue', we must bail out here,
duke@435 607 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@435 608 break;
duke@435 609 }
duke@435 610
duke@435 611 return NULL; // bail out
duke@435 612 }
duke@435 613
duke@435 614 //----------------------calculate_adr_type-------------------------------------
duke@435 615 // Helper function. Notices when the given type of address hits top or bottom.
duke@435 616 // Also, asserts a cross-check of the type against the expected address type.
duke@435 617 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@435 618 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@435 619 #ifdef PRODUCT
duke@435 620 cross_check = NULL;
duke@435 621 #else
duke@435 622 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@435 623 #endif
duke@435 624 const TypePtr* tp = t->isa_ptr();
duke@435 625 if (tp == NULL) {
duke@435 626 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@435 627 return TypePtr::BOTTOM; // touches lots of memory
duke@435 628 } else {
duke@435 629 #ifdef ASSERT
duke@435 630 // %%%% [phh] We don't check the alias index if cross_check is
duke@435 631 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@435 632 if (cross_check != NULL &&
duke@435 633 cross_check != TypePtr::BOTTOM &&
duke@435 634 cross_check != TypeRawPtr::BOTTOM) {
duke@435 635 // Recheck the alias index, to see if it has changed (due to a bug).
duke@435 636 Compile* C = Compile::current();
duke@435 637 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@435 638 "must stay in the original alias category");
duke@435 639 // The type of the address must be contained in the adr_type,
duke@435 640 // disregarding "null"-ness.
duke@435 641 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@435 642 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
roland@6313 643 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
duke@435 644 "real address must not escape from expected memory type");
duke@435 645 }
duke@435 646 #endif
duke@435 647 return tp;
duke@435 648 }
duke@435 649 }
duke@435 650
duke@435 651 //------------------------adr_phi_is_loop_invariant----------------------------
duke@435 652 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@435 653 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@435 654 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@435 655 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@435 656 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@435 657 // with the same data. This implies that any path into the loop already
duke@435 658 // includes such a CastPP, and so the original cast, whatever its input,
duke@435 659 // must be covered by an equivalent cast, with an earlier control input.
duke@435 660 ResourceMark rm;
duke@435 661
duke@435 662 // The loop entry input of the phi should be the unique dominating
duke@435 663 // node for every Phi/CastPP in the loop.
duke@435 664 Unique_Node_List closure;
duke@435 665 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@435 666
duke@435 667 // Add the phi node and the cast to the worklist.
duke@435 668 Unique_Node_List worklist;
duke@435 669 worklist.push(adr_phi);
duke@435 670 if( cast != NULL ){
duke@435 671 if( !cast->is_ConstraintCast() ) return false;
duke@435 672 worklist.push(cast);
duke@435 673 }
duke@435 674
duke@435 675 // Begin recursive walk of phi nodes.
duke@435 676 while( worklist.size() ){
duke@435 677 // Take a node off the worklist
duke@435 678 Node *n = worklist.pop();
duke@435 679 if( !closure.member(n) ){
duke@435 680 // Add it to the closure.
duke@435 681 closure.push(n);
duke@435 682 // Make a sanity check to ensure we don't waste too much time here.
duke@435 683 if( closure.size() > 20) return false;
duke@435 684 // This node is OK if:
duke@435 685 // - it is a cast of an identical value
duke@435 686 // - or it is a phi node (then we add its inputs to the worklist)
duke@435 687 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@435 688 if( n->is_ConstraintCast() ){
duke@435 689 worklist.push(n->in(1));
duke@435 690 } else if( n->is_Phi() ) {
duke@435 691 for( uint i = 1; i < n->req(); i++ ) {
duke@435 692 worklist.push(n->in(i));
duke@435 693 }
duke@435 694 } else {
duke@435 695 return false;
duke@435 696 }
duke@435 697 }
duke@435 698 }
duke@435 699
duke@435 700 // Quit when the worklist is empty, and we've found no offending nodes.
duke@435 701 return true;
duke@435 702 }
duke@435 703
duke@435 704 //------------------------------Ideal_DU_postCCP-------------------------------
duke@435 705 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@435 706 // going away in this pass and we need to make this memory op depend on the
duke@435 707 // gating null check.
kvn@598 708 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
kvn@598 709 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
kvn@598 710 }
duke@435 711
duke@435 712 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@435 713 // some sense; we get to keep around the knowledge that an oop is not-null
duke@435 714 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@435 715 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@435 716 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@435 717 // some of the more trivial cases in the optimizer. Removing more useless
duke@435 718 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@435 719 // up on this approach. CNC 10/20/2000
kvn@598 720 // This static method may be called not from MemNode (EncodePNode calls it).
kvn@598 721 // Only the control edge of the node 'n' might be updated.
kvn@598 722 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
duke@435 723 Node *skipped_cast = NULL;
duke@435 724 // Need a null check? Regular static accesses do not because they are
duke@435 725 // from constant addresses. Array ops are gated by the range check (which
duke@435 726 // always includes a NULL check). Just check field ops.
kvn@598 727 if( n->in(MemNode::Control) == NULL ) {
duke@435 728 // Scan upwards for the highest location we can place this memory op.
duke@435 729 while( true ) {
duke@435 730 switch( adr->Opcode() ) {
duke@435 731
duke@435 732 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@435 733 adr = adr->in(AddPNode::Base);
duke@435 734 continue;
duke@435 735
coleenp@548 736 case Op_DecodeN: // No change to NULL-ness, so peek thru
roland@4159 737 case Op_DecodeNKlass:
coleenp@548 738 adr = adr->in(1);
coleenp@548 739 continue;
coleenp@548 740
kvn@3842 741 case Op_EncodeP:
roland@4159 742 case Op_EncodePKlass:
kvn@3842 743 // EncodeP node's control edge could be set by this method
kvn@3842 744 // when EncodeP node depends on CastPP node.
kvn@3842 745 //
kvn@3842 746 // Use its control edge for memory op because EncodeP may go away
kvn@3842 747 // later when it is folded with following or preceding DecodeN node.
kvn@3842 748 if (adr->in(0) == NULL) {
kvn@3842 749 // Keep looking for cast nodes.
kvn@3842 750 adr = adr->in(1);
kvn@3842 751 continue;
kvn@3842 752 }
kvn@3842 753 ccp->hash_delete(n);
kvn@3842 754 n->set_req(MemNode::Control, adr->in(0));
kvn@3842 755 ccp->hash_insert(n);
kvn@3842 756 return n;
kvn@3842 757
duke@435 758 case Op_CastPP:
duke@435 759 // If the CastPP is useless, just peek on through it.
duke@435 760 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@435 761 // Remember the cast that we've peeked though. If we peek
duke@435 762 // through more than one, then we end up remembering the highest
duke@435 763 // one, that is, if in a loop, the one closest to the top.
duke@435 764 skipped_cast = adr;
duke@435 765 adr = adr->in(1);
duke@435 766 continue;
duke@435 767 }
duke@435 768 // CastPP is going away in this pass! We need this memory op to be
duke@435 769 // control-dependent on the test that is guarding the CastPP.
kvn@598 770 ccp->hash_delete(n);
kvn@598 771 n->set_req(MemNode::Control, adr->in(0));
kvn@598 772 ccp->hash_insert(n);
kvn@598 773 return n;
duke@435 774
duke@435 775 case Op_Phi:
duke@435 776 // Attempt to float above a Phi to some dominating point.
duke@435 777 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@435 778 // If we've already peeked through a Cast (which could have set the
duke@435 779 // control), we can't float above a Phi, because the skipped Cast
duke@435 780 // may not be loop invariant.
duke@435 781 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@435 782 adr = adr->in(1);
duke@435 783 continue;
duke@435 784 }
duke@435 785 }
duke@435 786
duke@435 787 // Intentional fallthrough!
duke@435 788
duke@435 789 // No obvious dominating point. The mem op is pinned below the Phi
duke@435 790 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@435 791 // then the mem op can float, but not indefinitely. It must be pinned
duke@435 792 // behind the controls leading to the Phi.
duke@435 793 case Op_CheckCastPP:
duke@435 794 // These usually stick around to change address type, however a
duke@435 795 // useless one can be elided and we still need to pick up a control edge
duke@435 796 if (adr->in(0) == NULL) {
duke@435 797 // This CheckCastPP node has NO control and is likely useless. But we
duke@435 798 // need check further up the ancestor chain for a control input to keep
duke@435 799 // the node in place. 4959717.
duke@435 800 skipped_cast = adr;
duke@435 801 adr = adr->in(1);
duke@435 802 continue;
duke@435 803 }
kvn@598 804 ccp->hash_delete(n);
kvn@598 805 n->set_req(MemNode::Control, adr->in(0));
kvn@598 806 ccp->hash_insert(n);
kvn@598 807 return n;
duke@435 808
duke@435 809 // List of "safe" opcodes; those that implicitly block the memory
duke@435 810 // op below any null check.
duke@435 811 case Op_CastX2P: // no null checks on native pointers
duke@435 812 case Op_Parm: // 'this' pointer is not null
duke@435 813 case Op_LoadP: // Loading from within a klass
coleenp@548 814 case Op_LoadN: // Loading from within a klass
duke@435 815 case Op_LoadKlass: // Loading from within a klass
kvn@599 816 case Op_LoadNKlass: // Loading from within a klass
duke@435 817 case Op_ConP: // Loading from a klass
kvn@598 818 case Op_ConN: // Loading from a klass
roland@4159 819 case Op_ConNKlass: // Loading from a klass
duke@435 820 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@435 821 case Op_Con: // Reading from TLS
duke@435 822 case Op_CMoveP: // CMoveP is pinned
kvn@599 823 case Op_CMoveN: // CMoveN is pinned
duke@435 824 break; // No progress
duke@435 825
duke@435 826 case Op_Proj: // Direct call to an allocation routine
duke@435 827 case Op_SCMemProj: // Memory state from store conditional ops
duke@435 828 #ifdef ASSERT
duke@435 829 {
duke@435 830 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@435 831 const Node* call = adr->in(0);
kvn@598 832 if (call->is_CallJava()) {
kvn@598 833 const CallJavaNode* call_java = call->as_CallJava();
kvn@499 834 const TypeTuple *r = call_java->tf()->range();
kvn@499 835 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@499 836 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@499 837 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@435 838 // We further presume that this is one of
duke@435 839 // new_instance_Java, new_array_Java, or
duke@435 840 // the like, but do not assert for this.
duke@435 841 } else if (call->is_Allocate()) {
duke@435 842 // similar case to new_instance_Java, etc.
duke@435 843 } else if (!call->is_CallLeaf()) {
duke@435 844 // Projections from fetch_oop (OSR) are allowed as well.
duke@435 845 ShouldNotReachHere();
duke@435 846 }
duke@435 847 }
duke@435 848 #endif
duke@435 849 break;
duke@435 850 default:
duke@435 851 ShouldNotReachHere();
duke@435 852 }
duke@435 853 break;
duke@435 854 }
duke@435 855 }
duke@435 856
duke@435 857 return NULL; // No progress
duke@435 858 }
duke@435 859
duke@435 860
duke@435 861 //=============================================================================
zmajo@7341 862 // Should LoadNode::Ideal() attempt to remove control edges?
zmajo@7341 863 bool LoadNode::can_remove_control() const {
zmajo@7341 864 return true;
zmajo@7341 865 }
duke@435 866 uint LoadNode::size_of() const { return sizeof(*this); }
duke@435 867 uint LoadNode::cmp( const Node &n ) const
duke@435 868 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@435 869 const Type *LoadNode::bottom_type() const { return _type; }
duke@435 870 uint LoadNode::ideal_reg() const {
coleenp@4037 871 return _type->ideal_reg();
duke@435 872 }
duke@435 873
duke@435 874 #ifndef PRODUCT
duke@435 875 void LoadNode::dump_spec(outputStream *st) const {
duke@435 876 MemNode::dump_spec(st);
duke@435 877 if( !Verbose && !WizardMode ) {
duke@435 878 // standard dump does this in Verbose and WizardMode
duke@435 879 st->print(" #"); _type->dump_on(st);
duke@435 880 }
roland@7859 881 if (!_depends_only_on_test) {
roland@7859 882 st->print(" (does not depend only on test)");
roland@7859 883 }
duke@435 884 }
duke@435 885 #endif
duke@435 886
kvn@1964 887 #ifdef ASSERT
kvn@1964 888 //----------------------------is_immutable_value-------------------------------
kvn@1964 889 // Helper function to allow a raw load without control edge for some cases
kvn@1964 890 bool LoadNode::is_immutable_value(Node* adr) {
kvn@1964 891 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
kvn@1964 892 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
kvn@1964 893 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
kvn@1964 894 in_bytes(JavaThread::osthread_offset())));
kvn@1964 895 }
kvn@1964 896 #endif
duke@435 897
duke@435 898 //----------------------------LoadNode::make-----------------------------------
duke@435 899 // Polymorphic factory method:
roland@7859 900 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, ControlDependency control_dependency) {
coleenp@548 901 Compile* C = gvn.C;
coleenp@548 902
duke@435 903 // sanity check the alias category against the created node type
duke@435 904 assert(!(adr_type->isa_oopptr() &&
duke@435 905 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@435 906 "use LoadKlassNode instead");
duke@435 907 assert(!(adr_type->isa_aryptr() &&
duke@435 908 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@435 909 "use LoadRangeNode instead");
kvn@1964 910 // Check control edge of raw loads
kvn@1964 911 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
kvn@1964 912 // oop will be recorded in oop map if load crosses safepoint
kvn@1964 913 rt->isa_oopptr() || is_immutable_value(adr),
kvn@1964 914 "raw memory operations should have control edge");
duke@435 915 switch (bt) {
roland@7859 916 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
roland@7859 917 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
roland@7859 918 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
roland@7859 919 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
roland@7859 920 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
roland@7859 921 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency);
roland@7859 922 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency);
roland@7859 923 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency);
roland@7859 924 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
coleenp@548 925 case T_OBJECT:
coleenp@548 926 #ifdef _LP64
kvn@598 927 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
roland@7859 928 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency));
kvn@4115 929 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
coleenp@548 930 } else
coleenp@548 931 #endif
kvn@598 932 {
roland@4159 933 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
roland@7859 934 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo, control_dependency);
kvn@598 935 }
duke@435 936 }
duke@435 937 ShouldNotReachHere();
duke@435 938 return (LoadNode*)NULL;
duke@435 939 }
duke@435 940
roland@7859 941 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) {
duke@435 942 bool require_atomic = true;
roland@7859 943 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
duke@435 944 }
duke@435 945
roland@7859 946 LoadDNode* LoadDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) {
anoll@7858 947 bool require_atomic = true;
roland@7859 948 return new (C) LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
anoll@7858 949 }
duke@435 950
duke@435 951
duke@435 952
duke@435 953 //------------------------------hash-------------------------------------------
duke@435 954 uint LoadNode::hash() const {
duke@435 955 // unroll addition of interesting fields
duke@435 956 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@435 957 }
duke@435 958
vlivanov@5658 959 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
vlivanov@5658 960 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
vlivanov@5658 961 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
vlivanov@5658 962 bool is_stable_ary = FoldStableValues &&
vlivanov@5658 963 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
vlivanov@5658 964 tp->isa_aryptr()->is_stable();
vlivanov@5658 965
vlivanov@5658 966 return (eliminate_boxing && non_volatile) || is_stable_ary;
vlivanov@5658 967 }
vlivanov@5658 968
vlivanov@5658 969 return false;
vlivanov@5658 970 }
vlivanov@5658 971
duke@435 972 //---------------------------can_see_stored_value------------------------------
duke@435 973 // This routine exists to make sure this set of tests is done the same
duke@435 974 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@435 975 // will change the graph shape in a way which makes memory alive twice at the
duke@435 976 // same time (uses the Oracle model of aliasing), then some
duke@435 977 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@435 978 // of aliasing.
duke@435 979 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@435 980 Node* ld_adr = in(MemNode::Address);
kvn@5110 981 intptr_t ld_off = 0;
kvn@5110 982 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
never@452 983 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
kvn@5110 984 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
kvn@5110 985 // This is more general than load from boxing objects.
vlivanov@5658 986 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
never@452 987 uint alias_idx = atp->index();
vlivanov@5658 988 bool final = !atp->is_rewritable();
never@452 989 Node* result = NULL;
never@452 990 Node* current = st;
never@452 991 // Skip through chains of MemBarNodes checking the MergeMems for
never@452 992 // new states for the slice of this load. Stop once any other
never@452 993 // kind of node is encountered. Loads from final memory can skip
never@452 994 // through any kind of MemBar but normal loads shouldn't skip
never@452 995 // through MemBarAcquire since the could allow them to move out of
never@452 996 // a synchronized region.
never@452 997 while (current->is_Proj()) {
never@452 998 int opc = current->in(0)->Opcode();
goetz@6489 999 if ((final && (opc == Op_MemBarAcquire ||
goetz@6489 1000 opc == Op_MemBarAcquireLock ||
goetz@6489 1001 opc == Op_LoadFence)) ||
goetz@6489 1002 opc == Op_MemBarRelease ||
goetz@6489 1003 opc == Op_StoreFence ||
goetz@6489 1004 opc == Op_MemBarReleaseLock ||
goetz@6489 1005 opc == Op_MemBarCPUOrder) {
never@452 1006 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@452 1007 if (mem->is_MergeMem()) {
never@452 1008 MergeMemNode* merge = mem->as_MergeMem();
never@452 1009 Node* new_st = merge->memory_at(alias_idx);
never@452 1010 if (new_st == merge->base_memory()) {
never@452 1011 // Keep searching
kvn@5110 1012 current = new_st;
never@452 1013 continue;
never@452 1014 }
never@452 1015 // Save the new memory state for the slice and fall through
never@452 1016 // to exit.
never@452 1017 result = new_st;
never@452 1018 }
never@452 1019 }
never@452 1020 break;
never@452 1021 }
never@452 1022 if (result != NULL) {
never@452 1023 st = result;
never@452 1024 }
never@452 1025 }
never@452 1026
duke@435 1027 // Loop around twice in the case Load -> Initialize -> Store.
duke@435 1028 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@435 1029 for (int trip = 0; trip <= 1; trip++) {
duke@435 1030
duke@435 1031 if (st->is_Store()) {
duke@435 1032 Node* st_adr = st->in(MemNode::Address);
duke@435 1033 if (!phase->eqv(st_adr, ld_adr)) {
duke@435 1034 // Try harder before giving up... Match raw and non-raw pointers.
duke@435 1035 intptr_t st_off = 0;
duke@435 1036 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@435 1037 if (alloc == NULL) return NULL;
kvn@5110 1038 if (alloc != ld_alloc) return NULL;
duke@435 1039 if (ld_off != st_off) return NULL;
duke@435 1040 // At this point we have proven something like this setup:
duke@435 1041 // A = Allocate(...)
duke@435 1042 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@435 1043 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@435 1044 // (Actually, we haven't yet proven the Q's are the same.)
duke@435 1045 // In other words, we are loading from a casted version of
duke@435 1046 // the same pointer-and-offset that we stored to.
duke@435 1047 // Thus, we are able to replace L by V.
duke@435 1048 }
duke@435 1049 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@435 1050 if (store_Opcode() != st->Opcode())
duke@435 1051 return NULL;
duke@435 1052 return st->in(MemNode::ValueIn);
duke@435 1053 }
duke@435 1054
duke@435 1055 // A load from a freshly-created object always returns zero.
duke@435 1056 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@435 1057 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@435 1058 if (st->is_Proj() && st->in(0)->is_Allocate() &&
kvn@5110 1059 (st->in(0) == ld_alloc) &&
kvn@5110 1060 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
duke@435 1061 // return a zero value for the load's basic type
duke@435 1062 // (This is one of the few places where a generic PhaseTransform
duke@435 1063 // can create new nodes. Think of it as lazily manifesting
duke@435 1064 // virtually pre-existing constants.)
duke@435 1065 return phase->zerocon(memory_type());
duke@435 1066 }
duke@435 1067
duke@435 1068 // A load from an initialization barrier can match a captured store.
duke@435 1069 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@435 1070 InitializeNode* init = st->in(0)->as_Initialize();
duke@435 1071 AllocateNode* alloc = init->allocation();
kvn@5110 1072 if ((alloc != NULL) && (alloc == ld_alloc)) {
duke@435 1073 // examine a captured store value
kvn@5110 1074 st = init->find_captured_store(ld_off, memory_size(), phase);
duke@435 1075 if (st != NULL)
duke@435 1076 continue; // take one more trip around
duke@435 1077 }
duke@435 1078 }
duke@435 1079
kvn@5110 1080 // Load boxed value from result of valueOf() call is input parameter.
kvn@5110 1081 if (this->is_Load() && ld_adr->is_AddP() &&
kvn@5110 1082 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
kvn@5110 1083 intptr_t ignore = 0;
kvn@5110 1084 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
kvn@5110 1085 if (base != NULL && base->is_Proj() &&
kvn@5110 1086 base->as_Proj()->_con == TypeFunc::Parms &&
kvn@5110 1087 base->in(0)->is_CallStaticJava() &&
kvn@5110 1088 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
kvn@5110 1089 return base->in(0)->in(TypeFunc::Parms);
kvn@5110 1090 }
kvn@5110 1091 }
kvn@5110 1092
duke@435 1093 break;
duke@435 1094 }
duke@435 1095
duke@435 1096 return NULL;
duke@435 1097 }
duke@435 1098
kvn@499 1099 //----------------------is_instance_field_load_with_local_phi------------------
kvn@499 1100 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@5110 1101 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
kvn@5110 1102 in(Address)->is_AddP() ) {
kvn@5110 1103 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
kvn@5110 1104 // Only instances and boxed values.
kvn@5110 1105 if( t_oop != NULL &&
kvn@5110 1106 (t_oop->is_ptr_to_boxed_value() ||
kvn@5110 1107 t_oop->is_known_instance_field()) &&
kvn@499 1108 t_oop->offset() != Type::OffsetBot &&
kvn@499 1109 t_oop->offset() != Type::OffsetTop) {
kvn@499 1110 return true;
kvn@499 1111 }
kvn@499 1112 }
kvn@499 1113 return false;
kvn@499 1114 }
kvn@499 1115
duke@435 1116 //------------------------------Identity---------------------------------------
duke@435 1117 // Loads are identity if previous store is to same address
duke@435 1118 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@435 1119 // If the previous store-maker is the right kind of Store, and the store is
duke@435 1120 // to the same address, then we are equal to the value stored.
kvn@5110 1121 Node* mem = in(Memory);
duke@435 1122 Node* value = can_see_stored_value(mem, phase);
duke@435 1123 if( value ) {
duke@435 1124 // byte, short & char stores truncate naturally.
duke@435 1125 // A load has to load the truncated value which requires
duke@435 1126 // some sort of masking operation and that requires an
duke@435 1127 // Ideal call instead of an Identity call.
duke@435 1128 if (memory_size() < BytesPerInt) {
duke@435 1129 // If the input to the store does not fit with the load's result type,
duke@435 1130 // it must be truncated via an Ideal call.
duke@435 1131 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@435 1132 return this;
duke@435 1133 }
duke@435 1134 // (This works even when value is a Con, but LoadNode::Value
duke@435 1135 // usually runs first, producing the singleton type of the Con.)
duke@435 1136 return value;
duke@435 1137 }
kvn@499 1138
kvn@499 1139 // Search for an existing data phi which was generated before for the same
twisti@1040 1140 // instance's field to avoid infinite generation of phis in a loop.
kvn@499 1141 Node *region = mem->in(0);
kvn@499 1142 if (is_instance_field_load_with_local_phi(region)) {
kvn@5110 1143 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
kvn@499 1144 int this_index = phase->C->get_alias_index(addr_t);
kvn@499 1145 int this_offset = addr_t->offset();
kvn@5110 1146 int this_iid = addr_t->instance_id();
kvn@5110 1147 if (!addr_t->is_known_instance() &&
kvn@5110 1148 addr_t->is_ptr_to_boxed_value()) {
kvn@5110 1149 // Use _idx of address base (could be Phi node) for boxed values.
kvn@5110 1150 intptr_t ignore = 0;
kvn@5110 1151 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
kvn@5110 1152 this_iid = base->_idx;
kvn@5110 1153 }
kvn@499 1154 const Type* this_type = bottom_type();
kvn@499 1155 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@499 1156 Node* phi = region->fast_out(i);
kvn@499 1157 if (phi->is_Phi() && phi != mem &&
kvn@5110 1158 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
kvn@499 1159 return phi;
kvn@499 1160 }
kvn@499 1161 }
kvn@499 1162 }
kvn@499 1163
duke@435 1164 return this;
duke@435 1165 }
duke@435 1166
never@452 1167 // We're loading from an object which has autobox behaviour.
never@452 1168 // If this object is result of a valueOf call we'll have a phi
never@452 1169 // merging a newly allocated object and a load from the cache.
never@452 1170 // We want to replace this load with the original incoming
never@452 1171 // argument to the valueOf call.
never@452 1172 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
kvn@5110 1173 assert(phase->C->eliminate_boxing(), "sanity");
kvn@5110 1174 intptr_t ignore = 0;
kvn@5110 1175 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
kvn@5110 1176 if ((base == NULL) || base->is_Phi()) {
kvn@5110 1177 // Push the loads from the phi that comes from valueOf up
kvn@5110 1178 // through it to allow elimination of the loads and the recovery
kvn@5110 1179 // of the original value. It is done in split_through_phi().
kvn@5110 1180 return NULL;
kvn@1018 1181 } else if (base->is_Load() ||
kvn@1018 1182 base->is_DecodeN() && base->in(1)->is_Load()) {
kvn@5110 1183 // Eliminate the load of boxed value for integer types from the cache
kvn@5110 1184 // array by deriving the value from the index into the array.
kvn@5110 1185 // Capture the offset of the load and then reverse the computation.
kvn@5110 1186
kvn@5110 1187 // Get LoadN node which loads a boxing object from 'cache' array.
kvn@1018 1188 if (base->is_DecodeN()) {
kvn@1018 1189 base = base->in(1);
kvn@1018 1190 }
kvn@5110 1191 if (!base->in(Address)->is_AddP()) {
kvn@5110 1192 return NULL; // Complex address
kvn@1018 1193 }
kvn@5110 1194 AddPNode* address = base->in(Address)->as_AddP();
kvn@5110 1195 Node* cache_base = address->in(AddPNode::Base);
kvn@5110 1196 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
kvn@5110 1197 // Get ConP node which is static 'cache' field.
kvn@5110 1198 cache_base = cache_base->in(1);
kvn@5110 1199 }
kvn@5110 1200 if ((cache_base != NULL) && cache_base->is_Con()) {
kvn@5110 1201 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
kvn@5110 1202 if ((base_type != NULL) && base_type->is_autobox_cache()) {
never@452 1203 Node* elements[4];
kvn@5110 1204 int shift = exact_log2(type2aelembytes(T_OBJECT));
kvn@5110 1205 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
kvn@5110 1206 if ((count > 0) && elements[0]->is_Con() &&
kvn@5110 1207 ((count == 1) ||
kvn@5110 1208 (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
kvn@5110 1209 elements[1]->in(2) == phase->intcon(shift))) {
kvn@5110 1210 ciObjArray* array = base_type->const_oop()->as_obj_array();
kvn@5110 1211 // Fetch the box object cache[0] at the base of the array and get its value
kvn@5110 1212 ciInstance* box = array->obj_at(0)->as_instance();
kvn@5110 1213 ciInstanceKlass* ik = box->klass()->as_instance_klass();
kvn@5110 1214 assert(ik->is_box_klass(), "sanity");
kvn@5110 1215 assert(ik->nof_nonstatic_fields() == 1, "change following code");
kvn@5110 1216 if (ik->nof_nonstatic_fields() == 1) {
kvn@5110 1217 // This should be true nonstatic_field_at requires calling
kvn@5110 1218 // nof_nonstatic_fields so check it anyway
kvn@5110 1219 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
kvn@5110 1220 BasicType bt = c.basic_type();
kvn@5110 1221 // Only integer types have boxing cache.
kvn@5110 1222 assert(bt == T_BOOLEAN || bt == T_CHAR ||
kvn@5110 1223 bt == T_BYTE || bt == T_SHORT ||
kvn@5110 1224 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
kvn@5110 1225 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
kvn@5110 1226 if (cache_low != (int)cache_low) {
kvn@5110 1227 return NULL; // should not happen since cache is array indexed by value
kvn@5110 1228 }
kvn@5110 1229 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
kvn@5110 1230 if (offset != (int)offset) {
kvn@5110 1231 return NULL; // should not happen since cache is array indexed by value
kvn@5110 1232 }
kvn@5110 1233 // Add up all the offsets making of the address of the load
kvn@5110 1234 Node* result = elements[0];
kvn@5110 1235 for (int i = 1; i < count; i++) {
kvn@5110 1236 result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
kvn@5110 1237 }
kvn@5110 1238 // Remove the constant offset from the address and then
kvn@5110 1239 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));
kvn@5110 1240 // remove the scaling of the offset to recover the original index.
kvn@5110 1241 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
kvn@5110 1242 // Peel the shift off directly but wrap it in a dummy node
kvn@5110 1243 // since Ideal can't return existing nodes
kvn@5110 1244 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
kvn@5110 1245 } else if (result->is_Add() && result->in(2)->is_Con() &&
kvn@5110 1246 result->in(1)->Opcode() == Op_LShiftX &&
kvn@5110 1247 result->in(1)->in(2) == phase->intcon(shift)) {
kvn@5110 1248 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
kvn@5110 1249 // but for boxing cache access we know that X<<Z will not overflow
kvn@5110 1250 // (there is range check) so we do this optimizatrion by hand here.
kvn@5110 1251 Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));
kvn@5110 1252 result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));
kvn@5110 1253 } else {
kvn@5110 1254 result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
kvn@5110 1255 }
kvn@5110 1256 #ifdef _LP64
kvn@5110 1257 if (bt != T_LONG) {
kvn@5110 1258 result = new (phase->C) ConvL2INode(phase->transform(result));
kvn@5110 1259 }
kvn@5110 1260 #else
kvn@5110 1261 if (bt == T_LONG) {
kvn@5110 1262 result = new (phase->C) ConvI2LNode(phase->transform(result));
kvn@5110 1263 }
kvn@5110 1264 #endif
vlivanov@7383 1265 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
vlivanov@7383 1266 // Need to preserve unboxing load type if it is unsigned.
vlivanov@7383 1267 switch(this->Opcode()) {
vlivanov@7383 1268 case Op_LoadUB:
vlivanov@7383 1269 result = new (phase->C) AndINode(phase->transform(result), phase->intcon(0xFF));
vlivanov@7383 1270 break;
vlivanov@7383 1271 case Op_LoadUS:
vlivanov@7383 1272 result = new (phase->C) AndINode(phase->transform(result), phase->intcon(0xFFFF));
vlivanov@7383 1273 break;
vlivanov@7383 1274 }
kvn@5110 1275 return result;
never@452 1276 }
never@452 1277 }
never@452 1278 }
never@452 1279 }
never@452 1280 }
never@452 1281 return NULL;
never@452 1282 }
never@452 1283
kvn@5110 1284 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
kvn@5110 1285 Node* region = phi->in(0);
kvn@598 1286 if (region == NULL) {
kvn@5110 1287 return false; // Wait stable graph
kvn@598 1288 }
kvn@5110 1289 uint cnt = phi->req();
kvn@2810 1290 for (uint i = 1; i < cnt; i++) {
kvn@2810 1291 Node* rc = region->in(i);
kvn@2810 1292 if (rc == NULL || phase->type(rc) == Type::TOP)
kvn@5110 1293 return false; // Wait stable graph
kvn@5110 1294 Node* in = phi->in(i);
kvn@5110 1295 if (in == NULL || phase->type(in) == Type::TOP)
kvn@5110 1296 return false; // Wait stable graph
kvn@5110 1297 }
kvn@5110 1298 return true;
kvn@5110 1299 }
kvn@5110 1300 //------------------------------split_through_phi------------------------------
kvn@5110 1301 // Split instance or boxed field load through Phi.
kvn@5110 1302 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
kvn@5110 1303 Node* mem = in(Memory);
kvn@5110 1304 Node* address = in(Address);
kvn@5110 1305 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
kvn@5110 1306
kvn@5110 1307 assert((t_oop != NULL) &&
kvn@5110 1308 (t_oop->is_known_instance_field() ||
kvn@5110 1309 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
kvn@5110 1310
kvn@5110 1311 Compile* C = phase->C;
kvn@5110 1312 intptr_t ignore = 0;
kvn@5110 1313 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
kvn@5110 1314 bool base_is_phi = (base != NULL) && base->is_Phi();
kvn@5110 1315 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
kvn@5110 1316 (base != NULL) && (base == address->in(AddPNode::Base)) &&
kvn@5110 1317 phase->type(base)->higher_equal(TypePtr::NOTNULL);
kvn@5110 1318
kvn@5110 1319 if (!((mem->is_Phi() || base_is_phi) &&
kvn@5110 1320 (load_boxed_values || t_oop->is_known_instance_field()))) {
kvn@5110 1321 return NULL; // memory is not Phi
kvn@5110 1322 }
kvn@5110 1323
kvn@5110 1324 if (mem->is_Phi()) {
kvn@5110 1325 if (!stable_phi(mem->as_Phi(), phase)) {
kvn@598 1326 return NULL; // Wait stable graph
kvn@598 1327 }
kvn@5110 1328 uint cnt = mem->req();
kvn@5110 1329 // Check for loop invariant memory.
kvn@5110 1330 if (cnt == 3) {
kvn@5110 1331 for (uint i = 1; i < cnt; i++) {
kvn@5110 1332 Node* in = mem->in(i);
kvn@5110 1333 Node* m = optimize_memory_chain(in, t_oop, this, phase);
kvn@5110 1334 if (m == mem) {
kvn@5110 1335 set_req(Memory, mem->in(cnt - i));
kvn@5110 1336 return this; // made change
kvn@5110 1337 }
kvn@598 1338 }
kvn@598 1339 }
kvn@598 1340 }
kvn@5110 1341 if (base_is_phi) {
kvn@5110 1342 if (!stable_phi(base->as_Phi(), phase)) {
kvn@5110 1343 return NULL; // Wait stable graph
kvn@5110 1344 }
kvn@5110 1345 uint cnt = base->req();
kvn@5110 1346 // Check for loop invariant memory.
kvn@5110 1347 if (cnt == 3) {
kvn@5110 1348 for (uint i = 1; i < cnt; i++) {
kvn@5110 1349 if (base->in(i) == base) {
kvn@5110 1350 return NULL; // Wait stable graph
kvn@5110 1351 }
kvn@5110 1352 }
kvn@5110 1353 }
kvn@5110 1354 }
kvn@5110 1355
kvn@5110 1356 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
kvn@5110 1357
kvn@598 1358 // Split through Phi (see original code in loopopts.cpp).
kvn@5110 1359 assert(C->have_alias_type(t_oop), "instance should have alias type");
kvn@598 1360
kvn@598 1361 // Do nothing here if Identity will find a value
kvn@598 1362 // (to avoid infinite chain of value phis generation).
kvn@2810 1363 if (!phase->eqv(this, this->Identity(phase)))
kvn@598 1364 return NULL;
kvn@598 1365
kvn@5110 1366 // Select Region to split through.
kvn@5110 1367 Node* region;
kvn@5110 1368 if (!base_is_phi) {
kvn@5110 1369 assert(mem->is_Phi(), "sanity");
kvn@5110 1370 region = mem->in(0);
kvn@5110 1371 // Skip if the region dominates some control edge of the address.
kvn@5110 1372 if (!MemNode::all_controls_dominate(address, region))
kvn@5110 1373 return NULL;
kvn@5110 1374 } else if (!mem->is_Phi()) {
kvn@5110 1375 assert(base_is_phi, "sanity");
kvn@5110 1376 region = base->in(0);
kvn@5110 1377 // Skip if the region dominates some control edge of the memory.
kvn@5110 1378 if (!MemNode::all_controls_dominate(mem, region))
kvn@5110 1379 return NULL;
kvn@5110 1380 } else if (base->in(0) != mem->in(0)) {
kvn@5110 1381 assert(base_is_phi && mem->is_Phi(), "sanity");
kvn@5110 1382 if (MemNode::all_controls_dominate(mem, base->in(0))) {
kvn@5110 1383 region = base->in(0);
kvn@5110 1384 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
kvn@5110 1385 region = mem->in(0);
kvn@5110 1386 } else {
kvn@5110 1387 return NULL; // complex graph
kvn@5110 1388 }
kvn@5110 1389 } else {
kvn@5110 1390 assert(base->in(0) == mem->in(0), "sanity");
kvn@5110 1391 region = mem->in(0);
kvn@5110 1392 }
kvn@598 1393
kvn@598 1394 const Type* this_type = this->bottom_type();
kvn@5110 1395 int this_index = C->get_alias_index(t_oop);
kvn@5110 1396 int this_offset = t_oop->offset();
kvn@5110 1397 int this_iid = t_oop->instance_id();
kvn@5110 1398 if (!t_oop->is_known_instance() && load_boxed_values) {
kvn@5110 1399 // Use _idx of address base for boxed values.
kvn@5110 1400 this_iid = base->_idx;
kvn@5110 1401 }
kvn@5110 1402 PhaseIterGVN* igvn = phase->is_IterGVN();
kvn@5110 1403 Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
kvn@2810 1404 for (uint i = 1; i < region->req(); i++) {
kvn@5110 1405 Node* x;
kvn@598 1406 Node* the_clone = NULL;
kvn@5110 1407 if (region->in(i) == C->top()) {
kvn@5110 1408 x = C->top(); // Dead path? Use a dead data op
kvn@598 1409 } else {
kvn@598 1410 x = this->clone(); // Else clone up the data op
kvn@598 1411 the_clone = x; // Remember for possible deletion.
kvn@598 1412 // Alter data node to use pre-phi inputs
kvn@2810 1413 if (this->in(0) == region) {
kvn@2810 1414 x->set_req(0, region->in(i));
kvn@598 1415 } else {
kvn@2810 1416 x->set_req(0, NULL);
kvn@598 1417 }
kvn@5110 1418 if (mem->is_Phi() && (mem->in(0) == region)) {
kvn@5110 1419 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
kvn@5110 1420 }
kvn@5110 1421 if (address->is_Phi() && address->in(0) == region) {
kvn@5110 1422 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
kvn@5110 1423 }
kvn@5110 1424 if (base_is_phi && (base->in(0) == region)) {
kvn@5110 1425 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
kvn@5110 1426 Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
kvn@5110 1427 x->set_req(Address, adr_x);
kvn@598 1428 }
kvn@598 1429 }
kvn@598 1430 // Check for a 'win' on some paths
kvn@598 1431 const Type *t = x->Value(igvn);
kvn@598 1432
kvn@598 1433 bool singleton = t->singleton();
kvn@598 1434
kvn@598 1435 // See comments in PhaseIdealLoop::split_thru_phi().
kvn@2810 1436 if (singleton && t == Type::TOP) {
kvn@598 1437 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
kvn@598 1438 }
kvn@598 1439
kvn@2810 1440 if (singleton) {
kvn@598 1441 x = igvn->makecon(t);
kvn@598 1442 } else {
kvn@598 1443 // We now call Identity to try to simplify the cloned node.
kvn@598 1444 // Note that some Identity methods call phase->type(this).
kvn@598 1445 // Make sure that the type array is big enough for
kvn@598 1446 // our new node, even though we may throw the node away.
kvn@598 1447 // (This tweaking with igvn only works because x is a new node.)
kvn@598 1448 igvn->set_type(x, t);
kvn@728 1449 // If x is a TypeNode, capture any more-precise type permanently into Node
twisti@1040 1450 // otherwise it will be not updated during igvn->transform since
kvn@728 1451 // igvn->type(x) is set to x->Value() already.
kvn@728 1452 x->raise_bottom_type(t);
kvn@598 1453 Node *y = x->Identity(igvn);
kvn@2810 1454 if (y != x) {
kvn@598 1455 x = y;
kvn@598 1456 } else {
kvn@5110 1457 y = igvn->hash_find_insert(x);
kvn@2810 1458 if (y) {
kvn@598 1459 x = y;
kvn@598 1460 } else {
kvn@598 1461 // Else x is a new node we are keeping
kvn@598 1462 // We do not need register_new_node_with_optimizer
kvn@598 1463 // because set_type has already been called.
kvn@598 1464 igvn->_worklist.push(x);
kvn@598 1465 }
kvn@598 1466 }
kvn@598 1467 }
kvn@5110 1468 if (x != the_clone && the_clone != NULL) {
kvn@598 1469 igvn->remove_dead_node(the_clone);
kvn@5110 1470 }
kvn@598 1471 phi->set_req(i, x);
kvn@598 1472 }
kvn@2810 1473 // Record Phi
kvn@2810 1474 igvn->register_new_node_with_optimizer(phi);
kvn@2810 1475 return phi;
kvn@598 1476 }
never@452 1477
duke@435 1478 //------------------------------Ideal------------------------------------------
zmajo@7341 1479 // If the load is from Field memory and the pointer is non-null, it might be possible to
duke@435 1480 // zero out the control input.
duke@435 1481 // If the offset is constant and the base is an object allocation,
duke@435 1482 // try to hook me up to the exact initializing store.
duke@435 1483 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@435 1484 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@435 1485 if (p) return (p == NodeSentinel) ? NULL : p;
duke@435 1486
duke@435 1487 Node* ctrl = in(MemNode::Control);
duke@435 1488 Node* address = in(MemNode::Address);
duke@435 1489
duke@435 1490 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@435 1491 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@435 1492 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@435 1493 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@435 1494 ctrl = ctrl->in(0);
duke@435 1495 set_req(MemNode::Control,ctrl);
duke@435 1496 }
duke@435 1497
kvn@1143 1498 intptr_t ignore = 0;
kvn@1143 1499 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
kvn@1143 1500 if (base != NULL
kvn@1143 1501 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
kvn@1143 1502 // Check for useless control edge in some common special cases
kvn@1143 1503 if (in(MemNode::Control) != NULL
zmajo@7341 1504 && can_remove_control()
duke@435 1505 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
kvn@554 1506 && all_controls_dominate(base, phase->C->start())) {
duke@435 1507 // A method-invariant, non-null address (constant or 'this' argument).
duke@435 1508 set_req(MemNode::Control, NULL);
duke@435 1509 }
never@452 1510 }
never@452 1511
kvn@509 1512 Node* mem = in(MemNode::Memory);
kvn@509 1513 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@509 1514
kvn@5110 1515 if (can_reshape && (addr_t != NULL)) {
kvn@509 1516 // try to optimize our memory input
kvn@5110 1517 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
kvn@509 1518 if (opt_mem != mem) {
kvn@509 1519 set_req(MemNode::Memory, opt_mem);
kvn@740 1520 if (phase->type( opt_mem ) == Type::TOP) return NULL;
kvn@509 1521 return this;
kvn@509 1522 }
kvn@509 1523 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@5110 1524 if ((t_oop != NULL) &&
kvn@5110 1525 (t_oop->is_known_instance_field() ||
kvn@5110 1526 t_oop->is_ptr_to_boxed_value())) {
kvn@3260 1527 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@3260 1528 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
kvn@3260 1529 // Delay this transformation until memory Phi is processed.
kvn@3260 1530 phase->is_IterGVN()->_worklist.push(this);
kvn@3260 1531 return NULL;
kvn@3260 1532 }
kvn@598 1533 // Split instance field load through Phi.
kvn@598 1534 Node* result = split_through_phi(phase);
kvn@598 1535 if (result != NULL) return result;
kvn@5110 1536
kvn@5110 1537 if (t_oop->is_ptr_to_boxed_value()) {
kvn@5110 1538 Node* result = eliminate_autobox(phase);
kvn@5110 1539 if (result != NULL) return result;
kvn@5110 1540 }
kvn@509 1541 }
kvn@509 1542 }
kvn@509 1543
duke@435 1544 // Check for prior store with a different base or offset; make Load
duke@435 1545 // independent. Skip through any number of them. Bail out if the stores
duke@435 1546 // are in an endless dead cycle and report no progress. This is a key
duke@435 1547 // transform for Reflection. However, if after skipping through the Stores
duke@435 1548 // we can't then fold up against a prior store do NOT do the transform as
duke@435 1549 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@435 1550 // array memory alive twice: once for the hoisted Load and again after the
duke@435 1551 // bypassed Store. This situation only works if EVERYBODY who does
duke@435 1552 // anti-dependence work knows how to bypass. I.e. we need all
duke@435 1553 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@435 1554 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@435 1555 // fold up, do so.
duke@435 1556 Node* prev_mem = find_previous_store(phase);
duke@435 1557 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@435 1558 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@435 1559 // (c) See if we can fold up on the spot, but don't fold up here.
twisti@993 1560 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
duke@435 1561 // just return a prior value, which is done by Identity calls.
duke@435 1562 if (can_see_stored_value(prev_mem, phase)) {
duke@435 1563 // Make ready for step (d):
duke@435 1564 set_req(MemNode::Memory, prev_mem);
duke@435 1565 return this;
duke@435 1566 }
duke@435 1567 }
duke@435 1568
duke@435 1569 return NULL; // No further progress
duke@435 1570 }
duke@435 1571
duke@435 1572 // Helper to recognize certain Klass fields which are invariant across
duke@435 1573 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@435 1574 const Type*
duke@435 1575 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@435 1576 ciKlass* klass) const {
stefank@3391 1577 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
duke@435 1578 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@435 1579 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@435 1580 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@435 1581 return TypeInt::make(klass->modifier_flags());
duke@435 1582 }
stefank@3391 1583 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
duke@435 1584 // The field is Klass::_access_flags. Return its (constant) value.
duke@435 1585 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@435 1586 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@435 1587 return TypeInt::make(klass->access_flags());
duke@435 1588 }
stefank@3391 1589 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
duke@435 1590 // The field is Klass::_layout_helper. Return its constant value if known.
duke@435 1591 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@435 1592 return TypeInt::make(klass->layout_helper());
duke@435 1593 }
duke@435 1594
duke@435 1595 // No match.
duke@435 1596 return NULL;
duke@435 1597 }
duke@435 1598
vlivanov@5658 1599 // Try to constant-fold a stable array element.
vlivanov@5658 1600 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
vlivanov@6526 1601 assert(ary->const_oop(), "array should be constant");
vlivanov@5658 1602 assert(ary->is_stable(), "array should be stable");
vlivanov@5658 1603
vlivanov@6526 1604 // Decode the results of GraphKit::array_element_address.
vlivanov@6526 1605 ciArray* aobj = ary->const_oop()->as_array();
vlivanov@6526 1606 ciConstant con = aobj->element_value_by_offset(off);
vlivanov@6526 1607
vlivanov@6526 1608 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
vlivanov@6526 1609 const Type* con_type = Type::make_from_constant(con);
vlivanov@6526 1610 if (con_type != NULL) {
vlivanov@6526 1611 if (con_type->isa_aryptr()) {
vlivanov@6526 1612 // Join with the array element type, in case it is also stable.
vlivanov@6526 1613 int dim = ary->stable_dimension();
vlivanov@6526 1614 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
vlivanov@6526 1615 }
vlivanov@6526 1616 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
vlivanov@6526 1617 con_type = con_type->make_narrowoop();
vlivanov@6526 1618 }
vlivanov@5658 1619 #ifndef PRODUCT
vlivanov@6526 1620 if (TraceIterativeGVN) {
vlivanov@6526 1621 tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
vlivanov@6526 1622 con_type->dump(); tty->cr();
vlivanov@6526 1623 }
vlivanov@5658 1624 #endif //PRODUCT
vlivanov@6526 1625 return con_type;
vlivanov@5658 1626 }
vlivanov@5658 1627 }
vlivanov@5658 1628 return NULL;
vlivanov@5658 1629 }
vlivanov@5658 1630
duke@435 1631 //------------------------------Value-----------------------------------------
duke@435 1632 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@435 1633 // Either input is TOP ==> the result is TOP
duke@435 1634 Node* mem = in(MemNode::Memory);
duke@435 1635 const Type *t1 = phase->type(mem);
duke@435 1636 if (t1 == Type::TOP) return Type::TOP;
duke@435 1637 Node* adr = in(MemNode::Address);
duke@435 1638 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@435 1639 if (tp == NULL || tp->empty()) return Type::TOP;
duke@435 1640 int off = tp->offset();
duke@435 1641 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
twisti@3102 1642 Compile* C = phase->C;
duke@435 1643
duke@435 1644 // Try to guess loaded type from pointer type
vlivanov@5658 1645 if (tp->isa_aryptr()) {
vlivanov@5658 1646 const TypeAryPtr* ary = tp->is_aryptr();
vlivanov@6526 1647 const Type* t = ary->elem();
vlivanov@5658 1648
vlivanov@5658 1649 // Determine whether the reference is beyond the header or not, by comparing
vlivanov@5658 1650 // the offset against the offset of the start of the array's data.
vlivanov@5658 1651 // Different array types begin at slightly different offsets (12 vs. 16).
vlivanov@5658 1652 // We choose T_BYTE as an example base type that is least restrictive
vlivanov@5658 1653 // as to alignment, which will therefore produce the smallest
vlivanov@5658 1654 // possible base offset.
vlivanov@5658 1655 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
vlivanov@5658 1656 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
vlivanov@5658 1657
vlivanov@5658 1658 // Try to constant-fold a stable array element.
vlivanov@6526 1659 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) {
vlivanov@6526 1660 // Make sure the reference is not into the header and the offset is constant
vlivanov@6526 1661 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
vlivanov@5658 1662 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
vlivanov@5658 1663 if (con_type != NULL) {
vlivanov@5658 1664 return con_type;
vlivanov@5658 1665 }
vlivanov@5658 1666 }
vlivanov@5658 1667 }
vlivanov@5658 1668
duke@435 1669 // Don't do this for integer types. There is only potential profit if
duke@435 1670 // the element type t is lower than _type; that is, for int types, if _type is
duke@435 1671 // more restrictive than t. This only happens here if one is short and the other
duke@435 1672 // char (both 16 bits), and in those cases we've made an intentional decision
duke@435 1673 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@435 1674 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@435 1675 //
duke@435 1676 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@435 1677 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@435 1678 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@435 1679 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@435 1680 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@435 1681 // In fact, that could have been the original type of p1, and p1 could have
duke@435 1682 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@435 1683 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@435 1684 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
kvn@3882 1685 && (_type->isa_vect() == NULL)
kvn@728 1686 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
duke@435 1687 // t might actually be lower than _type, if _type is a unique
duke@435 1688 // concrete subclass of abstract class t.
vlivanov@5658 1689 if (off_beyond_header) { // is the offset beyond the header?
roland@6313 1690 const Type* jt = t->join_speculative(_type);
duke@435 1691 // In any case, do not allow the join, per se, to empty out the type.
duke@435 1692 if (jt->empty() && !t->empty()) {
duke@435 1693 // This can happen if a interface-typed array narrows to a class type.
duke@435 1694 jt = _type;
duke@435 1695 }
kvn@5110 1696 #ifdef ASSERT
kvn@5110 1697 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
never@452 1698 // The pointers in the autobox arrays are always non-null
kvn@1143 1699 Node* base = adr->in(AddPNode::Base);
kvn@5110 1700 if ((base != NULL) && base->is_DecodeN()) {
kvn@5110 1701 // Get LoadN node which loads IntegerCache.cache field
kvn@5110 1702 base = base->in(1);
kvn@5110 1703 }
kvn@5110 1704 if ((base != NULL) && base->is_Con()) {
kvn@5110 1705 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
kvn@5110 1706 if ((base_type != NULL) && base_type->is_autobox_cache()) {
kvn@5110 1707 // It could be narrow oop
kvn@5110 1708 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
never@452 1709 }
never@452 1710 }
never@452 1711 }
kvn@5110 1712 #endif
duke@435 1713 return jt;
duke@435 1714 }
duke@435 1715 }
duke@435 1716 } else if (tp->base() == Type::InstPtr) {
twisti@3102 1717 ciEnv* env = C->env();
never@1515 1718 const TypeInstPtr* tinst = tp->is_instptr();
never@1515 1719 ciKlass* klass = tinst->klass();
duke@435 1720 assert( off != Type::OffsetBot ||
duke@435 1721 // arrays can be cast to Objects
duke@435 1722 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@435 1723 // unsafe field access may not have a constant offset
twisti@3102 1724 C->has_unsafe_access(),
duke@435 1725 "Field accesses must be precise" );
duke@435 1726 // For oop loads, we expect the _type to be precise
twisti@3102 1727 if (klass == env->String_klass() &&
never@1515 1728 adr->is_AddP() && off != Type::OffsetBot) {
kvn@2602 1729 // For constant Strings treat the final fields as compile time constants.
never@1515 1730 Node* base = adr->in(AddPNode::Base);
never@2121 1731 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
never@2121 1732 if (t != NULL && t->singleton()) {
twisti@3102 1733 ciField* field = env->String_klass()->get_field_by_offset(off, false);
kvn@2602 1734 if (field != NULL && field->is_final()) {
kvn@2602 1735 ciObject* string = t->const_oop();
kvn@2602 1736 ciConstant constant = string->as_instance()->field_value(field);
kvn@2602 1737 if (constant.basic_type() == T_INT) {
kvn@2602 1738 return TypeInt::make(constant.as_int());
kvn@2602 1739 } else if (constant.basic_type() == T_ARRAY) {
kvn@2602 1740 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
jcoomes@2661 1741 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
kvn@2602 1742 } else {
jcoomes@2661 1743 return TypeOopPtr::make_from_constant(constant.as_object(), true);
kvn@2602 1744 }
never@1515 1745 }
never@1515 1746 }
never@1515 1747 }
never@1515 1748 }
twisti@3102 1749 // Optimizations for constant objects
twisti@3102 1750 ciObject* const_oop = tinst->const_oop();
twisti@3102 1751 if (const_oop != NULL) {
kvn@5110 1752 // For constant Boxed value treat the target field as a compile time constant.
kvn@5110 1753 if (tinst->is_ptr_to_boxed_value()) {
kvn@5110 1754 return tinst->get_const_boxed_value();
kvn@5110 1755 } else
twisti@3102 1756 // For constant CallSites treat the target field as a compile time constant.
twisti@3102 1757 if (const_oop->is_call_site()) {
twisti@3102 1758 ciCallSite* call_site = const_oop->as_call_site();
twisti@3102 1759 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
twisti@3102 1760 if (field != NULL && field->is_call_site_target()) {
twisti@3102 1761 ciMethodHandle* target = call_site->get_target();
twisti@3102 1762 if (target != NULL) { // just in case
twisti@3102 1763 ciConstant constant(T_OBJECT, target);
twisti@3102 1764 const Type* t;
twisti@3102 1765 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
twisti@3102 1766 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
twisti@3102 1767 } else {
twisti@3102 1768 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
twisti@3102 1769 }
twisti@3102 1770 // Add a dependence for invalidation of the optimization.
twisti@3102 1771 if (!call_site->is_constant_call_site()) {
twisti@3102 1772 C->dependencies()->assert_call_site_target_value(call_site, target);
twisti@3102 1773 }
twisti@3102 1774 return t;
twisti@3102 1775 }
twisti@3102 1776 }
twisti@3102 1777 }
twisti@3102 1778 }
duke@435 1779 } else if (tp->base() == Type::KlassPtr) {
duke@435 1780 assert( off != Type::OffsetBot ||
duke@435 1781 // arrays can be cast to Objects
duke@435 1782 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@435 1783 // also allow array-loading from the primary supertype
duke@435 1784 // array during subtype checks
duke@435 1785 Opcode() == Op_LoadKlass,
duke@435 1786 "Field accesses must be precise" );
duke@435 1787 // For klass/static loads, we expect the _type to be precise
duke@435 1788 }
duke@435 1789
duke@435 1790 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@435 1791 if (tkls != NULL && !StressReflectiveCode) {
duke@435 1792 ciKlass* klass = tkls->klass();
duke@435 1793 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@435 1794 // We are loading a field from a Klass metaobject whose identity
duke@435 1795 // is known at compile time (the type is "exact" or "precise").
duke@435 1796 // Check for fields we know are maintained as constants by the VM.
stefank@3391 1797 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
duke@435 1798 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@435 1799 // (Folds up type checking code.)
duke@435 1800 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@435 1801 return TypeInt::make(klass->super_check_offset());
duke@435 1802 }
duke@435 1803 // Compute index into primary_supers array
coleenp@4037 1804 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
duke@435 1805 // Check for overflowing; use unsigned compare to handle the negative case.
duke@435 1806 if( depth < ciKlass::primary_super_limit() ) {
duke@435 1807 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@435 1808 // (Folds up type checking code.)
duke@435 1809 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@435 1810 ciKlass *ss = klass->super_of_depth(depth);
duke@435 1811 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@435 1812 }
duke@435 1813 const Type* aift = load_array_final_field(tkls, klass);
duke@435 1814 if (aift != NULL) return aift;
coleenp@4142 1815 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
duke@435 1816 && klass->is_array_klass()) {
coleenp@4142 1817 // The field is ArrayKlass::_component_mirror. Return its (constant) value.
duke@435 1818 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@435 1819 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@435 1820 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@435 1821 }
stefank@3391 1822 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
duke@435 1823 // The field is Klass::_java_mirror. Return its (constant) value.
duke@435 1824 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@435 1825 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@435 1826 return TypeInstPtr::make(klass->java_mirror());
duke@435 1827 }
duke@435 1828 }
duke@435 1829
duke@435 1830 // We can still check if we are loading from the primary_supers array at a
duke@435 1831 // shallow enough depth. Even though the klass is not exact, entries less
duke@435 1832 // than or equal to its super depth are correct.
duke@435 1833 if (klass->is_loaded() ) {
coleenp@4037 1834 ciType *inner = klass;
duke@435 1835 while( inner->is_obj_array_klass() )
duke@435 1836 inner = inner->as_obj_array_klass()->base_element_type();
duke@435 1837 if( inner->is_instance_klass() &&
duke@435 1838 !inner->as_instance_klass()->flags().is_interface() ) {
duke@435 1839 // Compute index into primary_supers array
coleenp@4037 1840 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
duke@435 1841 // Check for overflowing; use unsigned compare to handle the negative case.
duke@435 1842 if( depth < ciKlass::primary_super_limit() &&
duke@435 1843 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@435 1844 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@435 1845 // (Folds up type checking code.)
duke@435 1846 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@435 1847 ciKlass *ss = klass->super_of_depth(depth);
duke@435 1848 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@435 1849 }
duke@435 1850 }
duke@435 1851 }
duke@435 1852
duke@435 1853 // If the type is enough to determine that the thing is not an array,
duke@435 1854 // we can give the layout_helper a positive interval type.
duke@435 1855 // This will help short-circuit some reflective code.
stefank@3391 1856 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
duke@435 1857 && !klass->is_array_klass() // not directly typed as an array
duke@435 1858 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@435 1859 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@435 1860 ) {
duke@435 1861 // Note: When interfaces are reliable, we can narrow the interface
duke@435 1862 // test to (klass != Serializable && klass != Cloneable).
duke@435 1863 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@435 1864 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@435 1865 // The key property of this type is that it folds up tests
duke@435 1866 // for array-ness, since it proves that the layout_helper is positive.
duke@435 1867 // Thus, a generic value like the basic object layout helper works fine.
duke@435 1868 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@435 1869 }
duke@435 1870 }
duke@435 1871
duke@435 1872 // If we are loading from a freshly-allocated object, produce a zero,
duke@435 1873 // if the load is provably beyond the header of the object.
duke@435 1874 // (Also allow a variable load from a fresh array to produce zero.)
kvn@2810 1875 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@2810 1876 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
kvn@5110 1877 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
kvn@5110 1878 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
duke@435 1879 Node* value = can_see_stored_value(mem,phase);
kvn@3442 1880 if (value != NULL && value->is_Con()) {
kvn@3442 1881 assert(value->bottom_type()->higher_equal(_type),"sanity");
duke@435 1882 return value->bottom_type();
kvn@3442 1883 }
duke@435 1884 }
duke@435 1885
kvn@2810 1886 if (is_instance) {
kvn@499 1887 // If we have an instance type and our memory input is the
kvn@499 1888 // programs's initial memory state, there is no matching store,
kvn@499 1889 // so just return a zero of the appropriate type
kvn@499 1890 Node *mem = in(MemNode::Memory);
kvn@499 1891 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@499 1892 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@499 1893 return Type::get_zero_type(_type->basic_type());
kvn@499 1894 }
kvn@499 1895 }
duke@435 1896 return _type;
duke@435 1897 }
duke@435 1898
duke@435 1899 //------------------------------match_edge-------------------------------------
duke@435 1900 // Do we Match on this edge index or not? Match only the address.
duke@435 1901 uint LoadNode::match_edge(uint idx) const {
duke@435 1902 return idx == MemNode::Address;
duke@435 1903 }
duke@435 1904
duke@435 1905 //--------------------------LoadBNode::Ideal--------------------------------------
duke@435 1906 //
duke@435 1907 // If the previous store is to the same address as this load,
duke@435 1908 // and the value stored was larger than a byte, replace this load
duke@435 1909 // with the value stored truncated to a byte. If no truncation is
duke@435 1910 // needed, the replacement is done in LoadNode::Identity().
duke@435 1911 //
duke@435 1912 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@435 1913 Node* mem = in(MemNode::Memory);
duke@435 1914 Node* value = can_see_stored_value(mem,phase);
duke@435 1915 if( value && !phase->type(value)->higher_equal( _type ) ) {
kvn@4115 1916 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
kvn@4115 1917 return new (phase->C) RShiftINode(result, phase->intcon(24));
duke@435 1918 }
duke@435 1919 // Identity call will handle the case where truncation is not needed.
duke@435 1920 return LoadNode::Ideal(phase, can_reshape);
duke@435 1921 }
duke@435 1922
kvn@3442 1923 const Type* LoadBNode::Value(PhaseTransform *phase) const {
kvn@3442 1924 Node* mem = in(MemNode::Memory);
kvn@3442 1925 Node* value = can_see_stored_value(mem,phase);
kvn@3448 1926 if (value != NULL && value->is_Con() &&
kvn@3448 1927 !value->bottom_type()->higher_equal(_type)) {
kvn@3442 1928 // If the input to the store does not fit with the load's result type,
kvn@3442 1929 // it must be truncated. We can't delay until Ideal call since
kvn@3442 1930 // a singleton Value is needed for split_thru_phi optimization.
kvn@3442 1931 int con = value->get_int();
kvn@3442 1932 return TypeInt::make((con << 24) >> 24);
kvn@3442 1933 }
kvn@3442 1934 return LoadNode::Value(phase);
kvn@3442 1935 }
kvn@3442 1936
twisti@1059 1937 //--------------------------LoadUBNode::Ideal-------------------------------------
twisti@1059 1938 //
twisti@1059 1939 // If the previous store is to the same address as this load,
twisti@1059 1940 // and the value stored was larger than a byte, replace this load
twisti@1059 1941 // with the value stored truncated to a byte. If no truncation is
twisti@1059 1942 // needed, the replacement is done in LoadNode::Identity().
twisti@1059 1943 //
twisti@1059 1944 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
twisti@1059 1945 Node* mem = in(MemNode::Memory);
twisti@1059 1946 Node* value = can_see_stored_value(mem, phase);
twisti@1059 1947 if (value && !phase->type(value)->higher_equal(_type))
kvn@4115 1948 return new (phase->C) AndINode(value, phase->intcon(0xFF));
twisti@1059 1949 // Identity call will handle the case where truncation is not needed.
twisti@1059 1950 return LoadNode::Ideal(phase, can_reshape);
twisti@1059 1951 }
twisti@1059 1952
kvn@3442 1953 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
kvn@3442 1954 Node* mem = in(MemNode::Memory);
kvn@3442 1955 Node* value = can_see_stored_value(mem,phase);
kvn@3448 1956 if (value != NULL && value->is_Con() &&
kvn@3448 1957 !value->bottom_type()->higher_equal(_type)) {
kvn@3442 1958 // If the input to the store does not fit with the load's result type,
kvn@3442 1959 // it must be truncated. We can't delay until Ideal call since
kvn@3442 1960 // a singleton Value is needed for split_thru_phi optimization.
kvn@3442 1961 int con = value->get_int();
kvn@3442 1962 return TypeInt::make(con & 0xFF);
kvn@3442 1963 }
kvn@3442 1964 return LoadNode::Value(phase);
kvn@3442 1965 }
kvn@3442 1966
twisti@993 1967 //--------------------------LoadUSNode::Ideal-------------------------------------
duke@435 1968 //
duke@435 1969 // If the previous store is to the same address as this load,
duke@435 1970 // and the value stored was larger than a char, replace this load
duke@435 1971 // with the value stored truncated to a char. If no truncation is
duke@435 1972 // needed, the replacement is done in LoadNode::Identity().
duke@435 1973 //
twisti@993 1974 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@435 1975 Node* mem = in(MemNode::Memory);
duke@435 1976 Node* value = can_see_stored_value(mem,phase);
duke@435 1977 if( value && !phase->type(value)->higher_equal( _type ) )
kvn@4115 1978 return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
duke@435 1979 // Identity call will handle the case where truncation is not needed.
duke@435 1980 return LoadNode::Ideal(phase, can_reshape);
duke@435 1981 }
duke@435 1982
kvn@3442 1983 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
kvn@3442 1984 Node* mem = in(MemNode::Memory);
kvn@3442 1985 Node* value = can_see_stored_value(mem,phase);
kvn@3448 1986 if (value != NULL && value->is_Con() &&
kvn@3448 1987 !value->bottom_type()->higher_equal(_type)) {
kvn@3442 1988 // If the input to the store does not fit with the load's result type,
kvn@3442 1989 // it must be truncated. We can't delay until Ideal call since
kvn@3442 1990 // a singleton Value is needed for split_thru_phi optimization.
kvn@3442 1991 int con = value->get_int();
kvn@3442 1992 return TypeInt::make(con & 0xFFFF);
kvn@3442 1993 }
kvn@3442 1994 return LoadNode::Value(phase);
kvn@3442 1995 }
kvn@3442 1996
duke@435 1997 //--------------------------LoadSNode::Ideal--------------------------------------
duke@435 1998 //
duke@435 1999 // If the previous store is to the same address as this load,
duke@435 2000 // and the value stored was larger than a short, replace this load
duke@435 2001 // with the value stored truncated to a short. If no truncation is
duke@435 2002 // needed, the replacement is done in LoadNode::Identity().
duke@435 2003 //
duke@435 2004 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@435 2005 Node* mem = in(MemNode::Memory);
duke@435 2006 Node* value = can_see_stored_value(mem,phase);
duke@435 2007 if( value && !phase->type(value)->higher_equal( _type ) ) {
kvn@4115 2008 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
kvn@4115 2009 return new (phase->C) RShiftINode(result, phase->intcon(16));
duke@435 2010 }
duke@435 2011 // Identity call will handle the case where truncation is not needed.
duke@435 2012 return LoadNode::Ideal(phase, can_reshape);
duke@435 2013 }
duke@435 2014
kvn@3442 2015 const Type* LoadSNode::Value(PhaseTransform *phase) const {
kvn@3442 2016 Node* mem = in(MemNode::Memory);
kvn@3442 2017 Node* value = can_see_stored_value(mem,phase);
kvn@3448 2018 if (value != NULL && value->is_Con() &&
kvn@3448 2019 !value->bottom_type()->higher_equal(_type)) {
kvn@3442 2020 // If the input to the store does not fit with the load's result type,
kvn@3442 2021 // it must be truncated. We can't delay until Ideal call since
kvn@3442 2022 // a singleton Value is needed for split_thru_phi optimization.
kvn@3442 2023 int con = value->get_int();
kvn@3442 2024 return TypeInt::make((con << 16) >> 16);
kvn@3442 2025 }
kvn@3442 2026 return LoadNode::Value(phase);
kvn@3442 2027 }
kvn@3442 2028
duke@435 2029 //=============================================================================
kvn@599 2030 //----------------------------LoadKlassNode::make------------------------------
kvn@599 2031 // Polymorphic factory method:
zmajo@7341 2032 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk) {
kvn@599 2033 Compile* C = gvn.C;
kvn@599 2034 // sanity check the alias category against the created node type
coleenp@4037 2035 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
coleenp@4037 2036 assert(adr_type != NULL, "expecting TypeKlassPtr");
kvn@599 2037 #ifdef _LP64
roland@4159 2038 if (adr_type->is_ptr_to_narrowklass()) {
ehelin@5694 2039 assert(UseCompressedClassPointers, "no compressed klasses");
goetz@6479 2040 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
roland@4159 2041 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
kvn@603 2042 }
kvn@599 2043 #endif
roland@4159 2044 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
goetz@6479 2045 return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
kvn@599 2046 }
kvn@599 2047
duke@435 2048 //------------------------------Value------------------------------------------
duke@435 2049 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
kvn@599 2050 return klass_value_common(phase);
kvn@599 2051 }
kvn@599 2052
zmajo@7341 2053 // In most cases, LoadKlassNode does not have the control input set. If the control
zmajo@7341 2054 // input is set, it must not be removed (by LoadNode::Ideal()).
zmajo@7341 2055 bool LoadKlassNode::can_remove_control() const {
zmajo@7341 2056 return false;
zmajo@7341 2057 }
zmajo@7341 2058
kvn@599 2059 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
duke@435 2060 // Either input is TOP ==> the result is TOP
duke@435 2061 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@435 2062 if (t1 == Type::TOP) return Type::TOP;
duke@435 2063 Node *adr = in(MemNode::Address);
duke@435 2064 const Type *t2 = phase->type( adr );
duke@435 2065 if (t2 == Type::TOP) return Type::TOP;
duke@435 2066 const TypePtr *tp = t2->is_ptr();
duke@435 2067 if (TypePtr::above_centerline(tp->ptr()) ||
duke@435 2068 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@435 2069
duke@435 2070 // Return a more precise klass, if possible
duke@435 2071 const TypeInstPtr *tinst = tp->isa_instptr();
duke@435 2072 if (tinst != NULL) {
duke@435 2073 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@435 2074 int offset = tinst->offset();
duke@435 2075 if (ik == phase->C->env()->Class_klass()
duke@435 2076 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@435 2077 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@435 2078 // We are loading a special hidden field from a Class mirror object,
duke@435 2079 // the field which points to the VM's Klass metaobject.
duke@435 2080 ciType* t = tinst->java_mirror_type();
duke@435 2081 // java_mirror_type returns non-null for compile-time Class constants.
duke@435 2082 if (t != NULL) {
duke@435 2083 // constant oop => constant klass
duke@435 2084 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
twisti@6173 2085 if (t->is_void()) {
twisti@6173 2086 // We cannot create a void array. Since void is a primitive type return null
twisti@6173 2087 // klass. Users of this result need to do a null check on the returned klass.
twisti@6173 2088 return TypePtr::NULL_PTR;
twisti@6173 2089 }
duke@435 2090 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@435 2091 }
duke@435 2092 if (!t->is_klass()) {
duke@435 2093 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@435 2094 return TypePtr::NULL_PTR;
duke@435 2095 }
duke@435 2096 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@435 2097 return TypeKlassPtr::make(t->as_klass());
duke@435 2098 }
duke@435 2099 // non-constant mirror, so we can't tell what's going on
duke@435 2100 }
duke@435 2101 if( !ik->is_loaded() )
duke@435 2102 return _type; // Bail out if not loaded
duke@435 2103 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@435 2104 if (tinst->klass_is_exact()) {
duke@435 2105 return TypeKlassPtr::make(ik);
duke@435 2106 }
duke@435 2107 // See if we can become precise: no subklasses and no interface
duke@435 2108 // (Note: We need to support verified interfaces.)
duke@435 2109 if (!ik->is_interface() && !ik->has_subklass()) {
duke@435 2110 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@435 2111 // Add a dependence; if any subclass added we need to recompile
duke@435 2112 if (!ik->is_final()) {
duke@435 2113 // %%% should use stronger assert_unique_concrete_subtype instead
duke@435 2114 phase->C->dependencies()->assert_leaf_type(ik);
duke@435 2115 }
duke@435 2116 // Return precise klass
duke@435 2117 return TypeKlassPtr::make(ik);
duke@435 2118 }
duke@435 2119
duke@435 2120 // Return root of possible klass
duke@435 2121 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@435 2122 }
duke@435 2123 }
duke@435 2124
duke@435 2125 // Check for loading klass from an array
duke@435 2126 const TypeAryPtr *tary = tp->isa_aryptr();
duke@435 2127 if( tary != NULL ) {
duke@435 2128 ciKlass *tary_klass = tary->klass();
duke@435 2129 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@435 2130 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@435 2131 if (tary->klass_is_exact()) {
duke@435 2132 return TypeKlassPtr::make(tary_klass);
duke@435 2133 }
duke@435 2134 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@435 2135 // If the klass is an object array, we defer the question to the
duke@435 2136 // array component klass.
duke@435 2137 if( ak->is_obj_array_klass() ) {
duke@435 2138 assert( ak->is_loaded(), "" );
duke@435 2139 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@435 2140 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@435 2141 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@435 2142 // See if we can become precise: no subklasses and no interface
duke@435 2143 if (!ik->is_interface() && !ik->has_subklass()) {
duke@435 2144 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@435 2145 // Add a dependence; if any subclass added we need to recompile
duke@435 2146 if (!ik->is_final()) {
duke@435 2147 phase->C->dependencies()->assert_leaf_type(ik);
duke@435 2148 }
duke@435 2149 // Return precise array klass
duke@435 2150 return TypeKlassPtr::make(ak);
duke@435 2151 }
duke@435 2152 }
duke@435 2153 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@435 2154 } else { // Found a type-array?
duke@435 2155 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@435 2156 assert( ak->is_type_array_klass(), "" );
duke@435 2157 return TypeKlassPtr::make(ak); // These are always precise
duke@435 2158 }
duke@435 2159 }
duke@435 2160 }
duke@435 2161
duke@435 2162 // Check for loading klass from an array klass
duke@435 2163 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@435 2164 if (tkls != NULL && !StressReflectiveCode) {
duke@435 2165 ciKlass* klass = tkls->klass();
duke@435 2166 if( !klass->is_loaded() )
duke@435 2167 return _type; // Bail out if not loaded
duke@435 2168 if( klass->is_obj_array_klass() &&
coleenp@4142 2169 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
duke@435 2170 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@435 2171 // // Always returning precise element type is incorrect,
duke@435 2172 // // e.g., element type could be object and array may contain strings
duke@435 2173 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@435 2174
duke@435 2175 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@435 2176 // according to the element type's subclassing.
duke@435 2177 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@435 2178 }
duke@435 2179 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
stefank@3391 2180 tkls->offset() == in_bytes(Klass::super_offset())) {
duke@435 2181 ciKlass* sup = klass->as_instance_klass()->super();
duke@435 2182 // The field is Klass::_super. Return its (constant) value.
duke@435 2183 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@435 2184 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@435 2185 }
duke@435 2186 }
duke@435 2187
duke@435 2188 // Bailout case
duke@435 2189 return LoadNode::Value(phase);
duke@435 2190 }
duke@435 2191
duke@435 2192 //------------------------------Identity---------------------------------------
duke@435 2193 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@435 2194 // Also feed through the klass in Allocate(...klass...)._klass.
duke@435 2195 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
kvn@599 2196 return klass_identity_common(phase);
kvn@599 2197 }
kvn@599 2198
kvn@599 2199 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
duke@435 2200 Node* x = LoadNode::Identity(phase);
duke@435 2201 if (x != this) return x;
duke@435 2202
duke@435 2203 // Take apart the address into an oop and and offset.
duke@435 2204 // Return 'this' if we cannot.
duke@435 2205 Node* adr = in(MemNode::Address);
duke@435 2206 intptr_t offset = 0;
duke@435 2207 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@435 2208 if (base == NULL) return this;
duke@435 2209 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@435 2210 if (toop == NULL) return this;
duke@435 2211
duke@435 2212 // We can fetch the klass directly through an AllocateNode.
duke@435 2213 // This works even if the klass is not constant (clone or newArray).
duke@435 2214 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@435 2215 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@435 2216 if (allocated_klass != NULL) {
duke@435 2217 return allocated_klass;
duke@435 2218 }
duke@435 2219 }
duke@435 2220
coleenp@4037 2221 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
coleenp@4142 2222 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
duke@435 2223 // See inline_native_Class_query for occurrences of these patterns.
duke@435 2224 // Java Example: x.getClass().isAssignableFrom(y)
duke@435 2225 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@435 2226 //
duke@435 2227 // This improves reflective code, often making the Class
duke@435 2228 // mirror go completely dead. (Current exception: Class
duke@435 2229 // mirrors may appear in debug info, but we could clean them out by
coleenp@4037 2230 // introducing a new debug info operator for Klass*.java_mirror).
duke@435 2231 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@435 2232 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@435 2233 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@435 2234 // We are loading a special hidden field from a Class mirror,
coleenp@4142 2235 // the field which points to its Klass or ArrayKlass metaobject.
duke@435 2236 if (base->is_Load()) {
duke@435 2237 Node* adr2 = base->in(MemNode::Address);
duke@435 2238 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@435 2239 if (tkls != NULL && !tkls->empty()
duke@435 2240 && (tkls->klass()->is_instance_klass() ||
duke@435 2241 tkls->klass()->is_array_klass())
duke@435 2242 && adr2->is_AddP()
duke@435 2243 ) {
stefank@3391 2244 int mirror_field = in_bytes(Klass::java_mirror_offset());
duke@435 2245 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
coleenp@4142 2246 mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
duke@435 2247 }
stefank@3391 2248 if (tkls->offset() == mirror_field) {
duke@435 2249 return adr2->in(AddPNode::Base);
duke@435 2250 }
duke@435 2251 }
duke@435 2252 }
duke@435 2253 }
duke@435 2254
duke@435 2255 return this;
duke@435 2256 }
duke@435 2257
kvn@599 2258
kvn@599 2259 //------------------------------Value------------------------------------------
kvn@599 2260 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
kvn@599 2261 const Type *t = klass_value_common(phase);
kvn@656 2262 if (t == Type::TOP)
kvn@656 2263 return t;
kvn@656 2264
roland@4159 2265 return t->make_narrowklass();
kvn@599 2266 }
kvn@599 2267
kvn@599 2268 //------------------------------Identity---------------------------------------
kvn@599 2269 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
kvn@599 2270 // Also feed through the klass in Allocate(...klass...)._klass.
kvn@599 2271 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
kvn@599 2272 Node *x = klass_identity_common(phase);
kvn@599 2273
kvn@599 2274 const Type *t = phase->type( x );
kvn@599 2275 if( t == Type::TOP ) return x;
roland@4159 2276 if( t->isa_narrowklass()) return x;
roland@4159 2277 assert (!t->isa_narrowoop(), "no narrow oop here");
roland@4159 2278
roland@4159 2279 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
kvn@599 2280 }
kvn@599 2281
duke@435 2282 //------------------------------Value-----------------------------------------
duke@435 2283 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@435 2284 // Either input is TOP ==> the result is TOP
duke@435 2285 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@435 2286 if( t1 == Type::TOP ) return Type::TOP;
duke@435 2287 Node *adr = in(MemNode::Address);
duke@435 2288 const Type *t2 = phase->type( adr );
duke@435 2289 if( t2 == Type::TOP ) return Type::TOP;
duke@435 2290 const TypePtr *tp = t2->is_ptr();
duke@435 2291 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@435 2292 const TypeAryPtr *tap = tp->isa_aryptr();
duke@435 2293 if( !tap ) return _type;
duke@435 2294 return tap->size();
duke@435 2295 }
duke@435 2296
rasbold@801 2297 //-------------------------------Ideal---------------------------------------
rasbold@801 2298 // Feed through the length in AllocateArray(...length...)._length.
rasbold@801 2299 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
rasbold@801 2300 Node* p = MemNode::Ideal_common(phase, can_reshape);
rasbold@801 2301 if (p) return (p == NodeSentinel) ? NULL : p;
rasbold@801 2302
rasbold@801 2303 // Take apart the address into an oop and and offset.
rasbold@801 2304 // Return 'this' if we cannot.
rasbold@801 2305 Node* adr = in(MemNode::Address);
rasbold@801 2306 intptr_t offset = 0;
rasbold@801 2307 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
rasbold@801 2308 if (base == NULL) return NULL;
rasbold@801 2309 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
rasbold@801 2310 if (tary == NULL) return NULL;
rasbold@801 2311
rasbold@801 2312 // We can fetch the length directly through an AllocateArrayNode.
rasbold@801 2313 // This works even if the length is not constant (clone or newArray).
rasbold@801 2314 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@801 2315 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@801 2316 if (alloc != NULL) {
rasbold@801 2317 Node* allocated_length = alloc->Ideal_length();
rasbold@801 2318 Node* len = alloc->make_ideal_length(tary, phase);
rasbold@801 2319 if (allocated_length != len) {
rasbold@801 2320 // New CastII improves on this.
rasbold@801 2321 return len;
rasbold@801 2322 }
rasbold@801 2323 }
rasbold@801 2324 }
rasbold@801 2325
rasbold@801 2326 return NULL;
rasbold@801 2327 }
rasbold@801 2328
duke@435 2329 //------------------------------Identity---------------------------------------
duke@435 2330 // Feed through the length in AllocateArray(...length...)._length.
duke@435 2331 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@435 2332 Node* x = LoadINode::Identity(phase);
duke@435 2333 if (x != this) return x;
duke@435 2334
duke@435 2335 // Take apart the address into an oop and and offset.
duke@435 2336 // Return 'this' if we cannot.
duke@435 2337 Node* adr = in(MemNode::Address);
duke@435 2338 intptr_t offset = 0;
duke@435 2339 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@435 2340 if (base == NULL) return this;
duke@435 2341 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@435 2342 if (tary == NULL) return this;
duke@435 2343
duke@435 2344 // We can fetch the length directly through an AllocateArrayNode.
duke@435 2345 // This works even if the length is not constant (clone or newArray).
duke@435 2346 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@801 2347 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@801 2348 if (alloc != NULL) {
rasbold@801 2349 Node* allocated_length = alloc->Ideal_length();
rasbold@801 2350 // Do not allow make_ideal_length to allocate a CastII node.
rasbold@801 2351 Node* len = alloc->make_ideal_length(tary, phase, false);
rasbold@801 2352 if (allocated_length == len) {
rasbold@801 2353 // Return allocated_length only if it would not be improved by a CastII.
rasbold@801 2354 return allocated_length;
rasbold@801 2355 }
duke@435 2356 }
duke@435 2357 }
duke@435 2358
duke@435 2359 return this;
duke@435 2360
duke@435 2361 }
rasbold@801 2362
duke@435 2363 //=============================================================================
duke@435 2364 //---------------------------StoreNode::make-----------------------------------
duke@435 2365 // Polymorphic factory method:
goetz@6479 2366 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
goetz@6479 2367 assert((mo == unordered || mo == release), "unexpected");
coleenp@548 2368 Compile* C = gvn.C;
goetz@6479 2369 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
goetz@6479 2370 ctl != NULL, "raw memory operations should have control edge");
coleenp@548 2371
duke@435 2372 switch (bt) {
duke@435 2373 case T_BOOLEAN:
goetz@6479 2374 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo);
goetz@6479 2375 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo);
duke@435 2376 case T_CHAR:
goetz@6479 2377 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo);
goetz@6479 2378 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo);
goetz@6479 2379 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo);
goetz@6479 2380 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo);
coleenp@4037 2381 case T_METADATA:
duke@435 2382 case T_ADDRESS:
coleenp@548 2383 case T_OBJECT:
coleenp@548 2384 #ifdef _LP64
roland@4159 2385 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
kvn@4115 2386 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
goetz@6479 2387 return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo);
roland@4159 2388 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
ehelin@5694 2389 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
roland@4159 2390 adr->bottom_type()->isa_rawptr())) {
roland@4159 2391 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
goetz@6479 2392 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
roland@4159 2393 }
coleenp@548 2394 #endif
kvn@656 2395 {
goetz@6479 2396 return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo);
kvn@656 2397 }
duke@435 2398 }
duke@435 2399 ShouldNotReachHere();
duke@435 2400 return (StoreNode*)NULL;
duke@435 2401 }
duke@435 2402
goetz@6479 2403 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
duke@435 2404 bool require_atomic = true;
goetz@6479 2405 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
duke@435 2406 }
duke@435 2407
anoll@7858 2408 StoreDNode* StoreDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
anoll@7858 2409 bool require_atomic = true;
anoll@7858 2410 return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
anoll@7858 2411 }
anoll@7858 2412
duke@435 2413
duke@435 2414 //--------------------------bottom_type----------------------------------------
duke@435 2415 const Type *StoreNode::bottom_type() const {
duke@435 2416 return Type::MEMORY;
duke@435 2417 }
duke@435 2418
duke@435 2419 //------------------------------hash-------------------------------------------
duke@435 2420 uint StoreNode::hash() const {
duke@435 2421 // unroll addition of interesting fields
duke@435 2422 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@435 2423
duke@435 2424 // Since they are not commoned, do not hash them:
duke@435 2425 return NO_HASH;
duke@435 2426 }
duke@435 2427
duke@435 2428 //------------------------------Ideal------------------------------------------
duke@435 2429 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@435 2430 // When a store immediately follows a relevant allocation/initialization,
duke@435 2431 // try to capture it into the initialization, or hoist it above.
duke@435 2432 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@435 2433 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@435 2434 if (p) return (p == NodeSentinel) ? NULL : p;
duke@435 2435
duke@435 2436 Node* mem = in(MemNode::Memory);
duke@435 2437 Node* address = in(MemNode::Address);
duke@435 2438
never@2780 2439 // Back-to-back stores to same address? Fold em up. Generally
never@2780 2440 // unsafe if I have intervening uses... Also disallowed for StoreCM
never@2780 2441 // since they must follow each StoreP operation. Redundant StoreCMs
never@2780 2442 // are eliminated just before matching in final_graph_reshape.
kvn@3407 2443 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
never@2780 2444 mem->Opcode() != Op_StoreCM) {
duke@435 2445 // Looking at a dead closed cycle of memory?
duke@435 2446 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@435 2447
duke@435 2448 assert(Opcode() == mem->Opcode() ||
duke@435 2449 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@435 2450 "no mismatched stores, except on raw memory");
duke@435 2451
duke@435 2452 if (mem->outcnt() == 1 && // check for intervening uses
duke@435 2453 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@435 2454 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@435 2455 // For example, 'mem' might be the final state at a conditional return.
duke@435 2456 // Or, 'mem' might be used by some node which is live at the same time
duke@435 2457 // 'this' is live, which might be unschedulable. So, require exactly
duke@435 2458 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@435 2459 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@435 2460 if (can_reshape) { // (%%% is this an anachronism?)
duke@435 2461 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@435 2462 phase->is_IterGVN());
duke@435 2463 } else {
duke@435 2464 // It's OK to do this in the parser, since DU info is always accurate,
duke@435 2465 // and the parser always refers to nodes via SafePointNode maps.
duke@435 2466 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@435 2467 }
duke@435 2468 return this;
duke@435 2469 }
duke@435 2470 }
duke@435 2471
duke@435 2472 // Capture an unaliased, unconditional, simple store into an initializer.
duke@435 2473 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@435 2474 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@435 2475 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@435 2476 InitializeNode* init = mem->in(0)->as_Initialize();
roland@4657 2477 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
duke@435 2478 if (offset > 0) {
roland@4657 2479 Node* moved = init->capture_store(this, offset, phase, can_reshape);
duke@435 2480 // If the InitializeNode captured me, it made a raw copy of me,
duke@435 2481 // and I need to disappear.
duke@435 2482 if (moved != NULL) {
duke@435 2483 // %%% hack to ensure that Ideal returns a new node:
duke@435 2484 mem = MergeMemNode::make(phase->C, mem);
duke@435 2485 return mem; // fold me away
duke@435 2486 }
duke@435 2487 }
duke@435 2488 }
duke@435 2489
duke@435 2490 return NULL; // No further progress
duke@435 2491 }
duke@435 2492
duke@435 2493 //------------------------------Value-----------------------------------------
duke@435 2494 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@435 2495 // Either input is TOP ==> the result is TOP
duke@435 2496 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@435 2497 if( t1 == Type::TOP ) return Type::TOP;
duke@435 2498 const Type *t2 = phase->type( in(MemNode::Address) );
duke@435 2499 if( t2 == Type::TOP ) return Type::TOP;
duke@435 2500 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@435 2501 if( t3 == Type::TOP ) return Type::TOP;
duke@435 2502 return Type::MEMORY;
duke@435 2503 }
duke@435 2504
duke@435 2505 //------------------------------Identity---------------------------------------
duke@435 2506 // Remove redundant stores:
duke@435 2507 // Store(m, p, Load(m, p)) changes to m.
duke@435 2508 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@435 2509 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@435 2510 Node* mem = in(MemNode::Memory);
duke@435 2511 Node* adr = in(MemNode::Address);
duke@435 2512 Node* val = in(MemNode::ValueIn);
duke@435 2513
duke@435 2514 // Load then Store? Then the Store is useless
duke@435 2515 if (val->is_Load() &&
kvn@3407 2516 val->in(MemNode::Address)->eqv_uncast(adr) &&
kvn@3407 2517 val->in(MemNode::Memory )->eqv_uncast(mem) &&
duke@435 2518 val->as_Load()->store_Opcode() == Opcode()) {
duke@435 2519 return mem;
duke@435 2520 }
duke@435 2521
duke@435 2522 // Two stores in a row of the same value?
duke@435 2523 if (mem->is_Store() &&
kvn@3407 2524 mem->in(MemNode::Address)->eqv_uncast(adr) &&
kvn@3407 2525 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
duke@435 2526 mem->Opcode() == Opcode()) {
duke@435 2527 return mem;
duke@435 2528 }
duke@435 2529
duke@435 2530 // Store of zero anywhere into a freshly-allocated object?
duke@435 2531 // Then the store is useless.
duke@435 2532 // (It must already have been captured by the InitializeNode.)
duke@435 2533 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@435 2534 // a newly allocated object is already all-zeroes everywhere
duke@435 2535 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@435 2536 return mem;
duke@435 2537 }
duke@435 2538
duke@435 2539 // the store may also apply to zero-bits in an earlier object
duke@435 2540 Node* prev_mem = find_previous_store(phase);
duke@435 2541 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@435 2542 if (prev_mem != NULL) {
duke@435 2543 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@435 2544 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@435 2545 // prev_val and val might differ by a cast; it would be good
duke@435 2546 // to keep the more informative of the two.
duke@435 2547 return mem;
duke@435 2548 }
duke@435 2549 }
duke@435 2550 }
duke@435 2551
duke@435 2552 return this;
duke@435 2553 }
duke@435 2554
duke@435 2555 //------------------------------match_edge-------------------------------------
duke@435 2556 // Do we Match on this edge index or not? Match only memory & value
duke@435 2557 uint StoreNode::match_edge(uint idx) const {
duke@435 2558 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@435 2559 }
duke@435 2560
duke@435 2561 //------------------------------cmp--------------------------------------------
duke@435 2562 // Do not common stores up together. They generally have to be split
duke@435 2563 // back up anyways, so do not bother.
duke@435 2564 uint StoreNode::cmp( const Node &n ) const {
duke@435 2565 return (&n == this); // Always fail except on self
duke@435 2566 }
duke@435 2567
duke@435 2568 //------------------------------Ideal_masked_input-----------------------------
duke@435 2569 // Check for a useless mask before a partial-word store
duke@435 2570 // (StoreB ... (AndI valIn conIa) )
duke@435 2571 // If (conIa & mask == mask) this simplifies to
duke@435 2572 // (StoreB ... (valIn) )
duke@435 2573 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@435 2574 Node *val = in(MemNode::ValueIn);
duke@435 2575 if( val->Opcode() == Op_AndI ) {
duke@435 2576 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@435 2577 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@435 2578 set_req(MemNode::ValueIn, val->in(1));
duke@435 2579 return this;
duke@435 2580 }
duke@435 2581 }
duke@435 2582 return NULL;
duke@435 2583 }
duke@435 2584
duke@435 2585
duke@435 2586 //------------------------------Ideal_sign_extended_input----------------------
duke@435 2587 // Check for useless sign-extension before a partial-word store
duke@435 2588 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@435 2589 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@435 2590 // (StoreB ... (valIn) )
duke@435 2591 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@435 2592 Node *val = in(MemNode::ValueIn);
duke@435 2593 if( val->Opcode() == Op_RShiftI ) {
duke@435 2594 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@435 2595 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@435 2596 Node *shl = val->in(1);
duke@435 2597 if( shl->Opcode() == Op_LShiftI ) {
duke@435 2598 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@435 2599 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@435 2600 set_req(MemNode::ValueIn, shl->in(1));
duke@435 2601 return this;
duke@435 2602 }
duke@435 2603 }
duke@435 2604 }
duke@435 2605 }
duke@435 2606 return NULL;
duke@435 2607 }
duke@435 2608
duke@435 2609 //------------------------------value_never_loaded-----------------------------------
duke@435 2610 // Determine whether there are any possible loads of the value stored.
duke@435 2611 // For simplicity, we actually check if there are any loads from the
duke@435 2612 // address stored to, not just for loads of the value stored by this node.
duke@435 2613 //
duke@435 2614 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@435 2615 Node *adr = in(Address);
duke@435 2616 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@435 2617 if (adr_oop == NULL)
duke@435 2618 return false;
kvn@658 2619 if (!adr_oop->is_known_instance_field())
duke@435 2620 return false; // if not a distinct instance, there may be aliases of the address
duke@435 2621 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@435 2622 Node *use = adr->fast_out(i);
duke@435 2623 int opc = use->Opcode();
duke@435 2624 if (use->is_Load() || use->is_LoadStore()) {
duke@435 2625 return false;
duke@435 2626 }
duke@435 2627 }
duke@435 2628 return true;
duke@435 2629 }
duke@435 2630
duke@435 2631 //=============================================================================
duke@435 2632 //------------------------------Ideal------------------------------------------
duke@435 2633 // If the store is from an AND mask that leaves the low bits untouched, then
duke@435 2634 // we can skip the AND operation. If the store is from a sign-extension
duke@435 2635 // (a left shift, then right shift) we can skip both.
duke@435 2636 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@435 2637 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@435 2638 if( progress != NULL ) return progress;
duke@435 2639
duke@435 2640 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@435 2641 if( progress != NULL ) return progress;
duke@435 2642
duke@435 2643 // Finally check the default case
duke@435 2644 return StoreNode::Ideal(phase, can_reshape);
duke@435 2645 }
duke@435 2646
duke@435 2647 //=============================================================================
duke@435 2648 //------------------------------Ideal------------------------------------------
duke@435 2649 // If the store is from an AND mask that leaves the low bits untouched, then
duke@435 2650 // we can skip the AND operation
duke@435 2651 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@435 2652 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@435 2653 if( progress != NULL ) return progress;
duke@435 2654
duke@435 2655 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@435 2656 if( progress != NULL ) return progress;
duke@435 2657
duke@435 2658 // Finally check the default case
duke@435 2659 return StoreNode::Ideal(phase, can_reshape);
duke@435 2660 }
duke@435 2661
duke@435 2662 //=============================================================================
duke@435 2663 //------------------------------Identity---------------------------------------
duke@435 2664 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@435 2665 // No need to card mark when storing a null ptr
duke@435 2666 Node* my_store = in(MemNode::OopStore);
duke@435 2667 if (my_store->is_Store()) {
duke@435 2668 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@435 2669 if( t1 == TypePtr::NULL_PTR ) {
duke@435 2670 return in(MemNode::Memory);
duke@435 2671 }
duke@435 2672 }
duke@435 2673 return this;
duke@435 2674 }
duke@435 2675
cfang@1420 2676 //=============================================================================
cfang@1420 2677 //------------------------------Ideal---------------------------------------
cfang@1420 2678 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
cfang@1420 2679 Node* progress = StoreNode::Ideal(phase, can_reshape);
cfang@1420 2680 if (progress != NULL) return progress;
cfang@1420 2681
cfang@1420 2682 Node* my_store = in(MemNode::OopStore);
cfang@1420 2683 if (my_store->is_MergeMem()) {
cfang@1420 2684 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
cfang@1420 2685 set_req(MemNode::OopStore, mem);
cfang@1420 2686 return this;
cfang@1420 2687 }
cfang@1420 2688
cfang@1420 2689 return NULL;
cfang@1420 2690 }
cfang@1420 2691
duke@435 2692 //------------------------------Value-----------------------------------------
duke@435 2693 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@478 2694 // Either input is TOP ==> the result is TOP
kvn@478 2695 const Type *t = phase->type( in(MemNode::Memory) );
kvn@478 2696 if( t == Type::TOP ) return Type::TOP;
kvn@478 2697 t = phase->type( in(MemNode::Address) );
kvn@478 2698 if( t == Type::TOP ) return Type::TOP;
kvn@478 2699 t = phase->type( in(MemNode::ValueIn) );
kvn@478 2700 if( t == Type::TOP ) return Type::TOP;
duke@435 2701 // If extra input is TOP ==> the result is TOP
kvn@478 2702 t = phase->type( in(MemNode::OopStore) );
kvn@478 2703 if( t == Type::TOP ) return Type::TOP;
duke@435 2704
duke@435 2705 return StoreNode::Value( phase );
duke@435 2706 }
duke@435 2707
duke@435 2708
duke@435 2709 //=============================================================================
duke@435 2710 //----------------------------------SCMemProjNode------------------------------
duke@435 2711 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@435 2712 {
duke@435 2713 return bottom_type();
duke@435 2714 }
duke@435 2715
duke@435 2716 //=============================================================================
roland@4106 2717 //----------------------------------LoadStoreNode------------------------------
roland@4106 2718 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
roland@4106 2719 : Node(required),
roland@4106 2720 _type(rt),
roland@4106 2721 _adr_type(at)
roland@4106 2722 {
duke@435 2723 init_req(MemNode::Control, c );
duke@435 2724 init_req(MemNode::Memory , mem);
duke@435 2725 init_req(MemNode::Address, adr);
duke@435 2726 init_req(MemNode::ValueIn, val);
duke@435 2727 init_class_id(Class_LoadStore);
roland@4106 2728 }
roland@4106 2729
roland@4106 2730 uint LoadStoreNode::ideal_reg() const {
roland@4106 2731 return _type->ideal_reg();
roland@4106 2732 }
roland@4106 2733
roland@4106 2734 bool LoadStoreNode::result_not_used() const {
roland@4106 2735 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
roland@4106 2736 Node *x = fast_out(i);
roland@4106 2737 if (x->Opcode() == Op_SCMemProj) continue;
roland@4106 2738 return false;
roland@4106 2739 }
roland@4106 2740 return true;
roland@4106 2741 }
roland@4106 2742
roland@4106 2743 uint LoadStoreNode::size_of() const { return sizeof(*this); }
roland@4106 2744
roland@4106 2745 //=============================================================================
roland@4106 2746 //----------------------------------LoadStoreConditionalNode--------------------
roland@4106 2747 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
roland@4106 2748 init_req(ExpectedIn, ex );
duke@435 2749 }
duke@435 2750
duke@435 2751 //=============================================================================
duke@435 2752 //-------------------------------adr_type--------------------------------------
duke@435 2753 // Do we Match on this edge index or not? Do not match memory
duke@435 2754 const TypePtr* ClearArrayNode::adr_type() const {
duke@435 2755 Node *adr = in(3);
duke@435 2756 return MemNode::calculate_adr_type(adr->bottom_type());
duke@435 2757 }
duke@435 2758
duke@435 2759 //------------------------------match_edge-------------------------------------
duke@435 2760 // Do we Match on this edge index or not? Do not match memory
duke@435 2761 uint ClearArrayNode::match_edge(uint idx) const {
duke@435 2762 return idx > 1;
duke@435 2763 }
duke@435 2764
duke@435 2765 //------------------------------Identity---------------------------------------
duke@435 2766 // Clearing a zero length array does nothing
duke@435 2767 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@503 2768 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@435 2769 }
duke@435 2770
duke@435 2771 //------------------------------Idealize---------------------------------------
duke@435 2772 // Clearing a short array is faster with stores
duke@435 2773 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@435 2774 const int unit = BytesPerLong;
duke@435 2775 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@435 2776 if (!t) return NULL;
duke@435 2777 if (!t->is_con()) return NULL;
duke@435 2778 intptr_t raw_count = t->get_con();
duke@435 2779 intptr_t size = raw_count;
duke@435 2780 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@435 2781 // Clearing nothing uses the Identity call.
duke@435 2782 // Negative clears are possible on dead ClearArrays
duke@435 2783 // (see jck test stmt114.stmt11402.val).
duke@435 2784 if (size <= 0 || size % unit != 0) return NULL;
duke@435 2785 intptr_t count = size / unit;
duke@435 2786 // Length too long; use fast hardware clear
duke@435 2787 if (size > Matcher::init_array_short_size) return NULL;
duke@435 2788 Node *mem = in(1);
duke@435 2789 if( phase->type(mem)==Type::TOP ) return NULL;
duke@435 2790 Node *adr = in(3);
duke@435 2791 const Type* at = phase->type(adr);
duke@435 2792 if( at==Type::TOP ) return NULL;
duke@435 2793 const TypePtr* atp = at->isa_ptr();
duke@435 2794 // adjust atp to be the correct array element address type
duke@435 2795 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@435 2796 else atp = atp->add_offset(Type::OffsetBot);
duke@435 2797 // Get base for derived pointer purposes
duke@435 2798 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@435 2799 Node *base = adr->in(1);
duke@435 2800
duke@435 2801 Node *zero = phase->makecon(TypeLong::ZERO);
duke@435 2802 Node *off = phase->MakeConX(BytesPerLong);
goetz@6479 2803 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
duke@435 2804 count--;
duke@435 2805 while( count-- ) {
duke@435 2806 mem = phase->transform(mem);
kvn@4115 2807 adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
goetz@6479 2808 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
duke@435 2809 }
duke@435 2810 return mem;
duke@435 2811 }
duke@435 2812
kvn@1535 2813 //----------------------------step_through----------------------------------
kvn@1535 2814 // Return allocation input memory edge if it is different instance
kvn@1535 2815 // or itself if it is the one we are looking for.
kvn@1535 2816 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
kvn@1535 2817 Node* n = *np;
kvn@1535 2818 assert(n->is_ClearArray(), "sanity");
kvn@1535 2819 intptr_t offset;
kvn@1535 2820 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
kvn@1535 2821 // This method is called only before Allocate nodes are expanded during
kvn@1535 2822 // macro nodes expansion. Before that ClearArray nodes are only generated
kvn@1535 2823 // in LibraryCallKit::generate_arraycopy() which follows allocations.
kvn@1535 2824 assert(alloc != NULL, "should have allocation");
kvn@1535 2825 if (alloc->_idx == instance_id) {
kvn@1535 2826 // Can not bypass initialization of the instance we are looking for.
kvn@1535 2827 return false;
kvn@1535 2828 }
kvn@1535 2829 // Otherwise skip it.
kvn@1535 2830 InitializeNode* init = alloc->initialization();
kvn@1535 2831 if (init != NULL)
kvn@1535 2832 *np = init->in(TypeFunc::Memory);
kvn@1535 2833 else
kvn@1535 2834 *np = alloc->in(TypeFunc::Memory);
kvn@1535 2835 return true;
kvn@1535 2836 }
kvn@1535 2837
duke@435 2838 //----------------------------clear_memory-------------------------------------
duke@435 2839 // Generate code to initialize object storage to zero.
duke@435 2840 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@435 2841 intptr_t start_offset,
duke@435 2842 Node* end_offset,
duke@435 2843 PhaseGVN* phase) {
duke@435 2844 Compile* C = phase->C;
duke@435 2845 intptr_t offset = start_offset;
duke@435 2846
duke@435 2847 int unit = BytesPerLong;
duke@435 2848 if ((offset % unit) != 0) {
kvn@4115 2849 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
duke@435 2850 adr = phase->transform(adr);
duke@435 2851 const TypePtr* atp = TypeRawPtr::BOTTOM;
goetz@6479 2852 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
duke@435 2853 mem = phase->transform(mem);
duke@435 2854 offset += BytesPerInt;
duke@435 2855 }
duke@435 2856 assert((offset % unit) == 0, "");
duke@435 2857
duke@435 2858 // Initialize the remaining stuff, if any, with a ClearArray.
duke@435 2859 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@435 2860 }
duke@435 2861
duke@435 2862 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@435 2863 Node* start_offset,
duke@435 2864 Node* end_offset,
duke@435 2865 PhaseGVN* phase) {
never@503 2866 if (start_offset == end_offset) {
never@503 2867 // nothing to do
never@503 2868 return mem;
never@503 2869 }
never@503 2870
duke@435 2871 Compile* C = phase->C;
duke@435 2872 int unit = BytesPerLong;
duke@435 2873 Node* zbase = start_offset;
duke@435 2874 Node* zend = end_offset;
duke@435 2875
duke@435 2876 // Scale to the unit required by the CPU:
duke@435 2877 if (!Matcher::init_array_count_is_in_bytes) {
duke@435 2878 Node* shift = phase->intcon(exact_log2(unit));
kvn@4115 2879 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
kvn@4115 2880 zend = phase->transform( new(C) URShiftXNode(zend, shift) );
duke@435 2881 }
duke@435 2882
kvn@4410 2883 // Bulk clear double-words
kvn@4115 2884 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
kvn@4115 2885 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
kvn@4115 2886 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
duke@435 2887 return phase->transform(mem);
duke@435 2888 }
duke@435 2889
duke@435 2890 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@435 2891 intptr_t start_offset,
duke@435 2892 intptr_t end_offset,
duke@435 2893 PhaseGVN* phase) {
never@503 2894 if (start_offset == end_offset) {
never@503 2895 // nothing to do
never@503 2896 return mem;
never@503 2897 }
never@503 2898
duke@435 2899 Compile* C = phase->C;
duke@435 2900 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@435 2901 intptr_t done_offset = end_offset;
duke@435 2902 if ((done_offset % BytesPerLong) != 0) {
duke@435 2903 done_offset -= BytesPerInt;
duke@435 2904 }
duke@435 2905 if (done_offset > start_offset) {
duke@435 2906 mem = clear_memory(ctl, mem, dest,
duke@435 2907 start_offset, phase->MakeConX(done_offset), phase);
duke@435 2908 }
duke@435 2909 if (done_offset < end_offset) { // emit the final 32-bit store
kvn@4115 2910 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@435 2911 adr = phase->transform(adr);
duke@435 2912 const TypePtr* atp = TypeRawPtr::BOTTOM;
goetz@6479 2913 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
duke@435 2914 mem = phase->transform(mem);
duke@435 2915 done_offset += BytesPerInt;
duke@435 2916 }
duke@435 2917 assert(done_offset == end_offset, "");
duke@435 2918 return mem;
duke@435 2919 }
duke@435 2920
duke@435 2921 //=============================================================================
kvn@2694 2922 // Do not match memory edge.
kvn@2694 2923 uint StrIntrinsicNode::match_edge(uint idx) const {
kvn@2694 2924 return idx == 2 || idx == 3;
duke@435 2925 }
duke@435 2926
duke@435 2927 //------------------------------Ideal------------------------------------------
duke@435 2928 // Return a node which is more "ideal" than the current node. Strip out
duke@435 2929 // control copies
kvn@2694 2930 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@2694 2931 if (remove_dead_region(phase, can_reshape)) return this;
kvn@3311 2932 // Don't bother trying to transform a dead node
kvn@3311 2933 if (in(0) && in(0)->is_top()) return NULL;
kvn@2694 2934
kvn@2699 2935 if (can_reshape) {
kvn@2699 2936 Node* mem = phase->transform(in(MemNode::Memory));
kvn@2699 2937 // If transformed to a MergeMem, get the desired slice
kvn@2699 2938 uint alias_idx = phase->C->get_alias_index(adr_type());
kvn@2699 2939 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
kvn@2699 2940 if (mem != in(MemNode::Memory)) {
kvn@2699 2941 set_req(MemNode::Memory, mem);
kvn@2699 2942 return this;
kvn@2699 2943 }
kvn@2699 2944 }
kvn@2694 2945 return NULL;
rasbold@604 2946 }
rasbold@604 2947
kvn@3311 2948 //------------------------------Value------------------------------------------
kvn@3311 2949 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
kvn@3311 2950 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
kvn@3311 2951 return bottom_type();
kvn@3311 2952 }
kvn@3311 2953
duke@435 2954 //=============================================================================
kvn@4479 2955 //------------------------------match_edge-------------------------------------
kvn@4479 2956 // Do not match memory edge
kvn@4479 2957 uint EncodeISOArrayNode::match_edge(uint idx) const {
kvn@4479 2958 return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len)
kvn@4479 2959 }
kvn@4479 2960
kvn@4479 2961 //------------------------------Ideal------------------------------------------
kvn@4479 2962 // Return a node which is more "ideal" than the current node. Strip out
kvn@4479 2963 // control copies
kvn@4479 2964 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@4479 2965 return remove_dead_region(phase, can_reshape) ? this : NULL;
kvn@4479 2966 }
kvn@4479 2967
kvn@4479 2968 //------------------------------Value------------------------------------------
kvn@4479 2969 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const {
kvn@4479 2970 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
kvn@4479 2971 return bottom_type();
kvn@4479 2972 }
kvn@4479 2973
kvn@4479 2974 //=============================================================================
duke@435 2975 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@435 2976 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@435 2977 _adr_type(C->get_adr_type(alias_idx))
duke@435 2978 {
duke@435 2979 init_class_id(Class_MemBar);
duke@435 2980 Node* top = C->top();
duke@435 2981 init_req(TypeFunc::I_O,top);
duke@435 2982 init_req(TypeFunc::FramePtr,top);
duke@435 2983 init_req(TypeFunc::ReturnAdr,top);
duke@435 2984 if (precedent != NULL)
duke@435 2985 init_req(TypeFunc::Parms, precedent);
duke@435 2986 }
duke@435 2987
duke@435 2988 //------------------------------cmp--------------------------------------------
duke@435 2989 uint MemBarNode::hash() const { return NO_HASH; }
duke@435 2990 uint MemBarNode::cmp( const Node &n ) const {
duke@435 2991 return (&n == this); // Always fail except on self
duke@435 2992 }
duke@435 2993
duke@435 2994 //------------------------------make-------------------------------------------
duke@435 2995 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@435 2996 switch (opcode) {
goetz@6489 2997 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn);
goetz@6489 2998 case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn);
goetz@6489 2999 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn);
goetz@6489 3000 case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn);
goetz@6489 3001 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
goetz@6489 3002 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
goetz@6489 3003 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn);
goetz@6489 3004 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn);
goetz@6489 3005 case Op_Initialize: return new(C) InitializeNode(C, atp, pn);
goetz@6489 3006 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn);
goetz@6489 3007 default: ShouldNotReachHere(); return NULL;
duke@435 3008 }
duke@435 3009 }
duke@435 3010
duke@435 3011 //------------------------------Ideal------------------------------------------
duke@435 3012 // Return a node which is more "ideal" than the current node. Strip out
duke@435 3013 // control copies
duke@435 3014 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@1535 3015 if (remove_dead_region(phase, can_reshape)) return this;
kvn@3311 3016 // Don't bother trying to transform a dead node
adlertz@5291 3017 if (in(0) && in(0)->is_top()) {
adlertz@5291 3018 return NULL;
adlertz@5291 3019 }
kvn@1535 3020
kvn@1535 3021 // Eliminate volatile MemBars for scalar replaced objects.
kvn@5110 3022 if (can_reshape && req() == (Precedent+1)) {
kvn@5110 3023 bool eliminate = false;
kvn@5110 3024 int opc = Opcode();
kvn@5110 3025 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
kvn@5110 3026 // Volatile field loads and stores.
kvn@5110 3027 Node* my_mem = in(MemBarNode::Precedent);
adlertz@5291 3028 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
adlertz@5291 3029 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
adlertz@5317 3030 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
adlertz@5317 3031 // replace this Precedent (decodeN) with the Load instead.
adlertz@5317 3032 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
adlertz@5317 3033 Node* load_node = my_mem->in(1);
adlertz@5317 3034 set_req(MemBarNode::Precedent, load_node);
adlertz@5317 3035 phase->is_IterGVN()->_worklist.push(my_mem);
adlertz@5317 3036 my_mem = load_node;
adlertz@5317 3037 } else {
adlertz@5317 3038 assert(my_mem->unique_out() == this, "sanity");
adlertz@5317 3039 del_req(Precedent);
adlertz@5317 3040 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
adlertz@5317 3041 my_mem = NULL;
adlertz@5317 3042 }
adlertz@5291 3043 }
kvn@5110 3044 if (my_mem != NULL && my_mem->is_Mem()) {
kvn@5110 3045 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@5110 3046 // Check for scalar replaced object reference.
kvn@5110 3047 if( t_oop != NULL && t_oop->is_known_instance_field() &&
kvn@5110 3048 t_oop->offset() != Type::OffsetBot &&
kvn@5110 3049 t_oop->offset() != Type::OffsetTop) {
kvn@5110 3050 eliminate = true;
kvn@5110 3051 }
kvn@1535 3052 }
kvn@5110 3053 } else if (opc == Op_MemBarRelease) {
kvn@5110 3054 // Final field stores.
kvn@5110 3055 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
kvn@5110 3056 if ((alloc != NULL) && alloc->is_Allocate() &&
kvn@5110 3057 alloc->as_Allocate()->_is_non_escaping) {
kvn@5110 3058 // The allocated object does not escape.
kvn@5110 3059 eliminate = true;
kvn@5110 3060 }
kvn@5110 3061 }
kvn@5110 3062 if (eliminate) {
kvn@5110 3063 // Replace MemBar projections by its inputs.
kvn@5110 3064 PhaseIterGVN* igvn = phase->is_IterGVN();
kvn@5110 3065 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
kvn@5110 3066 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
kvn@5110 3067 // Must return either the original node (now dead) or a new node
kvn@5110 3068 // (Do not return a top here, since that would break the uniqueness of top.)
kvn@5110 3069 return new (phase->C) ConINode(TypeInt::ZERO);
kvn@1535 3070 }
kvn@1535 3071 }
kvn@1535 3072 return NULL;
duke@435 3073 }
duke@435 3074
duke@435 3075 //------------------------------Value------------------------------------------
duke@435 3076 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@435 3077 if( !in(0) ) return Type::TOP;
duke@435 3078 if( phase->type(in(0)) == Type::TOP )
duke@435 3079 return Type::TOP;
duke@435 3080 return TypeTuple::MEMBAR;
duke@435 3081 }
duke@435 3082
duke@435 3083 //------------------------------match------------------------------------------
duke@435 3084 // Construct projections for memory.
duke@435 3085 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@435 3086 switch (proj->_con) {
duke@435 3087 case TypeFunc::Control:
duke@435 3088 case TypeFunc::Memory:
kvn@4115 3089 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@435 3090 }
duke@435 3091 ShouldNotReachHere();
duke@435 3092 return NULL;
duke@435 3093 }
duke@435 3094
duke@435 3095 //===========================InitializeNode====================================
duke@435 3096 // SUMMARY:
duke@435 3097 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@435 3098 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@435 3099 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@435 3100 // It can coalesce related raw stores into larger units (called 'tiles').
duke@435 3101 // It can avoid zeroing new storage for memory units which have raw inits.
duke@435 3102 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@435 3103 //
duke@435 3104 // EXAMPLE:
duke@435 3105 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@435 3106 // ctl = incoming control; mem* = incoming memory
duke@435 3107 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@435 3108 // First allocate uninitialized memory and fill in the header:
duke@435 3109 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@435 3110 // ctl := alloc.Control; mem* := alloc.Memory*
duke@435 3111 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@435 3112 // Then initialize to zero the non-header parts of the raw memory block:
duke@435 3113 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@435 3114 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@435 3115 // After the initialize node executes, the object is ready for service:
duke@435 3116 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@435 3117 // Suppose its body is immediately initialized as {1,2}:
duke@435 3118 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@435 3119 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@435 3120 // mem.SLICE(#short[*]) := store2
duke@435 3121 //
duke@435 3122 // DETAILS:
duke@435 3123 // An InitializeNode collects and isolates object initialization after
duke@435 3124 // an AllocateNode and before the next possible safepoint. As a
duke@435 3125 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@435 3126 // down past any safepoint or any publication of the allocation.
duke@435 3127 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@435 3128 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@435 3129 //
duke@435 3130 // The semantics of the InitializeNode include an implicit zeroing of
duke@435 3131 // the new object from object header to the end of the object.
duke@435 3132 // (The object header and end are determined by the AllocateNode.)
duke@435 3133 //
duke@435 3134 // Certain stores may be added as direct inputs to the InitializeNode.
duke@435 3135 // These stores must update raw memory, and they must be to addresses
duke@435 3136 // derived from the raw address produced by AllocateNode, and with
duke@435 3137 // a constant offset. They must be ordered by increasing offset.
duke@435 3138 // The first one is at in(RawStores), the last at in(req()-1).
duke@435 3139 // Unlike most memory operations, they are not linked in a chain,
duke@435 3140 // but are displayed in parallel as users of the rawmem output of
duke@435 3141 // the allocation.
duke@435 3142 //
duke@435 3143 // (See comments in InitializeNode::capture_store, which continue
duke@435 3144 // the example given above.)
duke@435 3145 //
duke@435 3146 // When the associated Allocate is macro-expanded, the InitializeNode
duke@435 3147 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@435 3148 // may also be created at that point to represent any required zeroing.
duke@435 3149 // The InitializeNode is then marked 'complete', prohibiting further
duke@435 3150 // capturing of nearby memory operations.
duke@435 3151 //
duke@435 3152 // During macro-expansion, all captured initializations which store
twisti@1040 3153 // constant values of 32 bits or smaller are coalesced (if advantageous)
duke@435 3154 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@435 3155 // initialized in fewer memory operations. Memory words which are
duke@435 3156 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@435 3157 // by explicit stores of zero. (The code shape happens to do all
duke@435 3158 // zeroing first, then all other stores, with both sequences occurring
duke@435 3159 // in order of ascending offsets.)
duke@435 3160 //
duke@435 3161 // Alternatively, code may be inserted between an AllocateNode and its
duke@435 3162 // InitializeNode, to perform arbitrary initialization of the new object.
duke@435 3163 // E.g., the object copying intrinsics insert complex data transfers here.
duke@435 3164 // The initialization must then be marked as 'complete' disable the
duke@435 3165 // built-in zeroing semantics and the collection of initializing stores.
duke@435 3166 //
duke@435 3167 // While an InitializeNode is incomplete, reads from the memory state
duke@435 3168 // produced by it are optimizable if they match the control edge and
duke@435 3169 // new oop address associated with the allocation/initialization.
duke@435 3170 // They return a stored value (if the offset matches) or else zero.
duke@435 3171 // A write to the memory state, if it matches control and address,
duke@435 3172 // and if it is to a constant offset, may be 'captured' by the
duke@435 3173 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@435 3174 // inside the initialization, to the raw oop produced by the allocation.
duke@435 3175 // Operations on addresses which are provably distinct (e.g., to
duke@435 3176 // other AllocateNodes) are allowed to bypass the initialization.
duke@435 3177 //
duke@435 3178 // The effect of all this is to consolidate object initialization
duke@435 3179 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@435 3180 // single location, where it can be optimized as a unit.
duke@435 3181 //
duke@435 3182 // Only stores with an offset less than TrackedInitializationLimit words
duke@435 3183 // will be considered for capture by an InitializeNode. This puts a
duke@435 3184 // reasonable limit on the complexity of optimized initializations.
duke@435 3185
duke@435 3186 //---------------------------InitializeNode------------------------------------
duke@435 3187 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
roland@3392 3188 : _is_complete(Incomplete), _does_not_escape(false),
duke@435 3189 MemBarNode(C, adr_type, rawoop)
duke@435 3190 {
duke@435 3191 init_class_id(Class_Initialize);
duke@435 3192
duke@435 3193 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@435 3194 assert(in(RawAddress) == rawoop, "proper init");
duke@435 3195 // Note: allocation() can be NULL, for secondary initialization barriers
duke@435 3196 }
duke@435 3197
duke@435 3198 // Since this node is not matched, it will be processed by the
duke@435 3199 // register allocator. Declare that there are no constraints
duke@435 3200 // on the allocation of the RawAddress edge.
duke@435 3201 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@435 3202 // This edge should be set to top, by the set_complete. But be conservative.
duke@435 3203 if (idx == InitializeNode::RawAddress)
duke@435 3204 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@435 3205 return RegMask::Empty;
duke@435 3206 }
duke@435 3207
duke@435 3208 Node* InitializeNode::memory(uint alias_idx) {
duke@435 3209 Node* mem = in(Memory);
duke@435 3210 if (mem->is_MergeMem()) {
duke@435 3211 return mem->as_MergeMem()->memory_at(alias_idx);
duke@435 3212 } else {
duke@435 3213 // incoming raw memory is not split
duke@435 3214 return mem;
duke@435 3215 }
duke@435 3216 }
duke@435 3217
duke@435 3218 bool InitializeNode::is_non_zero() {
duke@435 3219 if (is_complete()) return false;
duke@435 3220 remove_extra_zeroes();
duke@435 3221 return (req() > RawStores);
duke@435 3222 }
duke@435 3223
duke@435 3224 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@435 3225 assert(!is_complete(), "caller responsibility");
kvn@3157 3226 _is_complete = Complete;
duke@435 3227
duke@435 3228 // After this node is complete, it contains a bunch of
duke@435 3229 // raw-memory initializations. There is no need for
duke@435 3230 // it to have anything to do with non-raw memory effects.
duke@435 3231 // Therefore, tell all non-raw users to re-optimize themselves,
duke@435 3232 // after skipping the memory effects of this initialization.
duke@435 3233 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@435 3234 if (igvn) igvn->add_users_to_worklist(this);
duke@435 3235 }
duke@435 3236
duke@435 3237 // convenience function
duke@435 3238 // return false if the init contains any stores already
duke@435 3239 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@435 3240 InitializeNode* init = initialization();
duke@435 3241 if (init == NULL || init->is_complete()) return false;
duke@435 3242 init->remove_extra_zeroes();
duke@435 3243 // for now, if this allocation has already collected any inits, bail:
duke@435 3244 if (init->is_non_zero()) return false;
duke@435 3245 init->set_complete(phase);
duke@435 3246 return true;
duke@435 3247 }
duke@435 3248
duke@435 3249 void InitializeNode::remove_extra_zeroes() {
duke@435 3250 if (req() == RawStores) return;
duke@435 3251 Node* zmem = zero_memory();
duke@435 3252 uint fill = RawStores;
duke@435 3253 for (uint i = fill; i < req(); i++) {
duke@435 3254 Node* n = in(i);
duke@435 3255 if (n->is_top() || n == zmem) continue; // skip
duke@435 3256 if (fill < i) set_req(fill, n); // compact
duke@435 3257 ++fill;
duke@435 3258 }
duke@435 3259 // delete any empty spaces created:
duke@435 3260 while (fill < req()) {
duke@435 3261 del_req(fill);
duke@435 3262 }
duke@435 3263 }
duke@435 3264
duke@435 3265 // Helper for remembering which stores go with which offsets.
duke@435 3266 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@435 3267 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@435 3268 intptr_t offset = -1;
duke@435 3269 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@435 3270 phase, offset);
duke@435 3271 if (base == NULL) return -1; // something is dead,
duke@435 3272 if (offset < 0) return -1; // dead, dead
duke@435 3273 return offset;
duke@435 3274 }
duke@435 3275
duke@435 3276 // Helper for proving that an initialization expression is
duke@435 3277 // "simple enough" to be folded into an object initialization.
duke@435 3278 // Attempts to prove that a store's initial value 'n' can be captured
duke@435 3279 // within the initialization without creating a vicious cycle, such as:
duke@435 3280 // { Foo p = new Foo(); p.next = p; }
duke@435 3281 // True for constants and parameters and small combinations thereof.
kvn@5110 3282 bool InitializeNode::detect_init_independence(Node* n, int& count) {
duke@435 3283 if (n == NULL) return true; // (can this really happen?)
duke@435 3284 if (n->is_Proj()) n = n->in(0);
duke@435 3285 if (n == this) return false; // found a cycle
duke@435 3286 if (n->is_Con()) return true;
duke@435 3287 if (n->is_Start()) return true; // params, etc., are OK
duke@435 3288 if (n->is_Root()) return true; // even better
duke@435 3289
duke@435 3290 Node* ctl = n->in(0);
duke@435 3291 if (ctl != NULL && !ctl->is_top()) {
duke@435 3292 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@435 3293 if (ctl == this) return false;
duke@435 3294
duke@435 3295 // If we already know that the enclosing memory op is pinned right after
duke@435 3296 // the init, then any control flow that the store has picked up
duke@435 3297 // must have preceded the init, or else be equal to the init.
duke@435 3298 // Even after loop optimizations (which might change control edges)
duke@435 3299 // a store is never pinned *before* the availability of its inputs.
kvn@554 3300 if (!MemNode::all_controls_dominate(n, this))
duke@435 3301 return false; // failed to prove a good control
duke@435 3302 }
duke@435 3303
duke@435 3304 // Check data edges for possible dependencies on 'this'.
duke@435 3305 if ((count += 1) > 20) return false; // complexity limit
duke@435 3306 for (uint i = 1; i < n->req(); i++) {
duke@435 3307 Node* m = n->in(i);
duke@435 3308 if (m == NULL || m == n || m->is_top()) continue;
duke@435 3309 uint first_i = n->find_edge(m);
duke@435 3310 if (i != first_i) continue; // process duplicate edge just once
kvn@5110 3311 if (!detect_init_independence(m, count)) {
duke@435 3312 return false;
duke@435 3313 }
duke@435 3314 }
duke@435 3315
duke@435 3316 return true;
duke@435 3317 }
duke@435 3318
duke@435 3319 // Here are all the checks a Store must pass before it can be moved into
duke@435 3320 // an initialization. Returns zero if a check fails.
duke@435 3321 // On success, returns the (constant) offset to which the store applies,
duke@435 3322 // within the initialized memory.
roland@4657 3323 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
duke@435 3324 const int FAIL = 0;
duke@435 3325 if (st->req() != MemNode::ValueIn + 1)
duke@435 3326 return FAIL; // an inscrutable StoreNode (card mark?)
duke@435 3327 Node* ctl = st->in(MemNode::Control);
duke@435 3328 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@435 3329 return FAIL; // must be unconditional after the initialization
duke@435 3330 Node* mem = st->in(MemNode::Memory);
duke@435 3331 if (!(mem->is_Proj() && mem->in(0) == this))
duke@435 3332 return FAIL; // must not be preceded by other stores
duke@435 3333 Node* adr = st->in(MemNode::Address);
duke@435 3334 intptr_t offset;
duke@435 3335 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@435 3336 if (alloc == NULL)
duke@435 3337 return FAIL; // inscrutable address
duke@435 3338 if (alloc != allocation())
duke@435 3339 return FAIL; // wrong allocation! (store needs to float up)
duke@435 3340 Node* val = st->in(MemNode::ValueIn);
duke@435 3341 int complexity_count = 0;
kvn@5110 3342 if (!detect_init_independence(val, complexity_count))
duke@435 3343 return FAIL; // stored value must be 'simple enough'
duke@435 3344
roland@4657 3345 // The Store can be captured only if nothing after the allocation
roland@4657 3346 // and before the Store is using the memory location that the store
roland@4657 3347 // overwrites.
roland@4657 3348 bool failed = false;
roland@4657 3349 // If is_complete_with_arraycopy() is true the shape of the graph is
roland@4657 3350 // well defined and is safe so no need for extra checks.
roland@4657 3351 if (!is_complete_with_arraycopy()) {
roland@4657 3352 // We are going to look at each use of the memory state following
roland@4657 3353 // the allocation to make sure nothing reads the memory that the
roland@4657 3354 // Store writes.
roland@4657 3355 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
roland@4657 3356 int alias_idx = phase->C->get_alias_index(t_adr);
roland@4657 3357 ResourceMark rm;
roland@4657 3358 Unique_Node_List mems;
roland@4657 3359 mems.push(mem);
roland@4657 3360 Node* unique_merge = NULL;
roland@4657 3361 for (uint next = 0; next < mems.size(); ++next) {
roland@4657 3362 Node *m = mems.at(next);
roland@4657 3363 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
roland@4657 3364 Node *n = m->fast_out(j);
roland@4657 3365 if (n->outcnt() == 0) {
roland@4657 3366 continue;
roland@4657 3367 }
roland@4657 3368 if (n == st) {
roland@4657 3369 continue;
roland@4657 3370 } else if (n->in(0) != NULL && n->in(0) != ctl) {
roland@4657 3371 // If the control of this use is different from the control
roland@4657 3372 // of the Store which is right after the InitializeNode then
roland@4657 3373 // this node cannot be between the InitializeNode and the
roland@4657 3374 // Store.
roland@4657 3375 continue;
roland@4657 3376 } else if (n->is_MergeMem()) {
roland@4657 3377 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
roland@4657 3378 // We can hit a MergeMemNode (that will likely go away
roland@4657 3379 // later) that is a direct use of the memory state
roland@4657 3380 // following the InitializeNode on the same slice as the
roland@4657 3381 // store node that we'd like to capture. We need to check
roland@4657 3382 // the uses of the MergeMemNode.
roland@4657 3383 mems.push(n);
roland@4657 3384 }
roland@4657 3385 } else if (n->is_Mem()) {
roland@4657 3386 Node* other_adr = n->in(MemNode::Address);
roland@4657 3387 if (other_adr == adr) {
roland@4657 3388 failed = true;
roland@4657 3389 break;
roland@4657 3390 } else {
roland@4657 3391 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
roland@4657 3392 if (other_t_adr != NULL) {
roland@4657 3393 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
roland@4657 3394 if (other_alias_idx == alias_idx) {
roland@4657 3395 // A load from the same memory slice as the store right
roland@4657 3396 // after the InitializeNode. We check the control of the
roland@4657 3397 // object/array that is loaded from. If it's the same as
roland@4657 3398 // the store control then we cannot capture the store.
roland@4657 3399 assert(!n->is_Store(), "2 stores to same slice on same control?");
roland@4657 3400 Node* base = other_adr;
roland@4657 3401 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
roland@4657 3402 base = base->in(AddPNode::Base);
roland@4657 3403 if (base != NULL) {
roland@4657 3404 base = base->uncast();
roland@4657 3405 if (base->is_Proj() && base->in(0) == alloc) {
roland@4657 3406 failed = true;
roland@4657 3407 break;
roland@4657 3408 }
roland@4657 3409 }
roland@4657 3410 }
roland@4657 3411 }
roland@4657 3412 }
roland@4657 3413 } else {
roland@4657 3414 failed = true;
roland@4657 3415 break;
roland@4657 3416 }
roland@4657 3417 }
roland@4657 3418 }
roland@4657 3419 }
roland@4657 3420 if (failed) {
roland@4657 3421 if (!can_reshape) {
roland@4657 3422 // We decided we couldn't capture the store during parsing. We
roland@4657 3423 // should try again during the next IGVN once the graph is
roland@4657 3424 // cleaner.
roland@4657 3425 phase->C->record_for_igvn(st);
roland@4657 3426 }
roland@4657 3427 return FAIL;
roland@4657 3428 }
roland@4657 3429
duke@435 3430 return offset; // success
duke@435 3431 }
duke@435 3432
duke@435 3433 // Find the captured store in(i) which corresponds to the range
duke@435 3434 // [start..start+size) in the initialized object.
duke@435 3435 // If there is one, return its index i. If there isn't, return the
duke@435 3436 // negative of the index where it should be inserted.
duke@435 3437 // Return 0 if the queried range overlaps an initialization boundary
duke@435 3438 // or if dead code is encountered.
duke@435 3439 // If size_in_bytes is zero, do not bother with overlap checks.
duke@435 3440 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@435 3441 int size_in_bytes,
duke@435 3442 PhaseTransform* phase) {
duke@435 3443 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@435 3444
duke@435 3445 if (is_complete())
duke@435 3446 return FAIL; // arraycopy got here first; punt
duke@435 3447
duke@435 3448 assert(allocation() != NULL, "must be present");
duke@435 3449
duke@435 3450 // no negatives, no header fields:
coleenp@548 3451 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@435 3452
duke@435 3453 // after a certain size, we bail out on tracking all the stores:
duke@435 3454 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@435 3455 if (start >= ti_limit) return FAIL;
duke@435 3456
duke@435 3457 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@435 3458 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@435 3459
duke@435 3460 Node* st = in(i);
duke@435 3461 intptr_t st_off = get_store_offset(st, phase);
duke@435 3462 if (st_off < 0) {
duke@435 3463 if (st != zero_memory()) {
duke@435 3464 return FAIL; // bail out if there is dead garbage
duke@435 3465 }
duke@435 3466 } else if (st_off > start) {
duke@435 3467 // ...we are done, since stores are ordered
duke@435 3468 if (st_off < start + size_in_bytes) {
duke@435 3469 return FAIL; // the next store overlaps
duke@435 3470 }
duke@435 3471 return -(int)i; // not found; here is where to put it
duke@435 3472 } else if (st_off < start) {
duke@435 3473 if (size_in_bytes != 0 &&
duke@435 3474 start < st_off + MAX_STORE &&
duke@435 3475 start < st_off + st->as_Store()->memory_size()) {
duke@435 3476 return FAIL; // the previous store overlaps
duke@435 3477 }
duke@435 3478 } else {
duke@435 3479 if (size_in_bytes != 0 &&
duke@435 3480 st->as_Store()->memory_size() != size_in_bytes) {
duke@435 3481 return FAIL; // mismatched store size
duke@435 3482 }
duke@435 3483 return i;
duke@435 3484 }
duke@435 3485
duke@435 3486 ++i;
duke@435 3487 }
duke@435 3488 }
duke@435 3489
duke@435 3490 // Look for a captured store which initializes at the offset 'start'
duke@435 3491 // with the given size. If there is no such store, and no other
duke@435 3492 // initialization interferes, then return zero_memory (the memory
duke@435 3493 // projection of the AllocateNode).
duke@435 3494 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@435 3495 PhaseTransform* phase) {
duke@435 3496 assert(stores_are_sane(phase), "");
duke@435 3497 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@435 3498 if (i == 0) {
duke@435 3499 return NULL; // something is dead
duke@435 3500 } else if (i < 0) {
duke@435 3501 return zero_memory(); // just primordial zero bits here
duke@435 3502 } else {
duke@435 3503 Node* st = in(i); // here is the store at this position
duke@435 3504 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@435 3505 return st;
duke@435 3506 }
duke@435 3507 }
duke@435 3508
duke@435 3509 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@435 3510 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@435 3511 PhaseTransform* phase) {
duke@435 3512 Node* addr = in(RawAddress);
duke@435 3513 if (offset != 0) {
duke@435 3514 Compile* C = phase->C;
kvn@4115 3515 addr = phase->transform( new (C) AddPNode(C->top(), addr,
duke@435 3516 phase->MakeConX(offset)) );
duke@435 3517 }
duke@435 3518 return addr;
duke@435 3519 }
duke@435 3520
duke@435 3521 // Clone the given store, converting it into a raw store
duke@435 3522 // initializing a field or element of my new object.
duke@435 3523 // Caller is responsible for retiring the original store,
duke@435 3524 // with subsume_node or the like.
duke@435 3525 //
duke@435 3526 // From the example above InitializeNode::InitializeNode,
duke@435 3527 // here are the old stores to be captured:
duke@435 3528 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@435 3529 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@435 3530 //
duke@435 3531 // Here is the changed code; note the extra edges on init:
duke@435 3532 // alloc = (Allocate ...)
duke@435 3533 // rawoop = alloc.RawAddress
duke@435 3534 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@435 3535 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@435 3536 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@435 3537 // rawstore1 rawstore2)
duke@435 3538 //
duke@435 3539 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
roland@4657 3540 PhaseTransform* phase, bool can_reshape) {
duke@435 3541 assert(stores_are_sane(phase), "");
duke@435 3542
duke@435 3543 if (start < 0) return NULL;
roland@4657 3544 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
duke@435 3545
duke@435 3546 Compile* C = phase->C;
duke@435 3547 int size_in_bytes = st->memory_size();
duke@435 3548 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@435 3549 if (i == 0) return NULL; // bail out
duke@435 3550 Node* prev_mem = NULL; // raw memory for the captured store
duke@435 3551 if (i > 0) {
duke@435 3552 prev_mem = in(i); // there is a pre-existing store under this one
duke@435 3553 set_req(i, C->top()); // temporarily disconnect it
duke@435 3554 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@435 3555 } else {
duke@435 3556 i = -i; // no pre-existing store
duke@435 3557 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@435 3558 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@435 3559 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@435 3560 else
duke@435 3561 ins_req(i, C->top()); // build a new edge
duke@435 3562 }
duke@435 3563 Node* new_st = st->clone();
duke@435 3564 new_st->set_req(MemNode::Control, in(Control));
duke@435 3565 new_st->set_req(MemNode::Memory, prev_mem);
duke@435 3566 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@435 3567 new_st = phase->transform(new_st);
duke@435 3568
duke@435 3569 // At this point, new_st might have swallowed a pre-existing store
duke@435 3570 // at the same offset, or perhaps new_st might have disappeared,
duke@435 3571 // if it redundantly stored the same value (or zero to fresh memory).
duke@435 3572
duke@435 3573 // In any case, wire it in:
duke@435 3574 set_req(i, new_st);
duke@435 3575
duke@435 3576 // The caller may now kill the old guy.
duke@435 3577 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@435 3578 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@435 3579 assert(!is_complete(), "");
duke@435 3580 return new_st;
duke@435 3581 }
duke@435 3582
duke@435 3583 static bool store_constant(jlong* tiles, int num_tiles,
duke@435 3584 intptr_t st_off, int st_size,
duke@435 3585 jlong con) {
duke@435 3586 if ((st_off & (st_size-1)) != 0)
duke@435 3587 return false; // strange store offset (assume size==2**N)
duke@435 3588 address addr = (address)tiles + st_off;
duke@435 3589 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@435 3590 switch (st_size) {
duke@435 3591 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@435 3592 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@435 3593 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@435 3594 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@435 3595 default: return false; // strange store size (detect size!=2**N here)
duke@435 3596 }
duke@435 3597 return true; // return success to caller
duke@435 3598 }
duke@435 3599
duke@435 3600 // Coalesce subword constants into int constants and possibly
duke@435 3601 // into long constants. The goal, if the CPU permits,
duke@435 3602 // is to initialize the object with a small number of 64-bit tiles.
duke@435 3603 // Also, convert floating-point constants to bit patterns.
duke@435 3604 // Non-constants are not relevant to this pass.
duke@435 3605 //
duke@435 3606 // In terms of the running example on InitializeNode::InitializeNode
duke@435 3607 // and InitializeNode::capture_store, here is the transformation
duke@435 3608 // of rawstore1 and rawstore2 into rawstore12:
duke@435 3609 // alloc = (Allocate ...)
duke@435 3610 // rawoop = alloc.RawAddress
duke@435 3611 // tile12 = 0x00010002
duke@435 3612 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@435 3613 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@435 3614 //
duke@435 3615 void
duke@435 3616 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@435 3617 Node* size_in_bytes,
duke@435 3618 PhaseGVN* phase) {
duke@435 3619 Compile* C = phase->C;
duke@435 3620
duke@435 3621 assert(stores_are_sane(phase), "");
duke@435 3622 // Note: After this pass, they are not completely sane,
duke@435 3623 // since there may be some overlaps.
duke@435 3624
duke@435 3625 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@435 3626
duke@435 3627 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@435 3628 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@435 3629 size_limit = MIN2(size_limit, ti_limit);
duke@435 3630 size_limit = align_size_up(size_limit, BytesPerLong);
duke@435 3631 int num_tiles = size_limit / BytesPerLong;
duke@435 3632
duke@435 3633 // allocate space for the tile map:
duke@435 3634 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@435 3635 jlong tiles_buf[small_len];
duke@435 3636 Node* nodes_buf[small_len];
duke@435 3637 jlong inits_buf[small_len];
duke@435 3638 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@435 3639 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@435 3640 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@435 3641 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@435 3642 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@435 3643 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@435 3644 // tiles: exact bitwise model of all primitive constants
duke@435 3645 // nodes: last constant-storing node subsumed into the tiles model
duke@435 3646 // inits: which bytes (in each tile) are touched by any initializations
duke@435 3647
duke@435 3648 //// Pass A: Fill in the tile model with any relevant stores.
duke@435 3649
duke@435 3650 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@435 3651 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@435 3652 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@435 3653 Node* zmem = zero_memory(); // initially zero memory state
duke@435 3654 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@435 3655 Node* st = in(i);
duke@435 3656 intptr_t st_off = get_store_offset(st, phase);
duke@435 3657
duke@435 3658 // Figure out the store's offset and constant value:
duke@435 3659 if (st_off < header_size) continue; //skip (ignore header)
duke@435 3660 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@435 3661 int st_size = st->as_Store()->memory_size();
duke@435 3662 if (st_off + st_size > size_limit) break;
duke@435 3663
duke@435 3664 // Record which bytes are touched, whether by constant or not.
duke@435 3665 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@435 3666 continue; // skip (strange store size)
duke@435 3667
duke@435 3668 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@435 3669 if (!val->singleton()) continue; //skip (non-con store)
duke@435 3670 BasicType type = val->basic_type();
duke@435 3671
duke@435 3672 jlong con = 0;
duke@435 3673 switch (type) {
duke@435 3674 case T_INT: con = val->is_int()->get_con(); break;
duke@435 3675 case T_LONG: con = val->is_long()->get_con(); break;
duke@435 3676 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@435 3677 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@435 3678 default: continue; //skip (odd store type)
duke@435 3679 }
duke@435 3680
duke@435 3681 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@435 3682 st->Opcode() == Op_StoreL) {
duke@435 3683 continue; // This StoreL is already optimal.
duke@435 3684 }
duke@435 3685
duke@435 3686 // Store down the constant.
duke@435 3687 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@435 3688
duke@435 3689 intptr_t j = st_off >> LogBytesPerLong;
duke@435 3690
duke@435 3691 if (type == T_INT && st_size == BytesPerInt
duke@435 3692 && (st_off & BytesPerInt) == BytesPerInt) {
duke@435 3693 jlong lcon = tiles[j];
duke@435 3694 if (!Matcher::isSimpleConstant64(lcon) &&
duke@435 3695 st->Opcode() == Op_StoreI) {
duke@435 3696 // This StoreI is already optimal by itself.
duke@435 3697 jint* intcon = (jint*) &tiles[j];
duke@435 3698 intcon[1] = 0; // undo the store_constant()
duke@435 3699
duke@435 3700 // If the previous store is also optimal by itself, back up and
duke@435 3701 // undo the action of the previous loop iteration... if we can.
duke@435 3702 // But if we can't, just let the previous half take care of itself.
duke@435 3703 st = nodes[j];
duke@435 3704 st_off -= BytesPerInt;
duke@435 3705 con = intcon[0];
duke@435 3706 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@435 3707 assert(st_off >= header_size, "still ignoring header");
duke@435 3708 assert(get_store_offset(st, phase) == st_off, "must be");
duke@435 3709 assert(in(i-1) == zmem, "must be");
duke@435 3710 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@435 3711 assert(con == tcon->is_int()->get_con(), "must be");
duke@435 3712 // Undo the effects of the previous loop trip, which swallowed st:
duke@435 3713 intcon[0] = 0; // undo store_constant()
duke@435 3714 set_req(i-1, st); // undo set_req(i, zmem)
duke@435 3715 nodes[j] = NULL; // undo nodes[j] = st
duke@435 3716 --old_subword; // undo ++old_subword
duke@435 3717 }
duke@435 3718 continue; // This StoreI is already optimal.
duke@435 3719 }
duke@435 3720 }
duke@435 3721
duke@435 3722 // This store is not needed.
duke@435 3723 set_req(i, zmem);
duke@435 3724 nodes[j] = st; // record for the moment
duke@435 3725 if (st_size < BytesPerLong) // something has changed
duke@435 3726 ++old_subword; // includes int/float, but who's counting...
duke@435 3727 else ++old_long;
duke@435 3728 }
duke@435 3729
duke@435 3730 if ((old_subword + old_long) == 0)
duke@435 3731 return; // nothing more to do
duke@435 3732
duke@435 3733 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@435 3734 // Be sure to insert them before overlapping non-constant stores.
duke@435 3735 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@435 3736 for (int j = 0; j < num_tiles; j++) {
duke@435 3737 jlong con = tiles[j];
duke@435 3738 jlong init = inits[j];
duke@435 3739 if (con == 0) continue;
duke@435 3740 jint con0, con1; // split the constant, address-wise
duke@435 3741 jint init0, init1; // split the init map, address-wise
duke@435 3742 { union { jlong con; jint intcon[2]; } u;
duke@435 3743 u.con = con;
duke@435 3744 con0 = u.intcon[0];
duke@435 3745 con1 = u.intcon[1];
duke@435 3746 u.con = init;
duke@435 3747 init0 = u.intcon[0];
duke@435 3748 init1 = u.intcon[1];
duke@435 3749 }
duke@435 3750
duke@435 3751 Node* old = nodes[j];
duke@435 3752 assert(old != NULL, "need the prior store");
duke@435 3753 intptr_t offset = (j * BytesPerLong);
duke@435 3754
duke@435 3755 bool split = !Matcher::isSimpleConstant64(con);
duke@435 3756
duke@435 3757 if (offset < header_size) {
duke@435 3758 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@435 3759 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@435 3760 split = true; // only the second word counts
duke@435 3761 // Example: int a[] = { 42 ... }
duke@435 3762 } else if (con0 == 0 && init0 == -1) {
duke@435 3763 split = true; // first word is covered by full inits
duke@435 3764 // Example: int a[] = { ... foo(), 42 ... }
duke@435 3765 } else if (con1 == 0 && init1 == -1) {
duke@435 3766 split = true; // second word is covered by full inits
duke@435 3767 // Example: int a[] = { ... 42, foo() ... }
duke@435 3768 }
duke@435 3769
duke@435 3770 // Here's a case where init0 is neither 0 nor -1:
duke@435 3771 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@435 3772 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@435 3773 // In this case the tile is not split; it is (jlong)42.
duke@435 3774 // The big tile is stored down, and then the foo() value is inserted.
duke@435 3775 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@435 3776
duke@435 3777 Node* ctl = old->in(MemNode::Control);
duke@435 3778 Node* adr = make_raw_address(offset, phase);
duke@435 3779 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@435 3780
duke@435 3781 // One or two coalesced stores to plop down.
duke@435 3782 Node* st[2];
duke@435 3783 intptr_t off[2];
duke@435 3784 int nst = 0;
duke@435 3785 if (!split) {
duke@435 3786 ++new_long;
duke@435 3787 off[nst] = offset;
coleenp@548 3788 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
goetz@6479 3789 phase->longcon(con), T_LONG, MemNode::unordered);
duke@435 3790 } else {
duke@435 3791 // Omit either if it is a zero.
duke@435 3792 if (con0 != 0) {
duke@435 3793 ++new_int;
duke@435 3794 off[nst] = offset;
coleenp@548 3795 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
goetz@6479 3796 phase->intcon(con0), T_INT, MemNode::unordered);
duke@435 3797 }
duke@435 3798 if (con1 != 0) {
duke@435 3799 ++new_int;
duke@435 3800 offset += BytesPerInt;
duke@435 3801 adr = make_raw_address(offset, phase);
duke@435 3802 off[nst] = offset;
coleenp@548 3803 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
goetz@6479 3804 phase->intcon(con1), T_INT, MemNode::unordered);
duke@435 3805 }
duke@435 3806 }
duke@435 3807
duke@435 3808 // Insert second store first, then the first before the second.
duke@435 3809 // Insert each one just before any overlapping non-constant stores.
duke@435 3810 while (nst > 0) {
duke@435 3811 Node* st1 = st[--nst];
duke@435 3812 C->copy_node_notes_to(st1, old);
duke@435 3813 st1 = phase->transform(st1);
duke@435 3814 offset = off[nst];
duke@435 3815 assert(offset >= header_size, "do not smash header");
duke@435 3816 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@435 3817 guarantee(ins_idx != 0, "must re-insert constant store");
duke@435 3818 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@435 3819 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@435 3820 set_req(--ins_idx, st1);
duke@435 3821 else
duke@435 3822 ins_req(ins_idx, st1);
duke@435 3823 }
duke@435 3824 }
duke@435 3825
duke@435 3826 if (PrintCompilation && WizardMode)
duke@435 3827 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@435 3828 old_subword, old_long, new_int, new_long);
duke@435 3829 if (C->log() != NULL)
duke@435 3830 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@435 3831 old_subword, old_long, new_int, new_long);
duke@435 3832
duke@435 3833 // Clean up any remaining occurrences of zmem:
duke@435 3834 remove_extra_zeroes();
duke@435 3835 }
duke@435 3836
duke@435 3837 // Explore forward from in(start) to find the first fully initialized
duke@435 3838 // word, and return its offset. Skip groups of subword stores which
duke@435 3839 // together initialize full words. If in(start) is itself part of a
duke@435 3840 // fully initialized word, return the offset of in(start). If there
duke@435 3841 // are no following full-word stores, or if something is fishy, return
duke@435 3842 // a negative value.
duke@435 3843 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@435 3844 int int_map = 0;
duke@435 3845 intptr_t int_map_off = 0;
duke@435 3846 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@435 3847
duke@435 3848 for (uint i = start, limit = req(); i < limit; i++) {
duke@435 3849 Node* st = in(i);
duke@435 3850
duke@435 3851 intptr_t st_off = get_store_offset(st, phase);
duke@435 3852 if (st_off < 0) break; // return conservative answer
duke@435 3853
duke@435 3854 int st_size = st->as_Store()->memory_size();
duke@435 3855 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@435 3856 return st_off; // we found a complete word init
duke@435 3857 }
duke@435 3858
duke@435 3859 // update the map:
duke@435 3860
duke@435 3861 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@435 3862 if (this_int_off != int_map_off) {
duke@435 3863 // reset the map:
duke@435 3864 int_map = 0;
duke@435 3865 int_map_off = this_int_off;
duke@435 3866 }
duke@435 3867
duke@435 3868 int subword_off = st_off - this_int_off;
duke@435 3869 int_map |= right_n_bits(st_size) << subword_off;
duke@435 3870 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@435 3871 return this_int_off; // we found a complete word init
duke@435 3872 }
duke@435 3873
duke@435 3874 // Did this store hit or cross the word boundary?
duke@435 3875 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@435 3876 if (next_int_off == this_int_off + BytesPerInt) {
duke@435 3877 // We passed the current int, without fully initializing it.
duke@435 3878 int_map_off = next_int_off;
duke@435 3879 int_map >>= BytesPerInt;
duke@435 3880 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@435 3881 // We passed the current and next int.
duke@435 3882 return this_int_off + BytesPerInt;
duke@435 3883 }
duke@435 3884 }
duke@435 3885
duke@435 3886 return -1;
duke@435 3887 }
duke@435 3888
duke@435 3889
duke@435 3890 // Called when the associated AllocateNode is expanded into CFG.
duke@435 3891 // At this point, we may perform additional optimizations.
duke@435 3892 // Linearize the stores by ascending offset, to make memory
duke@435 3893 // activity as coherent as possible.
duke@435 3894 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@435 3895 intptr_t header_size,
duke@435 3896 Node* size_in_bytes,
duke@435 3897 PhaseGVN* phase) {
duke@435 3898 assert(!is_complete(), "not already complete");
duke@435 3899 assert(stores_are_sane(phase), "");
duke@435 3900 assert(allocation() != NULL, "must be present");
duke@435 3901
duke@435 3902 remove_extra_zeroes();
duke@435 3903
duke@435 3904 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@435 3905 // reduce instruction count for common initialization patterns
duke@435 3906 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@435 3907
duke@435 3908 Node* zmem = zero_memory(); // initially zero memory state
duke@435 3909 Node* inits = zmem; // accumulating a linearized chain of inits
duke@435 3910 #ifdef ASSERT
coleenp@548 3911 intptr_t first_offset = allocation()->minimum_header_size();
coleenp@548 3912 intptr_t last_init_off = first_offset; // previous init offset
coleenp@548 3913 intptr_t last_init_end = first_offset; // previous init offset+size
coleenp@548 3914 intptr_t last_tile_end = first_offset; // previous tile offset+size
duke@435 3915 #endif
duke@435 3916 intptr_t zeroes_done = header_size;
duke@435 3917
duke@435 3918 bool do_zeroing = true; // we might give up if inits are very sparse
duke@435 3919 int big_init_gaps = 0; // how many large gaps have we seen?
duke@435 3920
duke@435 3921 if (ZeroTLAB) do_zeroing = false;
duke@435 3922 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@435 3923
duke@435 3924 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@435 3925 Node* st = in(i);
duke@435 3926 intptr_t st_off = get_store_offset(st, phase);
duke@435 3927 if (st_off < 0)
duke@435 3928 break; // unknown junk in the inits
duke@435 3929 if (st->in(MemNode::Memory) != zmem)
duke@435 3930 break; // complicated store chains somehow in list
duke@435 3931
duke@435 3932 int st_size = st->as_Store()->memory_size();
duke@435 3933 intptr_t next_init_off = st_off + st_size;
duke@435 3934
duke@435 3935 if (do_zeroing && zeroes_done < next_init_off) {
duke@435 3936 // See if this store needs a zero before it or under it.
duke@435 3937 intptr_t zeroes_needed = st_off;
duke@435 3938
duke@435 3939 if (st_size < BytesPerInt) {
duke@435 3940 // Look for subword stores which only partially initialize words.
duke@435 3941 // If we find some, we must lay down some word-level zeroes first,
duke@435 3942 // underneath the subword stores.
duke@435 3943 //
duke@435 3944 // Examples:
duke@435 3945 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
duke@435 3946 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
duke@435 3947 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
duke@435 3948 //
duke@435 3949 // Note: coalesce_subword_stores may have already done this,
duke@435 3950 // if it was prompted by constant non-zero subword initializers.
duke@435 3951 // But this case can still arise with non-constant stores.
duke@435 3952
duke@435 3953 intptr_t next_full_store = find_next_fullword_store(i, phase);
duke@435 3954
duke@435 3955 // In the examples above:
duke@435 3956 // in(i) p q r s x y z
duke@435 3957 // st_off 12 13 14 15 12 13 14
duke@435 3958 // st_size 1 1 1 1 1 1 1
duke@435 3959 // next_full_s. 12 16 16 16 16 16 16
duke@435 3960 // z's_done 12 16 16 16 12 16 12
duke@435 3961 // z's_needed 12 16 16 16 16 16 16
duke@435 3962 // zsize 0 0 0 0 4 0 4
duke@435 3963 if (next_full_store < 0) {
duke@435 3964 // Conservative tack: Zero to end of current word.
duke@435 3965 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
duke@435 3966 } else {
duke@435 3967 // Zero to beginning of next fully initialized word.
duke@435 3968 // Or, don't zero at all, if we are already in that word.
duke@435 3969 assert(next_full_store >= zeroes_needed, "must go forward");
duke@435 3970 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
duke@435 3971 zeroes_needed = next_full_store;
duke@435 3972 }
duke@435 3973 }
duke@435 3974
duke@435 3975 if (zeroes_needed > zeroes_done) {
duke@435 3976 intptr_t zsize = zeroes_needed - zeroes_done;
duke@435 3977 // Do some incremental zeroing on rawmem, in parallel with inits.
duke@435 3978 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@435 3979 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@435 3980 zeroes_done, zeroes_needed,
duke@435 3981 phase);
duke@435 3982 zeroes_done = zeroes_needed;
duke@435 3983 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
duke@435 3984 do_zeroing = false; // leave the hole, next time
duke@435 3985 }
duke@435 3986 }
duke@435 3987
duke@435 3988 // Collect the store and move on:
duke@435 3989 st->set_req(MemNode::Memory, inits);
duke@435 3990 inits = st; // put it on the linearized chain
duke@435 3991 set_req(i, zmem); // unhook from previous position
duke@435 3992
duke@435 3993 if (zeroes_done == st_off)
duke@435 3994 zeroes_done = next_init_off;
duke@435 3995
duke@435 3996 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
duke@435 3997
duke@435 3998 #ifdef ASSERT
duke@435 3999 // Various order invariants. Weaker than stores_are_sane because
duke@435 4000 // a large constant tile can be filled in by smaller non-constant stores.
duke@435 4001 assert(st_off >= last_init_off, "inits do not reverse");
duke@435 4002 last_init_off = st_off;
duke@435 4003 const Type* val = NULL;
duke@435 4004 if (st_size >= BytesPerInt &&
duke@435 4005 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
duke@435 4006 (int)val->basic_type() < (int)T_OBJECT) {
duke@435 4007 assert(st_off >= last_tile_end, "tiles do not overlap");
duke@435 4008 assert(st_off >= last_init_end, "tiles do not overwrite inits");
duke@435 4009 last_tile_end = MAX2(last_tile_end, next_init_off);
duke@435 4010 } else {
duke@435 4011 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
duke@435 4012 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
duke@435 4013 assert(st_off >= last_init_end, "inits do not overlap");
duke@435 4014 last_init_end = next_init_off; // it's a non-tile
duke@435 4015 }
duke@435 4016 #endif //ASSERT
duke@435 4017 }
duke@435 4018
duke@435 4019 remove_extra_zeroes(); // clear out all the zmems left over
duke@435 4020 add_req(inits);
duke@435 4021
duke@435 4022 if (!ZeroTLAB) {
duke@435 4023 // If anything remains to be zeroed, zero it all now.
duke@435 4024 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@435 4025 // if it is the last unused 4 bytes of an instance, forget about it
duke@435 4026 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
duke@435 4027 if (zeroes_done + BytesPerLong >= size_limit) {
duke@435 4028 assert(allocation() != NULL, "");
never@2346 4029 if (allocation()->Opcode() == Op_Allocate) {
never@2346 4030 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
never@2346 4031 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
never@2346 4032 if (zeroes_done == k->layout_helper())
never@2346 4033 zeroes_done = size_limit;
never@2346 4034 }
duke@435 4035 }
duke@435 4036 if (zeroes_done < size_limit) {
duke@435 4037 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@435 4038 zeroes_done, size_in_bytes, phase);
duke@435 4039 }
duke@435 4040 }
duke@435 4041
duke@435 4042 set_complete(phase);
duke@435 4043 return rawmem;
duke@435 4044 }
duke@435 4045
duke@435 4046
duke@435 4047 #ifdef ASSERT
duke@435 4048 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
duke@435 4049 if (is_complete())
duke@435 4050 return true; // stores could be anything at this point
coleenp@548 4051 assert(allocation() != NULL, "must be present");
coleenp@548 4052 intptr_t last_off = allocation()->minimum_header_size();
duke@435 4053 for (uint i = InitializeNode::RawStores; i < req(); i++) {
duke@435 4054 Node* st = in(i);
duke@435 4055 intptr_t st_off = get_store_offset(st, phase);
duke@435 4056 if (st_off < 0) continue; // ignore dead garbage
duke@435 4057 if (last_off > st_off) {
drchase@6680 4058 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
duke@435 4059 this->dump(2);
duke@435 4060 assert(false, "ascending store offsets");
duke@435 4061 return false;
duke@435 4062 }
duke@435 4063 last_off = st_off + st->as_Store()->memory_size();
duke@435 4064 }
duke@435 4065 return true;
duke@435 4066 }
duke@435 4067 #endif //ASSERT
duke@435 4068
duke@435 4069
duke@435 4070
duke@435 4071
duke@435 4072 //============================MergeMemNode=====================================
duke@435 4073 //
duke@435 4074 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
duke@435 4075 // contributing store or call operations. Each contributor provides the memory
duke@435 4076 // state for a particular "alias type" (see Compile::alias_type). For example,
duke@435 4077 // if a MergeMem has an input X for alias category #6, then any memory reference
duke@435 4078 // to alias category #6 may use X as its memory state input, as an exact equivalent
duke@435 4079 // to using the MergeMem as a whole.
duke@435 4080 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
duke@435 4081 //
duke@435 4082 // (Here, the <N> notation gives the index of the relevant adr_type.)
duke@435 4083 //
duke@435 4084 // In one special case (and more cases in the future), alias categories overlap.
duke@435 4085 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
duke@435 4086 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
duke@435 4087 // it is exactly equivalent to that state W:
duke@435 4088 // MergeMem(<Bot>: W) <==> W
duke@435 4089 //
duke@435 4090 // Usually, the merge has more than one input. In that case, where inputs
duke@435 4091 // overlap (i.e., one is Bot), the narrower alias type determines the memory
duke@435 4092 // state for that type, and the wider alias type (Bot) fills in everywhere else:
duke@435 4093 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
duke@435 4094 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
duke@435 4095 //
duke@435 4096 // A merge can take a "wide" memory state as one of its narrow inputs.
duke@435 4097 // This simply means that the merge observes out only the relevant parts of
duke@435 4098 // the wide input. That is, wide memory states arriving at narrow merge inputs
duke@435 4099 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
duke@435 4100 //
duke@435 4101 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
duke@435 4102 // and that memory slices "leak through":
duke@435 4103 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
duke@435 4104 //
duke@435 4105 // But, in such a cascade, repeated memory slices can "block the leak":
duke@435 4106 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
duke@435 4107 //
duke@435 4108 // In the last example, Y is not part of the combined memory state of the
duke@435 4109 // outermost MergeMem. The system must, of course, prevent unschedulable
duke@435 4110 // memory states from arising, so you can be sure that the state Y is somehow
duke@435 4111 // a precursor to state Y'.
duke@435 4112 //
duke@435 4113 //
duke@435 4114 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
duke@435 4115 // of each MergeMemNode array are exactly the numerical alias indexes, including
duke@435 4116 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
duke@435 4117 // Compile::alias_type (and kin) produce and manage these indexes.
duke@435 4118 //
duke@435 4119 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
duke@435 4120 // (Note that this provides quick access to the top node inside MergeMem methods,
duke@435 4121 // without the need to reach out via TLS to Compile::current.)
duke@435 4122 //
duke@435 4123 // As a consequence of what was just described, a MergeMem that represents a full
duke@435 4124 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
duke@435 4125 // containing all alias categories.
duke@435 4126 //
duke@435 4127 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
duke@435 4128 //
duke@435 4129 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
duke@435 4130 // a memory state for the alias type <N>, or else the top node, meaning that
duke@435 4131 // there is no particular input for that alias type. Note that the length of
duke@435 4132 // a MergeMem is variable, and may be extended at any time to accommodate new
duke@435 4133 // memory states at larger alias indexes. When merges grow, they are of course
duke@435 4134 // filled with "top" in the unused in() positions.
duke@435 4135 //
duke@435 4136 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
duke@435 4137 // (Top was chosen because it works smoothly with passes like GCM.)
duke@435 4138 //
duke@435 4139 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
duke@435 4140 // the type of random VM bits like TLS references.) Since it is always the
duke@435 4141 // first non-Bot memory slice, some low-level loops use it to initialize an
duke@435 4142 // index variable: for (i = AliasIdxRaw; i < req(); i++).
duke@435 4143 //
duke@435 4144 //
duke@435 4145 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
duke@435 4146 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
duke@435 4147 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
duke@435 4148 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
duke@435 4149 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
duke@435 4150 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
duke@435 4151 //
duke@435 4152 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
duke@435 4153 // really that different from the other memory inputs. An abbreviation called
duke@435 4154 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
duke@435 4155 //
duke@435 4156 //
duke@435 4157 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
duke@435 4158 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
duke@435 4159 // that "emerges though" the base memory will be marked as excluding the alias types
duke@435 4160 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
duke@435 4161 //
duke@435 4162 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
duke@435 4163 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
duke@435 4164 //
duke@435 4165 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
duke@435 4166 // (It is currently unimplemented.) As you can see, the resulting merge is
duke@435 4167 // actually a disjoint union of memory states, rather than an overlay.
duke@435 4168 //
duke@435 4169
duke@435 4170 //------------------------------MergeMemNode-----------------------------------
duke@435 4171 Node* MergeMemNode::make_empty_memory() {
duke@435 4172 Node* empty_memory = (Node*) Compile::current()->top();
duke@435 4173 assert(empty_memory->is_top(), "correct sentinel identity");
duke@435 4174 return empty_memory;
duke@435 4175 }
duke@435 4176
duke@435 4177 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
duke@435 4178 init_class_id(Class_MergeMem);
duke@435 4179 // all inputs are nullified in Node::Node(int)
duke@435 4180 // set_input(0, NULL); // no control input
duke@435 4181
duke@435 4182 // Initialize the edges uniformly to top, for starters.
duke@435 4183 Node* empty_mem = make_empty_memory();
duke@435 4184 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
duke@435 4185 init_req(i,empty_mem);
duke@435 4186 }
duke@435 4187 assert(empty_memory() == empty_mem, "");
duke@435 4188
duke@435 4189 if( new_base != NULL && new_base->is_MergeMem() ) {
duke@435 4190 MergeMemNode* mdef = new_base->as_MergeMem();
duke@435 4191 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
duke@435 4192 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
duke@435 4193 mms.set_memory(mms.memory2());
duke@435 4194 }
duke@435 4195 assert(base_memory() == mdef->base_memory(), "");
duke@435 4196 } else {
duke@435 4197 set_base_memory(new_base);
duke@435 4198 }
duke@435 4199 }
duke@435 4200
duke@435 4201 // Make a new, untransformed MergeMem with the same base as 'mem'.
duke@435 4202 // If mem is itself a MergeMem, populate the result with the same edges.
duke@435 4203 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
kvn@4115 4204 return new(C) MergeMemNode(mem);
duke@435 4205 }
duke@435 4206
duke@435 4207 //------------------------------cmp--------------------------------------------
duke@435 4208 uint MergeMemNode::hash() const { return NO_HASH; }
duke@435 4209 uint MergeMemNode::cmp( const Node &n ) const {
duke@435 4210 return (&n == this); // Always fail except on self
duke@435 4211 }
duke@435 4212
duke@435 4213 //------------------------------Identity---------------------------------------
duke@435 4214 Node* MergeMemNode::Identity(PhaseTransform *phase) {
duke@435 4215 // Identity if this merge point does not record any interesting memory
duke@435 4216 // disambiguations.
duke@435 4217 Node* base_mem = base_memory();
duke@435 4218 Node* empty_mem = empty_memory();
duke@435 4219 if (base_mem != empty_mem) { // Memory path is not dead?
duke@435 4220 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@435 4221 Node* mem = in(i);
duke@435 4222 if (mem != empty_mem && mem != base_mem) {
duke@435 4223 return this; // Many memory splits; no change
duke@435 4224 }
duke@435 4225 }
duke@435 4226 }
duke@435 4227 return base_mem; // No memory splits; ID on the one true input
duke@435 4228 }
duke@435 4229
duke@435 4230 //------------------------------Ideal------------------------------------------
duke@435 4231 // This method is invoked recursively on chains of MergeMem nodes
duke@435 4232 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@435 4233 // Remove chain'd MergeMems
duke@435 4234 //
duke@435 4235 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
duke@435 4236 // relative to the "in(Bot)". Since we are patching both at the same time,
duke@435 4237 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
duke@435 4238 // but rewrite each "in(i)" relative to the new "in(Bot)".
duke@435 4239 Node *progress = NULL;
duke@435 4240
duke@435 4241
duke@435 4242 Node* old_base = base_memory();
duke@435 4243 Node* empty_mem = empty_memory();
duke@435 4244 if (old_base == empty_mem)
duke@435 4245 return NULL; // Dead memory path.
duke@435 4246
duke@435 4247 MergeMemNode* old_mbase;
duke@435 4248 if (old_base != NULL && old_base->is_MergeMem())
duke@435 4249 old_mbase = old_base->as_MergeMem();
duke@435 4250 else
duke@435 4251 old_mbase = NULL;
duke@435 4252 Node* new_base = old_base;
duke@435 4253
duke@435 4254 // simplify stacked MergeMems in base memory
duke@435 4255 if (old_mbase) new_base = old_mbase->base_memory();
duke@435 4256
duke@435 4257 // the base memory might contribute new slices beyond my req()
duke@435 4258 if (old_mbase) grow_to_match(old_mbase);
duke@435 4259
duke@435 4260 // Look carefully at the base node if it is a phi.
duke@435 4261 PhiNode* phi_base;
duke@435 4262 if (new_base != NULL && new_base->is_Phi())
duke@435 4263 phi_base = new_base->as_Phi();
duke@435 4264 else
duke@435 4265 phi_base = NULL;
duke@435 4266
duke@435 4267 Node* phi_reg = NULL;
duke@435 4268 uint phi_len = (uint)-1;
duke@435 4269 if (phi_base != NULL && !phi_base->is_copy()) {
duke@435 4270 // do not examine phi if degraded to a copy
duke@435 4271 phi_reg = phi_base->region();
duke@435 4272 phi_len = phi_base->req();
duke@435 4273 // see if the phi is unfinished
duke@435 4274 for (uint i = 1; i < phi_len; i++) {
duke@435 4275 if (phi_base->in(i) == NULL) {
duke@435 4276 // incomplete phi; do not look at it yet!
duke@435 4277 phi_reg = NULL;
duke@435 4278 phi_len = (uint)-1;
duke@435 4279 break;
duke@435 4280 }
duke@435 4281 }
duke@435 4282 }
duke@435 4283
duke@435 4284 // Note: We do not call verify_sparse on entry, because inputs
duke@435 4285 // can normalize to the base_memory via subsume_node or similar
duke@435 4286 // mechanisms. This method repairs that damage.
duke@435 4287
duke@435 4288 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
duke@435 4289
duke@435 4290 // Look at each slice.
duke@435 4291 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@435 4292 Node* old_in = in(i);
duke@435 4293 // calculate the old memory value
duke@435 4294 Node* old_mem = old_in;
duke@435 4295 if (old_mem == empty_mem) old_mem = old_base;
duke@435 4296 assert(old_mem == memory_at(i), "");
duke@435 4297
duke@435 4298 // maybe update (reslice) the old memory value
duke@435 4299
duke@435 4300 // simplify stacked MergeMems
duke@435 4301 Node* new_mem = old_mem;
duke@435 4302 MergeMemNode* old_mmem;
duke@435 4303 if (old_mem != NULL && old_mem->is_MergeMem())
duke@435 4304 old_mmem = old_mem->as_MergeMem();
duke@435 4305 else
duke@435 4306 old_mmem = NULL;
duke@435 4307 if (old_mmem == this) {
duke@435 4308 // This can happen if loops break up and safepoints disappear.
duke@435 4309 // A merge of BotPtr (default) with a RawPtr memory derived from a
duke@435 4310 // safepoint can be rewritten to a merge of the same BotPtr with
duke@435 4311 // the BotPtr phi coming into the loop. If that phi disappears
duke@435 4312 // also, we can end up with a self-loop of the mergemem.
duke@435 4313 // In general, if loops degenerate and memory effects disappear,
duke@435 4314 // a mergemem can be left looking at itself. This simply means
duke@435 4315 // that the mergemem's default should be used, since there is
duke@435 4316 // no longer any apparent effect on this slice.
duke@435 4317 // Note: If a memory slice is a MergeMem cycle, it is unreachable
duke@435 4318 // from start. Update the input to TOP.
duke@435 4319 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
duke@435 4320 }
duke@435 4321 else if (old_mmem != NULL) {
duke@435 4322 new_mem = old_mmem->memory_at(i);
duke@435 4323 }
twisti@1040 4324 // else preceding memory was not a MergeMem
duke@435 4325
duke@435 4326 // replace equivalent phis (unfortunately, they do not GVN together)
duke@435 4327 if (new_mem != NULL && new_mem != new_base &&
duke@435 4328 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
duke@435 4329 if (new_mem->is_Phi()) {
duke@435 4330 PhiNode* phi_mem = new_mem->as_Phi();
duke@435 4331 for (uint i = 1; i < phi_len; i++) {
duke@435 4332 if (phi_base->in(i) != phi_mem->in(i)) {
duke@435 4333 phi_mem = NULL;
duke@435 4334 break;
duke@435 4335 }
duke@435 4336 }
duke@435 4337 if (phi_mem != NULL) {
duke@435 4338 // equivalent phi nodes; revert to the def
duke@435 4339 new_mem = new_base;
duke@435 4340 }
duke@435 4341 }
duke@435 4342 }
duke@435 4343
duke@435 4344 // maybe store down a new value
duke@435 4345 Node* new_in = new_mem;
duke@435 4346 if (new_in == new_base) new_in = empty_mem;
duke@435 4347
duke@435 4348 if (new_in != old_in) {
duke@435 4349 // Warning: Do not combine this "if" with the previous "if"
duke@435 4350 // A memory slice might have be be rewritten even if it is semantically
duke@435 4351 // unchanged, if the base_memory value has changed.
duke@435 4352 set_req(i, new_in);
duke@435 4353 progress = this; // Report progress
duke@435 4354 }
duke@435 4355 }
duke@435 4356
duke@435 4357 if (new_base != old_base) {
duke@435 4358 set_req(Compile::AliasIdxBot, new_base);
duke@435 4359 // Don't use set_base_memory(new_base), because we need to update du.
duke@435 4360 assert(base_memory() == new_base, "");
duke@435 4361 progress = this;
duke@435 4362 }
duke@435 4363
duke@435 4364 if( base_memory() == this ) {
duke@435 4365 // a self cycle indicates this memory path is dead
duke@435 4366 set_req(Compile::AliasIdxBot, empty_mem);
duke@435 4367 }
duke@435 4368
duke@435 4369 // Resolve external cycles by calling Ideal on a MergeMem base_memory
duke@435 4370 // Recursion must occur after the self cycle check above
duke@435 4371 if( base_memory()->is_MergeMem() ) {
duke@435 4372 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
duke@435 4373 Node *m = phase->transform(new_mbase); // Rollup any cycles
duke@435 4374 if( m != NULL && (m->is_top() ||
duke@435 4375 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
duke@435 4376 // propagate rollup of dead cycle to self
duke@435 4377 set_req(Compile::AliasIdxBot, empty_mem);
duke@435 4378 }
duke@435 4379 }
duke@435 4380
duke@435 4381 if( base_memory() == empty_mem ) {
duke@435 4382 progress = this;
duke@435 4383 // Cut inputs during Parse phase only.
duke@435 4384 // During Optimize phase a dead MergeMem node will be subsumed by Top.
duke@435 4385 if( !can_reshape ) {
duke@435 4386 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@435 4387 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
duke@435 4388 }
duke@435 4389 }
duke@435 4390 }
duke@435 4391
duke@435 4392 if( !progress && base_memory()->is_Phi() && can_reshape ) {
duke@435 4393 // Check if PhiNode::Ideal's "Split phis through memory merges"
duke@435 4394 // transform should be attempted. Look for this->phi->this cycle.
duke@435 4395 uint merge_width = req();
duke@435 4396 if (merge_width > Compile::AliasIdxRaw) {
duke@435 4397 PhiNode* phi = base_memory()->as_Phi();
duke@435 4398 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
duke@435 4399 if (phi->in(i) == this) {
duke@435 4400 phase->is_IterGVN()->_worklist.push(phi);
duke@435 4401 break;
duke@435 4402 }
duke@435 4403 }
duke@435 4404 }
duke@435 4405 }
duke@435 4406
kvn@499 4407 assert(progress || verify_sparse(), "please, no dups of base");
duke@435 4408 return progress;
duke@435 4409 }
duke@435 4410
duke@435 4411 //-------------------------set_base_memory-------------------------------------
duke@435 4412 void MergeMemNode::set_base_memory(Node *new_base) {
duke@435 4413 Node* empty_mem = empty_memory();
duke@435 4414 set_req(Compile::AliasIdxBot, new_base);
duke@435 4415 assert(memory_at(req()) == new_base, "must set default memory");
duke@435 4416 // Clear out other occurrences of new_base:
duke@435 4417 if (new_base != empty_mem) {
duke@435 4418 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@435 4419 if (in(i) == new_base) set_req(i, empty_mem);
duke@435 4420 }
duke@435 4421 }
duke@435 4422 }
duke@435 4423
duke@435 4424 //------------------------------out_RegMask------------------------------------
duke@435 4425 const RegMask &MergeMemNode::out_RegMask() const {
duke@435 4426 return RegMask::Empty;
duke@435 4427 }
duke@435 4428
duke@435 4429 //------------------------------dump_spec--------------------------------------
duke@435 4430 #ifndef PRODUCT
duke@435 4431 void MergeMemNode::dump_spec(outputStream *st) const {
duke@435 4432 st->print(" {");
duke@435 4433 Node* base_mem = base_memory();
duke@435 4434 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
duke@435 4435 Node* mem = memory_at(i);
duke@435 4436 if (mem == base_mem) { st->print(" -"); continue; }
duke@435 4437 st->print( " N%d:", mem->_idx );
duke@435 4438 Compile::current()->get_adr_type(i)->dump_on(st);
duke@435 4439 }
duke@435 4440 st->print(" }");
duke@435 4441 }
duke@435 4442 #endif // !PRODUCT
duke@435 4443
duke@435 4444
duke@435 4445 #ifdef ASSERT
duke@435 4446 static bool might_be_same(Node* a, Node* b) {
duke@435 4447 if (a == b) return true;
duke@435 4448 if (!(a->is_Phi() || b->is_Phi())) return false;
duke@435 4449 // phis shift around during optimization
duke@435 4450 return true; // pretty stupid...
duke@435 4451 }
duke@435 4452
duke@435 4453 // verify a narrow slice (either incoming or outgoing)
duke@435 4454 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
duke@435 4455 if (!VerifyAliases) return; // don't bother to verify unless requested
duke@435 4456 if (is_error_reported()) return; // muzzle asserts when debugging an error
duke@435 4457 if (Node::in_dump()) return; // muzzle asserts when printing
duke@435 4458 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
duke@435 4459 assert(n != NULL, "");
duke@435 4460 // Elide intervening MergeMem's
duke@435 4461 while (n->is_MergeMem()) {
duke@435 4462 n = n->as_MergeMem()->memory_at(alias_idx);
duke@435 4463 }
duke@435 4464 Compile* C = Compile::current();
duke@435 4465 const TypePtr* n_adr_type = n->adr_type();
duke@435 4466 if (n == m->empty_memory()) {
duke@435 4467 // Implicit copy of base_memory()
duke@435 4468 } else if (n_adr_type != TypePtr::BOTTOM) {
duke@435 4469 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
duke@435 4470 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
duke@435 4471 } else {
duke@435 4472 // A few places like make_runtime_call "know" that VM calls are narrow,
duke@435 4473 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
duke@435 4474 bool expected_wide_mem = false;
duke@435 4475 if (n == m->base_memory()) {
duke@435 4476 expected_wide_mem = true;
duke@435 4477 } else if (alias_idx == Compile::AliasIdxRaw ||
duke@435 4478 n == m->memory_at(Compile::AliasIdxRaw)) {
duke@435 4479 expected_wide_mem = true;
duke@435 4480 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
duke@435 4481 // memory can "leak through" calls on channels that
duke@435 4482 // are write-once. Allow this also.
duke@435 4483 expected_wide_mem = true;
duke@435 4484 }
duke@435 4485 assert(expected_wide_mem, "expected narrow slice replacement");
duke@435 4486 }
duke@435 4487 }
duke@435 4488 #else // !ASSERT
ccheung@5259 4489 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
duke@435 4490 #endif
duke@435 4491
duke@435 4492
duke@435 4493 //-----------------------------memory_at---------------------------------------
duke@435 4494 Node* MergeMemNode::memory_at(uint alias_idx) const {
duke@435 4495 assert(alias_idx >= Compile::AliasIdxRaw ||
duke@435 4496 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
duke@435 4497 "must avoid base_memory and AliasIdxTop");
duke@435 4498
duke@435 4499 // Otherwise, it is a narrow slice.
duke@435 4500 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
duke@435 4501 Compile *C = Compile::current();
duke@435 4502 if (is_empty_memory(n)) {
duke@435 4503 // the array is sparse; empty slots are the "top" node
duke@435 4504 n = base_memory();
duke@435 4505 assert(Node::in_dump()
duke@435 4506 || n == NULL || n->bottom_type() == Type::TOP
kvn@2599 4507 || n->adr_type() == NULL // address is TOP
duke@435 4508 || n->adr_type() == TypePtr::BOTTOM
duke@435 4509 || n->adr_type() == TypeRawPtr::BOTTOM
duke@435 4510 || Compile::current()->AliasLevel() == 0,
duke@435 4511 "must be a wide memory");
duke@435 4512 // AliasLevel == 0 if we are organizing the memory states manually.
duke@435 4513 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
duke@435 4514 } else {
duke@435 4515 // make sure the stored slice is sane
duke@435 4516 #ifdef ASSERT
duke@435 4517 if (is_error_reported() || Node::in_dump()) {
duke@435 4518 } else if (might_be_same(n, base_memory())) {
duke@435 4519 // Give it a pass: It is a mostly harmless repetition of the base.
duke@435 4520 // This can arise normally from node subsumption during optimization.
duke@435 4521 } else {
duke@435 4522 verify_memory_slice(this, alias_idx, n);
duke@435 4523 }
duke@435 4524 #endif
duke@435 4525 }
duke@435 4526 return n;
duke@435 4527 }
duke@435 4528
duke@435 4529 //---------------------------set_memory_at-------------------------------------
duke@435 4530 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
duke@435 4531 verify_memory_slice(this, alias_idx, n);
duke@435 4532 Node* empty_mem = empty_memory();
duke@435 4533 if (n == base_memory()) n = empty_mem; // collapse default
duke@435 4534 uint need_req = alias_idx+1;
duke@435 4535 if (req() < need_req) {
duke@435 4536 if (n == empty_mem) return; // already the default, so do not grow me
duke@435 4537 // grow the sparse array
duke@435 4538 do {
duke@435 4539 add_req(empty_mem);
duke@435 4540 } while (req() < need_req);
duke@435 4541 }
duke@435 4542 set_req( alias_idx, n );
duke@435 4543 }
duke@435 4544
duke@435 4545
duke@435 4546
duke@435 4547 //--------------------------iteration_setup------------------------------------
duke@435 4548 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
duke@435 4549 if (other != NULL) {
duke@435 4550 grow_to_match(other);
duke@435 4551 // invariant: the finite support of mm2 is within mm->req()
duke@435 4552 #ifdef ASSERT
duke@435 4553 for (uint i = req(); i < other->req(); i++) {
duke@435 4554 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
duke@435 4555 }
duke@435 4556 #endif
duke@435 4557 }
duke@435 4558 // Replace spurious copies of base_memory by top.
duke@435 4559 Node* base_mem = base_memory();
duke@435 4560 if (base_mem != NULL && !base_mem->is_top()) {
duke@435 4561 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
duke@435 4562 if (in(i) == base_mem)
duke@435 4563 set_req(i, empty_memory());
duke@435 4564 }
duke@435 4565 }
duke@435 4566 }
duke@435 4567
duke@435 4568 //---------------------------grow_to_match-------------------------------------
duke@435 4569 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
duke@435 4570 Node* empty_mem = empty_memory();
duke@435 4571 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
duke@435 4572 // look for the finite support of the other memory
duke@435 4573 for (uint i = other->req(); --i >= req(); ) {
duke@435 4574 if (other->in(i) != empty_mem) {
duke@435 4575 uint new_len = i+1;
duke@435 4576 while (req() < new_len) add_req(empty_mem);
duke@435 4577 break;
duke@435 4578 }
duke@435 4579 }
duke@435 4580 }
duke@435 4581
duke@435 4582 //---------------------------verify_sparse-------------------------------------
duke@435 4583 #ifndef PRODUCT
duke@435 4584 bool MergeMemNode::verify_sparse() const {
duke@435 4585 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
duke@435 4586 Node* base_mem = base_memory();
duke@435 4587 // The following can happen in degenerate cases, since empty==top.
duke@435 4588 if (is_empty_memory(base_mem)) return true;
duke@435 4589 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@435 4590 assert(in(i) != NULL, "sane slice");
duke@435 4591 if (in(i) == base_mem) return false; // should have been the sentinel value!
duke@435 4592 }
duke@435 4593 return true;
duke@435 4594 }
duke@435 4595
duke@435 4596 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
duke@435 4597 Node* n;
duke@435 4598 n = mm->in(idx);
duke@435 4599 if (mem == n) return true; // might be empty_memory()
duke@435 4600 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
duke@435 4601 if (mem == n) return true;
duke@435 4602 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
duke@435 4603 if (mem == n) return true;
duke@435 4604 if (n == NULL) break;
duke@435 4605 }
duke@435 4606 return false;
duke@435 4607 }
duke@435 4608 #endif // !PRODUCT

mercurial