Tue, 02 Sep 2014 12:48:45 -0700
8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
Summary: Add new C2 intrinsic for BigInteger::multiplyToLen() on x86 in 64-bit VM.
Reviewed-by: roland
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 | //============================================================================= |
duke@435 | 862 | uint LoadNode::size_of() const { return sizeof(*this); } |
duke@435 | 863 | uint LoadNode::cmp( const Node &n ) const |
duke@435 | 864 | { return !Type::cmp( _type, ((LoadNode&)n)._type ); } |
duke@435 | 865 | const Type *LoadNode::bottom_type() const { return _type; } |
duke@435 | 866 | uint LoadNode::ideal_reg() const { |
coleenp@4037 | 867 | return _type->ideal_reg(); |
duke@435 | 868 | } |
duke@435 | 869 | |
duke@435 | 870 | #ifndef PRODUCT |
duke@435 | 871 | void LoadNode::dump_spec(outputStream *st) const { |
duke@435 | 872 | MemNode::dump_spec(st); |
duke@435 | 873 | if( !Verbose && !WizardMode ) { |
duke@435 | 874 | // standard dump does this in Verbose and WizardMode |
duke@435 | 875 | st->print(" #"); _type->dump_on(st); |
duke@435 | 876 | } |
duke@435 | 877 | } |
duke@435 | 878 | #endif |
duke@435 | 879 | |
kvn@1964 | 880 | #ifdef ASSERT |
kvn@1964 | 881 | //----------------------------is_immutable_value------------------------------- |
kvn@1964 | 882 | // Helper function to allow a raw load without control edge for some cases |
kvn@1964 | 883 | bool LoadNode::is_immutable_value(Node* adr) { |
kvn@1964 | 884 | return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && |
kvn@1964 | 885 | adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && |
kvn@1964 | 886 | (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == |
kvn@1964 | 887 | in_bytes(JavaThread::osthread_offset()))); |
kvn@1964 | 888 | } |
kvn@1964 | 889 | #endif |
duke@435 | 890 | |
duke@435 | 891 | //----------------------------LoadNode::make----------------------------------- |
duke@435 | 892 | // Polymorphic factory method: |
goetz@6479 | 893 | Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) { |
coleenp@548 | 894 | Compile* C = gvn.C; |
coleenp@548 | 895 | |
duke@435 | 896 | // sanity check the alias category against the created node type |
duke@435 | 897 | assert(!(adr_type->isa_oopptr() && |
duke@435 | 898 | adr_type->offset() == oopDesc::klass_offset_in_bytes()), |
duke@435 | 899 | "use LoadKlassNode instead"); |
duke@435 | 900 | assert(!(adr_type->isa_aryptr() && |
duke@435 | 901 | adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), |
duke@435 | 902 | "use LoadRangeNode instead"); |
kvn@1964 | 903 | // Check control edge of raw loads |
kvn@1964 | 904 | assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || |
kvn@1964 | 905 | // oop will be recorded in oop map if load crosses safepoint |
kvn@1964 | 906 | rt->isa_oopptr() || is_immutable_value(adr), |
kvn@1964 | 907 | "raw memory operations should have control edge"); |
duke@435 | 908 | switch (bt) { |
goetz@6479 | 909 | case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo); |
goetz@6479 | 910 | case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo); |
goetz@6479 | 911 | case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo); |
goetz@6479 | 912 | case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo); |
goetz@6479 | 913 | case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo); |
goetz@6479 | 914 | case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo); |
goetz@6479 | 915 | case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo); |
goetz@6479 | 916 | case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo); |
goetz@6479 | 917 | case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo); |
coleenp@548 | 918 | case T_OBJECT: |
coleenp@548 | 919 | #ifdef _LP64 |
kvn@598 | 920 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
goetz@6479 | 921 | Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo)); |
kvn@4115 | 922 | return new (C) DecodeNNode(load, load->bottom_type()->make_ptr()); |
coleenp@548 | 923 | } else |
coleenp@548 | 924 | #endif |
kvn@598 | 925 | { |
roland@4159 | 926 | assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); |
goetz@6479 | 927 | return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo); |
kvn@598 | 928 | } |
duke@435 | 929 | } |
duke@435 | 930 | ShouldNotReachHere(); |
duke@435 | 931 | return (LoadNode*)NULL; |
duke@435 | 932 | } |
duke@435 | 933 | |
goetz@6479 | 934 | LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) { |
duke@435 | 935 | bool require_atomic = true; |
goetz@6479 | 936 | return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic); |
duke@435 | 937 | } |
duke@435 | 938 | |
duke@435 | 939 | |
duke@435 | 940 | |
duke@435 | 941 | |
duke@435 | 942 | //------------------------------hash------------------------------------------- |
duke@435 | 943 | uint LoadNode::hash() const { |
duke@435 | 944 | // unroll addition of interesting fields |
duke@435 | 945 | return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); |
duke@435 | 946 | } |
duke@435 | 947 | |
vlivanov@5658 | 948 | static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { |
vlivanov@5658 | 949 | if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { |
vlivanov@5658 | 950 | bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); |
vlivanov@5658 | 951 | bool is_stable_ary = FoldStableValues && |
vlivanov@5658 | 952 | (tp != NULL) && (tp->isa_aryptr() != NULL) && |
vlivanov@5658 | 953 | tp->isa_aryptr()->is_stable(); |
vlivanov@5658 | 954 | |
vlivanov@5658 | 955 | return (eliminate_boxing && non_volatile) || is_stable_ary; |
vlivanov@5658 | 956 | } |
vlivanov@5658 | 957 | |
vlivanov@5658 | 958 | return false; |
vlivanov@5658 | 959 | } |
vlivanov@5658 | 960 | |
duke@435 | 961 | //---------------------------can_see_stored_value------------------------------ |
duke@435 | 962 | // This routine exists to make sure this set of tests is done the same |
duke@435 | 963 | // everywhere. We need to make a coordinated change: first LoadNode::Ideal |
duke@435 | 964 | // will change the graph shape in a way which makes memory alive twice at the |
duke@435 | 965 | // same time (uses the Oracle model of aliasing), then some |
duke@435 | 966 | // LoadXNode::Identity will fold things back to the equivalence-class model |
duke@435 | 967 | // of aliasing. |
duke@435 | 968 | Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { |
duke@435 | 969 | Node* ld_adr = in(MemNode::Address); |
kvn@5110 | 970 | intptr_t ld_off = 0; |
kvn@5110 | 971 | AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); |
never@452 | 972 | const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); |
kvn@5110 | 973 | Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; |
kvn@5110 | 974 | // This is more general than load from boxing objects. |
vlivanov@5658 | 975 | if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { |
never@452 | 976 | uint alias_idx = atp->index(); |
vlivanov@5658 | 977 | bool final = !atp->is_rewritable(); |
never@452 | 978 | Node* result = NULL; |
never@452 | 979 | Node* current = st; |
never@452 | 980 | // Skip through chains of MemBarNodes checking the MergeMems for |
never@452 | 981 | // new states for the slice of this load. Stop once any other |
never@452 | 982 | // kind of node is encountered. Loads from final memory can skip |
never@452 | 983 | // through any kind of MemBar but normal loads shouldn't skip |
never@452 | 984 | // through MemBarAcquire since the could allow them to move out of |
never@452 | 985 | // a synchronized region. |
never@452 | 986 | while (current->is_Proj()) { |
never@452 | 987 | int opc = current->in(0)->Opcode(); |
goetz@6489 | 988 | if ((final && (opc == Op_MemBarAcquire || |
goetz@6489 | 989 | opc == Op_MemBarAcquireLock || |
goetz@6489 | 990 | opc == Op_LoadFence)) || |
goetz@6489 | 991 | opc == Op_MemBarRelease || |
goetz@6489 | 992 | opc == Op_StoreFence || |
goetz@6489 | 993 | opc == Op_MemBarReleaseLock || |
goetz@6489 | 994 | opc == Op_MemBarCPUOrder) { |
never@452 | 995 | Node* mem = current->in(0)->in(TypeFunc::Memory); |
never@452 | 996 | if (mem->is_MergeMem()) { |
never@452 | 997 | MergeMemNode* merge = mem->as_MergeMem(); |
never@452 | 998 | Node* new_st = merge->memory_at(alias_idx); |
never@452 | 999 | if (new_st == merge->base_memory()) { |
never@452 | 1000 | // Keep searching |
kvn@5110 | 1001 | current = new_st; |
never@452 | 1002 | continue; |
never@452 | 1003 | } |
never@452 | 1004 | // Save the new memory state for the slice and fall through |
never@452 | 1005 | // to exit. |
never@452 | 1006 | result = new_st; |
never@452 | 1007 | } |
never@452 | 1008 | } |
never@452 | 1009 | break; |
never@452 | 1010 | } |
never@452 | 1011 | if (result != NULL) { |
never@452 | 1012 | st = result; |
never@452 | 1013 | } |
never@452 | 1014 | } |
never@452 | 1015 | |
duke@435 | 1016 | // Loop around twice in the case Load -> Initialize -> Store. |
duke@435 | 1017 | // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) |
duke@435 | 1018 | for (int trip = 0; trip <= 1; trip++) { |
duke@435 | 1019 | |
duke@435 | 1020 | if (st->is_Store()) { |
duke@435 | 1021 | Node* st_adr = st->in(MemNode::Address); |
duke@435 | 1022 | if (!phase->eqv(st_adr, ld_adr)) { |
duke@435 | 1023 | // Try harder before giving up... Match raw and non-raw pointers. |
duke@435 | 1024 | intptr_t st_off = 0; |
duke@435 | 1025 | AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); |
duke@435 | 1026 | if (alloc == NULL) return NULL; |
kvn@5110 | 1027 | if (alloc != ld_alloc) return NULL; |
duke@435 | 1028 | if (ld_off != st_off) return NULL; |
duke@435 | 1029 | // At this point we have proven something like this setup: |
duke@435 | 1030 | // A = Allocate(...) |
duke@435 | 1031 | // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) |
duke@435 | 1032 | // S = StoreQ(, AddP(, A.Parm , #Off), V) |
duke@435 | 1033 | // (Actually, we haven't yet proven the Q's are the same.) |
duke@435 | 1034 | // In other words, we are loading from a casted version of |
duke@435 | 1035 | // the same pointer-and-offset that we stored to. |
duke@435 | 1036 | // Thus, we are able to replace L by V. |
duke@435 | 1037 | } |
duke@435 | 1038 | // Now prove that we have a LoadQ matched to a StoreQ, for some Q. |
duke@435 | 1039 | if (store_Opcode() != st->Opcode()) |
duke@435 | 1040 | return NULL; |
duke@435 | 1041 | return st->in(MemNode::ValueIn); |
duke@435 | 1042 | } |
duke@435 | 1043 | |
duke@435 | 1044 | // A load from a freshly-created object always returns zero. |
duke@435 | 1045 | // (This can happen after LoadNode::Ideal resets the load's memory input |
duke@435 | 1046 | // to find_captured_store, which returned InitializeNode::zero_memory.) |
duke@435 | 1047 | if (st->is_Proj() && st->in(0)->is_Allocate() && |
kvn@5110 | 1048 | (st->in(0) == ld_alloc) && |
kvn@5110 | 1049 | (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { |
duke@435 | 1050 | // return a zero value for the load's basic type |
duke@435 | 1051 | // (This is one of the few places where a generic PhaseTransform |
duke@435 | 1052 | // can create new nodes. Think of it as lazily manifesting |
duke@435 | 1053 | // virtually pre-existing constants.) |
duke@435 | 1054 | return phase->zerocon(memory_type()); |
duke@435 | 1055 | } |
duke@435 | 1056 | |
duke@435 | 1057 | // A load from an initialization barrier can match a captured store. |
duke@435 | 1058 | if (st->is_Proj() && st->in(0)->is_Initialize()) { |
duke@435 | 1059 | InitializeNode* init = st->in(0)->as_Initialize(); |
duke@435 | 1060 | AllocateNode* alloc = init->allocation(); |
kvn@5110 | 1061 | if ((alloc != NULL) && (alloc == ld_alloc)) { |
duke@435 | 1062 | // examine a captured store value |
kvn@5110 | 1063 | st = init->find_captured_store(ld_off, memory_size(), phase); |
duke@435 | 1064 | if (st != NULL) |
duke@435 | 1065 | continue; // take one more trip around |
duke@435 | 1066 | } |
duke@435 | 1067 | } |
duke@435 | 1068 | |
kvn@5110 | 1069 | // Load boxed value from result of valueOf() call is input parameter. |
kvn@5110 | 1070 | if (this->is_Load() && ld_adr->is_AddP() && |
kvn@5110 | 1071 | (tp != NULL) && tp->is_ptr_to_boxed_value()) { |
kvn@5110 | 1072 | intptr_t ignore = 0; |
kvn@5110 | 1073 | Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); |
kvn@5110 | 1074 | if (base != NULL && base->is_Proj() && |
kvn@5110 | 1075 | base->as_Proj()->_con == TypeFunc::Parms && |
kvn@5110 | 1076 | base->in(0)->is_CallStaticJava() && |
kvn@5110 | 1077 | base->in(0)->as_CallStaticJava()->is_boxing_method()) { |
kvn@5110 | 1078 | return base->in(0)->in(TypeFunc::Parms); |
kvn@5110 | 1079 | } |
kvn@5110 | 1080 | } |
kvn@5110 | 1081 | |
duke@435 | 1082 | break; |
duke@435 | 1083 | } |
duke@435 | 1084 | |
duke@435 | 1085 | return NULL; |
duke@435 | 1086 | } |
duke@435 | 1087 | |
kvn@499 | 1088 | //----------------------is_instance_field_load_with_local_phi------------------ |
kvn@499 | 1089 | bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { |
kvn@5110 | 1090 | if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && |
kvn@5110 | 1091 | in(Address)->is_AddP() ) { |
kvn@5110 | 1092 | const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); |
kvn@5110 | 1093 | // Only instances and boxed values. |
kvn@5110 | 1094 | if( t_oop != NULL && |
kvn@5110 | 1095 | (t_oop->is_ptr_to_boxed_value() || |
kvn@5110 | 1096 | t_oop->is_known_instance_field()) && |
kvn@499 | 1097 | t_oop->offset() != Type::OffsetBot && |
kvn@499 | 1098 | t_oop->offset() != Type::OffsetTop) { |
kvn@499 | 1099 | return true; |
kvn@499 | 1100 | } |
kvn@499 | 1101 | } |
kvn@499 | 1102 | return false; |
kvn@499 | 1103 | } |
kvn@499 | 1104 | |
duke@435 | 1105 | //------------------------------Identity--------------------------------------- |
duke@435 | 1106 | // Loads are identity if previous store is to same address |
duke@435 | 1107 | Node *LoadNode::Identity( PhaseTransform *phase ) { |
duke@435 | 1108 | // If the previous store-maker is the right kind of Store, and the store is |
duke@435 | 1109 | // to the same address, then we are equal to the value stored. |
kvn@5110 | 1110 | Node* mem = in(Memory); |
duke@435 | 1111 | Node* value = can_see_stored_value(mem, phase); |
duke@435 | 1112 | if( value ) { |
duke@435 | 1113 | // byte, short & char stores truncate naturally. |
duke@435 | 1114 | // A load has to load the truncated value which requires |
duke@435 | 1115 | // some sort of masking operation and that requires an |
duke@435 | 1116 | // Ideal call instead of an Identity call. |
duke@435 | 1117 | if (memory_size() < BytesPerInt) { |
duke@435 | 1118 | // If the input to the store does not fit with the load's result type, |
duke@435 | 1119 | // it must be truncated via an Ideal call. |
duke@435 | 1120 | if (!phase->type(value)->higher_equal(phase->type(this))) |
duke@435 | 1121 | return this; |
duke@435 | 1122 | } |
duke@435 | 1123 | // (This works even when value is a Con, but LoadNode::Value |
duke@435 | 1124 | // usually runs first, producing the singleton type of the Con.) |
duke@435 | 1125 | return value; |
duke@435 | 1126 | } |
kvn@499 | 1127 | |
kvn@499 | 1128 | // Search for an existing data phi which was generated before for the same |
twisti@1040 | 1129 | // instance's field to avoid infinite generation of phis in a loop. |
kvn@499 | 1130 | Node *region = mem->in(0); |
kvn@499 | 1131 | if (is_instance_field_load_with_local_phi(region)) { |
kvn@5110 | 1132 | const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); |
kvn@499 | 1133 | int this_index = phase->C->get_alias_index(addr_t); |
kvn@499 | 1134 | int this_offset = addr_t->offset(); |
kvn@5110 | 1135 | int this_iid = addr_t->instance_id(); |
kvn@5110 | 1136 | if (!addr_t->is_known_instance() && |
kvn@5110 | 1137 | addr_t->is_ptr_to_boxed_value()) { |
kvn@5110 | 1138 | // Use _idx of address base (could be Phi node) for boxed values. |
kvn@5110 | 1139 | intptr_t ignore = 0; |
kvn@5110 | 1140 | Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); |
kvn@5110 | 1141 | this_iid = base->_idx; |
kvn@5110 | 1142 | } |
kvn@499 | 1143 | const Type* this_type = bottom_type(); |
kvn@499 | 1144 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
kvn@499 | 1145 | Node* phi = region->fast_out(i); |
kvn@499 | 1146 | if (phi->is_Phi() && phi != mem && |
kvn@5110 | 1147 | phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) { |
kvn@499 | 1148 | return phi; |
kvn@499 | 1149 | } |
kvn@499 | 1150 | } |
kvn@499 | 1151 | } |
kvn@499 | 1152 | |
duke@435 | 1153 | return this; |
duke@435 | 1154 | } |
duke@435 | 1155 | |
never@452 | 1156 | // We're loading from an object which has autobox behaviour. |
never@452 | 1157 | // If this object is result of a valueOf call we'll have a phi |
never@452 | 1158 | // merging a newly allocated object and a load from the cache. |
never@452 | 1159 | // We want to replace this load with the original incoming |
never@452 | 1160 | // argument to the valueOf call. |
never@452 | 1161 | Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { |
kvn@5110 | 1162 | assert(phase->C->eliminate_boxing(), "sanity"); |
kvn@5110 | 1163 | intptr_t ignore = 0; |
kvn@5110 | 1164 | Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); |
kvn@5110 | 1165 | if ((base == NULL) || base->is_Phi()) { |
kvn@5110 | 1166 | // Push the loads from the phi that comes from valueOf up |
kvn@5110 | 1167 | // through it to allow elimination of the loads and the recovery |
kvn@5110 | 1168 | // of the original value. It is done in split_through_phi(). |
kvn@5110 | 1169 | return NULL; |
kvn@1018 | 1170 | } else if (base->is_Load() || |
kvn@1018 | 1171 | base->is_DecodeN() && base->in(1)->is_Load()) { |
kvn@5110 | 1172 | // Eliminate the load of boxed value for integer types from the cache |
kvn@5110 | 1173 | // array by deriving the value from the index into the array. |
kvn@5110 | 1174 | // Capture the offset of the load and then reverse the computation. |
kvn@5110 | 1175 | |
kvn@5110 | 1176 | // Get LoadN node which loads a boxing object from 'cache' array. |
kvn@1018 | 1177 | if (base->is_DecodeN()) { |
kvn@1018 | 1178 | base = base->in(1); |
kvn@1018 | 1179 | } |
kvn@5110 | 1180 | if (!base->in(Address)->is_AddP()) { |
kvn@5110 | 1181 | return NULL; // Complex address |
kvn@1018 | 1182 | } |
kvn@5110 | 1183 | AddPNode* address = base->in(Address)->as_AddP(); |
kvn@5110 | 1184 | Node* cache_base = address->in(AddPNode::Base); |
kvn@5110 | 1185 | if ((cache_base != NULL) && cache_base->is_DecodeN()) { |
kvn@5110 | 1186 | // Get ConP node which is static 'cache' field. |
kvn@5110 | 1187 | cache_base = cache_base->in(1); |
kvn@5110 | 1188 | } |
kvn@5110 | 1189 | if ((cache_base != NULL) && cache_base->is_Con()) { |
kvn@5110 | 1190 | const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); |
kvn@5110 | 1191 | if ((base_type != NULL) && base_type->is_autobox_cache()) { |
never@452 | 1192 | Node* elements[4]; |
kvn@5110 | 1193 | int shift = exact_log2(type2aelembytes(T_OBJECT)); |
kvn@5110 | 1194 | int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); |
kvn@5110 | 1195 | if ((count > 0) && elements[0]->is_Con() && |
kvn@5110 | 1196 | ((count == 1) || |
kvn@5110 | 1197 | (count == 2) && elements[1]->Opcode() == Op_LShiftX && |
kvn@5110 | 1198 | elements[1]->in(2) == phase->intcon(shift))) { |
kvn@5110 | 1199 | ciObjArray* array = base_type->const_oop()->as_obj_array(); |
kvn@5110 | 1200 | // Fetch the box object cache[0] at the base of the array and get its value |
kvn@5110 | 1201 | ciInstance* box = array->obj_at(0)->as_instance(); |
kvn@5110 | 1202 | ciInstanceKlass* ik = box->klass()->as_instance_klass(); |
kvn@5110 | 1203 | assert(ik->is_box_klass(), "sanity"); |
kvn@5110 | 1204 | assert(ik->nof_nonstatic_fields() == 1, "change following code"); |
kvn@5110 | 1205 | if (ik->nof_nonstatic_fields() == 1) { |
kvn@5110 | 1206 | // This should be true nonstatic_field_at requires calling |
kvn@5110 | 1207 | // nof_nonstatic_fields so check it anyway |
kvn@5110 | 1208 | ciConstant c = box->field_value(ik->nonstatic_field_at(0)); |
kvn@5110 | 1209 | BasicType bt = c.basic_type(); |
kvn@5110 | 1210 | // Only integer types have boxing cache. |
kvn@5110 | 1211 | assert(bt == T_BOOLEAN || bt == T_CHAR || |
kvn@5110 | 1212 | bt == T_BYTE || bt == T_SHORT || |
kvn@5110 | 1213 | bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt))); |
kvn@5110 | 1214 | jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); |
kvn@5110 | 1215 | if (cache_low != (int)cache_low) { |
kvn@5110 | 1216 | return NULL; // should not happen since cache is array indexed by value |
kvn@5110 | 1217 | } |
kvn@5110 | 1218 | jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); |
kvn@5110 | 1219 | if (offset != (int)offset) { |
kvn@5110 | 1220 | return NULL; // should not happen since cache is array indexed by value |
kvn@5110 | 1221 | } |
kvn@5110 | 1222 | // Add up all the offsets making of the address of the load |
kvn@5110 | 1223 | Node* result = elements[0]; |
kvn@5110 | 1224 | for (int i = 1; i < count; i++) { |
kvn@5110 | 1225 | result = phase->transform(new (phase->C) AddXNode(result, elements[i])); |
kvn@5110 | 1226 | } |
kvn@5110 | 1227 | // Remove the constant offset from the address and then |
kvn@5110 | 1228 | result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset))); |
kvn@5110 | 1229 | // remove the scaling of the offset to recover the original index. |
kvn@5110 | 1230 | if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { |
kvn@5110 | 1231 | // Peel the shift off directly but wrap it in a dummy node |
kvn@5110 | 1232 | // since Ideal can't return existing nodes |
kvn@5110 | 1233 | result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0)); |
kvn@5110 | 1234 | } else if (result->is_Add() && result->in(2)->is_Con() && |
kvn@5110 | 1235 | result->in(1)->Opcode() == Op_LShiftX && |
kvn@5110 | 1236 | result->in(1)->in(2) == phase->intcon(shift)) { |
kvn@5110 | 1237 | // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) |
kvn@5110 | 1238 | // but for boxing cache access we know that X<<Z will not overflow |
kvn@5110 | 1239 | // (there is range check) so we do this optimizatrion by hand here. |
kvn@5110 | 1240 | Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift)); |
kvn@5110 | 1241 | result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con)); |
kvn@5110 | 1242 | } else { |
kvn@5110 | 1243 | result = new (phase->C) RShiftXNode(result, phase->intcon(shift)); |
kvn@5110 | 1244 | } |
kvn@5110 | 1245 | #ifdef _LP64 |
kvn@5110 | 1246 | if (bt != T_LONG) { |
kvn@5110 | 1247 | result = new (phase->C) ConvL2INode(phase->transform(result)); |
kvn@5110 | 1248 | } |
kvn@5110 | 1249 | #else |
kvn@5110 | 1250 | if (bt == T_LONG) { |
kvn@5110 | 1251 | result = new (phase->C) ConvI2LNode(phase->transform(result)); |
kvn@5110 | 1252 | } |
kvn@5110 | 1253 | #endif |
kvn@5110 | 1254 | return result; |
never@452 | 1255 | } |
never@452 | 1256 | } |
never@452 | 1257 | } |
never@452 | 1258 | } |
never@452 | 1259 | } |
never@452 | 1260 | return NULL; |
never@452 | 1261 | } |
never@452 | 1262 | |
kvn@5110 | 1263 | static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { |
kvn@5110 | 1264 | Node* region = phi->in(0); |
kvn@598 | 1265 | if (region == NULL) { |
kvn@5110 | 1266 | return false; // Wait stable graph |
kvn@598 | 1267 | } |
kvn@5110 | 1268 | uint cnt = phi->req(); |
kvn@2810 | 1269 | for (uint i = 1; i < cnt; i++) { |
kvn@2810 | 1270 | Node* rc = region->in(i); |
kvn@2810 | 1271 | if (rc == NULL || phase->type(rc) == Type::TOP) |
kvn@5110 | 1272 | return false; // Wait stable graph |
kvn@5110 | 1273 | Node* in = phi->in(i); |
kvn@5110 | 1274 | if (in == NULL || phase->type(in) == Type::TOP) |
kvn@5110 | 1275 | return false; // Wait stable graph |
kvn@5110 | 1276 | } |
kvn@5110 | 1277 | return true; |
kvn@5110 | 1278 | } |
kvn@5110 | 1279 | //------------------------------split_through_phi------------------------------ |
kvn@5110 | 1280 | // Split instance or boxed field load through Phi. |
kvn@5110 | 1281 | Node *LoadNode::split_through_phi(PhaseGVN *phase) { |
kvn@5110 | 1282 | Node* mem = in(Memory); |
kvn@5110 | 1283 | Node* address = in(Address); |
kvn@5110 | 1284 | const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); |
kvn@5110 | 1285 | |
kvn@5110 | 1286 | assert((t_oop != NULL) && |
kvn@5110 | 1287 | (t_oop->is_known_instance_field() || |
kvn@5110 | 1288 | t_oop->is_ptr_to_boxed_value()), "invalide conditions"); |
kvn@5110 | 1289 | |
kvn@5110 | 1290 | Compile* C = phase->C; |
kvn@5110 | 1291 | intptr_t ignore = 0; |
kvn@5110 | 1292 | Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); |
kvn@5110 | 1293 | bool base_is_phi = (base != NULL) && base->is_Phi(); |
kvn@5110 | 1294 | bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && |
kvn@5110 | 1295 | (base != NULL) && (base == address->in(AddPNode::Base)) && |
kvn@5110 | 1296 | phase->type(base)->higher_equal(TypePtr::NOTNULL); |
kvn@5110 | 1297 | |
kvn@5110 | 1298 | if (!((mem->is_Phi() || base_is_phi) && |
kvn@5110 | 1299 | (load_boxed_values || t_oop->is_known_instance_field()))) { |
kvn@5110 | 1300 | return NULL; // memory is not Phi |
kvn@5110 | 1301 | } |
kvn@5110 | 1302 | |
kvn@5110 | 1303 | if (mem->is_Phi()) { |
kvn@5110 | 1304 | if (!stable_phi(mem->as_Phi(), phase)) { |
kvn@598 | 1305 | return NULL; // Wait stable graph |
kvn@598 | 1306 | } |
kvn@5110 | 1307 | uint cnt = mem->req(); |
kvn@5110 | 1308 | // Check for loop invariant memory. |
kvn@5110 | 1309 | if (cnt == 3) { |
kvn@5110 | 1310 | for (uint i = 1; i < cnt; i++) { |
kvn@5110 | 1311 | Node* in = mem->in(i); |
kvn@5110 | 1312 | Node* m = optimize_memory_chain(in, t_oop, this, phase); |
kvn@5110 | 1313 | if (m == mem) { |
kvn@5110 | 1314 | set_req(Memory, mem->in(cnt - i)); |
kvn@5110 | 1315 | return this; // made change |
kvn@5110 | 1316 | } |
kvn@598 | 1317 | } |
kvn@598 | 1318 | } |
kvn@598 | 1319 | } |
kvn@5110 | 1320 | if (base_is_phi) { |
kvn@5110 | 1321 | if (!stable_phi(base->as_Phi(), phase)) { |
kvn@5110 | 1322 | return NULL; // Wait stable graph |
kvn@5110 | 1323 | } |
kvn@5110 | 1324 | uint cnt = base->req(); |
kvn@5110 | 1325 | // Check for loop invariant memory. |
kvn@5110 | 1326 | if (cnt == 3) { |
kvn@5110 | 1327 | for (uint i = 1; i < cnt; i++) { |
kvn@5110 | 1328 | if (base->in(i) == base) { |
kvn@5110 | 1329 | return NULL; // Wait stable graph |
kvn@5110 | 1330 | } |
kvn@5110 | 1331 | } |
kvn@5110 | 1332 | } |
kvn@5110 | 1333 | } |
kvn@5110 | 1334 | |
kvn@5110 | 1335 | bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); |
kvn@5110 | 1336 | |
kvn@598 | 1337 | // Split through Phi (see original code in loopopts.cpp). |
kvn@5110 | 1338 | assert(C->have_alias_type(t_oop), "instance should have alias type"); |
kvn@598 | 1339 | |
kvn@598 | 1340 | // Do nothing here if Identity will find a value |
kvn@598 | 1341 | // (to avoid infinite chain of value phis generation). |
kvn@2810 | 1342 | if (!phase->eqv(this, this->Identity(phase))) |
kvn@598 | 1343 | return NULL; |
kvn@598 | 1344 | |
kvn@5110 | 1345 | // Select Region to split through. |
kvn@5110 | 1346 | Node* region; |
kvn@5110 | 1347 | if (!base_is_phi) { |
kvn@5110 | 1348 | assert(mem->is_Phi(), "sanity"); |
kvn@5110 | 1349 | region = mem->in(0); |
kvn@5110 | 1350 | // Skip if the region dominates some control edge of the address. |
kvn@5110 | 1351 | if (!MemNode::all_controls_dominate(address, region)) |
kvn@5110 | 1352 | return NULL; |
kvn@5110 | 1353 | } else if (!mem->is_Phi()) { |
kvn@5110 | 1354 | assert(base_is_phi, "sanity"); |
kvn@5110 | 1355 | region = base->in(0); |
kvn@5110 | 1356 | // Skip if the region dominates some control edge of the memory. |
kvn@5110 | 1357 | if (!MemNode::all_controls_dominate(mem, region)) |
kvn@5110 | 1358 | return NULL; |
kvn@5110 | 1359 | } else if (base->in(0) != mem->in(0)) { |
kvn@5110 | 1360 | assert(base_is_phi && mem->is_Phi(), "sanity"); |
kvn@5110 | 1361 | if (MemNode::all_controls_dominate(mem, base->in(0))) { |
kvn@5110 | 1362 | region = base->in(0); |
kvn@5110 | 1363 | } else if (MemNode::all_controls_dominate(address, mem->in(0))) { |
kvn@5110 | 1364 | region = mem->in(0); |
kvn@5110 | 1365 | } else { |
kvn@5110 | 1366 | return NULL; // complex graph |
kvn@5110 | 1367 | } |
kvn@5110 | 1368 | } else { |
kvn@5110 | 1369 | assert(base->in(0) == mem->in(0), "sanity"); |
kvn@5110 | 1370 | region = mem->in(0); |
kvn@5110 | 1371 | } |
kvn@598 | 1372 | |
kvn@598 | 1373 | const Type* this_type = this->bottom_type(); |
kvn@5110 | 1374 | int this_index = C->get_alias_index(t_oop); |
kvn@5110 | 1375 | int this_offset = t_oop->offset(); |
kvn@5110 | 1376 | int this_iid = t_oop->instance_id(); |
kvn@5110 | 1377 | if (!t_oop->is_known_instance() && load_boxed_values) { |
kvn@5110 | 1378 | // Use _idx of address base for boxed values. |
kvn@5110 | 1379 | this_iid = base->_idx; |
kvn@5110 | 1380 | } |
kvn@5110 | 1381 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
kvn@5110 | 1382 | Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); |
kvn@2810 | 1383 | for (uint i = 1; i < region->req(); i++) { |
kvn@5110 | 1384 | Node* x; |
kvn@598 | 1385 | Node* the_clone = NULL; |
kvn@5110 | 1386 | if (region->in(i) == C->top()) { |
kvn@5110 | 1387 | x = C->top(); // Dead path? Use a dead data op |
kvn@598 | 1388 | } else { |
kvn@598 | 1389 | x = this->clone(); // Else clone up the data op |
kvn@598 | 1390 | the_clone = x; // Remember for possible deletion. |
kvn@598 | 1391 | // Alter data node to use pre-phi inputs |
kvn@2810 | 1392 | if (this->in(0) == region) { |
kvn@2810 | 1393 | x->set_req(0, region->in(i)); |
kvn@598 | 1394 | } else { |
kvn@2810 | 1395 | x->set_req(0, NULL); |
kvn@598 | 1396 | } |
kvn@5110 | 1397 | if (mem->is_Phi() && (mem->in(0) == region)) { |
kvn@5110 | 1398 | x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. |
kvn@5110 | 1399 | } |
kvn@5110 | 1400 | if (address->is_Phi() && address->in(0) == region) { |
kvn@5110 | 1401 | x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone |
kvn@5110 | 1402 | } |
kvn@5110 | 1403 | if (base_is_phi && (base->in(0) == region)) { |
kvn@5110 | 1404 | Node* base_x = base->in(i); // Clone address for loads from boxed objects. |
kvn@5110 | 1405 | Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset))); |
kvn@5110 | 1406 | x->set_req(Address, adr_x); |
kvn@598 | 1407 | } |
kvn@598 | 1408 | } |
kvn@598 | 1409 | // Check for a 'win' on some paths |
kvn@598 | 1410 | const Type *t = x->Value(igvn); |
kvn@598 | 1411 | |
kvn@598 | 1412 | bool singleton = t->singleton(); |
kvn@598 | 1413 | |
kvn@598 | 1414 | // See comments in PhaseIdealLoop::split_thru_phi(). |
kvn@2810 | 1415 | if (singleton && t == Type::TOP) { |
kvn@598 | 1416 | singleton &= region->is_Loop() && (i != LoopNode::EntryControl); |
kvn@598 | 1417 | } |
kvn@598 | 1418 | |
kvn@2810 | 1419 | if (singleton) { |
kvn@598 | 1420 | x = igvn->makecon(t); |
kvn@598 | 1421 | } else { |
kvn@598 | 1422 | // We now call Identity to try to simplify the cloned node. |
kvn@598 | 1423 | // Note that some Identity methods call phase->type(this). |
kvn@598 | 1424 | // Make sure that the type array is big enough for |
kvn@598 | 1425 | // our new node, even though we may throw the node away. |
kvn@598 | 1426 | // (This tweaking with igvn only works because x is a new node.) |
kvn@598 | 1427 | igvn->set_type(x, t); |
kvn@728 | 1428 | // If x is a TypeNode, capture any more-precise type permanently into Node |
twisti@1040 | 1429 | // otherwise it will be not updated during igvn->transform since |
kvn@728 | 1430 | // igvn->type(x) is set to x->Value() already. |
kvn@728 | 1431 | x->raise_bottom_type(t); |
kvn@598 | 1432 | Node *y = x->Identity(igvn); |
kvn@2810 | 1433 | if (y != x) { |
kvn@598 | 1434 | x = y; |
kvn@598 | 1435 | } else { |
kvn@5110 | 1436 | y = igvn->hash_find_insert(x); |
kvn@2810 | 1437 | if (y) { |
kvn@598 | 1438 | x = y; |
kvn@598 | 1439 | } else { |
kvn@598 | 1440 | // Else x is a new node we are keeping |
kvn@598 | 1441 | // We do not need register_new_node_with_optimizer |
kvn@598 | 1442 | // because set_type has already been called. |
kvn@598 | 1443 | igvn->_worklist.push(x); |
kvn@598 | 1444 | } |
kvn@598 | 1445 | } |
kvn@598 | 1446 | } |
kvn@5110 | 1447 | if (x != the_clone && the_clone != NULL) { |
kvn@598 | 1448 | igvn->remove_dead_node(the_clone); |
kvn@5110 | 1449 | } |
kvn@598 | 1450 | phi->set_req(i, x); |
kvn@598 | 1451 | } |
kvn@2810 | 1452 | // Record Phi |
kvn@2810 | 1453 | igvn->register_new_node_with_optimizer(phi); |
kvn@2810 | 1454 | return phi; |
kvn@598 | 1455 | } |
never@452 | 1456 | |
duke@435 | 1457 | //------------------------------Ideal------------------------------------------ |
duke@435 | 1458 | // If the load is from Field memory and the pointer is non-null, we can |
duke@435 | 1459 | // zero out the control input. |
duke@435 | 1460 | // If the offset is constant and the base is an object allocation, |
duke@435 | 1461 | // try to hook me up to the exact initializing store. |
duke@435 | 1462 | Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 1463 | Node* p = MemNode::Ideal_common(phase, can_reshape); |
duke@435 | 1464 | if (p) return (p == NodeSentinel) ? NULL : p; |
duke@435 | 1465 | |
duke@435 | 1466 | Node* ctrl = in(MemNode::Control); |
duke@435 | 1467 | Node* address = in(MemNode::Address); |
duke@435 | 1468 | |
duke@435 | 1469 | // Skip up past a SafePoint control. Cannot do this for Stores because |
duke@435 | 1470 | // pointer stores & cardmarks must stay on the same side of a SafePoint. |
duke@435 | 1471 | if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && |
duke@435 | 1472 | phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { |
duke@435 | 1473 | ctrl = ctrl->in(0); |
duke@435 | 1474 | set_req(MemNode::Control,ctrl); |
duke@435 | 1475 | } |
duke@435 | 1476 | |
kvn@1143 | 1477 | intptr_t ignore = 0; |
kvn@1143 | 1478 | Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); |
kvn@1143 | 1479 | if (base != NULL |
kvn@1143 | 1480 | && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { |
kvn@1143 | 1481 | // Check for useless control edge in some common special cases |
kvn@1143 | 1482 | if (in(MemNode::Control) != NULL |
duke@435 | 1483 | && phase->type(base)->higher_equal(TypePtr::NOTNULL) |
kvn@554 | 1484 | && all_controls_dominate(base, phase->C->start())) { |
duke@435 | 1485 | // A method-invariant, non-null address (constant or 'this' argument). |
duke@435 | 1486 | set_req(MemNode::Control, NULL); |
duke@435 | 1487 | } |
never@452 | 1488 | } |
never@452 | 1489 | |
kvn@509 | 1490 | Node* mem = in(MemNode::Memory); |
kvn@509 | 1491 | const TypePtr *addr_t = phase->type(address)->isa_ptr(); |
kvn@509 | 1492 | |
kvn@5110 | 1493 | if (can_reshape && (addr_t != NULL)) { |
kvn@509 | 1494 | // try to optimize our memory input |
kvn@5110 | 1495 | Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); |
kvn@509 | 1496 | if (opt_mem != mem) { |
kvn@509 | 1497 | set_req(MemNode::Memory, opt_mem); |
kvn@740 | 1498 | if (phase->type( opt_mem ) == Type::TOP) return NULL; |
kvn@509 | 1499 | return this; |
kvn@509 | 1500 | } |
kvn@509 | 1501 | const TypeOopPtr *t_oop = addr_t->isa_oopptr(); |
kvn@5110 | 1502 | if ((t_oop != NULL) && |
kvn@5110 | 1503 | (t_oop->is_known_instance_field() || |
kvn@5110 | 1504 | t_oop->is_ptr_to_boxed_value())) { |
kvn@3260 | 1505 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
kvn@3260 | 1506 | if (igvn != NULL && igvn->_worklist.member(opt_mem)) { |
kvn@3260 | 1507 | // Delay this transformation until memory Phi is processed. |
kvn@3260 | 1508 | phase->is_IterGVN()->_worklist.push(this); |
kvn@3260 | 1509 | return NULL; |
kvn@3260 | 1510 | } |
kvn@598 | 1511 | // Split instance field load through Phi. |
kvn@598 | 1512 | Node* result = split_through_phi(phase); |
kvn@598 | 1513 | if (result != NULL) return result; |
kvn@5110 | 1514 | |
kvn@5110 | 1515 | if (t_oop->is_ptr_to_boxed_value()) { |
kvn@5110 | 1516 | Node* result = eliminate_autobox(phase); |
kvn@5110 | 1517 | if (result != NULL) return result; |
kvn@5110 | 1518 | } |
kvn@509 | 1519 | } |
kvn@509 | 1520 | } |
kvn@509 | 1521 | |
duke@435 | 1522 | // Check for prior store with a different base or offset; make Load |
duke@435 | 1523 | // independent. Skip through any number of them. Bail out if the stores |
duke@435 | 1524 | // are in an endless dead cycle and report no progress. This is a key |
duke@435 | 1525 | // transform for Reflection. However, if after skipping through the Stores |
duke@435 | 1526 | // we can't then fold up against a prior store do NOT do the transform as |
duke@435 | 1527 | // this amounts to using the 'Oracle' model of aliasing. It leaves the same |
duke@435 | 1528 | // array memory alive twice: once for the hoisted Load and again after the |
duke@435 | 1529 | // bypassed Store. This situation only works if EVERYBODY who does |
duke@435 | 1530 | // anti-dependence work knows how to bypass. I.e. we need all |
duke@435 | 1531 | // anti-dependence checks to ask the same Oracle. Right now, that Oracle is |
duke@435 | 1532 | // the alias index stuff. So instead, peek through Stores and IFF we can |
duke@435 | 1533 | // fold up, do so. |
duke@435 | 1534 | Node* prev_mem = find_previous_store(phase); |
duke@435 | 1535 | // Steps (a), (b): Walk past independent stores to find an exact match. |
duke@435 | 1536 | if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { |
duke@435 | 1537 | // (c) See if we can fold up on the spot, but don't fold up here. |
twisti@993 | 1538 | // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or |
duke@435 | 1539 | // just return a prior value, which is done by Identity calls. |
duke@435 | 1540 | if (can_see_stored_value(prev_mem, phase)) { |
duke@435 | 1541 | // Make ready for step (d): |
duke@435 | 1542 | set_req(MemNode::Memory, prev_mem); |
duke@435 | 1543 | return this; |
duke@435 | 1544 | } |
duke@435 | 1545 | } |
duke@435 | 1546 | |
duke@435 | 1547 | return NULL; // No further progress |
duke@435 | 1548 | } |
duke@435 | 1549 | |
duke@435 | 1550 | // Helper to recognize certain Klass fields which are invariant across |
duke@435 | 1551 | // some group of array types (e.g., int[] or all T[] where T < Object). |
duke@435 | 1552 | const Type* |
duke@435 | 1553 | LoadNode::load_array_final_field(const TypeKlassPtr *tkls, |
duke@435 | 1554 | ciKlass* klass) const { |
stefank@3391 | 1555 | if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { |
duke@435 | 1556 | // The field is Klass::_modifier_flags. Return its (constant) value. |
duke@435 | 1557 | // (Folds up the 2nd indirection in aClassConstant.getModifiers().) |
duke@435 | 1558 | assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); |
duke@435 | 1559 | return TypeInt::make(klass->modifier_flags()); |
duke@435 | 1560 | } |
stefank@3391 | 1561 | if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { |
duke@435 | 1562 | // The field is Klass::_access_flags. Return its (constant) value. |
duke@435 | 1563 | // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) |
duke@435 | 1564 | assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); |
duke@435 | 1565 | return TypeInt::make(klass->access_flags()); |
duke@435 | 1566 | } |
stefank@3391 | 1567 | if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { |
duke@435 | 1568 | // The field is Klass::_layout_helper. Return its constant value if known. |
duke@435 | 1569 | assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); |
duke@435 | 1570 | return TypeInt::make(klass->layout_helper()); |
duke@435 | 1571 | } |
duke@435 | 1572 | |
duke@435 | 1573 | // No match. |
duke@435 | 1574 | return NULL; |
duke@435 | 1575 | } |
duke@435 | 1576 | |
vlivanov@5658 | 1577 | // Try to constant-fold a stable array element. |
vlivanov@5658 | 1578 | static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) { |
vlivanov@6526 | 1579 | assert(ary->const_oop(), "array should be constant"); |
vlivanov@5658 | 1580 | assert(ary->is_stable(), "array should be stable"); |
vlivanov@5658 | 1581 | |
vlivanov@6526 | 1582 | // Decode the results of GraphKit::array_element_address. |
vlivanov@6526 | 1583 | ciArray* aobj = ary->const_oop()->as_array(); |
vlivanov@6526 | 1584 | ciConstant con = aobj->element_value_by_offset(off); |
vlivanov@6526 | 1585 | |
vlivanov@6526 | 1586 | if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) { |
vlivanov@6526 | 1587 | const Type* con_type = Type::make_from_constant(con); |
vlivanov@6526 | 1588 | if (con_type != NULL) { |
vlivanov@6526 | 1589 | if (con_type->isa_aryptr()) { |
vlivanov@6526 | 1590 | // Join with the array element type, in case it is also stable. |
vlivanov@6526 | 1591 | int dim = ary->stable_dimension(); |
vlivanov@6526 | 1592 | con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1); |
vlivanov@6526 | 1593 | } |
vlivanov@6526 | 1594 | if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) { |
vlivanov@6526 | 1595 | con_type = con_type->make_narrowoop(); |
vlivanov@6526 | 1596 | } |
vlivanov@5658 | 1597 | #ifndef PRODUCT |
vlivanov@6526 | 1598 | if (TraceIterativeGVN) { |
vlivanov@6526 | 1599 | tty->print("FoldStableValues: array element [off=%d]: con_type=", off); |
vlivanov@6526 | 1600 | con_type->dump(); tty->cr(); |
vlivanov@6526 | 1601 | } |
vlivanov@5658 | 1602 | #endif //PRODUCT |
vlivanov@6526 | 1603 | return con_type; |
vlivanov@5658 | 1604 | } |
vlivanov@5658 | 1605 | } |
vlivanov@5658 | 1606 | return NULL; |
vlivanov@5658 | 1607 | } |
vlivanov@5658 | 1608 | |
duke@435 | 1609 | //------------------------------Value----------------------------------------- |
duke@435 | 1610 | const Type *LoadNode::Value( PhaseTransform *phase ) const { |
duke@435 | 1611 | // Either input is TOP ==> the result is TOP |
duke@435 | 1612 | Node* mem = in(MemNode::Memory); |
duke@435 | 1613 | const Type *t1 = phase->type(mem); |
duke@435 | 1614 | if (t1 == Type::TOP) return Type::TOP; |
duke@435 | 1615 | Node* adr = in(MemNode::Address); |
duke@435 | 1616 | const TypePtr* tp = phase->type(adr)->isa_ptr(); |
duke@435 | 1617 | if (tp == NULL || tp->empty()) return Type::TOP; |
duke@435 | 1618 | int off = tp->offset(); |
duke@435 | 1619 | assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); |
twisti@3102 | 1620 | Compile* C = phase->C; |
duke@435 | 1621 | |
duke@435 | 1622 | // Try to guess loaded type from pointer type |
vlivanov@5658 | 1623 | if (tp->isa_aryptr()) { |
vlivanov@5658 | 1624 | const TypeAryPtr* ary = tp->is_aryptr(); |
vlivanov@6526 | 1625 | const Type* t = ary->elem(); |
vlivanov@5658 | 1626 | |
vlivanov@5658 | 1627 | // Determine whether the reference is beyond the header or not, by comparing |
vlivanov@5658 | 1628 | // the offset against the offset of the start of the array's data. |
vlivanov@5658 | 1629 | // Different array types begin at slightly different offsets (12 vs. 16). |
vlivanov@5658 | 1630 | // We choose T_BYTE as an example base type that is least restrictive |
vlivanov@5658 | 1631 | // as to alignment, which will therefore produce the smallest |
vlivanov@5658 | 1632 | // possible base offset. |
vlivanov@5658 | 1633 | const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
vlivanov@5658 | 1634 | const bool off_beyond_header = ((uint)off >= (uint)min_base_off); |
vlivanov@5658 | 1635 | |
vlivanov@5658 | 1636 | // Try to constant-fold a stable array element. |
vlivanov@6526 | 1637 | if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) { |
vlivanov@6526 | 1638 | // Make sure the reference is not into the header and the offset is constant |
vlivanov@6526 | 1639 | if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { |
vlivanov@5658 | 1640 | const Type* con_type = fold_stable_ary_elem(ary, off, memory_type()); |
vlivanov@5658 | 1641 | if (con_type != NULL) { |
vlivanov@5658 | 1642 | return con_type; |
vlivanov@5658 | 1643 | } |
vlivanov@5658 | 1644 | } |
vlivanov@5658 | 1645 | } |
vlivanov@5658 | 1646 | |
duke@435 | 1647 | // Don't do this for integer types. There is only potential profit if |
duke@435 | 1648 | // the element type t is lower than _type; that is, for int types, if _type is |
duke@435 | 1649 | // more restrictive than t. This only happens here if one is short and the other |
duke@435 | 1650 | // char (both 16 bits), and in those cases we've made an intentional decision |
duke@435 | 1651 | // to use one kind of load over the other. See AndINode::Ideal and 4965907. |
duke@435 | 1652 | // Also, do not try to narrow the type for a LoadKlass, regardless of offset. |
duke@435 | 1653 | // |
duke@435 | 1654 | // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) |
duke@435 | 1655 | // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier |
duke@435 | 1656 | // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been |
duke@435 | 1657 | // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, |
duke@435 | 1658 | // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. |
duke@435 | 1659 | // In fact, that could have been the original type of p1, and p1 could have |
duke@435 | 1660 | // had an original form like p1:(AddP x x (LShiftL quux 3)), where the |
duke@435 | 1661 | // expression (LShiftL quux 3) independently optimized to the constant 8. |
duke@435 | 1662 | if ((t->isa_int() == NULL) && (t->isa_long() == NULL) |
kvn@3882 | 1663 | && (_type->isa_vect() == NULL) |
kvn@728 | 1664 | && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { |
duke@435 | 1665 | // t might actually be lower than _type, if _type is a unique |
duke@435 | 1666 | // concrete subclass of abstract class t. |
vlivanov@5658 | 1667 | if (off_beyond_header) { // is the offset beyond the header? |
roland@6313 | 1668 | const Type* jt = t->join_speculative(_type); |
duke@435 | 1669 | // In any case, do not allow the join, per se, to empty out the type. |
duke@435 | 1670 | if (jt->empty() && !t->empty()) { |
duke@435 | 1671 | // This can happen if a interface-typed array narrows to a class type. |
duke@435 | 1672 | jt = _type; |
duke@435 | 1673 | } |
kvn@5110 | 1674 | #ifdef ASSERT |
kvn@5110 | 1675 | if (phase->C->eliminate_boxing() && adr->is_AddP()) { |
never@452 | 1676 | // The pointers in the autobox arrays are always non-null |
kvn@1143 | 1677 | Node* base = adr->in(AddPNode::Base); |
kvn@5110 | 1678 | if ((base != NULL) && base->is_DecodeN()) { |
kvn@5110 | 1679 | // Get LoadN node which loads IntegerCache.cache field |
kvn@5110 | 1680 | base = base->in(1); |
kvn@5110 | 1681 | } |
kvn@5110 | 1682 | if ((base != NULL) && base->is_Con()) { |
kvn@5110 | 1683 | const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); |
kvn@5110 | 1684 | if ((base_type != NULL) && base_type->is_autobox_cache()) { |
kvn@5110 | 1685 | // It could be narrow oop |
kvn@5110 | 1686 | assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); |
never@452 | 1687 | } |
never@452 | 1688 | } |
never@452 | 1689 | } |
kvn@5110 | 1690 | #endif |
duke@435 | 1691 | return jt; |
duke@435 | 1692 | } |
duke@435 | 1693 | } |
duke@435 | 1694 | } else if (tp->base() == Type::InstPtr) { |
twisti@3102 | 1695 | ciEnv* env = C->env(); |
never@1515 | 1696 | const TypeInstPtr* tinst = tp->is_instptr(); |
never@1515 | 1697 | ciKlass* klass = tinst->klass(); |
duke@435 | 1698 | assert( off != Type::OffsetBot || |
duke@435 | 1699 | // arrays can be cast to Objects |
duke@435 | 1700 | tp->is_oopptr()->klass()->is_java_lang_Object() || |
duke@435 | 1701 | // unsafe field access may not have a constant offset |
twisti@3102 | 1702 | C->has_unsafe_access(), |
duke@435 | 1703 | "Field accesses must be precise" ); |
duke@435 | 1704 | // For oop loads, we expect the _type to be precise |
twisti@3102 | 1705 | if (klass == env->String_klass() && |
never@1515 | 1706 | adr->is_AddP() && off != Type::OffsetBot) { |
kvn@2602 | 1707 | // For constant Strings treat the final fields as compile time constants. |
never@1515 | 1708 | Node* base = adr->in(AddPNode::Base); |
never@2121 | 1709 | const TypeOopPtr* t = phase->type(base)->isa_oopptr(); |
never@2121 | 1710 | if (t != NULL && t->singleton()) { |
twisti@3102 | 1711 | ciField* field = env->String_klass()->get_field_by_offset(off, false); |
kvn@2602 | 1712 | if (field != NULL && field->is_final()) { |
kvn@2602 | 1713 | ciObject* string = t->const_oop(); |
kvn@2602 | 1714 | ciConstant constant = string->as_instance()->field_value(field); |
kvn@2602 | 1715 | if (constant.basic_type() == T_INT) { |
kvn@2602 | 1716 | return TypeInt::make(constant.as_int()); |
kvn@2602 | 1717 | } else if (constant.basic_type() == T_ARRAY) { |
kvn@2602 | 1718 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
jcoomes@2661 | 1719 | return TypeNarrowOop::make_from_constant(constant.as_object(), true); |
kvn@2602 | 1720 | } else { |
jcoomes@2661 | 1721 | return TypeOopPtr::make_from_constant(constant.as_object(), true); |
kvn@2602 | 1722 | } |
never@1515 | 1723 | } |
never@1515 | 1724 | } |
never@1515 | 1725 | } |
never@1515 | 1726 | } |
twisti@3102 | 1727 | // Optimizations for constant objects |
twisti@3102 | 1728 | ciObject* const_oop = tinst->const_oop(); |
twisti@3102 | 1729 | if (const_oop != NULL) { |
kvn@5110 | 1730 | // For constant Boxed value treat the target field as a compile time constant. |
kvn@5110 | 1731 | if (tinst->is_ptr_to_boxed_value()) { |
kvn@5110 | 1732 | return tinst->get_const_boxed_value(); |
kvn@5110 | 1733 | } else |
twisti@3102 | 1734 | // For constant CallSites treat the target field as a compile time constant. |
twisti@3102 | 1735 | if (const_oop->is_call_site()) { |
twisti@3102 | 1736 | ciCallSite* call_site = const_oop->as_call_site(); |
twisti@3102 | 1737 | ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); |
twisti@3102 | 1738 | if (field != NULL && field->is_call_site_target()) { |
twisti@3102 | 1739 | ciMethodHandle* target = call_site->get_target(); |
twisti@3102 | 1740 | if (target != NULL) { // just in case |
twisti@3102 | 1741 | ciConstant constant(T_OBJECT, target); |
twisti@3102 | 1742 | const Type* t; |
twisti@3102 | 1743 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
twisti@3102 | 1744 | t = TypeNarrowOop::make_from_constant(constant.as_object(), true); |
twisti@3102 | 1745 | } else { |
twisti@3102 | 1746 | t = TypeOopPtr::make_from_constant(constant.as_object(), true); |
twisti@3102 | 1747 | } |
twisti@3102 | 1748 | // Add a dependence for invalidation of the optimization. |
twisti@3102 | 1749 | if (!call_site->is_constant_call_site()) { |
twisti@3102 | 1750 | C->dependencies()->assert_call_site_target_value(call_site, target); |
twisti@3102 | 1751 | } |
twisti@3102 | 1752 | return t; |
twisti@3102 | 1753 | } |
twisti@3102 | 1754 | } |
twisti@3102 | 1755 | } |
twisti@3102 | 1756 | } |
duke@435 | 1757 | } else if (tp->base() == Type::KlassPtr) { |
duke@435 | 1758 | assert( off != Type::OffsetBot || |
duke@435 | 1759 | // arrays can be cast to Objects |
duke@435 | 1760 | tp->is_klassptr()->klass()->is_java_lang_Object() || |
duke@435 | 1761 | // also allow array-loading from the primary supertype |
duke@435 | 1762 | // array during subtype checks |
duke@435 | 1763 | Opcode() == Op_LoadKlass, |
duke@435 | 1764 | "Field accesses must be precise" ); |
duke@435 | 1765 | // For klass/static loads, we expect the _type to be precise |
duke@435 | 1766 | } |
duke@435 | 1767 | |
duke@435 | 1768 | const TypeKlassPtr *tkls = tp->isa_klassptr(); |
duke@435 | 1769 | if (tkls != NULL && !StressReflectiveCode) { |
duke@435 | 1770 | ciKlass* klass = tkls->klass(); |
duke@435 | 1771 | if (klass->is_loaded() && tkls->klass_is_exact()) { |
duke@435 | 1772 | // We are loading a field from a Klass metaobject whose identity |
duke@435 | 1773 | // is known at compile time (the type is "exact" or "precise"). |
duke@435 | 1774 | // Check for fields we know are maintained as constants by the VM. |
stefank@3391 | 1775 | if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { |
duke@435 | 1776 | // The field is Klass::_super_check_offset. Return its (constant) value. |
duke@435 | 1777 | // (Folds up type checking code.) |
duke@435 | 1778 | assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); |
duke@435 | 1779 | return TypeInt::make(klass->super_check_offset()); |
duke@435 | 1780 | } |
duke@435 | 1781 | // Compute index into primary_supers array |
coleenp@4037 | 1782 | juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); |
duke@435 | 1783 | // Check for overflowing; use unsigned compare to handle the negative case. |
duke@435 | 1784 | if( depth < ciKlass::primary_super_limit() ) { |
duke@435 | 1785 | // The field is an element of Klass::_primary_supers. Return its (constant) value. |
duke@435 | 1786 | // (Folds up type checking code.) |
duke@435 | 1787 | assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); |
duke@435 | 1788 | ciKlass *ss = klass->super_of_depth(depth); |
duke@435 | 1789 | return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; |
duke@435 | 1790 | } |
duke@435 | 1791 | const Type* aift = load_array_final_field(tkls, klass); |
duke@435 | 1792 | if (aift != NULL) return aift; |
coleenp@4142 | 1793 | if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset()) |
duke@435 | 1794 | && klass->is_array_klass()) { |
coleenp@4142 | 1795 | // The field is ArrayKlass::_component_mirror. Return its (constant) value. |
duke@435 | 1796 | // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) |
duke@435 | 1797 | assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); |
duke@435 | 1798 | return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); |
duke@435 | 1799 | } |
stefank@3391 | 1800 | if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { |
duke@435 | 1801 | // The field is Klass::_java_mirror. Return its (constant) value. |
duke@435 | 1802 | // (Folds up the 2nd indirection in anObjConstant.getClass().) |
duke@435 | 1803 | assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); |
duke@435 | 1804 | return TypeInstPtr::make(klass->java_mirror()); |
duke@435 | 1805 | } |
duke@435 | 1806 | } |
duke@435 | 1807 | |
duke@435 | 1808 | // We can still check if we are loading from the primary_supers array at a |
duke@435 | 1809 | // shallow enough depth. Even though the klass is not exact, entries less |
duke@435 | 1810 | // than or equal to its super depth are correct. |
duke@435 | 1811 | if (klass->is_loaded() ) { |
coleenp@4037 | 1812 | ciType *inner = klass; |
duke@435 | 1813 | while( inner->is_obj_array_klass() ) |
duke@435 | 1814 | inner = inner->as_obj_array_klass()->base_element_type(); |
duke@435 | 1815 | if( inner->is_instance_klass() && |
duke@435 | 1816 | !inner->as_instance_klass()->flags().is_interface() ) { |
duke@435 | 1817 | // Compute index into primary_supers array |
coleenp@4037 | 1818 | juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); |
duke@435 | 1819 | // Check for overflowing; use unsigned compare to handle the negative case. |
duke@435 | 1820 | if( depth < ciKlass::primary_super_limit() && |
duke@435 | 1821 | depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case |
duke@435 | 1822 | // The field is an element of Klass::_primary_supers. Return its (constant) value. |
duke@435 | 1823 | // (Folds up type checking code.) |
duke@435 | 1824 | assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); |
duke@435 | 1825 | ciKlass *ss = klass->super_of_depth(depth); |
duke@435 | 1826 | return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; |
duke@435 | 1827 | } |
duke@435 | 1828 | } |
duke@435 | 1829 | } |
duke@435 | 1830 | |
duke@435 | 1831 | // If the type is enough to determine that the thing is not an array, |
duke@435 | 1832 | // we can give the layout_helper a positive interval type. |
duke@435 | 1833 | // This will help short-circuit some reflective code. |
stefank@3391 | 1834 | if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) |
duke@435 | 1835 | && !klass->is_array_klass() // not directly typed as an array |
duke@435 | 1836 | && !klass->is_interface() // specifically not Serializable & Cloneable |
duke@435 | 1837 | && !klass->is_java_lang_Object() // not the supertype of all T[] |
duke@435 | 1838 | ) { |
duke@435 | 1839 | // Note: When interfaces are reliable, we can narrow the interface |
duke@435 | 1840 | // test to (klass != Serializable && klass != Cloneable). |
duke@435 | 1841 | assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); |
duke@435 | 1842 | jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); |
duke@435 | 1843 | // The key property of this type is that it folds up tests |
duke@435 | 1844 | // for array-ness, since it proves that the layout_helper is positive. |
duke@435 | 1845 | // Thus, a generic value like the basic object layout helper works fine. |
duke@435 | 1846 | return TypeInt::make(min_size, max_jint, Type::WidenMin); |
duke@435 | 1847 | } |
duke@435 | 1848 | } |
duke@435 | 1849 | |
duke@435 | 1850 | // If we are loading from a freshly-allocated object, produce a zero, |
duke@435 | 1851 | // if the load is provably beyond the header of the object. |
duke@435 | 1852 | // (Also allow a variable load from a fresh array to produce zero.) |
kvn@2810 | 1853 | const TypeOopPtr *tinst = tp->isa_oopptr(); |
kvn@2810 | 1854 | bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); |
kvn@5110 | 1855 | bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); |
kvn@5110 | 1856 | if (ReduceFieldZeroing || is_instance || is_boxed_value) { |
duke@435 | 1857 | Node* value = can_see_stored_value(mem,phase); |
kvn@3442 | 1858 | if (value != NULL && value->is_Con()) { |
kvn@3442 | 1859 | assert(value->bottom_type()->higher_equal(_type),"sanity"); |
duke@435 | 1860 | return value->bottom_type(); |
kvn@3442 | 1861 | } |
duke@435 | 1862 | } |
duke@435 | 1863 | |
kvn@2810 | 1864 | if (is_instance) { |
kvn@499 | 1865 | // If we have an instance type and our memory input is the |
kvn@499 | 1866 | // programs's initial memory state, there is no matching store, |
kvn@499 | 1867 | // so just return a zero of the appropriate type |
kvn@499 | 1868 | Node *mem = in(MemNode::Memory); |
kvn@499 | 1869 | if (mem->is_Parm() && mem->in(0)->is_Start()) { |
kvn@499 | 1870 | assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); |
kvn@499 | 1871 | return Type::get_zero_type(_type->basic_type()); |
kvn@499 | 1872 | } |
kvn@499 | 1873 | } |
duke@435 | 1874 | return _type; |
duke@435 | 1875 | } |
duke@435 | 1876 | |
duke@435 | 1877 | //------------------------------match_edge------------------------------------- |
duke@435 | 1878 | // Do we Match on this edge index or not? Match only the address. |
duke@435 | 1879 | uint LoadNode::match_edge(uint idx) const { |
duke@435 | 1880 | return idx == MemNode::Address; |
duke@435 | 1881 | } |
duke@435 | 1882 | |
duke@435 | 1883 | //--------------------------LoadBNode::Ideal-------------------------------------- |
duke@435 | 1884 | // |
duke@435 | 1885 | // If the previous store is to the same address as this load, |
duke@435 | 1886 | // and the value stored was larger than a byte, replace this load |
duke@435 | 1887 | // with the value stored truncated to a byte. If no truncation is |
duke@435 | 1888 | // needed, the replacement is done in LoadNode::Identity(). |
duke@435 | 1889 | // |
duke@435 | 1890 | Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 1891 | Node* mem = in(MemNode::Memory); |
duke@435 | 1892 | Node* value = can_see_stored_value(mem,phase); |
duke@435 | 1893 | if( value && !phase->type(value)->higher_equal( _type ) ) { |
kvn@4115 | 1894 | Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) ); |
kvn@4115 | 1895 | return new (phase->C) RShiftINode(result, phase->intcon(24)); |
duke@435 | 1896 | } |
duke@435 | 1897 | // Identity call will handle the case where truncation is not needed. |
duke@435 | 1898 | return LoadNode::Ideal(phase, can_reshape); |
duke@435 | 1899 | } |
duke@435 | 1900 | |
kvn@3442 | 1901 | const Type* LoadBNode::Value(PhaseTransform *phase) const { |
kvn@3442 | 1902 | Node* mem = in(MemNode::Memory); |
kvn@3442 | 1903 | Node* value = can_see_stored_value(mem,phase); |
kvn@3448 | 1904 | if (value != NULL && value->is_Con() && |
kvn@3448 | 1905 | !value->bottom_type()->higher_equal(_type)) { |
kvn@3442 | 1906 | // If the input to the store does not fit with the load's result type, |
kvn@3442 | 1907 | // it must be truncated. We can't delay until Ideal call since |
kvn@3442 | 1908 | // a singleton Value is needed for split_thru_phi optimization. |
kvn@3442 | 1909 | int con = value->get_int(); |
kvn@3442 | 1910 | return TypeInt::make((con << 24) >> 24); |
kvn@3442 | 1911 | } |
kvn@3442 | 1912 | return LoadNode::Value(phase); |
kvn@3442 | 1913 | } |
kvn@3442 | 1914 | |
twisti@1059 | 1915 | //--------------------------LoadUBNode::Ideal------------------------------------- |
twisti@1059 | 1916 | // |
twisti@1059 | 1917 | // If the previous store is to the same address as this load, |
twisti@1059 | 1918 | // and the value stored was larger than a byte, replace this load |
twisti@1059 | 1919 | // with the value stored truncated to a byte. If no truncation is |
twisti@1059 | 1920 | // needed, the replacement is done in LoadNode::Identity(). |
twisti@1059 | 1921 | // |
twisti@1059 | 1922 | Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { |
twisti@1059 | 1923 | Node* mem = in(MemNode::Memory); |
twisti@1059 | 1924 | Node* value = can_see_stored_value(mem, phase); |
twisti@1059 | 1925 | if (value && !phase->type(value)->higher_equal(_type)) |
kvn@4115 | 1926 | return new (phase->C) AndINode(value, phase->intcon(0xFF)); |
twisti@1059 | 1927 | // Identity call will handle the case where truncation is not needed. |
twisti@1059 | 1928 | return LoadNode::Ideal(phase, can_reshape); |
twisti@1059 | 1929 | } |
twisti@1059 | 1930 | |
kvn@3442 | 1931 | const Type* LoadUBNode::Value(PhaseTransform *phase) const { |
kvn@3442 | 1932 | Node* mem = in(MemNode::Memory); |
kvn@3442 | 1933 | Node* value = can_see_stored_value(mem,phase); |
kvn@3448 | 1934 | if (value != NULL && value->is_Con() && |
kvn@3448 | 1935 | !value->bottom_type()->higher_equal(_type)) { |
kvn@3442 | 1936 | // If the input to the store does not fit with the load's result type, |
kvn@3442 | 1937 | // it must be truncated. We can't delay until Ideal call since |
kvn@3442 | 1938 | // a singleton Value is needed for split_thru_phi optimization. |
kvn@3442 | 1939 | int con = value->get_int(); |
kvn@3442 | 1940 | return TypeInt::make(con & 0xFF); |
kvn@3442 | 1941 | } |
kvn@3442 | 1942 | return LoadNode::Value(phase); |
kvn@3442 | 1943 | } |
kvn@3442 | 1944 | |
twisti@993 | 1945 | //--------------------------LoadUSNode::Ideal------------------------------------- |
duke@435 | 1946 | // |
duke@435 | 1947 | // If the previous store is to the same address as this load, |
duke@435 | 1948 | // and the value stored was larger than a char, replace this load |
duke@435 | 1949 | // with the value stored truncated to a char. If no truncation is |
duke@435 | 1950 | // needed, the replacement is done in LoadNode::Identity(). |
duke@435 | 1951 | // |
twisti@993 | 1952 | Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 1953 | Node* mem = in(MemNode::Memory); |
duke@435 | 1954 | Node* value = can_see_stored_value(mem,phase); |
duke@435 | 1955 | if( value && !phase->type(value)->higher_equal( _type ) ) |
kvn@4115 | 1956 | return new (phase->C) AndINode(value,phase->intcon(0xFFFF)); |
duke@435 | 1957 | // Identity call will handle the case where truncation is not needed. |
duke@435 | 1958 | return LoadNode::Ideal(phase, can_reshape); |
duke@435 | 1959 | } |
duke@435 | 1960 | |
kvn@3442 | 1961 | const Type* LoadUSNode::Value(PhaseTransform *phase) const { |
kvn@3442 | 1962 | Node* mem = in(MemNode::Memory); |
kvn@3442 | 1963 | Node* value = can_see_stored_value(mem,phase); |
kvn@3448 | 1964 | if (value != NULL && value->is_Con() && |
kvn@3448 | 1965 | !value->bottom_type()->higher_equal(_type)) { |
kvn@3442 | 1966 | // If the input to the store does not fit with the load's result type, |
kvn@3442 | 1967 | // it must be truncated. We can't delay until Ideal call since |
kvn@3442 | 1968 | // a singleton Value is needed for split_thru_phi optimization. |
kvn@3442 | 1969 | int con = value->get_int(); |
kvn@3442 | 1970 | return TypeInt::make(con & 0xFFFF); |
kvn@3442 | 1971 | } |
kvn@3442 | 1972 | return LoadNode::Value(phase); |
kvn@3442 | 1973 | } |
kvn@3442 | 1974 | |
duke@435 | 1975 | //--------------------------LoadSNode::Ideal-------------------------------------- |
duke@435 | 1976 | // |
duke@435 | 1977 | // If the previous store is to the same address as this load, |
duke@435 | 1978 | // and the value stored was larger than a short, replace this load |
duke@435 | 1979 | // with the value stored truncated to a short. If no truncation is |
duke@435 | 1980 | // needed, the replacement is done in LoadNode::Identity(). |
duke@435 | 1981 | // |
duke@435 | 1982 | Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 1983 | Node* mem = in(MemNode::Memory); |
duke@435 | 1984 | Node* value = can_see_stored_value(mem,phase); |
duke@435 | 1985 | if( value && !phase->type(value)->higher_equal( _type ) ) { |
kvn@4115 | 1986 | Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) ); |
kvn@4115 | 1987 | return new (phase->C) RShiftINode(result, phase->intcon(16)); |
duke@435 | 1988 | } |
duke@435 | 1989 | // Identity call will handle the case where truncation is not needed. |
duke@435 | 1990 | return LoadNode::Ideal(phase, can_reshape); |
duke@435 | 1991 | } |
duke@435 | 1992 | |
kvn@3442 | 1993 | const Type* LoadSNode::Value(PhaseTransform *phase) const { |
kvn@3442 | 1994 | Node* mem = in(MemNode::Memory); |
kvn@3442 | 1995 | Node* value = can_see_stored_value(mem,phase); |
kvn@3448 | 1996 | if (value != NULL && value->is_Con() && |
kvn@3448 | 1997 | !value->bottom_type()->higher_equal(_type)) { |
kvn@3442 | 1998 | // If the input to the store does not fit with the load's result type, |
kvn@3442 | 1999 | // it must be truncated. We can't delay until Ideal call since |
kvn@3442 | 2000 | // a singleton Value is needed for split_thru_phi optimization. |
kvn@3442 | 2001 | int con = value->get_int(); |
kvn@3442 | 2002 | return TypeInt::make((con << 16) >> 16); |
kvn@3442 | 2003 | } |
kvn@3442 | 2004 | return LoadNode::Value(phase); |
kvn@3442 | 2005 | } |
kvn@3442 | 2006 | |
duke@435 | 2007 | //============================================================================= |
kvn@599 | 2008 | //----------------------------LoadKlassNode::make------------------------------ |
kvn@599 | 2009 | // Polymorphic factory method: |
kvn@599 | 2010 | Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { |
kvn@599 | 2011 | Compile* C = gvn.C; |
kvn@599 | 2012 | Node *ctl = NULL; |
kvn@599 | 2013 | // sanity check the alias category against the created node type |
coleenp@4037 | 2014 | const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); |
coleenp@4037 | 2015 | assert(adr_type != NULL, "expecting TypeKlassPtr"); |
kvn@599 | 2016 | #ifdef _LP64 |
roland@4159 | 2017 | if (adr_type->is_ptr_to_narrowklass()) { |
ehelin@5694 | 2018 | assert(UseCompressedClassPointers, "no compressed klasses"); |
goetz@6479 | 2019 | Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); |
roland@4159 | 2020 | return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); |
kvn@603 | 2021 | } |
kvn@599 | 2022 | #endif |
roland@4159 | 2023 | assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); |
goetz@6479 | 2024 | return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); |
kvn@599 | 2025 | } |
kvn@599 | 2026 | |
duke@435 | 2027 | //------------------------------Value------------------------------------------ |
duke@435 | 2028 | const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { |
kvn@599 | 2029 | return klass_value_common(phase); |
kvn@599 | 2030 | } |
kvn@599 | 2031 | |
kvn@599 | 2032 | const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { |
duke@435 | 2033 | // Either input is TOP ==> the result is TOP |
duke@435 | 2034 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
duke@435 | 2035 | if (t1 == Type::TOP) return Type::TOP; |
duke@435 | 2036 | Node *adr = in(MemNode::Address); |
duke@435 | 2037 | const Type *t2 = phase->type( adr ); |
duke@435 | 2038 | if (t2 == Type::TOP) return Type::TOP; |
duke@435 | 2039 | const TypePtr *tp = t2->is_ptr(); |
duke@435 | 2040 | if (TypePtr::above_centerline(tp->ptr()) || |
duke@435 | 2041 | tp->ptr() == TypePtr::Null) return Type::TOP; |
duke@435 | 2042 | |
duke@435 | 2043 | // Return a more precise klass, if possible |
duke@435 | 2044 | const TypeInstPtr *tinst = tp->isa_instptr(); |
duke@435 | 2045 | if (tinst != NULL) { |
duke@435 | 2046 | ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); |
duke@435 | 2047 | int offset = tinst->offset(); |
duke@435 | 2048 | if (ik == phase->C->env()->Class_klass() |
duke@435 | 2049 | && (offset == java_lang_Class::klass_offset_in_bytes() || |
duke@435 | 2050 | offset == java_lang_Class::array_klass_offset_in_bytes())) { |
duke@435 | 2051 | // We are loading a special hidden field from a Class mirror object, |
duke@435 | 2052 | // the field which points to the VM's Klass metaobject. |
duke@435 | 2053 | ciType* t = tinst->java_mirror_type(); |
duke@435 | 2054 | // java_mirror_type returns non-null for compile-time Class constants. |
duke@435 | 2055 | if (t != NULL) { |
duke@435 | 2056 | // constant oop => constant klass |
duke@435 | 2057 | if (offset == java_lang_Class::array_klass_offset_in_bytes()) { |
twisti@6173 | 2058 | if (t->is_void()) { |
twisti@6173 | 2059 | // We cannot create a void array. Since void is a primitive type return null |
twisti@6173 | 2060 | // klass. Users of this result need to do a null check on the returned klass. |
twisti@6173 | 2061 | return TypePtr::NULL_PTR; |
twisti@6173 | 2062 | } |
duke@435 | 2063 | return TypeKlassPtr::make(ciArrayKlass::make(t)); |
duke@435 | 2064 | } |
duke@435 | 2065 | if (!t->is_klass()) { |
duke@435 | 2066 | // a primitive Class (e.g., int.class) has NULL for a klass field |
duke@435 | 2067 | return TypePtr::NULL_PTR; |
duke@435 | 2068 | } |
duke@435 | 2069 | // (Folds up the 1st indirection in aClassConstant.getModifiers().) |
duke@435 | 2070 | return TypeKlassPtr::make(t->as_klass()); |
duke@435 | 2071 | } |
duke@435 | 2072 | // non-constant mirror, so we can't tell what's going on |
duke@435 | 2073 | } |
duke@435 | 2074 | if( !ik->is_loaded() ) |
duke@435 | 2075 | return _type; // Bail out if not loaded |
duke@435 | 2076 | if (offset == oopDesc::klass_offset_in_bytes()) { |
duke@435 | 2077 | if (tinst->klass_is_exact()) { |
duke@435 | 2078 | return TypeKlassPtr::make(ik); |
duke@435 | 2079 | } |
duke@435 | 2080 | // See if we can become precise: no subklasses and no interface |
duke@435 | 2081 | // (Note: We need to support verified interfaces.) |
duke@435 | 2082 | if (!ik->is_interface() && !ik->has_subklass()) { |
duke@435 | 2083 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
duke@435 | 2084 | // Add a dependence; if any subclass added we need to recompile |
duke@435 | 2085 | if (!ik->is_final()) { |
duke@435 | 2086 | // %%% should use stronger assert_unique_concrete_subtype instead |
duke@435 | 2087 | phase->C->dependencies()->assert_leaf_type(ik); |
duke@435 | 2088 | } |
duke@435 | 2089 | // Return precise klass |
duke@435 | 2090 | return TypeKlassPtr::make(ik); |
duke@435 | 2091 | } |
duke@435 | 2092 | |
duke@435 | 2093 | // Return root of possible klass |
duke@435 | 2094 | return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); |
duke@435 | 2095 | } |
duke@435 | 2096 | } |
duke@435 | 2097 | |
duke@435 | 2098 | // Check for loading klass from an array |
duke@435 | 2099 | const TypeAryPtr *tary = tp->isa_aryptr(); |
duke@435 | 2100 | if( tary != NULL ) { |
duke@435 | 2101 | ciKlass *tary_klass = tary->klass(); |
duke@435 | 2102 | if (tary_klass != NULL // can be NULL when at BOTTOM or TOP |
duke@435 | 2103 | && tary->offset() == oopDesc::klass_offset_in_bytes()) { |
duke@435 | 2104 | if (tary->klass_is_exact()) { |
duke@435 | 2105 | return TypeKlassPtr::make(tary_klass); |
duke@435 | 2106 | } |
duke@435 | 2107 | ciArrayKlass *ak = tary->klass()->as_array_klass(); |
duke@435 | 2108 | // If the klass is an object array, we defer the question to the |
duke@435 | 2109 | // array component klass. |
duke@435 | 2110 | if( ak->is_obj_array_klass() ) { |
duke@435 | 2111 | assert( ak->is_loaded(), "" ); |
duke@435 | 2112 | ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); |
duke@435 | 2113 | if( base_k->is_loaded() && base_k->is_instance_klass() ) { |
duke@435 | 2114 | ciInstanceKlass* ik = base_k->as_instance_klass(); |
duke@435 | 2115 | // See if we can become precise: no subklasses and no interface |
duke@435 | 2116 | if (!ik->is_interface() && !ik->has_subklass()) { |
duke@435 | 2117 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
duke@435 | 2118 | // Add a dependence; if any subclass added we need to recompile |
duke@435 | 2119 | if (!ik->is_final()) { |
duke@435 | 2120 | phase->C->dependencies()->assert_leaf_type(ik); |
duke@435 | 2121 | } |
duke@435 | 2122 | // Return precise array klass |
duke@435 | 2123 | return TypeKlassPtr::make(ak); |
duke@435 | 2124 | } |
duke@435 | 2125 | } |
duke@435 | 2126 | return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); |
duke@435 | 2127 | } else { // Found a type-array? |
duke@435 | 2128 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
duke@435 | 2129 | assert( ak->is_type_array_klass(), "" ); |
duke@435 | 2130 | return TypeKlassPtr::make(ak); // These are always precise |
duke@435 | 2131 | } |
duke@435 | 2132 | } |
duke@435 | 2133 | } |
duke@435 | 2134 | |
duke@435 | 2135 | // Check for loading klass from an array klass |
duke@435 | 2136 | const TypeKlassPtr *tkls = tp->isa_klassptr(); |
duke@435 | 2137 | if (tkls != NULL && !StressReflectiveCode) { |
duke@435 | 2138 | ciKlass* klass = tkls->klass(); |
duke@435 | 2139 | if( !klass->is_loaded() ) |
duke@435 | 2140 | return _type; // Bail out if not loaded |
duke@435 | 2141 | if( klass->is_obj_array_klass() && |
coleenp@4142 | 2142 | tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { |
duke@435 | 2143 | ciKlass* elem = klass->as_obj_array_klass()->element_klass(); |
duke@435 | 2144 | // // Always returning precise element type is incorrect, |
duke@435 | 2145 | // // e.g., element type could be object and array may contain strings |
duke@435 | 2146 | // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); |
duke@435 | 2147 | |
duke@435 | 2148 | // The array's TypeKlassPtr was declared 'precise' or 'not precise' |
duke@435 | 2149 | // according to the element type's subclassing. |
duke@435 | 2150 | return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); |
duke@435 | 2151 | } |
duke@435 | 2152 | if( klass->is_instance_klass() && tkls->klass_is_exact() && |
stefank@3391 | 2153 | tkls->offset() == in_bytes(Klass::super_offset())) { |
duke@435 | 2154 | ciKlass* sup = klass->as_instance_klass()->super(); |
duke@435 | 2155 | // The field is Klass::_super. Return its (constant) value. |
duke@435 | 2156 | // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) |
duke@435 | 2157 | return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; |
duke@435 | 2158 | } |
duke@435 | 2159 | } |
duke@435 | 2160 | |
duke@435 | 2161 | // Bailout case |
duke@435 | 2162 | return LoadNode::Value(phase); |
duke@435 | 2163 | } |
duke@435 | 2164 | |
duke@435 | 2165 | //------------------------------Identity--------------------------------------- |
duke@435 | 2166 | // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. |
duke@435 | 2167 | // Also feed through the klass in Allocate(...klass...)._klass. |
duke@435 | 2168 | Node* LoadKlassNode::Identity( PhaseTransform *phase ) { |
kvn@599 | 2169 | return klass_identity_common(phase); |
kvn@599 | 2170 | } |
kvn@599 | 2171 | |
kvn@599 | 2172 | Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { |
duke@435 | 2173 | Node* x = LoadNode::Identity(phase); |
duke@435 | 2174 | if (x != this) return x; |
duke@435 | 2175 | |
duke@435 | 2176 | // Take apart the address into an oop and and offset. |
duke@435 | 2177 | // Return 'this' if we cannot. |
duke@435 | 2178 | Node* adr = in(MemNode::Address); |
duke@435 | 2179 | intptr_t offset = 0; |
duke@435 | 2180 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
duke@435 | 2181 | if (base == NULL) return this; |
duke@435 | 2182 | const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); |
duke@435 | 2183 | if (toop == NULL) return this; |
duke@435 | 2184 | |
duke@435 | 2185 | // We can fetch the klass directly through an AllocateNode. |
duke@435 | 2186 | // This works even if the klass is not constant (clone or newArray). |
duke@435 | 2187 | if (offset == oopDesc::klass_offset_in_bytes()) { |
duke@435 | 2188 | Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); |
duke@435 | 2189 | if (allocated_klass != NULL) { |
duke@435 | 2190 | return allocated_klass; |
duke@435 | 2191 | } |
duke@435 | 2192 | } |
duke@435 | 2193 | |
coleenp@4037 | 2194 | // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. |
coleenp@4142 | 2195 | // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass. |
duke@435 | 2196 | // See inline_native_Class_query for occurrences of these patterns. |
duke@435 | 2197 | // Java Example: x.getClass().isAssignableFrom(y) |
duke@435 | 2198 | // Java Example: Array.newInstance(x.getClass().getComponentType(), n) |
duke@435 | 2199 | // |
duke@435 | 2200 | // This improves reflective code, often making the Class |
duke@435 | 2201 | // mirror go completely dead. (Current exception: Class |
duke@435 | 2202 | // mirrors may appear in debug info, but we could clean them out by |
coleenp@4037 | 2203 | // introducing a new debug info operator for Klass*.java_mirror). |
duke@435 | 2204 | if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() |
duke@435 | 2205 | && (offset == java_lang_Class::klass_offset_in_bytes() || |
duke@435 | 2206 | offset == java_lang_Class::array_klass_offset_in_bytes())) { |
duke@435 | 2207 | // We are loading a special hidden field from a Class mirror, |
coleenp@4142 | 2208 | // the field which points to its Klass or ArrayKlass metaobject. |
duke@435 | 2209 | if (base->is_Load()) { |
duke@435 | 2210 | Node* adr2 = base->in(MemNode::Address); |
duke@435 | 2211 | const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); |
duke@435 | 2212 | if (tkls != NULL && !tkls->empty() |
duke@435 | 2213 | && (tkls->klass()->is_instance_klass() || |
duke@435 | 2214 | tkls->klass()->is_array_klass()) |
duke@435 | 2215 | && adr2->is_AddP() |
duke@435 | 2216 | ) { |
stefank@3391 | 2217 | int mirror_field = in_bytes(Klass::java_mirror_offset()); |
duke@435 | 2218 | if (offset == java_lang_Class::array_klass_offset_in_bytes()) { |
coleenp@4142 | 2219 | mirror_field = in_bytes(ArrayKlass::component_mirror_offset()); |
duke@435 | 2220 | } |
stefank@3391 | 2221 | if (tkls->offset() == mirror_field) { |
duke@435 | 2222 | return adr2->in(AddPNode::Base); |
duke@435 | 2223 | } |
duke@435 | 2224 | } |
duke@435 | 2225 | } |
duke@435 | 2226 | } |
duke@435 | 2227 | |
duke@435 | 2228 | return this; |
duke@435 | 2229 | } |
duke@435 | 2230 | |
kvn@599 | 2231 | |
kvn@599 | 2232 | //------------------------------Value------------------------------------------ |
kvn@599 | 2233 | const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { |
kvn@599 | 2234 | const Type *t = klass_value_common(phase); |
kvn@656 | 2235 | if (t == Type::TOP) |
kvn@656 | 2236 | return t; |
kvn@656 | 2237 | |
roland@4159 | 2238 | return t->make_narrowklass(); |
kvn@599 | 2239 | } |
kvn@599 | 2240 | |
kvn@599 | 2241 | //------------------------------Identity--------------------------------------- |
kvn@599 | 2242 | // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. |
kvn@599 | 2243 | // Also feed through the klass in Allocate(...klass...)._klass. |
kvn@599 | 2244 | Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { |
kvn@599 | 2245 | Node *x = klass_identity_common(phase); |
kvn@599 | 2246 | |
kvn@599 | 2247 | const Type *t = phase->type( x ); |
kvn@599 | 2248 | if( t == Type::TOP ) return x; |
roland@4159 | 2249 | if( t->isa_narrowklass()) return x; |
roland@4159 | 2250 | assert (!t->isa_narrowoop(), "no narrow oop here"); |
roland@4159 | 2251 | |
roland@4159 | 2252 | return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass())); |
kvn@599 | 2253 | } |
kvn@599 | 2254 | |
duke@435 | 2255 | //------------------------------Value----------------------------------------- |
duke@435 | 2256 | const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { |
duke@435 | 2257 | // Either input is TOP ==> the result is TOP |
duke@435 | 2258 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
duke@435 | 2259 | if( t1 == Type::TOP ) return Type::TOP; |
duke@435 | 2260 | Node *adr = in(MemNode::Address); |
duke@435 | 2261 | const Type *t2 = phase->type( adr ); |
duke@435 | 2262 | if( t2 == Type::TOP ) return Type::TOP; |
duke@435 | 2263 | const TypePtr *tp = t2->is_ptr(); |
duke@435 | 2264 | if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; |
duke@435 | 2265 | const TypeAryPtr *tap = tp->isa_aryptr(); |
duke@435 | 2266 | if( !tap ) return _type; |
duke@435 | 2267 | return tap->size(); |
duke@435 | 2268 | } |
duke@435 | 2269 | |
rasbold@801 | 2270 | //-------------------------------Ideal--------------------------------------- |
rasbold@801 | 2271 | // Feed through the length in AllocateArray(...length...)._length. |
rasbold@801 | 2272 | Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
rasbold@801 | 2273 | Node* p = MemNode::Ideal_common(phase, can_reshape); |
rasbold@801 | 2274 | if (p) return (p == NodeSentinel) ? NULL : p; |
rasbold@801 | 2275 | |
rasbold@801 | 2276 | // Take apart the address into an oop and and offset. |
rasbold@801 | 2277 | // Return 'this' if we cannot. |
rasbold@801 | 2278 | Node* adr = in(MemNode::Address); |
rasbold@801 | 2279 | intptr_t offset = 0; |
rasbold@801 | 2280 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
rasbold@801 | 2281 | if (base == NULL) return NULL; |
rasbold@801 | 2282 | const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); |
rasbold@801 | 2283 | if (tary == NULL) return NULL; |
rasbold@801 | 2284 | |
rasbold@801 | 2285 | // We can fetch the length directly through an AllocateArrayNode. |
rasbold@801 | 2286 | // This works even if the length is not constant (clone or newArray). |
rasbold@801 | 2287 | if (offset == arrayOopDesc::length_offset_in_bytes()) { |
rasbold@801 | 2288 | AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); |
rasbold@801 | 2289 | if (alloc != NULL) { |
rasbold@801 | 2290 | Node* allocated_length = alloc->Ideal_length(); |
rasbold@801 | 2291 | Node* len = alloc->make_ideal_length(tary, phase); |
rasbold@801 | 2292 | if (allocated_length != len) { |
rasbold@801 | 2293 | // New CastII improves on this. |
rasbold@801 | 2294 | return len; |
rasbold@801 | 2295 | } |
rasbold@801 | 2296 | } |
rasbold@801 | 2297 | } |
rasbold@801 | 2298 | |
rasbold@801 | 2299 | return NULL; |
rasbold@801 | 2300 | } |
rasbold@801 | 2301 | |
duke@435 | 2302 | //------------------------------Identity--------------------------------------- |
duke@435 | 2303 | // Feed through the length in AllocateArray(...length...)._length. |
duke@435 | 2304 | Node* LoadRangeNode::Identity( PhaseTransform *phase ) { |
duke@435 | 2305 | Node* x = LoadINode::Identity(phase); |
duke@435 | 2306 | if (x != this) return x; |
duke@435 | 2307 | |
duke@435 | 2308 | // Take apart the address into an oop and and offset. |
duke@435 | 2309 | // Return 'this' if we cannot. |
duke@435 | 2310 | Node* adr = in(MemNode::Address); |
duke@435 | 2311 | intptr_t offset = 0; |
duke@435 | 2312 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
duke@435 | 2313 | if (base == NULL) return this; |
duke@435 | 2314 | const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); |
duke@435 | 2315 | if (tary == NULL) return this; |
duke@435 | 2316 | |
duke@435 | 2317 | // We can fetch the length directly through an AllocateArrayNode. |
duke@435 | 2318 | // This works even if the length is not constant (clone or newArray). |
duke@435 | 2319 | if (offset == arrayOopDesc::length_offset_in_bytes()) { |
rasbold@801 | 2320 | AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); |
rasbold@801 | 2321 | if (alloc != NULL) { |
rasbold@801 | 2322 | Node* allocated_length = alloc->Ideal_length(); |
rasbold@801 | 2323 | // Do not allow make_ideal_length to allocate a CastII node. |
rasbold@801 | 2324 | Node* len = alloc->make_ideal_length(tary, phase, false); |
rasbold@801 | 2325 | if (allocated_length == len) { |
rasbold@801 | 2326 | // Return allocated_length only if it would not be improved by a CastII. |
rasbold@801 | 2327 | return allocated_length; |
rasbold@801 | 2328 | } |
duke@435 | 2329 | } |
duke@435 | 2330 | } |
duke@435 | 2331 | |
duke@435 | 2332 | return this; |
duke@435 | 2333 | |
duke@435 | 2334 | } |
rasbold@801 | 2335 | |
duke@435 | 2336 | //============================================================================= |
duke@435 | 2337 | //---------------------------StoreNode::make----------------------------------- |
duke@435 | 2338 | // Polymorphic factory method: |
goetz@6479 | 2339 | StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { |
goetz@6479 | 2340 | assert((mo == unordered || mo == release), "unexpected"); |
coleenp@548 | 2341 | Compile* C = gvn.C; |
goetz@6479 | 2342 | assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || |
goetz@6479 | 2343 | ctl != NULL, "raw memory operations should have control edge"); |
coleenp@548 | 2344 | |
duke@435 | 2345 | switch (bt) { |
duke@435 | 2346 | case T_BOOLEAN: |
goetz@6479 | 2347 | case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo); |
goetz@6479 | 2348 | case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo); |
duke@435 | 2349 | case T_CHAR: |
goetz@6479 | 2350 | case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo); |
goetz@6479 | 2351 | case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo); |
goetz@6479 | 2352 | case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo); |
goetz@6479 | 2353 | case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo); |
coleenp@4037 | 2354 | case T_METADATA: |
duke@435 | 2355 | case T_ADDRESS: |
coleenp@548 | 2356 | case T_OBJECT: |
coleenp@548 | 2357 | #ifdef _LP64 |
roland@4159 | 2358 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
kvn@4115 | 2359 | val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop())); |
goetz@6479 | 2360 | return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo); |
roland@4159 | 2361 | } else if (adr->bottom_type()->is_ptr_to_narrowklass() || |
ehelin@5694 | 2362 | (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && |
roland@4159 | 2363 | adr->bottom_type()->isa_rawptr())) { |
roland@4159 | 2364 | val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); |
goetz@6479 | 2365 | return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); |
roland@4159 | 2366 | } |
coleenp@548 | 2367 | #endif |
kvn@656 | 2368 | { |
goetz@6479 | 2369 | return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo); |
kvn@656 | 2370 | } |
duke@435 | 2371 | } |
duke@435 | 2372 | ShouldNotReachHere(); |
duke@435 | 2373 | return (StoreNode*)NULL; |
duke@435 | 2374 | } |
duke@435 | 2375 | |
goetz@6479 | 2376 | StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { |
duke@435 | 2377 | bool require_atomic = true; |
goetz@6479 | 2378 | return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); |
duke@435 | 2379 | } |
duke@435 | 2380 | |
duke@435 | 2381 | |
duke@435 | 2382 | //--------------------------bottom_type---------------------------------------- |
duke@435 | 2383 | const Type *StoreNode::bottom_type() const { |
duke@435 | 2384 | return Type::MEMORY; |
duke@435 | 2385 | } |
duke@435 | 2386 | |
duke@435 | 2387 | //------------------------------hash------------------------------------------- |
duke@435 | 2388 | uint StoreNode::hash() const { |
duke@435 | 2389 | // unroll addition of interesting fields |
duke@435 | 2390 | //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); |
duke@435 | 2391 | |
duke@435 | 2392 | // Since they are not commoned, do not hash them: |
duke@435 | 2393 | return NO_HASH; |
duke@435 | 2394 | } |
duke@435 | 2395 | |
duke@435 | 2396 | //------------------------------Ideal------------------------------------------ |
duke@435 | 2397 | // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). |
duke@435 | 2398 | // When a store immediately follows a relevant allocation/initialization, |
duke@435 | 2399 | // try to capture it into the initialization, or hoist it above. |
duke@435 | 2400 | Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 2401 | Node* p = MemNode::Ideal_common(phase, can_reshape); |
duke@435 | 2402 | if (p) return (p == NodeSentinel) ? NULL : p; |
duke@435 | 2403 | |
duke@435 | 2404 | Node* mem = in(MemNode::Memory); |
duke@435 | 2405 | Node* address = in(MemNode::Address); |
duke@435 | 2406 | |
never@2780 | 2407 | // Back-to-back stores to same address? Fold em up. Generally |
never@2780 | 2408 | // unsafe if I have intervening uses... Also disallowed for StoreCM |
never@2780 | 2409 | // since they must follow each StoreP operation. Redundant StoreCMs |
never@2780 | 2410 | // are eliminated just before matching in final_graph_reshape. |
kvn@3407 | 2411 | if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && |
never@2780 | 2412 | mem->Opcode() != Op_StoreCM) { |
duke@435 | 2413 | // Looking at a dead closed cycle of memory? |
duke@435 | 2414 | assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); |
duke@435 | 2415 | |
duke@435 | 2416 | assert(Opcode() == mem->Opcode() || |
duke@435 | 2417 | phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, |
duke@435 | 2418 | "no mismatched stores, except on raw memory"); |
duke@435 | 2419 | |
duke@435 | 2420 | if (mem->outcnt() == 1 && // check for intervening uses |
duke@435 | 2421 | mem->as_Store()->memory_size() <= this->memory_size()) { |
duke@435 | 2422 | // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. |
duke@435 | 2423 | // For example, 'mem' might be the final state at a conditional return. |
duke@435 | 2424 | // Or, 'mem' might be used by some node which is live at the same time |
duke@435 | 2425 | // 'this' is live, which might be unschedulable. So, require exactly |
duke@435 | 2426 | // ONE user, the 'this' store, until such time as we clone 'mem' for |
duke@435 | 2427 | // each of 'mem's uses (thus making the exactly-1-user-rule hold true). |
duke@435 | 2428 | if (can_reshape) { // (%%% is this an anachronism?) |
duke@435 | 2429 | set_req_X(MemNode::Memory, mem->in(MemNode::Memory), |
duke@435 | 2430 | phase->is_IterGVN()); |
duke@435 | 2431 | } else { |
duke@435 | 2432 | // It's OK to do this in the parser, since DU info is always accurate, |
duke@435 | 2433 | // and the parser always refers to nodes via SafePointNode maps. |
duke@435 | 2434 | set_req(MemNode::Memory, mem->in(MemNode::Memory)); |
duke@435 | 2435 | } |
duke@435 | 2436 | return this; |
duke@435 | 2437 | } |
duke@435 | 2438 | } |
duke@435 | 2439 | |
duke@435 | 2440 | // Capture an unaliased, unconditional, simple store into an initializer. |
duke@435 | 2441 | // Or, if it is independent of the allocation, hoist it above the allocation. |
duke@435 | 2442 | if (ReduceFieldZeroing && /*can_reshape &&*/ |
duke@435 | 2443 | mem->is_Proj() && mem->in(0)->is_Initialize()) { |
duke@435 | 2444 | InitializeNode* init = mem->in(0)->as_Initialize(); |
roland@4657 | 2445 | intptr_t offset = init->can_capture_store(this, phase, can_reshape); |
duke@435 | 2446 | if (offset > 0) { |
roland@4657 | 2447 | Node* moved = init->capture_store(this, offset, phase, can_reshape); |
duke@435 | 2448 | // If the InitializeNode captured me, it made a raw copy of me, |
duke@435 | 2449 | // and I need to disappear. |
duke@435 | 2450 | if (moved != NULL) { |
duke@435 | 2451 | // %%% hack to ensure that Ideal returns a new node: |
duke@435 | 2452 | mem = MergeMemNode::make(phase->C, mem); |
duke@435 | 2453 | return mem; // fold me away |
duke@435 | 2454 | } |
duke@435 | 2455 | } |
duke@435 | 2456 | } |
duke@435 | 2457 | |
duke@435 | 2458 | return NULL; // No further progress |
duke@435 | 2459 | } |
duke@435 | 2460 | |
duke@435 | 2461 | //------------------------------Value----------------------------------------- |
duke@435 | 2462 | const Type *StoreNode::Value( PhaseTransform *phase ) const { |
duke@435 | 2463 | // Either input is TOP ==> the result is TOP |
duke@435 | 2464 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
duke@435 | 2465 | if( t1 == Type::TOP ) return Type::TOP; |
duke@435 | 2466 | const Type *t2 = phase->type( in(MemNode::Address) ); |
duke@435 | 2467 | if( t2 == Type::TOP ) return Type::TOP; |
duke@435 | 2468 | const Type *t3 = phase->type( in(MemNode::ValueIn) ); |
duke@435 | 2469 | if( t3 == Type::TOP ) return Type::TOP; |
duke@435 | 2470 | return Type::MEMORY; |
duke@435 | 2471 | } |
duke@435 | 2472 | |
duke@435 | 2473 | //------------------------------Identity--------------------------------------- |
duke@435 | 2474 | // Remove redundant stores: |
duke@435 | 2475 | // Store(m, p, Load(m, p)) changes to m. |
duke@435 | 2476 | // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). |
duke@435 | 2477 | Node *StoreNode::Identity( PhaseTransform *phase ) { |
duke@435 | 2478 | Node* mem = in(MemNode::Memory); |
duke@435 | 2479 | Node* adr = in(MemNode::Address); |
duke@435 | 2480 | Node* val = in(MemNode::ValueIn); |
duke@435 | 2481 | |
duke@435 | 2482 | // Load then Store? Then the Store is useless |
duke@435 | 2483 | if (val->is_Load() && |
kvn@3407 | 2484 | val->in(MemNode::Address)->eqv_uncast(adr) && |
kvn@3407 | 2485 | val->in(MemNode::Memory )->eqv_uncast(mem) && |
duke@435 | 2486 | val->as_Load()->store_Opcode() == Opcode()) { |
duke@435 | 2487 | return mem; |
duke@435 | 2488 | } |
duke@435 | 2489 | |
duke@435 | 2490 | // Two stores in a row of the same value? |
duke@435 | 2491 | if (mem->is_Store() && |
kvn@3407 | 2492 | mem->in(MemNode::Address)->eqv_uncast(adr) && |
kvn@3407 | 2493 | mem->in(MemNode::ValueIn)->eqv_uncast(val) && |
duke@435 | 2494 | mem->Opcode() == Opcode()) { |
duke@435 | 2495 | return mem; |
duke@435 | 2496 | } |
duke@435 | 2497 | |
duke@435 | 2498 | // Store of zero anywhere into a freshly-allocated object? |
duke@435 | 2499 | // Then the store is useless. |
duke@435 | 2500 | // (It must already have been captured by the InitializeNode.) |
duke@435 | 2501 | if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { |
duke@435 | 2502 | // a newly allocated object is already all-zeroes everywhere |
duke@435 | 2503 | if (mem->is_Proj() && mem->in(0)->is_Allocate()) { |
duke@435 | 2504 | return mem; |
duke@435 | 2505 | } |
duke@435 | 2506 | |
duke@435 | 2507 | // the store may also apply to zero-bits in an earlier object |
duke@435 | 2508 | Node* prev_mem = find_previous_store(phase); |
duke@435 | 2509 | // Steps (a), (b): Walk past independent stores to find an exact match. |
duke@435 | 2510 | if (prev_mem != NULL) { |
duke@435 | 2511 | Node* prev_val = can_see_stored_value(prev_mem, phase); |
duke@435 | 2512 | if (prev_val != NULL && phase->eqv(prev_val, val)) { |
duke@435 | 2513 | // prev_val and val might differ by a cast; it would be good |
duke@435 | 2514 | // to keep the more informative of the two. |
duke@435 | 2515 | return mem; |
duke@435 | 2516 | } |
duke@435 | 2517 | } |
duke@435 | 2518 | } |
duke@435 | 2519 | |
duke@435 | 2520 | return this; |
duke@435 | 2521 | } |
duke@435 | 2522 | |
duke@435 | 2523 | //------------------------------match_edge------------------------------------- |
duke@435 | 2524 | // Do we Match on this edge index or not? Match only memory & value |
duke@435 | 2525 | uint StoreNode::match_edge(uint idx) const { |
duke@435 | 2526 | return idx == MemNode::Address || idx == MemNode::ValueIn; |
duke@435 | 2527 | } |
duke@435 | 2528 | |
duke@435 | 2529 | //------------------------------cmp-------------------------------------------- |
duke@435 | 2530 | // Do not common stores up together. They generally have to be split |
duke@435 | 2531 | // back up anyways, so do not bother. |
duke@435 | 2532 | uint StoreNode::cmp( const Node &n ) const { |
duke@435 | 2533 | return (&n == this); // Always fail except on self |
duke@435 | 2534 | } |
duke@435 | 2535 | |
duke@435 | 2536 | //------------------------------Ideal_masked_input----------------------------- |
duke@435 | 2537 | // Check for a useless mask before a partial-word store |
duke@435 | 2538 | // (StoreB ... (AndI valIn conIa) ) |
duke@435 | 2539 | // If (conIa & mask == mask) this simplifies to |
duke@435 | 2540 | // (StoreB ... (valIn) ) |
duke@435 | 2541 | Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { |
duke@435 | 2542 | Node *val = in(MemNode::ValueIn); |
duke@435 | 2543 | if( val->Opcode() == Op_AndI ) { |
duke@435 | 2544 | const TypeInt *t = phase->type( val->in(2) )->isa_int(); |
duke@435 | 2545 | if( t && t->is_con() && (t->get_con() & mask) == mask ) { |
duke@435 | 2546 | set_req(MemNode::ValueIn, val->in(1)); |
duke@435 | 2547 | return this; |
duke@435 | 2548 | } |
duke@435 | 2549 | } |
duke@435 | 2550 | return NULL; |
duke@435 | 2551 | } |
duke@435 | 2552 | |
duke@435 | 2553 | |
duke@435 | 2554 | //------------------------------Ideal_sign_extended_input---------------------- |
duke@435 | 2555 | // Check for useless sign-extension before a partial-word store |
duke@435 | 2556 | // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) |
duke@435 | 2557 | // If (conIL == conIR && conIR <= num_bits) this simplifies to |
duke@435 | 2558 | // (StoreB ... (valIn) ) |
duke@435 | 2559 | Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { |
duke@435 | 2560 | Node *val = in(MemNode::ValueIn); |
duke@435 | 2561 | if( val->Opcode() == Op_RShiftI ) { |
duke@435 | 2562 | const TypeInt *t = phase->type( val->in(2) )->isa_int(); |
duke@435 | 2563 | if( t && t->is_con() && (t->get_con() <= num_bits) ) { |
duke@435 | 2564 | Node *shl = val->in(1); |
duke@435 | 2565 | if( shl->Opcode() == Op_LShiftI ) { |
duke@435 | 2566 | const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); |
duke@435 | 2567 | if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { |
duke@435 | 2568 | set_req(MemNode::ValueIn, shl->in(1)); |
duke@435 | 2569 | return this; |
duke@435 | 2570 | } |
duke@435 | 2571 | } |
duke@435 | 2572 | } |
duke@435 | 2573 | } |
duke@435 | 2574 | return NULL; |
duke@435 | 2575 | } |
duke@435 | 2576 | |
duke@435 | 2577 | //------------------------------value_never_loaded----------------------------------- |
duke@435 | 2578 | // Determine whether there are any possible loads of the value stored. |
duke@435 | 2579 | // For simplicity, we actually check if there are any loads from the |
duke@435 | 2580 | // address stored to, not just for loads of the value stored by this node. |
duke@435 | 2581 | // |
duke@435 | 2582 | bool StoreNode::value_never_loaded( PhaseTransform *phase) const { |
duke@435 | 2583 | Node *adr = in(Address); |
duke@435 | 2584 | const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); |
duke@435 | 2585 | if (adr_oop == NULL) |
duke@435 | 2586 | return false; |
kvn@658 | 2587 | if (!adr_oop->is_known_instance_field()) |
duke@435 | 2588 | return false; // if not a distinct instance, there may be aliases of the address |
duke@435 | 2589 | for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { |
duke@435 | 2590 | Node *use = adr->fast_out(i); |
duke@435 | 2591 | int opc = use->Opcode(); |
duke@435 | 2592 | if (use->is_Load() || use->is_LoadStore()) { |
duke@435 | 2593 | return false; |
duke@435 | 2594 | } |
duke@435 | 2595 | } |
duke@435 | 2596 | return true; |
duke@435 | 2597 | } |
duke@435 | 2598 | |
duke@435 | 2599 | //============================================================================= |
duke@435 | 2600 | //------------------------------Ideal------------------------------------------ |
duke@435 | 2601 | // If the store is from an AND mask that leaves the low bits untouched, then |
duke@435 | 2602 | // we can skip the AND operation. If the store is from a sign-extension |
duke@435 | 2603 | // (a left shift, then right shift) we can skip both. |
duke@435 | 2604 | Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
duke@435 | 2605 | Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); |
duke@435 | 2606 | if( progress != NULL ) return progress; |
duke@435 | 2607 | |
duke@435 | 2608 | progress = StoreNode::Ideal_sign_extended_input(phase, 24); |
duke@435 | 2609 | if( progress != NULL ) return progress; |
duke@435 | 2610 | |
duke@435 | 2611 | // Finally check the default case |
duke@435 | 2612 | return StoreNode::Ideal(phase, can_reshape); |
duke@435 | 2613 | } |
duke@435 | 2614 | |
duke@435 | 2615 | //============================================================================= |
duke@435 | 2616 | //------------------------------Ideal------------------------------------------ |
duke@435 | 2617 | // If the store is from an AND mask that leaves the low bits untouched, then |
duke@435 | 2618 | // we can skip the AND operation |
duke@435 | 2619 | Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
duke@435 | 2620 | Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); |
duke@435 | 2621 | if( progress != NULL ) return progress; |
duke@435 | 2622 | |
duke@435 | 2623 | progress = StoreNode::Ideal_sign_extended_input(phase, 16); |
duke@435 | 2624 | if( progress != NULL ) return progress; |
duke@435 | 2625 | |
duke@435 | 2626 | // Finally check the default case |
duke@435 | 2627 | return StoreNode::Ideal(phase, can_reshape); |
duke@435 | 2628 | } |
duke@435 | 2629 | |
duke@435 | 2630 | //============================================================================= |
duke@435 | 2631 | //------------------------------Identity--------------------------------------- |
duke@435 | 2632 | Node *StoreCMNode::Identity( PhaseTransform *phase ) { |
duke@435 | 2633 | // No need to card mark when storing a null ptr |
duke@435 | 2634 | Node* my_store = in(MemNode::OopStore); |
duke@435 | 2635 | if (my_store->is_Store()) { |
duke@435 | 2636 | const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); |
duke@435 | 2637 | if( t1 == TypePtr::NULL_PTR ) { |
duke@435 | 2638 | return in(MemNode::Memory); |
duke@435 | 2639 | } |
duke@435 | 2640 | } |
duke@435 | 2641 | return this; |
duke@435 | 2642 | } |
duke@435 | 2643 | |
cfang@1420 | 2644 | //============================================================================= |
cfang@1420 | 2645 | //------------------------------Ideal--------------------------------------- |
cfang@1420 | 2646 | Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
cfang@1420 | 2647 | Node* progress = StoreNode::Ideal(phase, can_reshape); |
cfang@1420 | 2648 | if (progress != NULL) return progress; |
cfang@1420 | 2649 | |
cfang@1420 | 2650 | Node* my_store = in(MemNode::OopStore); |
cfang@1420 | 2651 | if (my_store->is_MergeMem()) { |
cfang@1420 | 2652 | Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); |
cfang@1420 | 2653 | set_req(MemNode::OopStore, mem); |
cfang@1420 | 2654 | return this; |
cfang@1420 | 2655 | } |
cfang@1420 | 2656 | |
cfang@1420 | 2657 | return NULL; |
cfang@1420 | 2658 | } |
cfang@1420 | 2659 | |
duke@435 | 2660 | //------------------------------Value----------------------------------------- |
duke@435 | 2661 | const Type *StoreCMNode::Value( PhaseTransform *phase ) const { |
kvn@478 | 2662 | // Either input is TOP ==> the result is TOP |
kvn@478 | 2663 | const Type *t = phase->type( in(MemNode::Memory) ); |
kvn@478 | 2664 | if( t == Type::TOP ) return Type::TOP; |
kvn@478 | 2665 | t = phase->type( in(MemNode::Address) ); |
kvn@478 | 2666 | if( t == Type::TOP ) return Type::TOP; |
kvn@478 | 2667 | t = phase->type( in(MemNode::ValueIn) ); |
kvn@478 | 2668 | if( t == Type::TOP ) return Type::TOP; |
duke@435 | 2669 | // If extra input is TOP ==> the result is TOP |
kvn@478 | 2670 | t = phase->type( in(MemNode::OopStore) ); |
kvn@478 | 2671 | if( t == Type::TOP ) return Type::TOP; |
duke@435 | 2672 | |
duke@435 | 2673 | return StoreNode::Value( phase ); |
duke@435 | 2674 | } |
duke@435 | 2675 | |
duke@435 | 2676 | |
duke@435 | 2677 | //============================================================================= |
duke@435 | 2678 | //----------------------------------SCMemProjNode------------------------------ |
duke@435 | 2679 | const Type * SCMemProjNode::Value( PhaseTransform *phase ) const |
duke@435 | 2680 | { |
duke@435 | 2681 | return bottom_type(); |
duke@435 | 2682 | } |
duke@435 | 2683 | |
duke@435 | 2684 | //============================================================================= |
roland@4106 | 2685 | //----------------------------------LoadStoreNode------------------------------ |
roland@4106 | 2686 | LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) |
roland@4106 | 2687 | : Node(required), |
roland@4106 | 2688 | _type(rt), |
roland@4106 | 2689 | _adr_type(at) |
roland@4106 | 2690 | { |
duke@435 | 2691 | init_req(MemNode::Control, c ); |
duke@435 | 2692 | init_req(MemNode::Memory , mem); |
duke@435 | 2693 | init_req(MemNode::Address, adr); |
duke@435 | 2694 | init_req(MemNode::ValueIn, val); |
duke@435 | 2695 | init_class_id(Class_LoadStore); |
roland@4106 | 2696 | } |
roland@4106 | 2697 | |
roland@4106 | 2698 | uint LoadStoreNode::ideal_reg() const { |
roland@4106 | 2699 | return _type->ideal_reg(); |
roland@4106 | 2700 | } |
roland@4106 | 2701 | |
roland@4106 | 2702 | bool LoadStoreNode::result_not_used() const { |
roland@4106 | 2703 | for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { |
roland@4106 | 2704 | Node *x = fast_out(i); |
roland@4106 | 2705 | if (x->Opcode() == Op_SCMemProj) continue; |
roland@4106 | 2706 | return false; |
roland@4106 | 2707 | } |
roland@4106 | 2708 | return true; |
roland@4106 | 2709 | } |
roland@4106 | 2710 | |
roland@4106 | 2711 | uint LoadStoreNode::size_of() const { return sizeof(*this); } |
roland@4106 | 2712 | |
roland@4106 | 2713 | //============================================================================= |
roland@4106 | 2714 | //----------------------------------LoadStoreConditionalNode-------------------- |
roland@4106 | 2715 | LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { |
roland@4106 | 2716 | init_req(ExpectedIn, ex ); |
duke@435 | 2717 | } |
duke@435 | 2718 | |
duke@435 | 2719 | //============================================================================= |
duke@435 | 2720 | //-------------------------------adr_type-------------------------------------- |
duke@435 | 2721 | // Do we Match on this edge index or not? Do not match memory |
duke@435 | 2722 | const TypePtr* ClearArrayNode::adr_type() const { |
duke@435 | 2723 | Node *adr = in(3); |
duke@435 | 2724 | return MemNode::calculate_adr_type(adr->bottom_type()); |
duke@435 | 2725 | } |
duke@435 | 2726 | |
duke@435 | 2727 | //------------------------------match_edge------------------------------------- |
duke@435 | 2728 | // Do we Match on this edge index or not? Do not match memory |
duke@435 | 2729 | uint ClearArrayNode::match_edge(uint idx) const { |
duke@435 | 2730 | return idx > 1; |
duke@435 | 2731 | } |
duke@435 | 2732 | |
duke@435 | 2733 | //------------------------------Identity--------------------------------------- |
duke@435 | 2734 | // Clearing a zero length array does nothing |
duke@435 | 2735 | Node *ClearArrayNode::Identity( PhaseTransform *phase ) { |
never@503 | 2736 | return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; |
duke@435 | 2737 | } |
duke@435 | 2738 | |
duke@435 | 2739 | //------------------------------Idealize--------------------------------------- |
duke@435 | 2740 | // Clearing a short array is faster with stores |
duke@435 | 2741 | Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
duke@435 | 2742 | const int unit = BytesPerLong; |
duke@435 | 2743 | const TypeX* t = phase->type(in(2))->isa_intptr_t(); |
duke@435 | 2744 | if (!t) return NULL; |
duke@435 | 2745 | if (!t->is_con()) return NULL; |
duke@435 | 2746 | intptr_t raw_count = t->get_con(); |
duke@435 | 2747 | intptr_t size = raw_count; |
duke@435 | 2748 | if (!Matcher::init_array_count_is_in_bytes) size *= unit; |
duke@435 | 2749 | // Clearing nothing uses the Identity call. |
duke@435 | 2750 | // Negative clears are possible on dead ClearArrays |
duke@435 | 2751 | // (see jck test stmt114.stmt11402.val). |
duke@435 | 2752 | if (size <= 0 || size % unit != 0) return NULL; |
duke@435 | 2753 | intptr_t count = size / unit; |
duke@435 | 2754 | // Length too long; use fast hardware clear |
duke@435 | 2755 | if (size > Matcher::init_array_short_size) return NULL; |
duke@435 | 2756 | Node *mem = in(1); |
duke@435 | 2757 | if( phase->type(mem)==Type::TOP ) return NULL; |
duke@435 | 2758 | Node *adr = in(3); |
duke@435 | 2759 | const Type* at = phase->type(adr); |
duke@435 | 2760 | if( at==Type::TOP ) return NULL; |
duke@435 | 2761 | const TypePtr* atp = at->isa_ptr(); |
duke@435 | 2762 | // adjust atp to be the correct array element address type |
duke@435 | 2763 | if (atp == NULL) atp = TypePtr::BOTTOM; |
duke@435 | 2764 | else atp = atp->add_offset(Type::OffsetBot); |
duke@435 | 2765 | // Get base for derived pointer purposes |
duke@435 | 2766 | if( adr->Opcode() != Op_AddP ) Unimplemented(); |
duke@435 | 2767 | Node *base = adr->in(1); |
duke@435 | 2768 | |
duke@435 | 2769 | Node *zero = phase->makecon(TypeLong::ZERO); |
duke@435 | 2770 | Node *off = phase->MakeConX(BytesPerLong); |
goetz@6479 | 2771 | mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); |
duke@435 | 2772 | count--; |
duke@435 | 2773 | while( count-- ) { |
duke@435 | 2774 | mem = phase->transform(mem); |
kvn@4115 | 2775 | adr = phase->transform(new (phase->C) AddPNode(base,adr,off)); |
goetz@6479 | 2776 | mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); |
duke@435 | 2777 | } |
duke@435 | 2778 | return mem; |
duke@435 | 2779 | } |
duke@435 | 2780 | |
kvn@1535 | 2781 | //----------------------------step_through---------------------------------- |
kvn@1535 | 2782 | // Return allocation input memory edge if it is different instance |
kvn@1535 | 2783 | // or itself if it is the one we are looking for. |
kvn@1535 | 2784 | bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { |
kvn@1535 | 2785 | Node* n = *np; |
kvn@1535 | 2786 | assert(n->is_ClearArray(), "sanity"); |
kvn@1535 | 2787 | intptr_t offset; |
kvn@1535 | 2788 | AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); |
kvn@1535 | 2789 | // This method is called only before Allocate nodes are expanded during |
kvn@1535 | 2790 | // macro nodes expansion. Before that ClearArray nodes are only generated |
kvn@1535 | 2791 | // in LibraryCallKit::generate_arraycopy() which follows allocations. |
kvn@1535 | 2792 | assert(alloc != NULL, "should have allocation"); |
kvn@1535 | 2793 | if (alloc->_idx == instance_id) { |
kvn@1535 | 2794 | // Can not bypass initialization of the instance we are looking for. |
kvn@1535 | 2795 | return false; |
kvn@1535 | 2796 | } |
kvn@1535 | 2797 | // Otherwise skip it. |
kvn@1535 | 2798 | InitializeNode* init = alloc->initialization(); |
kvn@1535 | 2799 | if (init != NULL) |
kvn@1535 | 2800 | *np = init->in(TypeFunc::Memory); |
kvn@1535 | 2801 | else |
kvn@1535 | 2802 | *np = alloc->in(TypeFunc::Memory); |
kvn@1535 | 2803 | return true; |
kvn@1535 | 2804 | } |
kvn@1535 | 2805 | |
duke@435 | 2806 | //----------------------------clear_memory------------------------------------- |
duke@435 | 2807 | // Generate code to initialize object storage to zero. |
duke@435 | 2808 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
duke@435 | 2809 | intptr_t start_offset, |
duke@435 | 2810 | Node* end_offset, |
duke@435 | 2811 | PhaseGVN* phase) { |
duke@435 | 2812 | Compile* C = phase->C; |
duke@435 | 2813 | intptr_t offset = start_offset; |
duke@435 | 2814 | |
duke@435 | 2815 | int unit = BytesPerLong; |
duke@435 | 2816 | if ((offset % unit) != 0) { |
kvn@4115 | 2817 | Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset)); |
duke@435 | 2818 | adr = phase->transform(adr); |
duke@435 | 2819 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
goetz@6479 | 2820 | mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); |
duke@435 | 2821 | mem = phase->transform(mem); |
duke@435 | 2822 | offset += BytesPerInt; |
duke@435 | 2823 | } |
duke@435 | 2824 | assert((offset % unit) == 0, ""); |
duke@435 | 2825 | |
duke@435 | 2826 | // Initialize the remaining stuff, if any, with a ClearArray. |
duke@435 | 2827 | return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); |
duke@435 | 2828 | } |
duke@435 | 2829 | |
duke@435 | 2830 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
duke@435 | 2831 | Node* start_offset, |
duke@435 | 2832 | Node* end_offset, |
duke@435 | 2833 | PhaseGVN* phase) { |
never@503 | 2834 | if (start_offset == end_offset) { |
never@503 | 2835 | // nothing to do |
never@503 | 2836 | return mem; |
never@503 | 2837 | } |
never@503 | 2838 | |
duke@435 | 2839 | Compile* C = phase->C; |
duke@435 | 2840 | int unit = BytesPerLong; |
duke@435 | 2841 | Node* zbase = start_offset; |
duke@435 | 2842 | Node* zend = end_offset; |
duke@435 | 2843 | |
duke@435 | 2844 | // Scale to the unit required by the CPU: |
duke@435 | 2845 | if (!Matcher::init_array_count_is_in_bytes) { |
duke@435 | 2846 | Node* shift = phase->intcon(exact_log2(unit)); |
kvn@4115 | 2847 | zbase = phase->transform( new(C) URShiftXNode(zbase, shift) ); |
kvn@4115 | 2848 | zend = phase->transform( new(C) URShiftXNode(zend, shift) ); |
duke@435 | 2849 | } |
duke@435 | 2850 | |
kvn@4410 | 2851 | // Bulk clear double-words |
kvn@4115 | 2852 | Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) ); |
kvn@4115 | 2853 | Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) ); |
kvn@4115 | 2854 | mem = new (C) ClearArrayNode(ctl, mem, zsize, adr); |
duke@435 | 2855 | return phase->transform(mem); |
duke@435 | 2856 | } |
duke@435 | 2857 | |
duke@435 | 2858 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
duke@435 | 2859 | intptr_t start_offset, |
duke@435 | 2860 | intptr_t end_offset, |
duke@435 | 2861 | PhaseGVN* phase) { |
never@503 | 2862 | if (start_offset == end_offset) { |
never@503 | 2863 | // nothing to do |
never@503 | 2864 | return mem; |
never@503 | 2865 | } |
never@503 | 2866 | |
duke@435 | 2867 | Compile* C = phase->C; |
duke@435 | 2868 | assert((end_offset % BytesPerInt) == 0, "odd end offset"); |
duke@435 | 2869 | intptr_t done_offset = end_offset; |
duke@435 | 2870 | if ((done_offset % BytesPerLong) != 0) { |
duke@435 | 2871 | done_offset -= BytesPerInt; |
duke@435 | 2872 | } |
duke@435 | 2873 | if (done_offset > start_offset) { |
duke@435 | 2874 | mem = clear_memory(ctl, mem, dest, |
duke@435 | 2875 | start_offset, phase->MakeConX(done_offset), phase); |
duke@435 | 2876 | } |
duke@435 | 2877 | if (done_offset < end_offset) { // emit the final 32-bit store |
kvn@4115 | 2878 | Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset)); |
duke@435 | 2879 | adr = phase->transform(adr); |
duke@435 | 2880 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
goetz@6479 | 2881 | mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); |
duke@435 | 2882 | mem = phase->transform(mem); |
duke@435 | 2883 | done_offset += BytesPerInt; |
duke@435 | 2884 | } |
duke@435 | 2885 | assert(done_offset == end_offset, ""); |
duke@435 | 2886 | return mem; |
duke@435 | 2887 | } |
duke@435 | 2888 | |
duke@435 | 2889 | //============================================================================= |
kvn@2694 | 2890 | // Do not match memory edge. |
kvn@2694 | 2891 | uint StrIntrinsicNode::match_edge(uint idx) const { |
kvn@2694 | 2892 | return idx == 2 || idx == 3; |
duke@435 | 2893 | } |
duke@435 | 2894 | |
duke@435 | 2895 | //------------------------------Ideal------------------------------------------ |
duke@435 | 2896 | // Return a node which is more "ideal" than the current node. Strip out |
duke@435 | 2897 | // control copies |
kvn@2694 | 2898 | Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
kvn@2694 | 2899 | if (remove_dead_region(phase, can_reshape)) return this; |
kvn@3311 | 2900 | // Don't bother trying to transform a dead node |
kvn@3311 | 2901 | if (in(0) && in(0)->is_top()) return NULL; |
kvn@2694 | 2902 | |
kvn@2699 | 2903 | if (can_reshape) { |
kvn@2699 | 2904 | Node* mem = phase->transform(in(MemNode::Memory)); |
kvn@2699 | 2905 | // If transformed to a MergeMem, get the desired slice |
kvn@2699 | 2906 | uint alias_idx = phase->C->get_alias_index(adr_type()); |
kvn@2699 | 2907 | mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; |
kvn@2699 | 2908 | if (mem != in(MemNode::Memory)) { |
kvn@2699 | 2909 | set_req(MemNode::Memory, mem); |
kvn@2699 | 2910 | return this; |
kvn@2699 | 2911 | } |
kvn@2699 | 2912 | } |
kvn@2694 | 2913 | return NULL; |
rasbold@604 | 2914 | } |
rasbold@604 | 2915 | |
kvn@3311 | 2916 | //------------------------------Value------------------------------------------ |
kvn@3311 | 2917 | const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { |
kvn@3311 | 2918 | if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; |
kvn@3311 | 2919 | return bottom_type(); |
kvn@3311 | 2920 | } |
kvn@3311 | 2921 | |
duke@435 | 2922 | //============================================================================= |
kvn@4479 | 2923 | //------------------------------match_edge------------------------------------- |
kvn@4479 | 2924 | // Do not match memory edge |
kvn@4479 | 2925 | uint EncodeISOArrayNode::match_edge(uint idx) const { |
kvn@4479 | 2926 | return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len) |
kvn@4479 | 2927 | } |
kvn@4479 | 2928 | |
kvn@4479 | 2929 | //------------------------------Ideal------------------------------------------ |
kvn@4479 | 2930 | // Return a node which is more "ideal" than the current node. Strip out |
kvn@4479 | 2931 | // control copies |
kvn@4479 | 2932 | Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
kvn@4479 | 2933 | return remove_dead_region(phase, can_reshape) ? this : NULL; |
kvn@4479 | 2934 | } |
kvn@4479 | 2935 | |
kvn@4479 | 2936 | //------------------------------Value------------------------------------------ |
kvn@4479 | 2937 | const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const { |
kvn@4479 | 2938 | if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; |
kvn@4479 | 2939 | return bottom_type(); |
kvn@4479 | 2940 | } |
kvn@4479 | 2941 | |
kvn@4479 | 2942 | //============================================================================= |
duke@435 | 2943 | MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) |
duke@435 | 2944 | : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), |
duke@435 | 2945 | _adr_type(C->get_adr_type(alias_idx)) |
duke@435 | 2946 | { |
duke@435 | 2947 | init_class_id(Class_MemBar); |
duke@435 | 2948 | Node* top = C->top(); |
duke@435 | 2949 | init_req(TypeFunc::I_O,top); |
duke@435 | 2950 | init_req(TypeFunc::FramePtr,top); |
duke@435 | 2951 | init_req(TypeFunc::ReturnAdr,top); |
duke@435 | 2952 | if (precedent != NULL) |
duke@435 | 2953 | init_req(TypeFunc::Parms, precedent); |
duke@435 | 2954 | } |
duke@435 | 2955 | |
duke@435 | 2956 | //------------------------------cmp-------------------------------------------- |
duke@435 | 2957 | uint MemBarNode::hash() const { return NO_HASH; } |
duke@435 | 2958 | uint MemBarNode::cmp( const Node &n ) const { |
duke@435 | 2959 | return (&n == this); // Always fail except on self |
duke@435 | 2960 | } |
duke@435 | 2961 | |
duke@435 | 2962 | //------------------------------make------------------------------------------- |
duke@435 | 2963 | MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { |
duke@435 | 2964 | switch (opcode) { |
goetz@6489 | 2965 | case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn); |
goetz@6489 | 2966 | case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn); |
goetz@6489 | 2967 | case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn); |
goetz@6489 | 2968 | case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn); |
goetz@6489 | 2969 | case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn); |
goetz@6489 | 2970 | case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn); |
goetz@6489 | 2971 | case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn); |
goetz@6489 | 2972 | case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn); |
goetz@6489 | 2973 | case Op_Initialize: return new(C) InitializeNode(C, atp, pn); |
goetz@6489 | 2974 | case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn); |
goetz@6489 | 2975 | default: ShouldNotReachHere(); return NULL; |
duke@435 | 2976 | } |
duke@435 | 2977 | } |
duke@435 | 2978 | |
duke@435 | 2979 | //------------------------------Ideal------------------------------------------ |
duke@435 | 2980 | // Return a node which is more "ideal" than the current node. Strip out |
duke@435 | 2981 | // control copies |
duke@435 | 2982 | Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
kvn@1535 | 2983 | if (remove_dead_region(phase, can_reshape)) return this; |
kvn@3311 | 2984 | // Don't bother trying to transform a dead node |
adlertz@5291 | 2985 | if (in(0) && in(0)->is_top()) { |
adlertz@5291 | 2986 | return NULL; |
adlertz@5291 | 2987 | } |
kvn@1535 | 2988 | |
kvn@1535 | 2989 | // Eliminate volatile MemBars for scalar replaced objects. |
kvn@5110 | 2990 | if (can_reshape && req() == (Precedent+1)) { |
kvn@5110 | 2991 | bool eliminate = false; |
kvn@5110 | 2992 | int opc = Opcode(); |
kvn@5110 | 2993 | if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { |
kvn@5110 | 2994 | // Volatile field loads and stores. |
kvn@5110 | 2995 | Node* my_mem = in(MemBarNode::Precedent); |
adlertz@5291 | 2996 | // The MembarAquire may keep an unused LoadNode alive through the Precedent edge |
adlertz@5291 | 2997 | if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { |
adlertz@5317 | 2998 | // if the Precedent is a decodeN and its input (a Load) is used at more than one place, |
adlertz@5317 | 2999 | // replace this Precedent (decodeN) with the Load instead. |
adlertz@5317 | 3000 | if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { |
adlertz@5317 | 3001 | Node* load_node = my_mem->in(1); |
adlertz@5317 | 3002 | set_req(MemBarNode::Precedent, load_node); |
adlertz@5317 | 3003 | phase->is_IterGVN()->_worklist.push(my_mem); |
adlertz@5317 | 3004 | my_mem = load_node; |
adlertz@5317 | 3005 | } else { |
adlertz@5317 | 3006 | assert(my_mem->unique_out() == this, "sanity"); |
adlertz@5317 | 3007 | del_req(Precedent); |
adlertz@5317 | 3008 | phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later |
adlertz@5317 | 3009 | my_mem = NULL; |
adlertz@5317 | 3010 | } |
adlertz@5291 | 3011 | } |
kvn@5110 | 3012 | if (my_mem != NULL && my_mem->is_Mem()) { |
kvn@5110 | 3013 | const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); |
kvn@5110 | 3014 | // Check for scalar replaced object reference. |
kvn@5110 | 3015 | if( t_oop != NULL && t_oop->is_known_instance_field() && |
kvn@5110 | 3016 | t_oop->offset() != Type::OffsetBot && |
kvn@5110 | 3017 | t_oop->offset() != Type::OffsetTop) { |
kvn@5110 | 3018 | eliminate = true; |
kvn@5110 | 3019 | } |
kvn@1535 | 3020 | } |
kvn@5110 | 3021 | } else if (opc == Op_MemBarRelease) { |
kvn@5110 | 3022 | // Final field stores. |
kvn@5110 | 3023 | Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); |
kvn@5110 | 3024 | if ((alloc != NULL) && alloc->is_Allocate() && |
kvn@5110 | 3025 | alloc->as_Allocate()->_is_non_escaping) { |
kvn@5110 | 3026 | // The allocated object does not escape. |
kvn@5110 | 3027 | eliminate = true; |
kvn@5110 | 3028 | } |
kvn@5110 | 3029 | } |
kvn@5110 | 3030 | if (eliminate) { |
kvn@5110 | 3031 | // Replace MemBar projections by its inputs. |
kvn@5110 | 3032 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
kvn@5110 | 3033 | igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); |
kvn@5110 | 3034 | igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); |
kvn@5110 | 3035 | // Must return either the original node (now dead) or a new node |
kvn@5110 | 3036 | // (Do not return a top here, since that would break the uniqueness of top.) |
kvn@5110 | 3037 | return new (phase->C) ConINode(TypeInt::ZERO); |
kvn@1535 | 3038 | } |
kvn@1535 | 3039 | } |
kvn@1535 | 3040 | return NULL; |
duke@435 | 3041 | } |
duke@435 | 3042 | |
duke@435 | 3043 | //------------------------------Value------------------------------------------ |
duke@435 | 3044 | const Type *MemBarNode::Value( PhaseTransform *phase ) const { |
duke@435 | 3045 | if( !in(0) ) return Type::TOP; |
duke@435 | 3046 | if( phase->type(in(0)) == Type::TOP ) |
duke@435 | 3047 | return Type::TOP; |
duke@435 | 3048 | return TypeTuple::MEMBAR; |
duke@435 | 3049 | } |
duke@435 | 3050 | |
duke@435 | 3051 | //------------------------------match------------------------------------------ |
duke@435 | 3052 | // Construct projections for memory. |
duke@435 | 3053 | Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { |
duke@435 | 3054 | switch (proj->_con) { |
duke@435 | 3055 | case TypeFunc::Control: |
duke@435 | 3056 | case TypeFunc::Memory: |
kvn@4115 | 3057 | return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); |
duke@435 | 3058 | } |
duke@435 | 3059 | ShouldNotReachHere(); |
duke@435 | 3060 | return NULL; |
duke@435 | 3061 | } |
duke@435 | 3062 | |
duke@435 | 3063 | //===========================InitializeNode==================================== |
duke@435 | 3064 | // SUMMARY: |
duke@435 | 3065 | // This node acts as a memory barrier on raw memory, after some raw stores. |
duke@435 | 3066 | // The 'cooked' oop value feeds from the Initialize, not the Allocation. |
duke@435 | 3067 | // The Initialize can 'capture' suitably constrained stores as raw inits. |
duke@435 | 3068 | // It can coalesce related raw stores into larger units (called 'tiles'). |
duke@435 | 3069 | // It can avoid zeroing new storage for memory units which have raw inits. |
duke@435 | 3070 | // At macro-expansion, it is marked 'complete', and does not optimize further. |
duke@435 | 3071 | // |
duke@435 | 3072 | // EXAMPLE: |
duke@435 | 3073 | // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. |
duke@435 | 3074 | // ctl = incoming control; mem* = incoming memory |
duke@435 | 3075 | // (Note: A star * on a memory edge denotes I/O and other standard edges.) |
duke@435 | 3076 | // First allocate uninitialized memory and fill in the header: |
duke@435 | 3077 | // alloc = (Allocate ctl mem* 16 #short[].klass ...) |
duke@435 | 3078 | // ctl := alloc.Control; mem* := alloc.Memory* |
duke@435 | 3079 | // rawmem = alloc.Memory; rawoop = alloc.RawAddress |
duke@435 | 3080 | // Then initialize to zero the non-header parts of the raw memory block: |
duke@435 | 3081 | // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) |
duke@435 | 3082 | // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory |
duke@435 | 3083 | // After the initialize node executes, the object is ready for service: |
duke@435 | 3084 | // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) |
duke@435 | 3085 | // Suppose its body is immediately initialized as {1,2}: |
duke@435 | 3086 | // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) |
duke@435 | 3087 | // store2 = (StoreC init.Control store1 (+ oop 14) 2) |
duke@435 | 3088 | // mem.SLICE(#short[*]) := store2 |
duke@435 | 3089 | // |
duke@435 | 3090 | // DETAILS: |
duke@435 | 3091 | // An InitializeNode collects and isolates object initialization after |
duke@435 | 3092 | // an AllocateNode and before the next possible safepoint. As a |
duke@435 | 3093 | // memory barrier (MemBarNode), it keeps critical stores from drifting |
duke@435 | 3094 | // down past any safepoint or any publication of the allocation. |
duke@435 | 3095 | // Before this barrier, a newly-allocated object may have uninitialized bits. |
duke@435 | 3096 | // After this barrier, it may be treated as a real oop, and GC is allowed. |
duke@435 | 3097 | // |
duke@435 | 3098 | // The semantics of the InitializeNode include an implicit zeroing of |
duke@435 | 3099 | // the new object from object header to the end of the object. |
duke@435 | 3100 | // (The object header and end are determined by the AllocateNode.) |
duke@435 | 3101 | // |
duke@435 | 3102 | // Certain stores may be added as direct inputs to the InitializeNode. |
duke@435 | 3103 | // These stores must update raw memory, and they must be to addresses |
duke@435 | 3104 | // derived from the raw address produced by AllocateNode, and with |
duke@435 | 3105 | // a constant offset. They must be ordered by increasing offset. |
duke@435 | 3106 | // The first one is at in(RawStores), the last at in(req()-1). |
duke@435 | 3107 | // Unlike most memory operations, they are not linked in a chain, |
duke@435 | 3108 | // but are displayed in parallel as users of the rawmem output of |
duke@435 | 3109 | // the allocation. |
duke@435 | 3110 | // |
duke@435 | 3111 | // (See comments in InitializeNode::capture_store, which continue |
duke@435 | 3112 | // the example given above.) |
duke@435 | 3113 | // |
duke@435 | 3114 | // When the associated Allocate is macro-expanded, the InitializeNode |
duke@435 | 3115 | // may be rewritten to optimize collected stores. A ClearArrayNode |
duke@435 | 3116 | // may also be created at that point to represent any required zeroing. |
duke@435 | 3117 | // The InitializeNode is then marked 'complete', prohibiting further |
duke@435 | 3118 | // capturing of nearby memory operations. |
duke@435 | 3119 | // |
duke@435 | 3120 | // During macro-expansion, all captured initializations which store |
twisti@1040 | 3121 | // constant values of 32 bits or smaller are coalesced (if advantageous) |
duke@435 | 3122 | // into larger 'tiles' 32 or 64 bits. This allows an object to be |
duke@435 | 3123 | // initialized in fewer memory operations. Memory words which are |
duke@435 | 3124 | // covered by neither tiles nor non-constant stores are pre-zeroed |
duke@435 | 3125 | // by explicit stores of zero. (The code shape happens to do all |
duke@435 | 3126 | // zeroing first, then all other stores, with both sequences occurring |
duke@435 | 3127 | // in order of ascending offsets.) |
duke@435 | 3128 | // |
duke@435 | 3129 | // Alternatively, code may be inserted between an AllocateNode and its |
duke@435 | 3130 | // InitializeNode, to perform arbitrary initialization of the new object. |
duke@435 | 3131 | // E.g., the object copying intrinsics insert complex data transfers here. |
duke@435 | 3132 | // The initialization must then be marked as 'complete' disable the |
duke@435 | 3133 | // built-in zeroing semantics and the collection of initializing stores. |
duke@435 | 3134 | // |
duke@435 | 3135 | // While an InitializeNode is incomplete, reads from the memory state |
duke@435 | 3136 | // produced by it are optimizable if they match the control edge and |
duke@435 | 3137 | // new oop address associated with the allocation/initialization. |
duke@435 | 3138 | // They return a stored value (if the offset matches) or else zero. |
duke@435 | 3139 | // A write to the memory state, if it matches control and address, |
duke@435 | 3140 | // and if it is to a constant offset, may be 'captured' by the |
duke@435 | 3141 | // InitializeNode. It is cloned as a raw memory operation and rewired |
duke@435 | 3142 | // inside the initialization, to the raw oop produced by the allocation. |
duke@435 | 3143 | // Operations on addresses which are provably distinct (e.g., to |
duke@435 | 3144 | // other AllocateNodes) are allowed to bypass the initialization. |
duke@435 | 3145 | // |
duke@435 | 3146 | // The effect of all this is to consolidate object initialization |
duke@435 | 3147 | // (both arrays and non-arrays, both piecewise and bulk) into a |
duke@435 | 3148 | // single location, where it can be optimized as a unit. |
duke@435 | 3149 | // |
duke@435 | 3150 | // Only stores with an offset less than TrackedInitializationLimit words |
duke@435 | 3151 | // will be considered for capture by an InitializeNode. This puts a |
duke@435 | 3152 | // reasonable limit on the complexity of optimized initializations. |
duke@435 | 3153 | |
duke@435 | 3154 | //---------------------------InitializeNode------------------------------------ |
duke@435 | 3155 | InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) |
roland@3392 | 3156 | : _is_complete(Incomplete), _does_not_escape(false), |
duke@435 | 3157 | MemBarNode(C, adr_type, rawoop) |
duke@435 | 3158 | { |
duke@435 | 3159 | init_class_id(Class_Initialize); |
duke@435 | 3160 | |
duke@435 | 3161 | assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); |
duke@435 | 3162 | assert(in(RawAddress) == rawoop, "proper init"); |
duke@435 | 3163 | // Note: allocation() can be NULL, for secondary initialization barriers |
duke@435 | 3164 | } |
duke@435 | 3165 | |
duke@435 | 3166 | // Since this node is not matched, it will be processed by the |
duke@435 | 3167 | // register allocator. Declare that there are no constraints |
duke@435 | 3168 | // on the allocation of the RawAddress edge. |
duke@435 | 3169 | const RegMask &InitializeNode::in_RegMask(uint idx) const { |
duke@435 | 3170 | // This edge should be set to top, by the set_complete. But be conservative. |
duke@435 | 3171 | if (idx == InitializeNode::RawAddress) |
duke@435 | 3172 | return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); |
duke@435 | 3173 | return RegMask::Empty; |
duke@435 | 3174 | } |
duke@435 | 3175 | |
duke@435 | 3176 | Node* InitializeNode::memory(uint alias_idx) { |
duke@435 | 3177 | Node* mem = in(Memory); |
duke@435 | 3178 | if (mem->is_MergeMem()) { |
duke@435 | 3179 | return mem->as_MergeMem()->memory_at(alias_idx); |
duke@435 | 3180 | } else { |
duke@435 | 3181 | // incoming raw memory is not split |
duke@435 | 3182 | return mem; |
duke@435 | 3183 | } |
duke@435 | 3184 | } |
duke@435 | 3185 | |
duke@435 | 3186 | bool InitializeNode::is_non_zero() { |
duke@435 | 3187 | if (is_complete()) return false; |
duke@435 | 3188 | remove_extra_zeroes(); |
duke@435 | 3189 | return (req() > RawStores); |
duke@435 | 3190 | } |
duke@435 | 3191 | |
duke@435 | 3192 | void InitializeNode::set_complete(PhaseGVN* phase) { |
duke@435 | 3193 | assert(!is_complete(), "caller responsibility"); |
kvn@3157 | 3194 | _is_complete = Complete; |
duke@435 | 3195 | |
duke@435 | 3196 | // After this node is complete, it contains a bunch of |
duke@435 | 3197 | // raw-memory initializations. There is no need for |
duke@435 | 3198 | // it to have anything to do with non-raw memory effects. |
duke@435 | 3199 | // Therefore, tell all non-raw users to re-optimize themselves, |
duke@435 | 3200 | // after skipping the memory effects of this initialization. |
duke@435 | 3201 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
duke@435 | 3202 | if (igvn) igvn->add_users_to_worklist(this); |
duke@435 | 3203 | } |
duke@435 | 3204 | |
duke@435 | 3205 | // convenience function |
duke@435 | 3206 | // return false if the init contains any stores already |
duke@435 | 3207 | bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { |
duke@435 | 3208 | InitializeNode* init = initialization(); |
duke@435 | 3209 | if (init == NULL || init->is_complete()) return false; |
duke@435 | 3210 | init->remove_extra_zeroes(); |
duke@435 | 3211 | // for now, if this allocation has already collected any inits, bail: |
duke@435 | 3212 | if (init->is_non_zero()) return false; |
duke@435 | 3213 | init->set_complete(phase); |
duke@435 | 3214 | return true; |
duke@435 | 3215 | } |
duke@435 | 3216 | |
duke@435 | 3217 | void InitializeNode::remove_extra_zeroes() { |
duke@435 | 3218 | if (req() == RawStores) return; |
duke@435 | 3219 | Node* zmem = zero_memory(); |
duke@435 | 3220 | uint fill = RawStores; |
duke@435 | 3221 | for (uint i = fill; i < req(); i++) { |
duke@435 | 3222 | Node* n = in(i); |
duke@435 | 3223 | if (n->is_top() || n == zmem) continue; // skip |
duke@435 | 3224 | if (fill < i) set_req(fill, n); // compact |
duke@435 | 3225 | ++fill; |
duke@435 | 3226 | } |
duke@435 | 3227 | // delete any empty spaces created: |
duke@435 | 3228 | while (fill < req()) { |
duke@435 | 3229 | del_req(fill); |
duke@435 | 3230 | } |
duke@435 | 3231 | } |
duke@435 | 3232 | |
duke@435 | 3233 | // Helper for remembering which stores go with which offsets. |
duke@435 | 3234 | intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { |
duke@435 | 3235 | if (!st->is_Store()) return -1; // can happen to dead code via subsume_node |
duke@435 | 3236 | intptr_t offset = -1; |
duke@435 | 3237 | Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), |
duke@435 | 3238 | phase, offset); |
duke@435 | 3239 | if (base == NULL) return -1; // something is dead, |
duke@435 | 3240 | if (offset < 0) return -1; // dead, dead |
duke@435 | 3241 | return offset; |
duke@435 | 3242 | } |
duke@435 | 3243 | |
duke@435 | 3244 | // Helper for proving that an initialization expression is |
duke@435 | 3245 | // "simple enough" to be folded into an object initialization. |
duke@435 | 3246 | // Attempts to prove that a store's initial value 'n' can be captured |
duke@435 | 3247 | // within the initialization without creating a vicious cycle, such as: |
duke@435 | 3248 | // { Foo p = new Foo(); p.next = p; } |
duke@435 | 3249 | // True for constants and parameters and small combinations thereof. |
kvn@5110 | 3250 | bool InitializeNode::detect_init_independence(Node* n, int& count) { |
duke@435 | 3251 | if (n == NULL) return true; // (can this really happen?) |
duke@435 | 3252 | if (n->is_Proj()) n = n->in(0); |
duke@435 | 3253 | if (n == this) return false; // found a cycle |
duke@435 | 3254 | if (n->is_Con()) return true; |
duke@435 | 3255 | if (n->is_Start()) return true; // params, etc., are OK |
duke@435 | 3256 | if (n->is_Root()) return true; // even better |
duke@435 | 3257 | |
duke@435 | 3258 | Node* ctl = n->in(0); |
duke@435 | 3259 | if (ctl != NULL && !ctl->is_top()) { |
duke@435 | 3260 | if (ctl->is_Proj()) ctl = ctl->in(0); |
duke@435 | 3261 | if (ctl == this) return false; |
duke@435 | 3262 | |
duke@435 | 3263 | // If we already know that the enclosing memory op is pinned right after |
duke@435 | 3264 | // the init, then any control flow that the store has picked up |
duke@435 | 3265 | // must have preceded the init, or else be equal to the init. |
duke@435 | 3266 | // Even after loop optimizations (which might change control edges) |
duke@435 | 3267 | // a store is never pinned *before* the availability of its inputs. |
kvn@554 | 3268 | if (!MemNode::all_controls_dominate(n, this)) |
duke@435 | 3269 | return false; // failed to prove a good control |
duke@435 | 3270 | } |
duke@435 | 3271 | |
duke@435 | 3272 | // Check data edges for possible dependencies on 'this'. |
duke@435 | 3273 | if ((count += 1) > 20) return false; // complexity limit |
duke@435 | 3274 | for (uint i = 1; i < n->req(); i++) { |
duke@435 | 3275 | Node* m = n->in(i); |
duke@435 | 3276 | if (m == NULL || m == n || m->is_top()) continue; |
duke@435 | 3277 | uint first_i = n->find_edge(m); |
duke@435 | 3278 | if (i != first_i) continue; // process duplicate edge just once |
kvn@5110 | 3279 | if (!detect_init_independence(m, count)) { |
duke@435 | 3280 | return false; |
duke@435 | 3281 | } |
duke@435 | 3282 | } |
duke@435 | 3283 | |
duke@435 | 3284 | return true; |
duke@435 | 3285 | } |
duke@435 | 3286 | |
duke@435 | 3287 | // Here are all the checks a Store must pass before it can be moved into |
duke@435 | 3288 | // an initialization. Returns zero if a check fails. |
duke@435 | 3289 | // On success, returns the (constant) offset to which the store applies, |
duke@435 | 3290 | // within the initialized memory. |
roland@4657 | 3291 | intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { |
duke@435 | 3292 | const int FAIL = 0; |
duke@435 | 3293 | if (st->req() != MemNode::ValueIn + 1) |
duke@435 | 3294 | return FAIL; // an inscrutable StoreNode (card mark?) |
duke@435 | 3295 | Node* ctl = st->in(MemNode::Control); |
duke@435 | 3296 | if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) |
duke@435 | 3297 | return FAIL; // must be unconditional after the initialization |
duke@435 | 3298 | Node* mem = st->in(MemNode::Memory); |
duke@435 | 3299 | if (!(mem->is_Proj() && mem->in(0) == this)) |
duke@435 | 3300 | return FAIL; // must not be preceded by other stores |
duke@435 | 3301 | Node* adr = st->in(MemNode::Address); |
duke@435 | 3302 | intptr_t offset; |
duke@435 | 3303 | AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); |
duke@435 | 3304 | if (alloc == NULL) |
duke@435 | 3305 | return FAIL; // inscrutable address |
duke@435 | 3306 | if (alloc != allocation()) |
duke@435 | 3307 | return FAIL; // wrong allocation! (store needs to float up) |
duke@435 | 3308 | Node* val = st->in(MemNode::ValueIn); |
duke@435 | 3309 | int complexity_count = 0; |
kvn@5110 | 3310 | if (!detect_init_independence(val, complexity_count)) |
duke@435 | 3311 | return FAIL; // stored value must be 'simple enough' |
duke@435 | 3312 | |
roland@4657 | 3313 | // The Store can be captured only if nothing after the allocation |
roland@4657 | 3314 | // and before the Store is using the memory location that the store |
roland@4657 | 3315 | // overwrites. |
roland@4657 | 3316 | bool failed = false; |
roland@4657 | 3317 | // If is_complete_with_arraycopy() is true the shape of the graph is |
roland@4657 | 3318 | // well defined and is safe so no need for extra checks. |
roland@4657 | 3319 | if (!is_complete_with_arraycopy()) { |
roland@4657 | 3320 | // We are going to look at each use of the memory state following |
roland@4657 | 3321 | // the allocation to make sure nothing reads the memory that the |
roland@4657 | 3322 | // Store writes. |
roland@4657 | 3323 | const TypePtr* t_adr = phase->type(adr)->isa_ptr(); |
roland@4657 | 3324 | int alias_idx = phase->C->get_alias_index(t_adr); |
roland@4657 | 3325 | ResourceMark rm; |
roland@4657 | 3326 | Unique_Node_List mems; |
roland@4657 | 3327 | mems.push(mem); |
roland@4657 | 3328 | Node* unique_merge = NULL; |
roland@4657 | 3329 | for (uint next = 0; next < mems.size(); ++next) { |
roland@4657 | 3330 | Node *m = mems.at(next); |
roland@4657 | 3331 | for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { |
roland@4657 | 3332 | Node *n = m->fast_out(j); |
roland@4657 | 3333 | if (n->outcnt() == 0) { |
roland@4657 | 3334 | continue; |
roland@4657 | 3335 | } |
roland@4657 | 3336 | if (n == st) { |
roland@4657 | 3337 | continue; |
roland@4657 | 3338 | } else if (n->in(0) != NULL && n->in(0) != ctl) { |
roland@4657 | 3339 | // If the control of this use is different from the control |
roland@4657 | 3340 | // of the Store which is right after the InitializeNode then |
roland@4657 | 3341 | // this node cannot be between the InitializeNode and the |
roland@4657 | 3342 | // Store. |
roland@4657 | 3343 | continue; |
roland@4657 | 3344 | } else if (n->is_MergeMem()) { |
roland@4657 | 3345 | if (n->as_MergeMem()->memory_at(alias_idx) == m) { |
roland@4657 | 3346 | // We can hit a MergeMemNode (that will likely go away |
roland@4657 | 3347 | // later) that is a direct use of the memory state |
roland@4657 | 3348 | // following the InitializeNode on the same slice as the |
roland@4657 | 3349 | // store node that we'd like to capture. We need to check |
roland@4657 | 3350 | // the uses of the MergeMemNode. |
roland@4657 | 3351 | mems.push(n); |
roland@4657 | 3352 | } |
roland@4657 | 3353 | } else if (n->is_Mem()) { |
roland@4657 | 3354 | Node* other_adr = n->in(MemNode::Address); |
roland@4657 | 3355 | if (other_adr == adr) { |
roland@4657 | 3356 | failed = true; |
roland@4657 | 3357 | break; |
roland@4657 | 3358 | } else { |
roland@4657 | 3359 | const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); |
roland@4657 | 3360 | if (other_t_adr != NULL) { |
roland@4657 | 3361 | int other_alias_idx = phase->C->get_alias_index(other_t_adr); |
roland@4657 | 3362 | if (other_alias_idx == alias_idx) { |
roland@4657 | 3363 | // A load from the same memory slice as the store right |
roland@4657 | 3364 | // after the InitializeNode. We check the control of the |
roland@4657 | 3365 | // object/array that is loaded from. If it's the same as |
roland@4657 | 3366 | // the store control then we cannot capture the store. |
roland@4657 | 3367 | assert(!n->is_Store(), "2 stores to same slice on same control?"); |
roland@4657 | 3368 | Node* base = other_adr; |
roland@4657 | 3369 | assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); |
roland@4657 | 3370 | base = base->in(AddPNode::Base); |
roland@4657 | 3371 | if (base != NULL) { |
roland@4657 | 3372 | base = base->uncast(); |
roland@4657 | 3373 | if (base->is_Proj() && base->in(0) == alloc) { |
roland@4657 | 3374 | failed = true; |
roland@4657 | 3375 | break; |
roland@4657 | 3376 | } |
roland@4657 | 3377 | } |
roland@4657 | 3378 | } |
roland@4657 | 3379 | } |
roland@4657 | 3380 | } |
roland@4657 | 3381 | } else { |
roland@4657 | 3382 | failed = true; |
roland@4657 | 3383 | break; |
roland@4657 | 3384 | } |
roland@4657 | 3385 | } |
roland@4657 | 3386 | } |
roland@4657 | 3387 | } |
roland@4657 | 3388 | if (failed) { |
roland@4657 | 3389 | if (!can_reshape) { |
roland@4657 | 3390 | // We decided we couldn't capture the store during parsing. We |
roland@4657 | 3391 | // should try again during the next IGVN once the graph is |
roland@4657 | 3392 | // cleaner. |
roland@4657 | 3393 | phase->C->record_for_igvn(st); |
roland@4657 | 3394 | } |
roland@4657 | 3395 | return FAIL; |
roland@4657 | 3396 | } |
roland@4657 | 3397 | |
duke@435 | 3398 | return offset; // success |
duke@435 | 3399 | } |
duke@435 | 3400 | |
duke@435 | 3401 | // Find the captured store in(i) which corresponds to the range |
duke@435 | 3402 | // [start..start+size) in the initialized object. |
duke@435 | 3403 | // If there is one, return its index i. If there isn't, return the |
duke@435 | 3404 | // negative of the index where it should be inserted. |
duke@435 | 3405 | // Return 0 if the queried range overlaps an initialization boundary |
duke@435 | 3406 | // or if dead code is encountered. |
duke@435 | 3407 | // If size_in_bytes is zero, do not bother with overlap checks. |
duke@435 | 3408 | int InitializeNode::captured_store_insertion_point(intptr_t start, |
duke@435 | 3409 | int size_in_bytes, |
duke@435 | 3410 | PhaseTransform* phase) { |
duke@435 | 3411 | const int FAIL = 0, MAX_STORE = BytesPerLong; |
duke@435 | 3412 | |
duke@435 | 3413 | if (is_complete()) |
duke@435 | 3414 | return FAIL; // arraycopy got here first; punt |
duke@435 | 3415 | |
duke@435 | 3416 | assert(allocation() != NULL, "must be present"); |
duke@435 | 3417 | |
duke@435 | 3418 | // no negatives, no header fields: |
coleenp@548 | 3419 | if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; |
duke@435 | 3420 | |
duke@435 | 3421 | // after a certain size, we bail out on tracking all the stores: |
duke@435 | 3422 | intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); |
duke@435 | 3423 | if (start >= ti_limit) return FAIL; |
duke@435 | 3424 | |
duke@435 | 3425 | for (uint i = InitializeNode::RawStores, limit = req(); ; ) { |
duke@435 | 3426 | if (i >= limit) return -(int)i; // not found; here is where to put it |
duke@435 | 3427 | |
duke@435 | 3428 | Node* st = in(i); |
duke@435 | 3429 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 3430 | if (st_off < 0) { |
duke@435 | 3431 | if (st != zero_memory()) { |
duke@435 | 3432 | return FAIL; // bail out if there is dead garbage |
duke@435 | 3433 | } |
duke@435 | 3434 | } else if (st_off > start) { |
duke@435 | 3435 | // ...we are done, since stores are ordered |
duke@435 | 3436 | if (st_off < start + size_in_bytes) { |
duke@435 | 3437 | return FAIL; // the next store overlaps |
duke@435 | 3438 | } |
duke@435 | 3439 | return -(int)i; // not found; here is where to put it |
duke@435 | 3440 | } else if (st_off < start) { |
duke@435 | 3441 | if (size_in_bytes != 0 && |
duke@435 | 3442 | start < st_off + MAX_STORE && |
duke@435 | 3443 | start < st_off + st->as_Store()->memory_size()) { |
duke@435 | 3444 | return FAIL; // the previous store overlaps |
duke@435 | 3445 | } |
duke@435 | 3446 | } else { |
duke@435 | 3447 | if (size_in_bytes != 0 && |
duke@435 | 3448 | st->as_Store()->memory_size() != size_in_bytes) { |
duke@435 | 3449 | return FAIL; // mismatched store size |
duke@435 | 3450 | } |
duke@435 | 3451 | return i; |
duke@435 | 3452 | } |
duke@435 | 3453 | |
duke@435 | 3454 | ++i; |
duke@435 | 3455 | } |
duke@435 | 3456 | } |
duke@435 | 3457 | |
duke@435 | 3458 | // Look for a captured store which initializes at the offset 'start' |
duke@435 | 3459 | // with the given size. If there is no such store, and no other |
duke@435 | 3460 | // initialization interferes, then return zero_memory (the memory |
duke@435 | 3461 | // projection of the AllocateNode). |
duke@435 | 3462 | Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, |
duke@435 | 3463 | PhaseTransform* phase) { |
duke@435 | 3464 | assert(stores_are_sane(phase), ""); |
duke@435 | 3465 | int i = captured_store_insertion_point(start, size_in_bytes, phase); |
duke@435 | 3466 | if (i == 0) { |
duke@435 | 3467 | return NULL; // something is dead |
duke@435 | 3468 | } else if (i < 0) { |
duke@435 | 3469 | return zero_memory(); // just primordial zero bits here |
duke@435 | 3470 | } else { |
duke@435 | 3471 | Node* st = in(i); // here is the store at this position |
duke@435 | 3472 | assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); |
duke@435 | 3473 | return st; |
duke@435 | 3474 | } |
duke@435 | 3475 | } |
duke@435 | 3476 | |
duke@435 | 3477 | // Create, as a raw pointer, an address within my new object at 'offset'. |
duke@435 | 3478 | Node* InitializeNode::make_raw_address(intptr_t offset, |
duke@435 | 3479 | PhaseTransform* phase) { |
duke@435 | 3480 | Node* addr = in(RawAddress); |
duke@435 | 3481 | if (offset != 0) { |
duke@435 | 3482 | Compile* C = phase->C; |
kvn@4115 | 3483 | addr = phase->transform( new (C) AddPNode(C->top(), addr, |
duke@435 | 3484 | phase->MakeConX(offset)) ); |
duke@435 | 3485 | } |
duke@435 | 3486 | return addr; |
duke@435 | 3487 | } |
duke@435 | 3488 | |
duke@435 | 3489 | // Clone the given store, converting it into a raw store |
duke@435 | 3490 | // initializing a field or element of my new object. |
duke@435 | 3491 | // Caller is responsible for retiring the original store, |
duke@435 | 3492 | // with subsume_node or the like. |
duke@435 | 3493 | // |
duke@435 | 3494 | // From the example above InitializeNode::InitializeNode, |
duke@435 | 3495 | // here are the old stores to be captured: |
duke@435 | 3496 | // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) |
duke@435 | 3497 | // store2 = (StoreC init.Control store1 (+ oop 14) 2) |
duke@435 | 3498 | // |
duke@435 | 3499 | // Here is the changed code; note the extra edges on init: |
duke@435 | 3500 | // alloc = (Allocate ...) |
duke@435 | 3501 | // rawoop = alloc.RawAddress |
duke@435 | 3502 | // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) |
duke@435 | 3503 | // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) |
duke@435 | 3504 | // init = (Initialize alloc.Control alloc.Memory rawoop |
duke@435 | 3505 | // rawstore1 rawstore2) |
duke@435 | 3506 | // |
duke@435 | 3507 | Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, |
roland@4657 | 3508 | PhaseTransform* phase, bool can_reshape) { |
duke@435 | 3509 | assert(stores_are_sane(phase), ""); |
duke@435 | 3510 | |
duke@435 | 3511 | if (start < 0) return NULL; |
roland@4657 | 3512 | assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); |
duke@435 | 3513 | |
duke@435 | 3514 | Compile* C = phase->C; |
duke@435 | 3515 | int size_in_bytes = st->memory_size(); |
duke@435 | 3516 | int i = captured_store_insertion_point(start, size_in_bytes, phase); |
duke@435 | 3517 | if (i == 0) return NULL; // bail out |
duke@435 | 3518 | Node* prev_mem = NULL; // raw memory for the captured store |
duke@435 | 3519 | if (i > 0) { |
duke@435 | 3520 | prev_mem = in(i); // there is a pre-existing store under this one |
duke@435 | 3521 | set_req(i, C->top()); // temporarily disconnect it |
duke@435 | 3522 | // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. |
duke@435 | 3523 | } else { |
duke@435 | 3524 | i = -i; // no pre-existing store |
duke@435 | 3525 | prev_mem = zero_memory(); // a slice of the newly allocated object |
duke@435 | 3526 | if (i > InitializeNode::RawStores && in(i-1) == prev_mem) |
duke@435 | 3527 | set_req(--i, C->top()); // reuse this edge; it has been folded away |
duke@435 | 3528 | else |
duke@435 | 3529 | ins_req(i, C->top()); // build a new edge |
duke@435 | 3530 | } |
duke@435 | 3531 | Node* new_st = st->clone(); |
duke@435 | 3532 | new_st->set_req(MemNode::Control, in(Control)); |
duke@435 | 3533 | new_st->set_req(MemNode::Memory, prev_mem); |
duke@435 | 3534 | new_st->set_req(MemNode::Address, make_raw_address(start, phase)); |
duke@435 | 3535 | new_st = phase->transform(new_st); |
duke@435 | 3536 | |
duke@435 | 3537 | // At this point, new_st might have swallowed a pre-existing store |
duke@435 | 3538 | // at the same offset, or perhaps new_st might have disappeared, |
duke@435 | 3539 | // if it redundantly stored the same value (or zero to fresh memory). |
duke@435 | 3540 | |
duke@435 | 3541 | // In any case, wire it in: |
duke@435 | 3542 | set_req(i, new_st); |
duke@435 | 3543 | |
duke@435 | 3544 | // The caller may now kill the old guy. |
duke@435 | 3545 | DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); |
duke@435 | 3546 | assert(check_st == new_st || check_st == NULL, "must be findable"); |
duke@435 | 3547 | assert(!is_complete(), ""); |
duke@435 | 3548 | return new_st; |
duke@435 | 3549 | } |
duke@435 | 3550 | |
duke@435 | 3551 | static bool store_constant(jlong* tiles, int num_tiles, |
duke@435 | 3552 | intptr_t st_off, int st_size, |
duke@435 | 3553 | jlong con) { |
duke@435 | 3554 | if ((st_off & (st_size-1)) != 0) |
duke@435 | 3555 | return false; // strange store offset (assume size==2**N) |
duke@435 | 3556 | address addr = (address)tiles + st_off; |
duke@435 | 3557 | assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); |
duke@435 | 3558 | switch (st_size) { |
duke@435 | 3559 | case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; |
duke@435 | 3560 | case sizeof(jchar): *(jchar*) addr = (jchar) con; break; |
duke@435 | 3561 | case sizeof(jint): *(jint*) addr = (jint) con; break; |
duke@435 | 3562 | case sizeof(jlong): *(jlong*) addr = (jlong) con; break; |
duke@435 | 3563 | default: return false; // strange store size (detect size!=2**N here) |
duke@435 | 3564 | } |
duke@435 | 3565 | return true; // return success to caller |
duke@435 | 3566 | } |
duke@435 | 3567 | |
duke@435 | 3568 | // Coalesce subword constants into int constants and possibly |
duke@435 | 3569 | // into long constants. The goal, if the CPU permits, |
duke@435 | 3570 | // is to initialize the object with a small number of 64-bit tiles. |
duke@435 | 3571 | // Also, convert floating-point constants to bit patterns. |
duke@435 | 3572 | // Non-constants are not relevant to this pass. |
duke@435 | 3573 | // |
duke@435 | 3574 | // In terms of the running example on InitializeNode::InitializeNode |
duke@435 | 3575 | // and InitializeNode::capture_store, here is the transformation |
duke@435 | 3576 | // of rawstore1 and rawstore2 into rawstore12: |
duke@435 | 3577 | // alloc = (Allocate ...) |
duke@435 | 3578 | // rawoop = alloc.RawAddress |
duke@435 | 3579 | // tile12 = 0x00010002 |
duke@435 | 3580 | // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) |
duke@435 | 3581 | // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) |
duke@435 | 3582 | // |
duke@435 | 3583 | void |
duke@435 | 3584 | InitializeNode::coalesce_subword_stores(intptr_t header_size, |
duke@435 | 3585 | Node* size_in_bytes, |
duke@435 | 3586 | PhaseGVN* phase) { |
duke@435 | 3587 | Compile* C = phase->C; |
duke@435 | 3588 | |
duke@435 | 3589 | assert(stores_are_sane(phase), ""); |
duke@435 | 3590 | // Note: After this pass, they are not completely sane, |
duke@435 | 3591 | // since there may be some overlaps. |
duke@435 | 3592 | |
duke@435 | 3593 | int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; |
duke@435 | 3594 | |
duke@435 | 3595 | intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); |
duke@435 | 3596 | intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); |
duke@435 | 3597 | size_limit = MIN2(size_limit, ti_limit); |
duke@435 | 3598 | size_limit = align_size_up(size_limit, BytesPerLong); |
duke@435 | 3599 | int num_tiles = size_limit / BytesPerLong; |
duke@435 | 3600 | |
duke@435 | 3601 | // allocate space for the tile map: |
duke@435 | 3602 | const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small |
duke@435 | 3603 | jlong tiles_buf[small_len]; |
duke@435 | 3604 | Node* nodes_buf[small_len]; |
duke@435 | 3605 | jlong inits_buf[small_len]; |
duke@435 | 3606 | jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] |
duke@435 | 3607 | : NEW_RESOURCE_ARRAY(jlong, num_tiles)); |
duke@435 | 3608 | Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] |
duke@435 | 3609 | : NEW_RESOURCE_ARRAY(Node*, num_tiles)); |
duke@435 | 3610 | jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] |
duke@435 | 3611 | : NEW_RESOURCE_ARRAY(jlong, num_tiles)); |
duke@435 | 3612 | // tiles: exact bitwise model of all primitive constants |
duke@435 | 3613 | // nodes: last constant-storing node subsumed into the tiles model |
duke@435 | 3614 | // inits: which bytes (in each tile) are touched by any initializations |
duke@435 | 3615 | |
duke@435 | 3616 | //// Pass A: Fill in the tile model with any relevant stores. |
duke@435 | 3617 | |
duke@435 | 3618 | Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); |
duke@435 | 3619 | Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); |
duke@435 | 3620 | Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); |
duke@435 | 3621 | Node* zmem = zero_memory(); // initially zero memory state |
duke@435 | 3622 | for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { |
duke@435 | 3623 | Node* st = in(i); |
duke@435 | 3624 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 3625 | |
duke@435 | 3626 | // Figure out the store's offset and constant value: |
duke@435 | 3627 | if (st_off < header_size) continue; //skip (ignore header) |
duke@435 | 3628 | if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) |
duke@435 | 3629 | int st_size = st->as_Store()->memory_size(); |
duke@435 | 3630 | if (st_off + st_size > size_limit) break; |
duke@435 | 3631 | |
duke@435 | 3632 | // Record which bytes are touched, whether by constant or not. |
duke@435 | 3633 | if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) |
duke@435 | 3634 | continue; // skip (strange store size) |
duke@435 | 3635 | |
duke@435 | 3636 | const Type* val = phase->type(st->in(MemNode::ValueIn)); |
duke@435 | 3637 | if (!val->singleton()) continue; //skip (non-con store) |
duke@435 | 3638 | BasicType type = val->basic_type(); |
duke@435 | 3639 | |
duke@435 | 3640 | jlong con = 0; |
duke@435 | 3641 | switch (type) { |
duke@435 | 3642 | case T_INT: con = val->is_int()->get_con(); break; |
duke@435 | 3643 | case T_LONG: con = val->is_long()->get_con(); break; |
duke@435 | 3644 | case T_FLOAT: con = jint_cast(val->getf()); break; |
duke@435 | 3645 | case T_DOUBLE: con = jlong_cast(val->getd()); break; |
duke@435 | 3646 | default: continue; //skip (odd store type) |
duke@435 | 3647 | } |
duke@435 | 3648 | |
duke@435 | 3649 | if (type == T_LONG && Matcher::isSimpleConstant64(con) && |
duke@435 | 3650 | st->Opcode() == Op_StoreL) { |
duke@435 | 3651 | continue; // This StoreL is already optimal. |
duke@435 | 3652 | } |
duke@435 | 3653 | |
duke@435 | 3654 | // Store down the constant. |
duke@435 | 3655 | store_constant(tiles, num_tiles, st_off, st_size, con); |
duke@435 | 3656 | |
duke@435 | 3657 | intptr_t j = st_off >> LogBytesPerLong; |
duke@435 | 3658 | |
duke@435 | 3659 | if (type == T_INT && st_size == BytesPerInt |
duke@435 | 3660 | && (st_off & BytesPerInt) == BytesPerInt) { |
duke@435 | 3661 | jlong lcon = tiles[j]; |
duke@435 | 3662 | if (!Matcher::isSimpleConstant64(lcon) && |
duke@435 | 3663 | st->Opcode() == Op_StoreI) { |
duke@435 | 3664 | // This StoreI is already optimal by itself. |
duke@435 | 3665 | jint* intcon = (jint*) &tiles[j]; |
duke@435 | 3666 | intcon[1] = 0; // undo the store_constant() |
duke@435 | 3667 | |
duke@435 | 3668 | // If the previous store is also optimal by itself, back up and |
duke@435 | 3669 | // undo the action of the previous loop iteration... if we can. |
duke@435 | 3670 | // But if we can't, just let the previous half take care of itself. |
duke@435 | 3671 | st = nodes[j]; |
duke@435 | 3672 | st_off -= BytesPerInt; |
duke@435 | 3673 | con = intcon[0]; |
duke@435 | 3674 | if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { |
duke@435 | 3675 | assert(st_off >= header_size, "still ignoring header"); |
duke@435 | 3676 | assert(get_store_offset(st, phase) == st_off, "must be"); |
duke@435 | 3677 | assert(in(i-1) == zmem, "must be"); |
duke@435 | 3678 | DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); |
duke@435 | 3679 | assert(con == tcon->is_int()->get_con(), "must be"); |
duke@435 | 3680 | // Undo the effects of the previous loop trip, which swallowed st: |
duke@435 | 3681 | intcon[0] = 0; // undo store_constant() |
duke@435 | 3682 | set_req(i-1, st); // undo set_req(i, zmem) |
duke@435 | 3683 | nodes[j] = NULL; // undo nodes[j] = st |
duke@435 | 3684 | --old_subword; // undo ++old_subword |
duke@435 | 3685 | } |
duke@435 | 3686 | continue; // This StoreI is already optimal. |
duke@435 | 3687 | } |
duke@435 | 3688 | } |
duke@435 | 3689 | |
duke@435 | 3690 | // This store is not needed. |
duke@435 | 3691 | set_req(i, zmem); |
duke@435 | 3692 | nodes[j] = st; // record for the moment |
duke@435 | 3693 | if (st_size < BytesPerLong) // something has changed |
duke@435 | 3694 | ++old_subword; // includes int/float, but who's counting... |
duke@435 | 3695 | else ++old_long; |
duke@435 | 3696 | } |
duke@435 | 3697 | |
duke@435 | 3698 | if ((old_subword + old_long) == 0) |
duke@435 | 3699 | return; // nothing more to do |
duke@435 | 3700 | |
duke@435 | 3701 | //// Pass B: Convert any non-zero tiles into optimal constant stores. |
duke@435 | 3702 | // Be sure to insert them before overlapping non-constant stores. |
duke@435 | 3703 | // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) |
duke@435 | 3704 | for (int j = 0; j < num_tiles; j++) { |
duke@435 | 3705 | jlong con = tiles[j]; |
duke@435 | 3706 | jlong init = inits[j]; |
duke@435 | 3707 | if (con == 0) continue; |
duke@435 | 3708 | jint con0, con1; // split the constant, address-wise |
duke@435 | 3709 | jint init0, init1; // split the init map, address-wise |
duke@435 | 3710 | { union { jlong con; jint intcon[2]; } u; |
duke@435 | 3711 | u.con = con; |
duke@435 | 3712 | con0 = u.intcon[0]; |
duke@435 | 3713 | con1 = u.intcon[1]; |
duke@435 | 3714 | u.con = init; |
duke@435 | 3715 | init0 = u.intcon[0]; |
duke@435 | 3716 | init1 = u.intcon[1]; |
duke@435 | 3717 | } |
duke@435 | 3718 | |
duke@435 | 3719 | Node* old = nodes[j]; |
duke@435 | 3720 | assert(old != NULL, "need the prior store"); |
duke@435 | 3721 | intptr_t offset = (j * BytesPerLong); |
duke@435 | 3722 | |
duke@435 | 3723 | bool split = !Matcher::isSimpleConstant64(con); |
duke@435 | 3724 | |
duke@435 | 3725 | if (offset < header_size) { |
duke@435 | 3726 | assert(offset + BytesPerInt >= header_size, "second int counts"); |
duke@435 | 3727 | assert(*(jint*)&tiles[j] == 0, "junk in header"); |
duke@435 | 3728 | split = true; // only the second word counts |
duke@435 | 3729 | // Example: int a[] = { 42 ... } |
duke@435 | 3730 | } else if (con0 == 0 && init0 == -1) { |
duke@435 | 3731 | split = true; // first word is covered by full inits |
duke@435 | 3732 | // Example: int a[] = { ... foo(), 42 ... } |
duke@435 | 3733 | } else if (con1 == 0 && init1 == -1) { |
duke@435 | 3734 | split = true; // second word is covered by full inits |
duke@435 | 3735 | // Example: int a[] = { ... 42, foo() ... } |
duke@435 | 3736 | } |
duke@435 | 3737 | |
duke@435 | 3738 | // Here's a case where init0 is neither 0 nor -1: |
duke@435 | 3739 | // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } |
duke@435 | 3740 | // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. |
duke@435 | 3741 | // In this case the tile is not split; it is (jlong)42. |
duke@435 | 3742 | // The big tile is stored down, and then the foo() value is inserted. |
duke@435 | 3743 | // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) |
duke@435 | 3744 | |
duke@435 | 3745 | Node* ctl = old->in(MemNode::Control); |
duke@435 | 3746 | Node* adr = make_raw_address(offset, phase); |
duke@435 | 3747 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
duke@435 | 3748 | |
duke@435 | 3749 | // One or two coalesced stores to plop down. |
duke@435 | 3750 | Node* st[2]; |
duke@435 | 3751 | intptr_t off[2]; |
duke@435 | 3752 | int nst = 0; |
duke@435 | 3753 | if (!split) { |
duke@435 | 3754 | ++new_long; |
duke@435 | 3755 | off[nst] = offset; |
coleenp@548 | 3756 | st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
goetz@6479 | 3757 | phase->longcon(con), T_LONG, MemNode::unordered); |
duke@435 | 3758 | } else { |
duke@435 | 3759 | // Omit either if it is a zero. |
duke@435 | 3760 | if (con0 != 0) { |
duke@435 | 3761 | ++new_int; |
duke@435 | 3762 | off[nst] = offset; |
coleenp@548 | 3763 | st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
goetz@6479 | 3764 | phase->intcon(con0), T_INT, MemNode::unordered); |
duke@435 | 3765 | } |
duke@435 | 3766 | if (con1 != 0) { |
duke@435 | 3767 | ++new_int; |
duke@435 | 3768 | offset += BytesPerInt; |
duke@435 | 3769 | adr = make_raw_address(offset, phase); |
duke@435 | 3770 | off[nst] = offset; |
coleenp@548 | 3771 | st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
goetz@6479 | 3772 | phase->intcon(con1), T_INT, MemNode::unordered); |
duke@435 | 3773 | } |
duke@435 | 3774 | } |
duke@435 | 3775 | |
duke@435 | 3776 | // Insert second store first, then the first before the second. |
duke@435 | 3777 | // Insert each one just before any overlapping non-constant stores. |
duke@435 | 3778 | while (nst > 0) { |
duke@435 | 3779 | Node* st1 = st[--nst]; |
duke@435 | 3780 | C->copy_node_notes_to(st1, old); |
duke@435 | 3781 | st1 = phase->transform(st1); |
duke@435 | 3782 | offset = off[nst]; |
duke@435 | 3783 | assert(offset >= header_size, "do not smash header"); |
duke@435 | 3784 | int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); |
duke@435 | 3785 | guarantee(ins_idx != 0, "must re-insert constant store"); |
duke@435 | 3786 | if (ins_idx < 0) ins_idx = -ins_idx; // never overlap |
duke@435 | 3787 | if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) |
duke@435 | 3788 | set_req(--ins_idx, st1); |
duke@435 | 3789 | else |
duke@435 | 3790 | ins_req(ins_idx, st1); |
duke@435 | 3791 | } |
duke@435 | 3792 | } |
duke@435 | 3793 | |
duke@435 | 3794 | if (PrintCompilation && WizardMode) |
duke@435 | 3795 | tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", |
duke@435 | 3796 | old_subword, old_long, new_int, new_long); |
duke@435 | 3797 | if (C->log() != NULL) |
duke@435 | 3798 | C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", |
duke@435 | 3799 | old_subword, old_long, new_int, new_long); |
duke@435 | 3800 | |
duke@435 | 3801 | // Clean up any remaining occurrences of zmem: |
duke@435 | 3802 | remove_extra_zeroes(); |
duke@435 | 3803 | } |
duke@435 | 3804 | |
duke@435 | 3805 | // Explore forward from in(start) to find the first fully initialized |
duke@435 | 3806 | // word, and return its offset. Skip groups of subword stores which |
duke@435 | 3807 | // together initialize full words. If in(start) is itself part of a |
duke@435 | 3808 | // fully initialized word, return the offset of in(start). If there |
duke@435 | 3809 | // are no following full-word stores, or if something is fishy, return |
duke@435 | 3810 | // a negative value. |
duke@435 | 3811 | intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { |
duke@435 | 3812 | int int_map = 0; |
duke@435 | 3813 | intptr_t int_map_off = 0; |
duke@435 | 3814 | const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for |
duke@435 | 3815 | |
duke@435 | 3816 | for (uint i = start, limit = req(); i < limit; i++) { |
duke@435 | 3817 | Node* st = in(i); |
duke@435 | 3818 | |
duke@435 | 3819 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 3820 | if (st_off < 0) break; // return conservative answer |
duke@435 | 3821 | |
duke@435 | 3822 | int st_size = st->as_Store()->memory_size(); |
duke@435 | 3823 | if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { |
duke@435 | 3824 | return st_off; // we found a complete word init |
duke@435 | 3825 | } |
duke@435 | 3826 | |
duke@435 | 3827 | // update the map: |
duke@435 | 3828 | |
duke@435 | 3829 | intptr_t this_int_off = align_size_down(st_off, BytesPerInt); |
duke@435 | 3830 | if (this_int_off != int_map_off) { |
duke@435 | 3831 | // reset the map: |
duke@435 | 3832 | int_map = 0; |
duke@435 | 3833 | int_map_off = this_int_off; |
duke@435 | 3834 | } |
duke@435 | 3835 | |
duke@435 | 3836 | int subword_off = st_off - this_int_off; |
duke@435 | 3837 | int_map |= right_n_bits(st_size) << subword_off; |
duke@435 | 3838 | if ((int_map & FULL_MAP) == FULL_MAP) { |
duke@435 | 3839 | return this_int_off; // we found a complete word init |
duke@435 | 3840 | } |
duke@435 | 3841 | |
duke@435 | 3842 | // Did this store hit or cross the word boundary? |
duke@435 | 3843 | intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); |
duke@435 | 3844 | if (next_int_off == this_int_off + BytesPerInt) { |
duke@435 | 3845 | // We passed the current int, without fully initializing it. |
duke@435 | 3846 | int_map_off = next_int_off; |
duke@435 | 3847 | int_map >>= BytesPerInt; |
duke@435 | 3848 | } else if (next_int_off > this_int_off + BytesPerInt) { |
duke@435 | 3849 | // We passed the current and next int. |
duke@435 | 3850 | return this_int_off + BytesPerInt; |
duke@435 | 3851 | } |
duke@435 | 3852 | } |
duke@435 | 3853 | |
duke@435 | 3854 | return -1; |
duke@435 | 3855 | } |
duke@435 | 3856 | |
duke@435 | 3857 | |
duke@435 | 3858 | // Called when the associated AllocateNode is expanded into CFG. |
duke@435 | 3859 | // At this point, we may perform additional optimizations. |
duke@435 | 3860 | // Linearize the stores by ascending offset, to make memory |
duke@435 | 3861 | // activity as coherent as possible. |
duke@435 | 3862 | Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, |
duke@435 | 3863 | intptr_t header_size, |
duke@435 | 3864 | Node* size_in_bytes, |
duke@435 | 3865 | PhaseGVN* phase) { |
duke@435 | 3866 | assert(!is_complete(), "not already complete"); |
duke@435 | 3867 | assert(stores_are_sane(phase), ""); |
duke@435 | 3868 | assert(allocation() != NULL, "must be present"); |
duke@435 | 3869 | |
duke@435 | 3870 | remove_extra_zeroes(); |
duke@435 | 3871 | |
duke@435 | 3872 | if (ReduceFieldZeroing || ReduceBulkZeroing) |
duke@435 | 3873 | // reduce instruction count for common initialization patterns |
duke@435 | 3874 | coalesce_subword_stores(header_size, size_in_bytes, phase); |
duke@435 | 3875 | |
duke@435 | 3876 | Node* zmem = zero_memory(); // initially zero memory state |
duke@435 | 3877 | Node* inits = zmem; // accumulating a linearized chain of inits |
duke@435 | 3878 | #ifdef ASSERT |
coleenp@548 | 3879 | intptr_t first_offset = allocation()->minimum_header_size(); |
coleenp@548 | 3880 | intptr_t last_init_off = first_offset; // previous init offset |
coleenp@548 | 3881 | intptr_t last_init_end = first_offset; // previous init offset+size |
coleenp@548 | 3882 | intptr_t last_tile_end = first_offset; // previous tile offset+size |
duke@435 | 3883 | #endif |
duke@435 | 3884 | intptr_t zeroes_done = header_size; |
duke@435 | 3885 | |
duke@435 | 3886 | bool do_zeroing = true; // we might give up if inits are very sparse |
duke@435 | 3887 | int big_init_gaps = 0; // how many large gaps have we seen? |
duke@435 | 3888 | |
duke@435 | 3889 | if (ZeroTLAB) do_zeroing = false; |
duke@435 | 3890 | if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; |
duke@435 | 3891 | |
duke@435 | 3892 | for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { |
duke@435 | 3893 | Node* st = in(i); |
duke@435 | 3894 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 3895 | if (st_off < 0) |
duke@435 | 3896 | break; // unknown junk in the inits |
duke@435 | 3897 | if (st->in(MemNode::Memory) != zmem) |
duke@435 | 3898 | break; // complicated store chains somehow in list |
duke@435 | 3899 | |
duke@435 | 3900 | int st_size = st->as_Store()->memory_size(); |
duke@435 | 3901 | intptr_t next_init_off = st_off + st_size; |
duke@435 | 3902 | |
duke@435 | 3903 | if (do_zeroing && zeroes_done < next_init_off) { |
duke@435 | 3904 | // See if this store needs a zero before it or under it. |
duke@435 | 3905 | intptr_t zeroes_needed = st_off; |
duke@435 | 3906 | |
duke@435 | 3907 | if (st_size < BytesPerInt) { |
duke@435 | 3908 | // Look for subword stores which only partially initialize words. |
duke@435 | 3909 | // If we find some, we must lay down some word-level zeroes first, |
duke@435 | 3910 | // underneath the subword stores. |
duke@435 | 3911 | // |
duke@435 | 3912 | // Examples: |
duke@435 | 3913 | // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s |
duke@435 | 3914 | // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y |
duke@435 | 3915 | // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z |
duke@435 | 3916 | // |
duke@435 | 3917 | // Note: coalesce_subword_stores may have already done this, |
duke@435 | 3918 | // if it was prompted by constant non-zero subword initializers. |
duke@435 | 3919 | // But this case can still arise with non-constant stores. |
duke@435 | 3920 | |
duke@435 | 3921 | intptr_t next_full_store = find_next_fullword_store(i, phase); |
duke@435 | 3922 | |
duke@435 | 3923 | // In the examples above: |
duke@435 | 3924 | // in(i) p q r s x y z |
duke@435 | 3925 | // st_off 12 13 14 15 12 13 14 |
duke@435 | 3926 | // st_size 1 1 1 1 1 1 1 |
duke@435 | 3927 | // next_full_s. 12 16 16 16 16 16 16 |
duke@435 | 3928 | // z's_done 12 16 16 16 12 16 12 |
duke@435 | 3929 | // z's_needed 12 16 16 16 16 16 16 |
duke@435 | 3930 | // zsize 0 0 0 0 4 0 4 |
duke@435 | 3931 | if (next_full_store < 0) { |
duke@435 | 3932 | // Conservative tack: Zero to end of current word. |
duke@435 | 3933 | zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); |
duke@435 | 3934 | } else { |
duke@435 | 3935 | // Zero to beginning of next fully initialized word. |
duke@435 | 3936 | // Or, don't zero at all, if we are already in that word. |
duke@435 | 3937 | assert(next_full_store >= zeroes_needed, "must go forward"); |
duke@435 | 3938 | assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); |
duke@435 | 3939 | zeroes_needed = next_full_store; |
duke@435 | 3940 | } |
duke@435 | 3941 | } |
duke@435 | 3942 | |
duke@435 | 3943 | if (zeroes_needed > zeroes_done) { |
duke@435 | 3944 | intptr_t zsize = zeroes_needed - zeroes_done; |
duke@435 | 3945 | // Do some incremental zeroing on rawmem, in parallel with inits. |
duke@435 | 3946 | zeroes_done = align_size_down(zeroes_done, BytesPerInt); |
duke@435 | 3947 | rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, |
duke@435 | 3948 | zeroes_done, zeroes_needed, |
duke@435 | 3949 | phase); |
duke@435 | 3950 | zeroes_done = zeroes_needed; |
duke@435 | 3951 | if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) |
duke@435 | 3952 | do_zeroing = false; // leave the hole, next time |
duke@435 | 3953 | } |
duke@435 | 3954 | } |
duke@435 | 3955 | |
duke@435 | 3956 | // Collect the store and move on: |
duke@435 | 3957 | st->set_req(MemNode::Memory, inits); |
duke@435 | 3958 | inits = st; // put it on the linearized chain |
duke@435 | 3959 | set_req(i, zmem); // unhook from previous position |
duke@435 | 3960 | |
duke@435 | 3961 | if (zeroes_done == st_off) |
duke@435 | 3962 | zeroes_done = next_init_off; |
duke@435 | 3963 | |
duke@435 | 3964 | assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); |
duke@435 | 3965 | |
duke@435 | 3966 | #ifdef ASSERT |
duke@435 | 3967 | // Various order invariants. Weaker than stores_are_sane because |
duke@435 | 3968 | // a large constant tile can be filled in by smaller non-constant stores. |
duke@435 | 3969 | assert(st_off >= last_init_off, "inits do not reverse"); |
duke@435 | 3970 | last_init_off = st_off; |
duke@435 | 3971 | const Type* val = NULL; |
duke@435 | 3972 | if (st_size >= BytesPerInt && |
duke@435 | 3973 | (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && |
duke@435 | 3974 | (int)val->basic_type() < (int)T_OBJECT) { |
duke@435 | 3975 | assert(st_off >= last_tile_end, "tiles do not overlap"); |
duke@435 | 3976 | assert(st_off >= last_init_end, "tiles do not overwrite inits"); |
duke@435 | 3977 | last_tile_end = MAX2(last_tile_end, next_init_off); |
duke@435 | 3978 | } else { |
duke@435 | 3979 | intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); |
duke@435 | 3980 | assert(st_tile_end >= last_tile_end, "inits stay with tiles"); |
duke@435 | 3981 | assert(st_off >= last_init_end, "inits do not overlap"); |
duke@435 | 3982 | last_init_end = next_init_off; // it's a non-tile |
duke@435 | 3983 | } |
duke@435 | 3984 | #endif //ASSERT |
duke@435 | 3985 | } |
duke@435 | 3986 | |
duke@435 | 3987 | remove_extra_zeroes(); // clear out all the zmems left over |
duke@435 | 3988 | add_req(inits); |
duke@435 | 3989 | |
duke@435 | 3990 | if (!ZeroTLAB) { |
duke@435 | 3991 | // If anything remains to be zeroed, zero it all now. |
duke@435 | 3992 | zeroes_done = align_size_down(zeroes_done, BytesPerInt); |
duke@435 | 3993 | // if it is the last unused 4 bytes of an instance, forget about it |
duke@435 | 3994 | intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); |
duke@435 | 3995 | if (zeroes_done + BytesPerLong >= size_limit) { |
duke@435 | 3996 | assert(allocation() != NULL, ""); |
never@2346 | 3997 | if (allocation()->Opcode() == Op_Allocate) { |
never@2346 | 3998 | Node* klass_node = allocation()->in(AllocateNode::KlassNode); |
never@2346 | 3999 | ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); |
never@2346 | 4000 | if (zeroes_done == k->layout_helper()) |
never@2346 | 4001 | zeroes_done = size_limit; |
never@2346 | 4002 | } |
duke@435 | 4003 | } |
duke@435 | 4004 | if (zeroes_done < size_limit) { |
duke@435 | 4005 | rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, |
duke@435 | 4006 | zeroes_done, size_in_bytes, phase); |
duke@435 | 4007 | } |
duke@435 | 4008 | } |
duke@435 | 4009 | |
duke@435 | 4010 | set_complete(phase); |
duke@435 | 4011 | return rawmem; |
duke@435 | 4012 | } |
duke@435 | 4013 | |
duke@435 | 4014 | |
duke@435 | 4015 | #ifdef ASSERT |
duke@435 | 4016 | bool InitializeNode::stores_are_sane(PhaseTransform* phase) { |
duke@435 | 4017 | if (is_complete()) |
duke@435 | 4018 | return true; // stores could be anything at this point |
coleenp@548 | 4019 | assert(allocation() != NULL, "must be present"); |
coleenp@548 | 4020 | intptr_t last_off = allocation()->minimum_header_size(); |
duke@435 | 4021 | for (uint i = InitializeNode::RawStores; i < req(); i++) { |
duke@435 | 4022 | Node* st = in(i); |
duke@435 | 4023 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 4024 | if (st_off < 0) continue; // ignore dead garbage |
duke@435 | 4025 | if (last_off > st_off) { |
drchase@6680 | 4026 | tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); |
duke@435 | 4027 | this->dump(2); |
duke@435 | 4028 | assert(false, "ascending store offsets"); |
duke@435 | 4029 | return false; |
duke@435 | 4030 | } |
duke@435 | 4031 | last_off = st_off + st->as_Store()->memory_size(); |
duke@435 | 4032 | } |
duke@435 | 4033 | return true; |
duke@435 | 4034 | } |
duke@435 | 4035 | #endif //ASSERT |
duke@435 | 4036 | |
duke@435 | 4037 | |
duke@435 | 4038 | |
duke@435 | 4039 | |
duke@435 | 4040 | //============================MergeMemNode===================================== |
duke@435 | 4041 | // |
duke@435 | 4042 | // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several |
duke@435 | 4043 | // contributing store or call operations. Each contributor provides the memory |
duke@435 | 4044 | // state for a particular "alias type" (see Compile::alias_type). For example, |
duke@435 | 4045 | // if a MergeMem has an input X for alias category #6, then any memory reference |
duke@435 | 4046 | // to alias category #6 may use X as its memory state input, as an exact equivalent |
duke@435 | 4047 | // to using the MergeMem as a whole. |
duke@435 | 4048 | // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) |
duke@435 | 4049 | // |
duke@435 | 4050 | // (Here, the <N> notation gives the index of the relevant adr_type.) |
duke@435 | 4051 | // |
duke@435 | 4052 | // In one special case (and more cases in the future), alias categories overlap. |
duke@435 | 4053 | // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory |
duke@435 | 4054 | // states. Therefore, if a MergeMem has only one contributing input W for Bot, |
duke@435 | 4055 | // it is exactly equivalent to that state W: |
duke@435 | 4056 | // MergeMem(<Bot>: W) <==> W |
duke@435 | 4057 | // |
duke@435 | 4058 | // Usually, the merge has more than one input. In that case, where inputs |
duke@435 | 4059 | // overlap (i.e., one is Bot), the narrower alias type determines the memory |
duke@435 | 4060 | // state for that type, and the wider alias type (Bot) fills in everywhere else: |
duke@435 | 4061 | // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) |
duke@435 | 4062 | // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) |
duke@435 | 4063 | // |
duke@435 | 4064 | // A merge can take a "wide" memory state as one of its narrow inputs. |
duke@435 | 4065 | // This simply means that the merge observes out only the relevant parts of |
duke@435 | 4066 | // the wide input. That is, wide memory states arriving at narrow merge inputs |
duke@435 | 4067 | // are implicitly "filtered" or "sliced" as necessary. (This is rare.) |
duke@435 | 4068 | // |
duke@435 | 4069 | // These rules imply that MergeMem nodes may cascade (via their <Bot> links), |
duke@435 | 4070 | // and that memory slices "leak through": |
duke@435 | 4071 | // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) |
duke@435 | 4072 | // |
duke@435 | 4073 | // But, in such a cascade, repeated memory slices can "block the leak": |
duke@435 | 4074 | // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') |
duke@435 | 4075 | // |
duke@435 | 4076 | // In the last example, Y is not part of the combined memory state of the |
duke@435 | 4077 | // outermost MergeMem. The system must, of course, prevent unschedulable |
duke@435 | 4078 | // memory states from arising, so you can be sure that the state Y is somehow |
duke@435 | 4079 | // a precursor to state Y'. |
duke@435 | 4080 | // |
duke@435 | 4081 | // |
duke@435 | 4082 | // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array |
duke@435 | 4083 | // of each MergeMemNode array are exactly the numerical alias indexes, including |
duke@435 | 4084 | // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions |
duke@435 | 4085 | // Compile::alias_type (and kin) produce and manage these indexes. |
duke@435 | 4086 | // |
duke@435 | 4087 | // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. |
duke@435 | 4088 | // (Note that this provides quick access to the top node inside MergeMem methods, |
duke@435 | 4089 | // without the need to reach out via TLS to Compile::current.) |
duke@435 | 4090 | // |
duke@435 | 4091 | // As a consequence of what was just described, a MergeMem that represents a full |
duke@435 | 4092 | // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, |
duke@435 | 4093 | // containing all alias categories. |
duke@435 | 4094 | // |
duke@435 | 4095 | // MergeMem nodes never (?) have control inputs, so in(0) is NULL. |
duke@435 | 4096 | // |
duke@435 | 4097 | // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either |
duke@435 | 4098 | // a memory state for the alias type <N>, or else the top node, meaning that |
duke@435 | 4099 | // there is no particular input for that alias type. Note that the length of |
duke@435 | 4100 | // a MergeMem is variable, and may be extended at any time to accommodate new |
duke@435 | 4101 | // memory states at larger alias indexes. When merges grow, they are of course |
duke@435 | 4102 | // filled with "top" in the unused in() positions. |
duke@435 | 4103 | // |
duke@435 | 4104 | // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. |
duke@435 | 4105 | // (Top was chosen because it works smoothly with passes like GCM.) |
duke@435 | 4106 | // |
duke@435 | 4107 | // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is |
duke@435 | 4108 | // the type of random VM bits like TLS references.) Since it is always the |
duke@435 | 4109 | // first non-Bot memory slice, some low-level loops use it to initialize an |
duke@435 | 4110 | // index variable: for (i = AliasIdxRaw; i < req(); i++). |
duke@435 | 4111 | // |
duke@435 | 4112 | // |
duke@435 | 4113 | // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns |
duke@435 | 4114 | // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns |
duke@435 | 4115 | // the memory state for alias type <N>, or (if there is no particular slice at <N>, |
duke@435 | 4116 | // it returns the base memory. To prevent bugs, memory_at does not accept <Top> |
duke@435 | 4117 | // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over |
duke@435 | 4118 | // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. |
duke@435 | 4119 | // |
duke@435 | 4120 | // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't |
duke@435 | 4121 | // really that different from the other memory inputs. An abbreviation called |
duke@435 | 4122 | // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. |
duke@435 | 4123 | // |
duke@435 | 4124 | // |
duke@435 | 4125 | // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent |
duke@435 | 4126 | // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi |
duke@435 | 4127 | // that "emerges though" the base memory will be marked as excluding the alias types |
duke@435 | 4128 | // of the other (narrow-memory) copies which "emerged through" the narrow edges: |
duke@435 | 4129 | // |
duke@435 | 4130 | // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) |
duke@435 | 4131 | // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) |
duke@435 | 4132 | // |
duke@435 | 4133 | // This strange "subtraction" effect is necessary to ensure IGVN convergence. |
duke@435 | 4134 | // (It is currently unimplemented.) As you can see, the resulting merge is |
duke@435 | 4135 | // actually a disjoint union of memory states, rather than an overlay. |
duke@435 | 4136 | // |
duke@435 | 4137 | |
duke@435 | 4138 | //------------------------------MergeMemNode----------------------------------- |
duke@435 | 4139 | Node* MergeMemNode::make_empty_memory() { |
duke@435 | 4140 | Node* empty_memory = (Node*) Compile::current()->top(); |
duke@435 | 4141 | assert(empty_memory->is_top(), "correct sentinel identity"); |
duke@435 | 4142 | return empty_memory; |
duke@435 | 4143 | } |
duke@435 | 4144 | |
duke@435 | 4145 | MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { |
duke@435 | 4146 | init_class_id(Class_MergeMem); |
duke@435 | 4147 | // all inputs are nullified in Node::Node(int) |
duke@435 | 4148 | // set_input(0, NULL); // no control input |
duke@435 | 4149 | |
duke@435 | 4150 | // Initialize the edges uniformly to top, for starters. |
duke@435 | 4151 | Node* empty_mem = make_empty_memory(); |
duke@435 | 4152 | for (uint i = Compile::AliasIdxTop; i < req(); i++) { |
duke@435 | 4153 | init_req(i,empty_mem); |
duke@435 | 4154 | } |
duke@435 | 4155 | assert(empty_memory() == empty_mem, ""); |
duke@435 | 4156 | |
duke@435 | 4157 | if( new_base != NULL && new_base->is_MergeMem() ) { |
duke@435 | 4158 | MergeMemNode* mdef = new_base->as_MergeMem(); |
duke@435 | 4159 | assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); |
duke@435 | 4160 | for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { |
duke@435 | 4161 | mms.set_memory(mms.memory2()); |
duke@435 | 4162 | } |
duke@435 | 4163 | assert(base_memory() == mdef->base_memory(), ""); |
duke@435 | 4164 | } else { |
duke@435 | 4165 | set_base_memory(new_base); |
duke@435 | 4166 | } |
duke@435 | 4167 | } |
duke@435 | 4168 | |
duke@435 | 4169 | // Make a new, untransformed MergeMem with the same base as 'mem'. |
duke@435 | 4170 | // If mem is itself a MergeMem, populate the result with the same edges. |
duke@435 | 4171 | MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { |
kvn@4115 | 4172 | return new(C) MergeMemNode(mem); |
duke@435 | 4173 | } |
duke@435 | 4174 | |
duke@435 | 4175 | //------------------------------cmp-------------------------------------------- |
duke@435 | 4176 | uint MergeMemNode::hash() const { return NO_HASH; } |
duke@435 | 4177 | uint MergeMemNode::cmp( const Node &n ) const { |
duke@435 | 4178 | return (&n == this); // Always fail except on self |
duke@435 | 4179 | } |
duke@435 | 4180 | |
duke@435 | 4181 | //------------------------------Identity--------------------------------------- |
duke@435 | 4182 | Node* MergeMemNode::Identity(PhaseTransform *phase) { |
duke@435 | 4183 | // Identity if this merge point does not record any interesting memory |
duke@435 | 4184 | // disambiguations. |
duke@435 | 4185 | Node* base_mem = base_memory(); |
duke@435 | 4186 | Node* empty_mem = empty_memory(); |
duke@435 | 4187 | if (base_mem != empty_mem) { // Memory path is not dead? |
duke@435 | 4188 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 4189 | Node* mem = in(i); |
duke@435 | 4190 | if (mem != empty_mem && mem != base_mem) { |
duke@435 | 4191 | return this; // Many memory splits; no change |
duke@435 | 4192 | } |
duke@435 | 4193 | } |
duke@435 | 4194 | } |
duke@435 | 4195 | return base_mem; // No memory splits; ID on the one true input |
duke@435 | 4196 | } |
duke@435 | 4197 | |
duke@435 | 4198 | //------------------------------Ideal------------------------------------------ |
duke@435 | 4199 | // This method is invoked recursively on chains of MergeMem nodes |
duke@435 | 4200 | Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 4201 | // Remove chain'd MergeMems |
duke@435 | 4202 | // |
duke@435 | 4203 | // This is delicate, because the each "in(i)" (i >= Raw) is interpreted |
duke@435 | 4204 | // relative to the "in(Bot)". Since we are patching both at the same time, |
duke@435 | 4205 | // we have to be careful to read each "in(i)" relative to the old "in(Bot)", |
duke@435 | 4206 | // but rewrite each "in(i)" relative to the new "in(Bot)". |
duke@435 | 4207 | Node *progress = NULL; |
duke@435 | 4208 | |
duke@435 | 4209 | |
duke@435 | 4210 | Node* old_base = base_memory(); |
duke@435 | 4211 | Node* empty_mem = empty_memory(); |
duke@435 | 4212 | if (old_base == empty_mem) |
duke@435 | 4213 | return NULL; // Dead memory path. |
duke@435 | 4214 | |
duke@435 | 4215 | MergeMemNode* old_mbase; |
duke@435 | 4216 | if (old_base != NULL && old_base->is_MergeMem()) |
duke@435 | 4217 | old_mbase = old_base->as_MergeMem(); |
duke@435 | 4218 | else |
duke@435 | 4219 | old_mbase = NULL; |
duke@435 | 4220 | Node* new_base = old_base; |
duke@435 | 4221 | |
duke@435 | 4222 | // simplify stacked MergeMems in base memory |
duke@435 | 4223 | if (old_mbase) new_base = old_mbase->base_memory(); |
duke@435 | 4224 | |
duke@435 | 4225 | // the base memory might contribute new slices beyond my req() |
duke@435 | 4226 | if (old_mbase) grow_to_match(old_mbase); |
duke@435 | 4227 | |
duke@435 | 4228 | // Look carefully at the base node if it is a phi. |
duke@435 | 4229 | PhiNode* phi_base; |
duke@435 | 4230 | if (new_base != NULL && new_base->is_Phi()) |
duke@435 | 4231 | phi_base = new_base->as_Phi(); |
duke@435 | 4232 | else |
duke@435 | 4233 | phi_base = NULL; |
duke@435 | 4234 | |
duke@435 | 4235 | Node* phi_reg = NULL; |
duke@435 | 4236 | uint phi_len = (uint)-1; |
duke@435 | 4237 | if (phi_base != NULL && !phi_base->is_copy()) { |
duke@435 | 4238 | // do not examine phi if degraded to a copy |
duke@435 | 4239 | phi_reg = phi_base->region(); |
duke@435 | 4240 | phi_len = phi_base->req(); |
duke@435 | 4241 | // see if the phi is unfinished |
duke@435 | 4242 | for (uint i = 1; i < phi_len; i++) { |
duke@435 | 4243 | if (phi_base->in(i) == NULL) { |
duke@435 | 4244 | // incomplete phi; do not look at it yet! |
duke@435 | 4245 | phi_reg = NULL; |
duke@435 | 4246 | phi_len = (uint)-1; |
duke@435 | 4247 | break; |
duke@435 | 4248 | } |
duke@435 | 4249 | } |
duke@435 | 4250 | } |
duke@435 | 4251 | |
duke@435 | 4252 | // Note: We do not call verify_sparse on entry, because inputs |
duke@435 | 4253 | // can normalize to the base_memory via subsume_node or similar |
duke@435 | 4254 | // mechanisms. This method repairs that damage. |
duke@435 | 4255 | |
duke@435 | 4256 | assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); |
duke@435 | 4257 | |
duke@435 | 4258 | // Look at each slice. |
duke@435 | 4259 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 4260 | Node* old_in = in(i); |
duke@435 | 4261 | // calculate the old memory value |
duke@435 | 4262 | Node* old_mem = old_in; |
duke@435 | 4263 | if (old_mem == empty_mem) old_mem = old_base; |
duke@435 | 4264 | assert(old_mem == memory_at(i), ""); |
duke@435 | 4265 | |
duke@435 | 4266 | // maybe update (reslice) the old memory value |
duke@435 | 4267 | |
duke@435 | 4268 | // simplify stacked MergeMems |
duke@435 | 4269 | Node* new_mem = old_mem; |
duke@435 | 4270 | MergeMemNode* old_mmem; |
duke@435 | 4271 | if (old_mem != NULL && old_mem->is_MergeMem()) |
duke@435 | 4272 | old_mmem = old_mem->as_MergeMem(); |
duke@435 | 4273 | else |
duke@435 | 4274 | old_mmem = NULL; |
duke@435 | 4275 | if (old_mmem == this) { |
duke@435 | 4276 | // This can happen if loops break up and safepoints disappear. |
duke@435 | 4277 | // A merge of BotPtr (default) with a RawPtr memory derived from a |
duke@435 | 4278 | // safepoint can be rewritten to a merge of the same BotPtr with |
duke@435 | 4279 | // the BotPtr phi coming into the loop. If that phi disappears |
duke@435 | 4280 | // also, we can end up with a self-loop of the mergemem. |
duke@435 | 4281 | // In general, if loops degenerate and memory effects disappear, |
duke@435 | 4282 | // a mergemem can be left looking at itself. This simply means |
duke@435 | 4283 | // that the mergemem's default should be used, since there is |
duke@435 | 4284 | // no longer any apparent effect on this slice. |
duke@435 | 4285 | // Note: If a memory slice is a MergeMem cycle, it is unreachable |
duke@435 | 4286 | // from start. Update the input to TOP. |
duke@435 | 4287 | new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; |
duke@435 | 4288 | } |
duke@435 | 4289 | else if (old_mmem != NULL) { |
duke@435 | 4290 | new_mem = old_mmem->memory_at(i); |
duke@435 | 4291 | } |
twisti@1040 | 4292 | // else preceding memory was not a MergeMem |
duke@435 | 4293 | |
duke@435 | 4294 | // replace equivalent phis (unfortunately, they do not GVN together) |
duke@435 | 4295 | if (new_mem != NULL && new_mem != new_base && |
duke@435 | 4296 | new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { |
duke@435 | 4297 | if (new_mem->is_Phi()) { |
duke@435 | 4298 | PhiNode* phi_mem = new_mem->as_Phi(); |
duke@435 | 4299 | for (uint i = 1; i < phi_len; i++) { |
duke@435 | 4300 | if (phi_base->in(i) != phi_mem->in(i)) { |
duke@435 | 4301 | phi_mem = NULL; |
duke@435 | 4302 | break; |
duke@435 | 4303 | } |
duke@435 | 4304 | } |
duke@435 | 4305 | if (phi_mem != NULL) { |
duke@435 | 4306 | // equivalent phi nodes; revert to the def |
duke@435 | 4307 | new_mem = new_base; |
duke@435 | 4308 | } |
duke@435 | 4309 | } |
duke@435 | 4310 | } |
duke@435 | 4311 | |
duke@435 | 4312 | // maybe store down a new value |
duke@435 | 4313 | Node* new_in = new_mem; |
duke@435 | 4314 | if (new_in == new_base) new_in = empty_mem; |
duke@435 | 4315 | |
duke@435 | 4316 | if (new_in != old_in) { |
duke@435 | 4317 | // Warning: Do not combine this "if" with the previous "if" |
duke@435 | 4318 | // A memory slice might have be be rewritten even if it is semantically |
duke@435 | 4319 | // unchanged, if the base_memory value has changed. |
duke@435 | 4320 | set_req(i, new_in); |
duke@435 | 4321 | progress = this; // Report progress |
duke@435 | 4322 | } |
duke@435 | 4323 | } |
duke@435 | 4324 | |
duke@435 | 4325 | if (new_base != old_base) { |
duke@435 | 4326 | set_req(Compile::AliasIdxBot, new_base); |
duke@435 | 4327 | // Don't use set_base_memory(new_base), because we need to update du. |
duke@435 | 4328 | assert(base_memory() == new_base, ""); |
duke@435 | 4329 | progress = this; |
duke@435 | 4330 | } |
duke@435 | 4331 | |
duke@435 | 4332 | if( base_memory() == this ) { |
duke@435 | 4333 | // a self cycle indicates this memory path is dead |
duke@435 | 4334 | set_req(Compile::AliasIdxBot, empty_mem); |
duke@435 | 4335 | } |
duke@435 | 4336 | |
duke@435 | 4337 | // Resolve external cycles by calling Ideal on a MergeMem base_memory |
duke@435 | 4338 | // Recursion must occur after the self cycle check above |
duke@435 | 4339 | if( base_memory()->is_MergeMem() ) { |
duke@435 | 4340 | MergeMemNode *new_mbase = base_memory()->as_MergeMem(); |
duke@435 | 4341 | Node *m = phase->transform(new_mbase); // Rollup any cycles |
duke@435 | 4342 | if( m != NULL && (m->is_top() || |
duke@435 | 4343 | m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { |
duke@435 | 4344 | // propagate rollup of dead cycle to self |
duke@435 | 4345 | set_req(Compile::AliasIdxBot, empty_mem); |
duke@435 | 4346 | } |
duke@435 | 4347 | } |
duke@435 | 4348 | |
duke@435 | 4349 | if( base_memory() == empty_mem ) { |
duke@435 | 4350 | progress = this; |
duke@435 | 4351 | // Cut inputs during Parse phase only. |
duke@435 | 4352 | // During Optimize phase a dead MergeMem node will be subsumed by Top. |
duke@435 | 4353 | if( !can_reshape ) { |
duke@435 | 4354 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 4355 | if( in(i) != empty_mem ) { set_req(i, empty_mem); } |
duke@435 | 4356 | } |
duke@435 | 4357 | } |
duke@435 | 4358 | } |
duke@435 | 4359 | |
duke@435 | 4360 | if( !progress && base_memory()->is_Phi() && can_reshape ) { |
duke@435 | 4361 | // Check if PhiNode::Ideal's "Split phis through memory merges" |
duke@435 | 4362 | // transform should be attempted. Look for this->phi->this cycle. |
duke@435 | 4363 | uint merge_width = req(); |
duke@435 | 4364 | if (merge_width > Compile::AliasIdxRaw) { |
duke@435 | 4365 | PhiNode* phi = base_memory()->as_Phi(); |
duke@435 | 4366 | for( uint i = 1; i < phi->req(); ++i ) {// For all paths in |
duke@435 | 4367 | if (phi->in(i) == this) { |
duke@435 | 4368 | phase->is_IterGVN()->_worklist.push(phi); |
duke@435 | 4369 | break; |
duke@435 | 4370 | } |
duke@435 | 4371 | } |
duke@435 | 4372 | } |
duke@435 | 4373 | } |
duke@435 | 4374 | |
kvn@499 | 4375 | assert(progress || verify_sparse(), "please, no dups of base"); |
duke@435 | 4376 | return progress; |
duke@435 | 4377 | } |
duke@435 | 4378 | |
duke@435 | 4379 | //-------------------------set_base_memory------------------------------------- |
duke@435 | 4380 | void MergeMemNode::set_base_memory(Node *new_base) { |
duke@435 | 4381 | Node* empty_mem = empty_memory(); |
duke@435 | 4382 | set_req(Compile::AliasIdxBot, new_base); |
duke@435 | 4383 | assert(memory_at(req()) == new_base, "must set default memory"); |
duke@435 | 4384 | // Clear out other occurrences of new_base: |
duke@435 | 4385 | if (new_base != empty_mem) { |
duke@435 | 4386 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 4387 | if (in(i) == new_base) set_req(i, empty_mem); |
duke@435 | 4388 | } |
duke@435 | 4389 | } |
duke@435 | 4390 | } |
duke@435 | 4391 | |
duke@435 | 4392 | //------------------------------out_RegMask------------------------------------ |
duke@435 | 4393 | const RegMask &MergeMemNode::out_RegMask() const { |
duke@435 | 4394 | return RegMask::Empty; |
duke@435 | 4395 | } |
duke@435 | 4396 | |
duke@435 | 4397 | //------------------------------dump_spec-------------------------------------- |
duke@435 | 4398 | #ifndef PRODUCT |
duke@435 | 4399 | void MergeMemNode::dump_spec(outputStream *st) const { |
duke@435 | 4400 | st->print(" {"); |
duke@435 | 4401 | Node* base_mem = base_memory(); |
duke@435 | 4402 | for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { |
duke@435 | 4403 | Node* mem = memory_at(i); |
duke@435 | 4404 | if (mem == base_mem) { st->print(" -"); continue; } |
duke@435 | 4405 | st->print( " N%d:", mem->_idx ); |
duke@435 | 4406 | Compile::current()->get_adr_type(i)->dump_on(st); |
duke@435 | 4407 | } |
duke@435 | 4408 | st->print(" }"); |
duke@435 | 4409 | } |
duke@435 | 4410 | #endif // !PRODUCT |
duke@435 | 4411 | |
duke@435 | 4412 | |
duke@435 | 4413 | #ifdef ASSERT |
duke@435 | 4414 | static bool might_be_same(Node* a, Node* b) { |
duke@435 | 4415 | if (a == b) return true; |
duke@435 | 4416 | if (!(a->is_Phi() || b->is_Phi())) return false; |
duke@435 | 4417 | // phis shift around during optimization |
duke@435 | 4418 | return true; // pretty stupid... |
duke@435 | 4419 | } |
duke@435 | 4420 | |
duke@435 | 4421 | // verify a narrow slice (either incoming or outgoing) |
duke@435 | 4422 | static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { |
duke@435 | 4423 | if (!VerifyAliases) return; // don't bother to verify unless requested |
duke@435 | 4424 | if (is_error_reported()) return; // muzzle asserts when debugging an error |
duke@435 | 4425 | if (Node::in_dump()) return; // muzzle asserts when printing |
duke@435 | 4426 | assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); |
duke@435 | 4427 | assert(n != NULL, ""); |
duke@435 | 4428 | // Elide intervening MergeMem's |
duke@435 | 4429 | while (n->is_MergeMem()) { |
duke@435 | 4430 | n = n->as_MergeMem()->memory_at(alias_idx); |
duke@435 | 4431 | } |
duke@435 | 4432 | Compile* C = Compile::current(); |
duke@435 | 4433 | const TypePtr* n_adr_type = n->adr_type(); |
duke@435 | 4434 | if (n == m->empty_memory()) { |
duke@435 | 4435 | // Implicit copy of base_memory() |
duke@435 | 4436 | } else if (n_adr_type != TypePtr::BOTTOM) { |
duke@435 | 4437 | assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); |
duke@435 | 4438 | assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); |
duke@435 | 4439 | } else { |
duke@435 | 4440 | // A few places like make_runtime_call "know" that VM calls are narrow, |
duke@435 | 4441 | // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. |
duke@435 | 4442 | bool expected_wide_mem = false; |
duke@435 | 4443 | if (n == m->base_memory()) { |
duke@435 | 4444 | expected_wide_mem = true; |
duke@435 | 4445 | } else if (alias_idx == Compile::AliasIdxRaw || |
duke@435 | 4446 | n == m->memory_at(Compile::AliasIdxRaw)) { |
duke@435 | 4447 | expected_wide_mem = true; |
duke@435 | 4448 | } else if (!C->alias_type(alias_idx)->is_rewritable()) { |
duke@435 | 4449 | // memory can "leak through" calls on channels that |
duke@435 | 4450 | // are write-once. Allow this also. |
duke@435 | 4451 | expected_wide_mem = true; |
duke@435 | 4452 | } |
duke@435 | 4453 | assert(expected_wide_mem, "expected narrow slice replacement"); |
duke@435 | 4454 | } |
duke@435 | 4455 | } |
duke@435 | 4456 | #else // !ASSERT |
ccheung@5259 | 4457 | #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op |
duke@435 | 4458 | #endif |
duke@435 | 4459 | |
duke@435 | 4460 | |
duke@435 | 4461 | //-----------------------------memory_at--------------------------------------- |
duke@435 | 4462 | Node* MergeMemNode::memory_at(uint alias_idx) const { |
duke@435 | 4463 | assert(alias_idx >= Compile::AliasIdxRaw || |
duke@435 | 4464 | alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, |
duke@435 | 4465 | "must avoid base_memory and AliasIdxTop"); |
duke@435 | 4466 | |
duke@435 | 4467 | // Otherwise, it is a narrow slice. |
duke@435 | 4468 | Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); |
duke@435 | 4469 | Compile *C = Compile::current(); |
duke@435 | 4470 | if (is_empty_memory(n)) { |
duke@435 | 4471 | // the array is sparse; empty slots are the "top" node |
duke@435 | 4472 | n = base_memory(); |
duke@435 | 4473 | assert(Node::in_dump() |
duke@435 | 4474 | || n == NULL || n->bottom_type() == Type::TOP |
kvn@2599 | 4475 | || n->adr_type() == NULL // address is TOP |
duke@435 | 4476 | || n->adr_type() == TypePtr::BOTTOM |
duke@435 | 4477 | || n->adr_type() == TypeRawPtr::BOTTOM |
duke@435 | 4478 | || Compile::current()->AliasLevel() == 0, |
duke@435 | 4479 | "must be a wide memory"); |
duke@435 | 4480 | // AliasLevel == 0 if we are organizing the memory states manually. |
duke@435 | 4481 | // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. |
duke@435 | 4482 | } else { |
duke@435 | 4483 | // make sure the stored slice is sane |
duke@435 | 4484 | #ifdef ASSERT |
duke@435 | 4485 | if (is_error_reported() || Node::in_dump()) { |
duke@435 | 4486 | } else if (might_be_same(n, base_memory())) { |
duke@435 | 4487 | // Give it a pass: It is a mostly harmless repetition of the base. |
duke@435 | 4488 | // This can arise normally from node subsumption during optimization. |
duke@435 | 4489 | } else { |
duke@435 | 4490 | verify_memory_slice(this, alias_idx, n); |
duke@435 | 4491 | } |
duke@435 | 4492 | #endif |
duke@435 | 4493 | } |
duke@435 | 4494 | return n; |
duke@435 | 4495 | } |
duke@435 | 4496 | |
duke@435 | 4497 | //---------------------------set_memory_at------------------------------------- |
duke@435 | 4498 | void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { |
duke@435 | 4499 | verify_memory_slice(this, alias_idx, n); |
duke@435 | 4500 | Node* empty_mem = empty_memory(); |
duke@435 | 4501 | if (n == base_memory()) n = empty_mem; // collapse default |
duke@435 | 4502 | uint need_req = alias_idx+1; |
duke@435 | 4503 | if (req() < need_req) { |
duke@435 | 4504 | if (n == empty_mem) return; // already the default, so do not grow me |
duke@435 | 4505 | // grow the sparse array |
duke@435 | 4506 | do { |
duke@435 | 4507 | add_req(empty_mem); |
duke@435 | 4508 | } while (req() < need_req); |
duke@435 | 4509 | } |
duke@435 | 4510 | set_req( alias_idx, n ); |
duke@435 | 4511 | } |
duke@435 | 4512 | |
duke@435 | 4513 | |
duke@435 | 4514 | |
duke@435 | 4515 | //--------------------------iteration_setup------------------------------------ |
duke@435 | 4516 | void MergeMemNode::iteration_setup(const MergeMemNode* other) { |
duke@435 | 4517 | if (other != NULL) { |
duke@435 | 4518 | grow_to_match(other); |
duke@435 | 4519 | // invariant: the finite support of mm2 is within mm->req() |
duke@435 | 4520 | #ifdef ASSERT |
duke@435 | 4521 | for (uint i = req(); i < other->req(); i++) { |
duke@435 | 4522 | assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); |
duke@435 | 4523 | } |
duke@435 | 4524 | #endif |
duke@435 | 4525 | } |
duke@435 | 4526 | // Replace spurious copies of base_memory by top. |
duke@435 | 4527 | Node* base_mem = base_memory(); |
duke@435 | 4528 | if (base_mem != NULL && !base_mem->is_top()) { |
duke@435 | 4529 | for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { |
duke@435 | 4530 | if (in(i) == base_mem) |
duke@435 | 4531 | set_req(i, empty_memory()); |
duke@435 | 4532 | } |
duke@435 | 4533 | } |
duke@435 | 4534 | } |
duke@435 | 4535 | |
duke@435 | 4536 | //---------------------------grow_to_match------------------------------------- |
duke@435 | 4537 | void MergeMemNode::grow_to_match(const MergeMemNode* other) { |
duke@435 | 4538 | Node* empty_mem = empty_memory(); |
duke@435 | 4539 | assert(other->is_empty_memory(empty_mem), "consistent sentinels"); |
duke@435 | 4540 | // look for the finite support of the other memory |
duke@435 | 4541 | for (uint i = other->req(); --i >= req(); ) { |
duke@435 | 4542 | if (other->in(i) != empty_mem) { |
duke@435 | 4543 | uint new_len = i+1; |
duke@435 | 4544 | while (req() < new_len) add_req(empty_mem); |
duke@435 | 4545 | break; |
duke@435 | 4546 | } |
duke@435 | 4547 | } |
duke@435 | 4548 | } |
duke@435 | 4549 | |
duke@435 | 4550 | //---------------------------verify_sparse------------------------------------- |
duke@435 | 4551 | #ifndef PRODUCT |
duke@435 | 4552 | bool MergeMemNode::verify_sparse() const { |
duke@435 | 4553 | assert(is_empty_memory(make_empty_memory()), "sane sentinel"); |
duke@435 | 4554 | Node* base_mem = base_memory(); |
duke@435 | 4555 | // The following can happen in degenerate cases, since empty==top. |
duke@435 | 4556 | if (is_empty_memory(base_mem)) return true; |
duke@435 | 4557 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 4558 | assert(in(i) != NULL, "sane slice"); |
duke@435 | 4559 | if (in(i) == base_mem) return false; // should have been the sentinel value! |
duke@435 | 4560 | } |
duke@435 | 4561 | return true; |
duke@435 | 4562 | } |
duke@435 | 4563 | |
duke@435 | 4564 | bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { |
duke@435 | 4565 | Node* n; |
duke@435 | 4566 | n = mm->in(idx); |
duke@435 | 4567 | if (mem == n) return true; // might be empty_memory() |
duke@435 | 4568 | n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); |
duke@435 | 4569 | if (mem == n) return true; |
duke@435 | 4570 | while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { |
duke@435 | 4571 | if (mem == n) return true; |
duke@435 | 4572 | if (n == NULL) break; |
duke@435 | 4573 | } |
duke@435 | 4574 | return false; |
duke@435 | 4575 | } |
duke@435 | 4576 | #endif // !PRODUCT |