Sat, 01 Dec 2007 00:00:00 +0000
Initial load
duke@435 | 1 | /* |
duke@435 | 2 | * Copyright 1997-2007 Sun Microsystems, Inc. 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 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | // Portions of code courtesy of Clifford Click |
duke@435 | 26 | |
duke@435 | 27 | // Optimization - Graph Style |
duke@435 | 28 | |
duke@435 | 29 | #include "incls/_precompiled.incl" |
duke@435 | 30 | #include "incls/_memnode.cpp.incl" |
duke@435 | 31 | |
duke@435 | 32 | //============================================================================= |
duke@435 | 33 | uint MemNode::size_of() const { return sizeof(*this); } |
duke@435 | 34 | |
duke@435 | 35 | const TypePtr *MemNode::adr_type() const { |
duke@435 | 36 | Node* adr = in(Address); |
duke@435 | 37 | const TypePtr* cross_check = NULL; |
duke@435 | 38 | DEBUG_ONLY(cross_check = _adr_type); |
duke@435 | 39 | return calculate_adr_type(adr->bottom_type(), cross_check); |
duke@435 | 40 | } |
duke@435 | 41 | |
duke@435 | 42 | #ifndef PRODUCT |
duke@435 | 43 | void MemNode::dump_spec(outputStream *st) const { |
duke@435 | 44 | if (in(Address) == NULL) return; // node is dead |
duke@435 | 45 | #ifndef ASSERT |
duke@435 | 46 | // fake the missing field |
duke@435 | 47 | const TypePtr* _adr_type = NULL; |
duke@435 | 48 | if (in(Address) != NULL) |
duke@435 | 49 | _adr_type = in(Address)->bottom_type()->isa_ptr(); |
duke@435 | 50 | #endif |
duke@435 | 51 | dump_adr_type(this, _adr_type, st); |
duke@435 | 52 | |
duke@435 | 53 | Compile* C = Compile::current(); |
duke@435 | 54 | if( C->alias_type(_adr_type)->is_volatile() ) |
duke@435 | 55 | st->print(" Volatile!"); |
duke@435 | 56 | } |
duke@435 | 57 | |
duke@435 | 58 | void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { |
duke@435 | 59 | st->print(" @"); |
duke@435 | 60 | if (adr_type == NULL) { |
duke@435 | 61 | st->print("NULL"); |
duke@435 | 62 | } else { |
duke@435 | 63 | adr_type->dump_on(st); |
duke@435 | 64 | Compile* C = Compile::current(); |
duke@435 | 65 | Compile::AliasType* atp = NULL; |
duke@435 | 66 | if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); |
duke@435 | 67 | if (atp == NULL) |
duke@435 | 68 | st->print(", idx=?\?;"); |
duke@435 | 69 | else if (atp->index() == Compile::AliasIdxBot) |
duke@435 | 70 | st->print(", idx=Bot;"); |
duke@435 | 71 | else if (atp->index() == Compile::AliasIdxTop) |
duke@435 | 72 | st->print(", idx=Top;"); |
duke@435 | 73 | else if (atp->index() == Compile::AliasIdxRaw) |
duke@435 | 74 | st->print(", idx=Raw;"); |
duke@435 | 75 | else { |
duke@435 | 76 | ciField* field = atp->field(); |
duke@435 | 77 | if (field) { |
duke@435 | 78 | st->print(", name="); |
duke@435 | 79 | field->print_name_on(st); |
duke@435 | 80 | } |
duke@435 | 81 | st->print(", idx=%d;", atp->index()); |
duke@435 | 82 | } |
duke@435 | 83 | } |
duke@435 | 84 | } |
duke@435 | 85 | |
duke@435 | 86 | extern void print_alias_types(); |
duke@435 | 87 | |
duke@435 | 88 | #endif |
duke@435 | 89 | |
duke@435 | 90 | //--------------------------Ideal_common--------------------------------------- |
duke@435 | 91 | // Look for degenerate control and memory inputs. Bypass MergeMem inputs. |
duke@435 | 92 | // Unhook non-raw memories from complete (macro-expanded) initializations. |
duke@435 | 93 | Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 94 | // If our control input is a dead region, kill all below the region |
duke@435 | 95 | Node *ctl = in(MemNode::Control); |
duke@435 | 96 | if (ctl && remove_dead_region(phase, can_reshape)) |
duke@435 | 97 | return this; |
duke@435 | 98 | |
duke@435 | 99 | // Ignore if memory is dead, or self-loop |
duke@435 | 100 | Node *mem = in(MemNode::Memory); |
duke@435 | 101 | if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL |
duke@435 | 102 | assert( mem != this, "dead loop in MemNode::Ideal" ); |
duke@435 | 103 | |
duke@435 | 104 | Node *address = in(MemNode::Address); |
duke@435 | 105 | const Type *t_adr = phase->type( address ); |
duke@435 | 106 | if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL |
duke@435 | 107 | |
duke@435 | 108 | // Avoid independent memory operations |
duke@435 | 109 | Node* old_mem = mem; |
duke@435 | 110 | |
duke@435 | 111 | if (mem->is_Proj() && mem->in(0)->is_Initialize()) { |
duke@435 | 112 | InitializeNode* init = mem->in(0)->as_Initialize(); |
duke@435 | 113 | if (init->is_complete()) { // i.e., after macro expansion |
duke@435 | 114 | const TypePtr* tp = t_adr->is_ptr(); |
duke@435 | 115 | uint alias_idx = phase->C->get_alias_index(tp); |
duke@435 | 116 | // Free this slice from the init. It was hooked, temporarily, |
duke@435 | 117 | // by GraphKit::set_output_for_allocation. |
duke@435 | 118 | if (alias_idx > Compile::AliasIdxRaw) { |
duke@435 | 119 | mem = init->memory(alias_idx); |
duke@435 | 120 | // ...but not with the raw-pointer slice. |
duke@435 | 121 | } |
duke@435 | 122 | } |
duke@435 | 123 | } |
duke@435 | 124 | |
duke@435 | 125 | if (mem->is_MergeMem()) { |
duke@435 | 126 | MergeMemNode* mmem = mem->as_MergeMem(); |
duke@435 | 127 | const TypePtr *tp = t_adr->is_ptr(); |
duke@435 | 128 | uint alias_idx = phase->C->get_alias_index(tp); |
duke@435 | 129 | #ifdef ASSERT |
duke@435 | 130 | { |
duke@435 | 131 | // Check that current type is consistent with the alias index used during graph construction |
duke@435 | 132 | assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); |
duke@435 | 133 | const TypePtr *adr_t = adr_type(); |
duke@435 | 134 | bool consistent = adr_t == NULL || adr_t->empty() || phase->C->must_alias(adr_t, alias_idx ); |
duke@435 | 135 | // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] |
duke@435 | 136 | if( !consistent && adr_t != NULL && !adr_t->empty() && |
duke@435 | 137 | tp->isa_aryptr() && tp->offset() == Type::OffsetBot && |
duke@435 | 138 | adr_t->isa_aryptr() && adr_t->offset() != Type::OffsetBot && |
duke@435 | 139 | ( adr_t->offset() == arrayOopDesc::length_offset_in_bytes() || |
duke@435 | 140 | adr_t->offset() == oopDesc::klass_offset_in_bytes() || |
duke@435 | 141 | adr_t->offset() == oopDesc::mark_offset_in_bytes() ) ) { |
duke@435 | 142 | // don't assert if it is dead code. |
duke@435 | 143 | consistent = true; |
duke@435 | 144 | } |
duke@435 | 145 | if( !consistent ) { |
duke@435 | 146 | tty->print("alias_idx==%d, adr_type()==", alias_idx); if( adr_t == NULL ) { tty->print("NULL"); } else { adr_t->dump(); } |
duke@435 | 147 | tty->cr(); |
duke@435 | 148 | print_alias_types(); |
duke@435 | 149 | assert(consistent, "adr_type must match alias idx"); |
duke@435 | 150 | } |
duke@435 | 151 | } |
duke@435 | 152 | #endif |
duke@435 | 153 | // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally |
duke@435 | 154 | // means an array I have not precisely typed yet. Do not do any |
duke@435 | 155 | // alias stuff with it any time soon. |
duke@435 | 156 | const TypeInstPtr *tinst = tp->isa_instptr(); |
duke@435 | 157 | if( tp->base() != Type::AnyPtr && |
duke@435 | 158 | !(tinst && |
duke@435 | 159 | tinst->klass()->is_java_lang_Object() && |
duke@435 | 160 | tinst->offset() == Type::OffsetBot) ) { |
duke@435 | 161 | // compress paths and change unreachable cycles to TOP |
duke@435 | 162 | // If not, we can update the input infinitely along a MergeMem cycle |
duke@435 | 163 | // Equivalent code in PhiNode::Ideal |
duke@435 | 164 | Node* m = phase->transform(mmem); |
duke@435 | 165 | // If tranformed to a MergeMem, get the desired slice |
duke@435 | 166 | // Otherwise the returned node represents memory for every slice |
duke@435 | 167 | mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; |
duke@435 | 168 | // Update input if it is progress over what we have now |
duke@435 | 169 | } |
duke@435 | 170 | } |
duke@435 | 171 | |
duke@435 | 172 | if (mem != old_mem) { |
duke@435 | 173 | set_req(MemNode::Memory, mem); |
duke@435 | 174 | return this; |
duke@435 | 175 | } |
duke@435 | 176 | |
duke@435 | 177 | // let the subclass continue analyzing... |
duke@435 | 178 | return NULL; |
duke@435 | 179 | } |
duke@435 | 180 | |
duke@435 | 181 | // Helper function for proving some simple control dominations. |
duke@435 | 182 | // Attempt to prove that control input 'dom' dominates (or equals) 'sub'. |
duke@435 | 183 | // Already assumes that 'dom' is available at 'sub', and that 'sub' |
duke@435 | 184 | // is not a constant (dominated by the method's StartNode). |
duke@435 | 185 | // Used by MemNode::find_previous_store to prove that the |
duke@435 | 186 | // control input of a memory operation predates (dominates) |
duke@435 | 187 | // an allocation it wants to look past. |
duke@435 | 188 | bool MemNode::detect_dominating_control(Node* dom, Node* sub) { |
duke@435 | 189 | if (dom == NULL) return false; |
duke@435 | 190 | if (dom->is_Proj()) dom = dom->in(0); |
duke@435 | 191 | if (dom->is_Start()) return true; // anything inside the method |
duke@435 | 192 | if (dom->is_Root()) return true; // dom 'controls' a constant |
duke@435 | 193 | int cnt = 20; // detect cycle or too much effort |
duke@435 | 194 | while (sub != NULL) { // walk 'sub' up the chain to 'dom' |
duke@435 | 195 | if (--cnt < 0) return false; // in a cycle or too complex |
duke@435 | 196 | if (sub == dom) return true; |
duke@435 | 197 | if (sub->is_Start()) return false; |
duke@435 | 198 | if (sub->is_Root()) return false; |
duke@435 | 199 | Node* up = sub->in(0); |
duke@435 | 200 | if (sub == up && sub->is_Region()) { |
duke@435 | 201 | for (uint i = 1; i < sub->req(); i++) { |
duke@435 | 202 | Node* in = sub->in(i); |
duke@435 | 203 | if (in != NULL && !in->is_top() && in != sub) { |
duke@435 | 204 | up = in; break; // take any path on the way up to 'dom' |
duke@435 | 205 | } |
duke@435 | 206 | } |
duke@435 | 207 | } |
duke@435 | 208 | if (sub == up) return false; // some kind of tight cycle |
duke@435 | 209 | sub = up; |
duke@435 | 210 | } |
duke@435 | 211 | return false; |
duke@435 | 212 | } |
duke@435 | 213 | |
duke@435 | 214 | //---------------------detect_ptr_independence--------------------------------- |
duke@435 | 215 | // Used by MemNode::find_previous_store to prove that two base |
duke@435 | 216 | // pointers are never equal. |
duke@435 | 217 | // The pointers are accompanied by their associated allocations, |
duke@435 | 218 | // if any, which have been previously discovered by the caller. |
duke@435 | 219 | bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, |
duke@435 | 220 | Node* p2, AllocateNode* a2, |
duke@435 | 221 | PhaseTransform* phase) { |
duke@435 | 222 | // Attempt to prove that these two pointers cannot be aliased. |
duke@435 | 223 | // They may both manifestly be allocations, and they should differ. |
duke@435 | 224 | // Or, if they are not both allocations, they can be distinct constants. |
duke@435 | 225 | // Otherwise, one is an allocation and the other a pre-existing value. |
duke@435 | 226 | if (a1 == NULL && a2 == NULL) { // neither an allocation |
duke@435 | 227 | return (p1 != p2) && p1->is_Con() && p2->is_Con(); |
duke@435 | 228 | } else if (a1 != NULL && a2 != NULL) { // both allocations |
duke@435 | 229 | return (a1 != a2); |
duke@435 | 230 | } else if (a1 != NULL) { // one allocation a1 |
duke@435 | 231 | // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) |
duke@435 | 232 | return detect_dominating_control(p2->in(0), a1->in(0)); |
duke@435 | 233 | } else { //(a2 != NULL) // one allocation a2 |
duke@435 | 234 | return detect_dominating_control(p1->in(0), a2->in(0)); |
duke@435 | 235 | } |
duke@435 | 236 | return false; |
duke@435 | 237 | } |
duke@435 | 238 | |
duke@435 | 239 | |
duke@435 | 240 | // The logic for reordering loads and stores uses four steps: |
duke@435 | 241 | // (a) Walk carefully past stores and initializations which we |
duke@435 | 242 | // can prove are independent of this load. |
duke@435 | 243 | // (b) Observe that the next memory state makes an exact match |
duke@435 | 244 | // with self (load or store), and locate the relevant store. |
duke@435 | 245 | // (c) Ensure that, if we were to wire self directly to the store, |
duke@435 | 246 | // the optimizer would fold it up somehow. |
duke@435 | 247 | // (d) Do the rewiring, and return, depending on some other part of |
duke@435 | 248 | // the optimizer to fold up the load. |
duke@435 | 249 | // This routine handles steps (a) and (b). Steps (c) and (d) are |
duke@435 | 250 | // specific to loads and stores, so they are handled by the callers. |
duke@435 | 251 | // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) |
duke@435 | 252 | // |
duke@435 | 253 | Node* MemNode::find_previous_store(PhaseTransform* phase) { |
duke@435 | 254 | Node* ctrl = in(MemNode::Control); |
duke@435 | 255 | Node* adr = in(MemNode::Address); |
duke@435 | 256 | intptr_t offset = 0; |
duke@435 | 257 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
duke@435 | 258 | AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); |
duke@435 | 259 | |
duke@435 | 260 | if (offset == Type::OffsetBot) |
duke@435 | 261 | return NULL; // cannot unalias unless there are precise offsets |
duke@435 | 262 | |
duke@435 | 263 | intptr_t size_in_bytes = memory_size(); |
duke@435 | 264 | |
duke@435 | 265 | Node* mem = in(MemNode::Memory); // start searching here... |
duke@435 | 266 | |
duke@435 | 267 | int cnt = 50; // Cycle limiter |
duke@435 | 268 | for (;;) { // While we can dance past unrelated stores... |
duke@435 | 269 | if (--cnt < 0) break; // Caught in cycle or a complicated dance? |
duke@435 | 270 | |
duke@435 | 271 | if (mem->is_Store()) { |
duke@435 | 272 | Node* st_adr = mem->in(MemNode::Address); |
duke@435 | 273 | intptr_t st_offset = 0; |
duke@435 | 274 | Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); |
duke@435 | 275 | if (st_base == NULL) |
duke@435 | 276 | break; // inscrutable pointer |
duke@435 | 277 | if (st_offset != offset && st_offset != Type::OffsetBot) { |
duke@435 | 278 | const int MAX_STORE = BytesPerLong; |
duke@435 | 279 | if (st_offset >= offset + size_in_bytes || |
duke@435 | 280 | st_offset <= offset - MAX_STORE || |
duke@435 | 281 | st_offset <= offset - mem->as_Store()->memory_size()) { |
duke@435 | 282 | // Success: The offsets are provably independent. |
duke@435 | 283 | // (You may ask, why not just test st_offset != offset and be done? |
duke@435 | 284 | // The answer is that stores of different sizes can co-exist |
duke@435 | 285 | // in the same sequence of RawMem effects. We sometimes initialize |
duke@435 | 286 | // a whole 'tile' of array elements with a single jint or jlong.) |
duke@435 | 287 | mem = mem->in(MemNode::Memory); |
duke@435 | 288 | continue; // (a) advance through independent store memory |
duke@435 | 289 | } |
duke@435 | 290 | } |
duke@435 | 291 | if (st_base != base && |
duke@435 | 292 | detect_ptr_independence(base, alloc, |
duke@435 | 293 | st_base, |
duke@435 | 294 | AllocateNode::Ideal_allocation(st_base, phase), |
duke@435 | 295 | phase)) { |
duke@435 | 296 | // Success: The bases are provably independent. |
duke@435 | 297 | mem = mem->in(MemNode::Memory); |
duke@435 | 298 | continue; // (a) advance through independent store memory |
duke@435 | 299 | } |
duke@435 | 300 | |
duke@435 | 301 | // (b) At this point, if the bases or offsets do not agree, we lose, |
duke@435 | 302 | // since we have not managed to prove 'this' and 'mem' independent. |
duke@435 | 303 | if (st_base == base && st_offset == offset) { |
duke@435 | 304 | return mem; // let caller handle steps (c), (d) |
duke@435 | 305 | } |
duke@435 | 306 | |
duke@435 | 307 | } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { |
duke@435 | 308 | InitializeNode* st_init = mem->in(0)->as_Initialize(); |
duke@435 | 309 | AllocateNode* st_alloc = st_init->allocation(); |
duke@435 | 310 | if (st_alloc == NULL) |
duke@435 | 311 | break; // something degenerated |
duke@435 | 312 | bool known_identical = false; |
duke@435 | 313 | bool known_independent = false; |
duke@435 | 314 | if (alloc == st_alloc) |
duke@435 | 315 | known_identical = true; |
duke@435 | 316 | else if (alloc != NULL) |
duke@435 | 317 | known_independent = true; |
duke@435 | 318 | else if (ctrl != NULL && |
duke@435 | 319 | detect_dominating_control(ctrl, st_alloc->in(0))) |
duke@435 | 320 | known_independent = true; |
duke@435 | 321 | |
duke@435 | 322 | if (known_independent) { |
duke@435 | 323 | // The bases are provably independent: Either they are |
duke@435 | 324 | // manifestly distinct allocations, or else the control |
duke@435 | 325 | // of this load dominates the store's allocation. |
duke@435 | 326 | int alias_idx = phase->C->get_alias_index(adr_type()); |
duke@435 | 327 | if (alias_idx == Compile::AliasIdxRaw) { |
duke@435 | 328 | mem = st_alloc->in(TypeFunc::Memory); |
duke@435 | 329 | } else { |
duke@435 | 330 | mem = st_init->memory(alias_idx); |
duke@435 | 331 | } |
duke@435 | 332 | continue; // (a) advance through independent store memory |
duke@435 | 333 | } |
duke@435 | 334 | |
duke@435 | 335 | // (b) at this point, if we are not looking at a store initializing |
duke@435 | 336 | // the same allocation we are loading from, we lose. |
duke@435 | 337 | if (known_identical) { |
duke@435 | 338 | // From caller, can_see_stored_value will consult find_captured_store. |
duke@435 | 339 | return mem; // let caller handle steps (c), (d) |
duke@435 | 340 | } |
duke@435 | 341 | |
duke@435 | 342 | } |
duke@435 | 343 | |
duke@435 | 344 | // Unless there is an explicit 'continue', we must bail out here, |
duke@435 | 345 | // because 'mem' is an inscrutable memory state (e.g., a call). |
duke@435 | 346 | break; |
duke@435 | 347 | } |
duke@435 | 348 | |
duke@435 | 349 | return NULL; // bail out |
duke@435 | 350 | } |
duke@435 | 351 | |
duke@435 | 352 | //----------------------calculate_adr_type------------------------------------- |
duke@435 | 353 | // Helper function. Notices when the given type of address hits top or bottom. |
duke@435 | 354 | // Also, asserts a cross-check of the type against the expected address type. |
duke@435 | 355 | const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { |
duke@435 | 356 | if (t == Type::TOP) return NULL; // does not touch memory any more? |
duke@435 | 357 | #ifdef PRODUCT |
duke@435 | 358 | cross_check = NULL; |
duke@435 | 359 | #else |
duke@435 | 360 | if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; |
duke@435 | 361 | #endif |
duke@435 | 362 | const TypePtr* tp = t->isa_ptr(); |
duke@435 | 363 | if (tp == NULL) { |
duke@435 | 364 | assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); |
duke@435 | 365 | return TypePtr::BOTTOM; // touches lots of memory |
duke@435 | 366 | } else { |
duke@435 | 367 | #ifdef ASSERT |
duke@435 | 368 | // %%%% [phh] We don't check the alias index if cross_check is |
duke@435 | 369 | // TypeRawPtr::BOTTOM. Needs to be investigated. |
duke@435 | 370 | if (cross_check != NULL && |
duke@435 | 371 | cross_check != TypePtr::BOTTOM && |
duke@435 | 372 | cross_check != TypeRawPtr::BOTTOM) { |
duke@435 | 373 | // Recheck the alias index, to see if it has changed (due to a bug). |
duke@435 | 374 | Compile* C = Compile::current(); |
duke@435 | 375 | assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), |
duke@435 | 376 | "must stay in the original alias category"); |
duke@435 | 377 | // The type of the address must be contained in the adr_type, |
duke@435 | 378 | // disregarding "null"-ness. |
duke@435 | 379 | // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) |
duke@435 | 380 | const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); |
duke@435 | 381 | assert(cross_check->meet(tp_notnull) == cross_check, |
duke@435 | 382 | "real address must not escape from expected memory type"); |
duke@435 | 383 | } |
duke@435 | 384 | #endif |
duke@435 | 385 | return tp; |
duke@435 | 386 | } |
duke@435 | 387 | } |
duke@435 | 388 | |
duke@435 | 389 | //------------------------adr_phi_is_loop_invariant---------------------------- |
duke@435 | 390 | // A helper function for Ideal_DU_postCCP to check if a Phi in a counted |
duke@435 | 391 | // loop is loop invariant. Make a quick traversal of Phi and associated |
duke@435 | 392 | // CastPP nodes, looking to see if they are a closed group within the loop. |
duke@435 | 393 | bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { |
duke@435 | 394 | // The idea is that the phi-nest must boil down to only CastPP nodes |
duke@435 | 395 | // with the same data. This implies that any path into the loop already |
duke@435 | 396 | // includes such a CastPP, and so the original cast, whatever its input, |
duke@435 | 397 | // must be covered by an equivalent cast, with an earlier control input. |
duke@435 | 398 | ResourceMark rm; |
duke@435 | 399 | |
duke@435 | 400 | // The loop entry input of the phi should be the unique dominating |
duke@435 | 401 | // node for every Phi/CastPP in the loop. |
duke@435 | 402 | Unique_Node_List closure; |
duke@435 | 403 | closure.push(adr_phi->in(LoopNode::EntryControl)); |
duke@435 | 404 | |
duke@435 | 405 | // Add the phi node and the cast to the worklist. |
duke@435 | 406 | Unique_Node_List worklist; |
duke@435 | 407 | worklist.push(adr_phi); |
duke@435 | 408 | if( cast != NULL ){ |
duke@435 | 409 | if( !cast->is_ConstraintCast() ) return false; |
duke@435 | 410 | worklist.push(cast); |
duke@435 | 411 | } |
duke@435 | 412 | |
duke@435 | 413 | // Begin recursive walk of phi nodes. |
duke@435 | 414 | while( worklist.size() ){ |
duke@435 | 415 | // Take a node off the worklist |
duke@435 | 416 | Node *n = worklist.pop(); |
duke@435 | 417 | if( !closure.member(n) ){ |
duke@435 | 418 | // Add it to the closure. |
duke@435 | 419 | closure.push(n); |
duke@435 | 420 | // Make a sanity check to ensure we don't waste too much time here. |
duke@435 | 421 | if( closure.size() > 20) return false; |
duke@435 | 422 | // This node is OK if: |
duke@435 | 423 | // - it is a cast of an identical value |
duke@435 | 424 | // - or it is a phi node (then we add its inputs to the worklist) |
duke@435 | 425 | // Otherwise, the node is not OK, and we presume the cast is not invariant |
duke@435 | 426 | if( n->is_ConstraintCast() ){ |
duke@435 | 427 | worklist.push(n->in(1)); |
duke@435 | 428 | } else if( n->is_Phi() ) { |
duke@435 | 429 | for( uint i = 1; i < n->req(); i++ ) { |
duke@435 | 430 | worklist.push(n->in(i)); |
duke@435 | 431 | } |
duke@435 | 432 | } else { |
duke@435 | 433 | return false; |
duke@435 | 434 | } |
duke@435 | 435 | } |
duke@435 | 436 | } |
duke@435 | 437 | |
duke@435 | 438 | // Quit when the worklist is empty, and we've found no offending nodes. |
duke@435 | 439 | return true; |
duke@435 | 440 | } |
duke@435 | 441 | |
duke@435 | 442 | //------------------------------Ideal_DU_postCCP------------------------------- |
duke@435 | 443 | // Find any cast-away of null-ness and keep its control. Null cast-aways are |
duke@435 | 444 | // going away in this pass and we need to make this memory op depend on the |
duke@435 | 445 | // gating null check. |
duke@435 | 446 | |
duke@435 | 447 | // I tried to leave the CastPP's in. This makes the graph more accurate in |
duke@435 | 448 | // some sense; we get to keep around the knowledge that an oop is not-null |
duke@435 | 449 | // after some test. Alas, the CastPP's interfere with GVN (some values are |
duke@435 | 450 | // the regular oop, some are the CastPP of the oop, all merge at Phi's which |
duke@435 | 451 | // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed |
duke@435 | 452 | // some of the more trivial cases in the optimizer. Removing more useless |
duke@435 | 453 | // Phi's started allowing Loads to illegally float above null checks. I gave |
duke@435 | 454 | // up on this approach. CNC 10/20/2000 |
duke@435 | 455 | Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { |
duke@435 | 456 | Node *ctr = in(MemNode::Control); |
duke@435 | 457 | Node *mem = in(MemNode::Memory); |
duke@435 | 458 | Node *adr = in(MemNode::Address); |
duke@435 | 459 | Node *skipped_cast = NULL; |
duke@435 | 460 | // Need a null check? Regular static accesses do not because they are |
duke@435 | 461 | // from constant addresses. Array ops are gated by the range check (which |
duke@435 | 462 | // always includes a NULL check). Just check field ops. |
duke@435 | 463 | if( !ctr ) { |
duke@435 | 464 | // Scan upwards for the highest location we can place this memory op. |
duke@435 | 465 | while( true ) { |
duke@435 | 466 | switch( adr->Opcode() ) { |
duke@435 | 467 | |
duke@435 | 468 | case Op_AddP: // No change to NULL-ness, so peek thru AddP's |
duke@435 | 469 | adr = adr->in(AddPNode::Base); |
duke@435 | 470 | continue; |
duke@435 | 471 | |
duke@435 | 472 | case Op_CastPP: |
duke@435 | 473 | // If the CastPP is useless, just peek on through it. |
duke@435 | 474 | if( ccp->type(adr) == ccp->type(adr->in(1)) ) { |
duke@435 | 475 | // Remember the cast that we've peeked though. If we peek |
duke@435 | 476 | // through more than one, then we end up remembering the highest |
duke@435 | 477 | // one, that is, if in a loop, the one closest to the top. |
duke@435 | 478 | skipped_cast = adr; |
duke@435 | 479 | adr = adr->in(1); |
duke@435 | 480 | continue; |
duke@435 | 481 | } |
duke@435 | 482 | // CastPP is going away in this pass! We need this memory op to be |
duke@435 | 483 | // control-dependent on the test that is guarding the CastPP. |
duke@435 | 484 | ccp->hash_delete(this); |
duke@435 | 485 | set_req(MemNode::Control, adr->in(0)); |
duke@435 | 486 | ccp->hash_insert(this); |
duke@435 | 487 | return this; |
duke@435 | 488 | |
duke@435 | 489 | case Op_Phi: |
duke@435 | 490 | // Attempt to float above a Phi to some dominating point. |
duke@435 | 491 | if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { |
duke@435 | 492 | // If we've already peeked through a Cast (which could have set the |
duke@435 | 493 | // control), we can't float above a Phi, because the skipped Cast |
duke@435 | 494 | // may not be loop invariant. |
duke@435 | 495 | if (adr_phi_is_loop_invariant(adr, skipped_cast)) { |
duke@435 | 496 | adr = adr->in(1); |
duke@435 | 497 | continue; |
duke@435 | 498 | } |
duke@435 | 499 | } |
duke@435 | 500 | |
duke@435 | 501 | // Intentional fallthrough! |
duke@435 | 502 | |
duke@435 | 503 | // No obvious dominating point. The mem op is pinned below the Phi |
duke@435 | 504 | // by the Phi itself. If the Phi goes away (no true value is merged) |
duke@435 | 505 | // then the mem op can float, but not indefinitely. It must be pinned |
duke@435 | 506 | // behind the controls leading to the Phi. |
duke@435 | 507 | case Op_CheckCastPP: |
duke@435 | 508 | // These usually stick around to change address type, however a |
duke@435 | 509 | // useless one can be elided and we still need to pick up a control edge |
duke@435 | 510 | if (adr->in(0) == NULL) { |
duke@435 | 511 | // This CheckCastPP node has NO control and is likely useless. But we |
duke@435 | 512 | // need check further up the ancestor chain for a control input to keep |
duke@435 | 513 | // the node in place. 4959717. |
duke@435 | 514 | skipped_cast = adr; |
duke@435 | 515 | adr = adr->in(1); |
duke@435 | 516 | continue; |
duke@435 | 517 | } |
duke@435 | 518 | ccp->hash_delete(this); |
duke@435 | 519 | set_req(MemNode::Control, adr->in(0)); |
duke@435 | 520 | ccp->hash_insert(this); |
duke@435 | 521 | return this; |
duke@435 | 522 | |
duke@435 | 523 | // List of "safe" opcodes; those that implicitly block the memory |
duke@435 | 524 | // op below any null check. |
duke@435 | 525 | case Op_CastX2P: // no null checks on native pointers |
duke@435 | 526 | case Op_Parm: // 'this' pointer is not null |
duke@435 | 527 | case Op_LoadP: // Loading from within a klass |
duke@435 | 528 | case Op_LoadKlass: // Loading from within a klass |
duke@435 | 529 | case Op_ConP: // Loading from a klass |
duke@435 | 530 | case Op_CreateEx: // Sucking up the guts of an exception oop |
duke@435 | 531 | case Op_Con: // Reading from TLS |
duke@435 | 532 | case Op_CMoveP: // CMoveP is pinned |
duke@435 | 533 | break; // No progress |
duke@435 | 534 | |
duke@435 | 535 | case Op_Proj: // Direct call to an allocation routine |
duke@435 | 536 | case Op_SCMemProj: // Memory state from store conditional ops |
duke@435 | 537 | #ifdef ASSERT |
duke@435 | 538 | { |
duke@435 | 539 | assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); |
duke@435 | 540 | const Node* call = adr->in(0); |
duke@435 | 541 | if (call->is_CallStaticJava()) { |
duke@435 | 542 | const CallStaticJavaNode* call_java = call->as_CallStaticJava(); |
duke@435 | 543 | assert(call_java && call_java->method() == NULL, "must be runtime call"); |
duke@435 | 544 | // We further presume that this is one of |
duke@435 | 545 | // new_instance_Java, new_array_Java, or |
duke@435 | 546 | // the like, but do not assert for this. |
duke@435 | 547 | } else if (call->is_Allocate()) { |
duke@435 | 548 | // similar case to new_instance_Java, etc. |
duke@435 | 549 | } else if (!call->is_CallLeaf()) { |
duke@435 | 550 | // Projections from fetch_oop (OSR) are allowed as well. |
duke@435 | 551 | ShouldNotReachHere(); |
duke@435 | 552 | } |
duke@435 | 553 | } |
duke@435 | 554 | #endif |
duke@435 | 555 | break; |
duke@435 | 556 | default: |
duke@435 | 557 | ShouldNotReachHere(); |
duke@435 | 558 | } |
duke@435 | 559 | break; |
duke@435 | 560 | } |
duke@435 | 561 | } |
duke@435 | 562 | |
duke@435 | 563 | return NULL; // No progress |
duke@435 | 564 | } |
duke@435 | 565 | |
duke@435 | 566 | |
duke@435 | 567 | //============================================================================= |
duke@435 | 568 | uint LoadNode::size_of() const { return sizeof(*this); } |
duke@435 | 569 | uint LoadNode::cmp( const Node &n ) const |
duke@435 | 570 | { return !Type::cmp( _type, ((LoadNode&)n)._type ); } |
duke@435 | 571 | const Type *LoadNode::bottom_type() const { return _type; } |
duke@435 | 572 | uint LoadNode::ideal_reg() const { |
duke@435 | 573 | return Matcher::base2reg[_type->base()]; |
duke@435 | 574 | } |
duke@435 | 575 | |
duke@435 | 576 | #ifndef PRODUCT |
duke@435 | 577 | void LoadNode::dump_spec(outputStream *st) const { |
duke@435 | 578 | MemNode::dump_spec(st); |
duke@435 | 579 | if( !Verbose && !WizardMode ) { |
duke@435 | 580 | // standard dump does this in Verbose and WizardMode |
duke@435 | 581 | st->print(" #"); _type->dump_on(st); |
duke@435 | 582 | } |
duke@435 | 583 | } |
duke@435 | 584 | #endif |
duke@435 | 585 | |
duke@435 | 586 | |
duke@435 | 587 | //----------------------------LoadNode::make----------------------------------- |
duke@435 | 588 | // Polymorphic factory method: |
duke@435 | 589 | LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { |
duke@435 | 590 | // sanity check the alias category against the created node type |
duke@435 | 591 | assert(!(adr_type->isa_oopptr() && |
duke@435 | 592 | adr_type->offset() == oopDesc::klass_offset_in_bytes()), |
duke@435 | 593 | "use LoadKlassNode instead"); |
duke@435 | 594 | assert(!(adr_type->isa_aryptr() && |
duke@435 | 595 | adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), |
duke@435 | 596 | "use LoadRangeNode instead"); |
duke@435 | 597 | switch (bt) { |
duke@435 | 598 | case T_BOOLEAN: |
duke@435 | 599 | case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() ); |
duke@435 | 600 | case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() ); |
duke@435 | 601 | case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() ); |
duke@435 | 602 | case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() ); |
duke@435 | 603 | case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() ); |
duke@435 | 604 | case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt ); |
duke@435 | 605 | case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt ); |
duke@435 | 606 | case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() ); |
duke@435 | 607 | case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); |
duke@435 | 608 | } |
duke@435 | 609 | ShouldNotReachHere(); |
duke@435 | 610 | return (LoadNode*)NULL; |
duke@435 | 611 | } |
duke@435 | 612 | |
duke@435 | 613 | LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { |
duke@435 | 614 | bool require_atomic = true; |
duke@435 | 615 | return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); |
duke@435 | 616 | } |
duke@435 | 617 | |
duke@435 | 618 | |
duke@435 | 619 | |
duke@435 | 620 | |
duke@435 | 621 | //------------------------------hash------------------------------------------- |
duke@435 | 622 | uint LoadNode::hash() const { |
duke@435 | 623 | // unroll addition of interesting fields |
duke@435 | 624 | return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); |
duke@435 | 625 | } |
duke@435 | 626 | |
duke@435 | 627 | //---------------------------can_see_stored_value------------------------------ |
duke@435 | 628 | // This routine exists to make sure this set of tests is done the same |
duke@435 | 629 | // everywhere. We need to make a coordinated change: first LoadNode::Ideal |
duke@435 | 630 | // will change the graph shape in a way which makes memory alive twice at the |
duke@435 | 631 | // same time (uses the Oracle model of aliasing), then some |
duke@435 | 632 | // LoadXNode::Identity will fold things back to the equivalence-class model |
duke@435 | 633 | // of aliasing. |
duke@435 | 634 | Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { |
duke@435 | 635 | Node* ld_adr = in(MemNode::Address); |
duke@435 | 636 | |
duke@435 | 637 | // Loop around twice in the case Load -> Initialize -> Store. |
duke@435 | 638 | // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) |
duke@435 | 639 | for (int trip = 0; trip <= 1; trip++) { |
duke@435 | 640 | |
duke@435 | 641 | if (st->is_Store()) { |
duke@435 | 642 | Node* st_adr = st->in(MemNode::Address); |
duke@435 | 643 | if (!phase->eqv(st_adr, ld_adr)) { |
duke@435 | 644 | // Try harder before giving up... Match raw and non-raw pointers. |
duke@435 | 645 | intptr_t st_off = 0; |
duke@435 | 646 | AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); |
duke@435 | 647 | if (alloc == NULL) return NULL; |
duke@435 | 648 | intptr_t ld_off = 0; |
duke@435 | 649 | AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); |
duke@435 | 650 | if (alloc != allo2) return NULL; |
duke@435 | 651 | if (ld_off != st_off) return NULL; |
duke@435 | 652 | // At this point we have proven something like this setup: |
duke@435 | 653 | // A = Allocate(...) |
duke@435 | 654 | // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) |
duke@435 | 655 | // S = StoreQ(, AddP(, A.Parm , #Off), V) |
duke@435 | 656 | // (Actually, we haven't yet proven the Q's are the same.) |
duke@435 | 657 | // In other words, we are loading from a casted version of |
duke@435 | 658 | // the same pointer-and-offset that we stored to. |
duke@435 | 659 | // Thus, we are able to replace L by V. |
duke@435 | 660 | } |
duke@435 | 661 | // Now prove that we have a LoadQ matched to a StoreQ, for some Q. |
duke@435 | 662 | if (store_Opcode() != st->Opcode()) |
duke@435 | 663 | return NULL; |
duke@435 | 664 | return st->in(MemNode::ValueIn); |
duke@435 | 665 | } |
duke@435 | 666 | |
duke@435 | 667 | intptr_t offset = 0; // scratch |
duke@435 | 668 | |
duke@435 | 669 | // A load from a freshly-created object always returns zero. |
duke@435 | 670 | // (This can happen after LoadNode::Ideal resets the load's memory input |
duke@435 | 671 | // to find_captured_store, which returned InitializeNode::zero_memory.) |
duke@435 | 672 | if (st->is_Proj() && st->in(0)->is_Allocate() && |
duke@435 | 673 | st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && |
duke@435 | 674 | offset >= st->in(0)->as_Allocate()->minimum_header_size()) { |
duke@435 | 675 | // return a zero value for the load's basic type |
duke@435 | 676 | // (This is one of the few places where a generic PhaseTransform |
duke@435 | 677 | // can create new nodes. Think of it as lazily manifesting |
duke@435 | 678 | // virtually pre-existing constants.) |
duke@435 | 679 | return phase->zerocon(memory_type()); |
duke@435 | 680 | } |
duke@435 | 681 | |
duke@435 | 682 | // A load from an initialization barrier can match a captured store. |
duke@435 | 683 | if (st->is_Proj() && st->in(0)->is_Initialize()) { |
duke@435 | 684 | InitializeNode* init = st->in(0)->as_Initialize(); |
duke@435 | 685 | AllocateNode* alloc = init->allocation(); |
duke@435 | 686 | if (alloc != NULL && |
duke@435 | 687 | alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { |
duke@435 | 688 | // examine a captured store value |
duke@435 | 689 | st = init->find_captured_store(offset, memory_size(), phase); |
duke@435 | 690 | if (st != NULL) |
duke@435 | 691 | continue; // take one more trip around |
duke@435 | 692 | } |
duke@435 | 693 | } |
duke@435 | 694 | |
duke@435 | 695 | break; |
duke@435 | 696 | } |
duke@435 | 697 | |
duke@435 | 698 | return NULL; |
duke@435 | 699 | } |
duke@435 | 700 | |
duke@435 | 701 | //------------------------------Identity--------------------------------------- |
duke@435 | 702 | // Loads are identity if previous store is to same address |
duke@435 | 703 | Node *LoadNode::Identity( PhaseTransform *phase ) { |
duke@435 | 704 | // If the previous store-maker is the right kind of Store, and the store is |
duke@435 | 705 | // to the same address, then we are equal to the value stored. |
duke@435 | 706 | Node* mem = in(MemNode::Memory); |
duke@435 | 707 | Node* value = can_see_stored_value(mem, phase); |
duke@435 | 708 | if( value ) { |
duke@435 | 709 | // byte, short & char stores truncate naturally. |
duke@435 | 710 | // A load has to load the truncated value which requires |
duke@435 | 711 | // some sort of masking operation and that requires an |
duke@435 | 712 | // Ideal call instead of an Identity call. |
duke@435 | 713 | if (memory_size() < BytesPerInt) { |
duke@435 | 714 | // If the input to the store does not fit with the load's result type, |
duke@435 | 715 | // it must be truncated via an Ideal call. |
duke@435 | 716 | if (!phase->type(value)->higher_equal(phase->type(this))) |
duke@435 | 717 | return this; |
duke@435 | 718 | } |
duke@435 | 719 | // (This works even when value is a Con, but LoadNode::Value |
duke@435 | 720 | // usually runs first, producing the singleton type of the Con.) |
duke@435 | 721 | return value; |
duke@435 | 722 | } |
duke@435 | 723 | return this; |
duke@435 | 724 | } |
duke@435 | 725 | |
duke@435 | 726 | //------------------------------Ideal------------------------------------------ |
duke@435 | 727 | // If the load is from Field memory and the pointer is non-null, we can |
duke@435 | 728 | // zero out the control input. |
duke@435 | 729 | // If the offset is constant and the base is an object allocation, |
duke@435 | 730 | // try to hook me up to the exact initializing store. |
duke@435 | 731 | Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 732 | Node* p = MemNode::Ideal_common(phase, can_reshape); |
duke@435 | 733 | if (p) return (p == NodeSentinel) ? NULL : p; |
duke@435 | 734 | |
duke@435 | 735 | Node* ctrl = in(MemNode::Control); |
duke@435 | 736 | Node* address = in(MemNode::Address); |
duke@435 | 737 | |
duke@435 | 738 | // Skip up past a SafePoint control. Cannot do this for Stores because |
duke@435 | 739 | // pointer stores & cardmarks must stay on the same side of a SafePoint. |
duke@435 | 740 | if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && |
duke@435 | 741 | phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { |
duke@435 | 742 | ctrl = ctrl->in(0); |
duke@435 | 743 | set_req(MemNode::Control,ctrl); |
duke@435 | 744 | } |
duke@435 | 745 | |
duke@435 | 746 | // Check for useless control edge in some common special cases |
duke@435 | 747 | if (in(MemNode::Control) != NULL) { |
duke@435 | 748 | intptr_t ignore = 0; |
duke@435 | 749 | Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); |
duke@435 | 750 | if (base != NULL |
duke@435 | 751 | && phase->type(base)->higher_equal(TypePtr::NOTNULL) |
duke@435 | 752 | && detect_dominating_control(base->in(0), phase->C->start())) { |
duke@435 | 753 | // A method-invariant, non-null address (constant or 'this' argument). |
duke@435 | 754 | set_req(MemNode::Control, NULL); |
duke@435 | 755 | } |
duke@435 | 756 | } |
duke@435 | 757 | |
duke@435 | 758 | // Check for prior store with a different base or offset; make Load |
duke@435 | 759 | // independent. Skip through any number of them. Bail out if the stores |
duke@435 | 760 | // are in an endless dead cycle and report no progress. This is a key |
duke@435 | 761 | // transform for Reflection. However, if after skipping through the Stores |
duke@435 | 762 | // we can't then fold up against a prior store do NOT do the transform as |
duke@435 | 763 | // this amounts to using the 'Oracle' model of aliasing. It leaves the same |
duke@435 | 764 | // array memory alive twice: once for the hoisted Load and again after the |
duke@435 | 765 | // bypassed Store. This situation only works if EVERYBODY who does |
duke@435 | 766 | // anti-dependence work knows how to bypass. I.e. we need all |
duke@435 | 767 | // anti-dependence checks to ask the same Oracle. Right now, that Oracle is |
duke@435 | 768 | // the alias index stuff. So instead, peek through Stores and IFF we can |
duke@435 | 769 | // fold up, do so. |
duke@435 | 770 | Node* prev_mem = find_previous_store(phase); |
duke@435 | 771 | // Steps (a), (b): Walk past independent stores to find an exact match. |
duke@435 | 772 | if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { |
duke@435 | 773 | // (c) See if we can fold up on the spot, but don't fold up here. |
duke@435 | 774 | // Fold-up might require truncation (for LoadB/LoadS/LoadC) or |
duke@435 | 775 | // just return a prior value, which is done by Identity calls. |
duke@435 | 776 | if (can_see_stored_value(prev_mem, phase)) { |
duke@435 | 777 | // Make ready for step (d): |
duke@435 | 778 | set_req(MemNode::Memory, prev_mem); |
duke@435 | 779 | return this; |
duke@435 | 780 | } |
duke@435 | 781 | } |
duke@435 | 782 | |
duke@435 | 783 | return NULL; // No further progress |
duke@435 | 784 | } |
duke@435 | 785 | |
duke@435 | 786 | // Helper to recognize certain Klass fields which are invariant across |
duke@435 | 787 | // some group of array types (e.g., int[] or all T[] where T < Object). |
duke@435 | 788 | const Type* |
duke@435 | 789 | LoadNode::load_array_final_field(const TypeKlassPtr *tkls, |
duke@435 | 790 | ciKlass* klass) const { |
duke@435 | 791 | if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { |
duke@435 | 792 | // The field is Klass::_modifier_flags. Return its (constant) value. |
duke@435 | 793 | // (Folds up the 2nd indirection in aClassConstant.getModifiers().) |
duke@435 | 794 | assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); |
duke@435 | 795 | return TypeInt::make(klass->modifier_flags()); |
duke@435 | 796 | } |
duke@435 | 797 | if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { |
duke@435 | 798 | // The field is Klass::_access_flags. Return its (constant) value. |
duke@435 | 799 | // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) |
duke@435 | 800 | assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); |
duke@435 | 801 | return TypeInt::make(klass->access_flags()); |
duke@435 | 802 | } |
duke@435 | 803 | if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { |
duke@435 | 804 | // The field is Klass::_layout_helper. Return its constant value if known. |
duke@435 | 805 | assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); |
duke@435 | 806 | return TypeInt::make(klass->layout_helper()); |
duke@435 | 807 | } |
duke@435 | 808 | |
duke@435 | 809 | // No match. |
duke@435 | 810 | return NULL; |
duke@435 | 811 | } |
duke@435 | 812 | |
duke@435 | 813 | //------------------------------Value----------------------------------------- |
duke@435 | 814 | const Type *LoadNode::Value( PhaseTransform *phase ) const { |
duke@435 | 815 | // Either input is TOP ==> the result is TOP |
duke@435 | 816 | Node* mem = in(MemNode::Memory); |
duke@435 | 817 | const Type *t1 = phase->type(mem); |
duke@435 | 818 | if (t1 == Type::TOP) return Type::TOP; |
duke@435 | 819 | Node* adr = in(MemNode::Address); |
duke@435 | 820 | const TypePtr* tp = phase->type(adr)->isa_ptr(); |
duke@435 | 821 | if (tp == NULL || tp->empty()) return Type::TOP; |
duke@435 | 822 | int off = tp->offset(); |
duke@435 | 823 | assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); |
duke@435 | 824 | |
duke@435 | 825 | // Try to guess loaded type from pointer type |
duke@435 | 826 | if (tp->base() == Type::AryPtr) { |
duke@435 | 827 | const Type *t = tp->is_aryptr()->elem(); |
duke@435 | 828 | // Don't do this for integer types. There is only potential profit if |
duke@435 | 829 | // the element type t is lower than _type; that is, for int types, if _type is |
duke@435 | 830 | // more restrictive than t. This only happens here if one is short and the other |
duke@435 | 831 | // char (both 16 bits), and in those cases we've made an intentional decision |
duke@435 | 832 | // to use one kind of load over the other. See AndINode::Ideal and 4965907. |
duke@435 | 833 | // Also, do not try to narrow the type for a LoadKlass, regardless of offset. |
duke@435 | 834 | // |
duke@435 | 835 | // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) |
duke@435 | 836 | // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier |
duke@435 | 837 | // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been |
duke@435 | 838 | // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, |
duke@435 | 839 | // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. |
duke@435 | 840 | // In fact, that could have been the original type of p1, and p1 could have |
duke@435 | 841 | // had an original form like p1:(AddP x x (LShiftL quux 3)), where the |
duke@435 | 842 | // expression (LShiftL quux 3) independently optimized to the constant 8. |
duke@435 | 843 | if ((t->isa_int() == NULL) && (t->isa_long() == NULL) |
duke@435 | 844 | && Opcode() != Op_LoadKlass) { |
duke@435 | 845 | // t might actually be lower than _type, if _type is a unique |
duke@435 | 846 | // concrete subclass of abstract class t. |
duke@435 | 847 | // Make sure the reference is not into the header, by comparing |
duke@435 | 848 | // the offset against the offset of the start of the array's data. |
duke@435 | 849 | // Different array types begin at slightly different offsets (12 vs. 16). |
duke@435 | 850 | // We choose T_BYTE as an example base type that is least restrictive |
duke@435 | 851 | // as to alignment, which will therefore produce the smallest |
duke@435 | 852 | // possible base offset. |
duke@435 | 853 | const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
duke@435 | 854 | if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? |
duke@435 | 855 | const Type* jt = t->join(_type); |
duke@435 | 856 | // In any case, do not allow the join, per se, to empty out the type. |
duke@435 | 857 | if (jt->empty() && !t->empty()) { |
duke@435 | 858 | // This can happen if a interface-typed array narrows to a class type. |
duke@435 | 859 | jt = _type; |
duke@435 | 860 | } |
duke@435 | 861 | return jt; |
duke@435 | 862 | } |
duke@435 | 863 | } |
duke@435 | 864 | } else if (tp->base() == Type::InstPtr) { |
duke@435 | 865 | assert( off != Type::OffsetBot || |
duke@435 | 866 | // arrays can be cast to Objects |
duke@435 | 867 | tp->is_oopptr()->klass()->is_java_lang_Object() || |
duke@435 | 868 | // unsafe field access may not have a constant offset |
duke@435 | 869 | phase->C->has_unsafe_access(), |
duke@435 | 870 | "Field accesses must be precise" ); |
duke@435 | 871 | // For oop loads, we expect the _type to be precise |
duke@435 | 872 | } else if (tp->base() == Type::KlassPtr) { |
duke@435 | 873 | assert( off != Type::OffsetBot || |
duke@435 | 874 | // arrays can be cast to Objects |
duke@435 | 875 | tp->is_klassptr()->klass()->is_java_lang_Object() || |
duke@435 | 876 | // also allow array-loading from the primary supertype |
duke@435 | 877 | // array during subtype checks |
duke@435 | 878 | Opcode() == Op_LoadKlass, |
duke@435 | 879 | "Field accesses must be precise" ); |
duke@435 | 880 | // For klass/static loads, we expect the _type to be precise |
duke@435 | 881 | } |
duke@435 | 882 | |
duke@435 | 883 | const TypeKlassPtr *tkls = tp->isa_klassptr(); |
duke@435 | 884 | if (tkls != NULL && !StressReflectiveCode) { |
duke@435 | 885 | ciKlass* klass = tkls->klass(); |
duke@435 | 886 | if (klass->is_loaded() && tkls->klass_is_exact()) { |
duke@435 | 887 | // We are loading a field from a Klass metaobject whose identity |
duke@435 | 888 | // is known at compile time (the type is "exact" or "precise"). |
duke@435 | 889 | // Check for fields we know are maintained as constants by the VM. |
duke@435 | 890 | if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { |
duke@435 | 891 | // The field is Klass::_super_check_offset. Return its (constant) value. |
duke@435 | 892 | // (Folds up type checking code.) |
duke@435 | 893 | assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); |
duke@435 | 894 | return TypeInt::make(klass->super_check_offset()); |
duke@435 | 895 | } |
duke@435 | 896 | // Compute index into primary_supers array |
duke@435 | 897 | juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); |
duke@435 | 898 | // Check for overflowing; use unsigned compare to handle the negative case. |
duke@435 | 899 | if( depth < ciKlass::primary_super_limit() ) { |
duke@435 | 900 | // The field is an element of Klass::_primary_supers. Return its (constant) value. |
duke@435 | 901 | // (Folds up type checking code.) |
duke@435 | 902 | assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); |
duke@435 | 903 | ciKlass *ss = klass->super_of_depth(depth); |
duke@435 | 904 | return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; |
duke@435 | 905 | } |
duke@435 | 906 | const Type* aift = load_array_final_field(tkls, klass); |
duke@435 | 907 | if (aift != NULL) return aift; |
duke@435 | 908 | if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) |
duke@435 | 909 | && klass->is_array_klass()) { |
duke@435 | 910 | // The field is arrayKlass::_component_mirror. Return its (constant) value. |
duke@435 | 911 | // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) |
duke@435 | 912 | assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); |
duke@435 | 913 | return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); |
duke@435 | 914 | } |
duke@435 | 915 | if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { |
duke@435 | 916 | // The field is Klass::_java_mirror. Return its (constant) value. |
duke@435 | 917 | // (Folds up the 2nd indirection in anObjConstant.getClass().) |
duke@435 | 918 | assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); |
duke@435 | 919 | return TypeInstPtr::make(klass->java_mirror()); |
duke@435 | 920 | } |
duke@435 | 921 | } |
duke@435 | 922 | |
duke@435 | 923 | // We can still check if we are loading from the primary_supers array at a |
duke@435 | 924 | // shallow enough depth. Even though the klass is not exact, entries less |
duke@435 | 925 | // than or equal to its super depth are correct. |
duke@435 | 926 | if (klass->is_loaded() ) { |
duke@435 | 927 | ciType *inner = klass->klass(); |
duke@435 | 928 | while( inner->is_obj_array_klass() ) |
duke@435 | 929 | inner = inner->as_obj_array_klass()->base_element_type(); |
duke@435 | 930 | if( inner->is_instance_klass() && |
duke@435 | 931 | !inner->as_instance_klass()->flags().is_interface() ) { |
duke@435 | 932 | // Compute index into primary_supers array |
duke@435 | 933 | juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); |
duke@435 | 934 | // Check for overflowing; use unsigned compare to handle the negative case. |
duke@435 | 935 | if( depth < ciKlass::primary_super_limit() && |
duke@435 | 936 | depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case |
duke@435 | 937 | // The field is an element of Klass::_primary_supers. Return its (constant) value. |
duke@435 | 938 | // (Folds up type checking code.) |
duke@435 | 939 | assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); |
duke@435 | 940 | ciKlass *ss = klass->super_of_depth(depth); |
duke@435 | 941 | return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; |
duke@435 | 942 | } |
duke@435 | 943 | } |
duke@435 | 944 | } |
duke@435 | 945 | |
duke@435 | 946 | // If the type is enough to determine that the thing is not an array, |
duke@435 | 947 | // we can give the layout_helper a positive interval type. |
duke@435 | 948 | // This will help short-circuit some reflective code. |
duke@435 | 949 | if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) |
duke@435 | 950 | && !klass->is_array_klass() // not directly typed as an array |
duke@435 | 951 | && !klass->is_interface() // specifically not Serializable & Cloneable |
duke@435 | 952 | && !klass->is_java_lang_Object() // not the supertype of all T[] |
duke@435 | 953 | ) { |
duke@435 | 954 | // Note: When interfaces are reliable, we can narrow the interface |
duke@435 | 955 | // test to (klass != Serializable && klass != Cloneable). |
duke@435 | 956 | assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); |
duke@435 | 957 | jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); |
duke@435 | 958 | // The key property of this type is that it folds up tests |
duke@435 | 959 | // for array-ness, since it proves that the layout_helper is positive. |
duke@435 | 960 | // Thus, a generic value like the basic object layout helper works fine. |
duke@435 | 961 | return TypeInt::make(min_size, max_jint, Type::WidenMin); |
duke@435 | 962 | } |
duke@435 | 963 | } |
duke@435 | 964 | |
duke@435 | 965 | // If we are loading from a freshly-allocated object, produce a zero, |
duke@435 | 966 | // if the load is provably beyond the header of the object. |
duke@435 | 967 | // (Also allow a variable load from a fresh array to produce zero.) |
duke@435 | 968 | if (ReduceFieldZeroing) { |
duke@435 | 969 | Node* value = can_see_stored_value(mem,phase); |
duke@435 | 970 | if (value != NULL && value->is_Con()) |
duke@435 | 971 | return value->bottom_type(); |
duke@435 | 972 | } |
duke@435 | 973 | |
duke@435 | 974 | return _type; |
duke@435 | 975 | } |
duke@435 | 976 | |
duke@435 | 977 | //------------------------------match_edge------------------------------------- |
duke@435 | 978 | // Do we Match on this edge index or not? Match only the address. |
duke@435 | 979 | uint LoadNode::match_edge(uint idx) const { |
duke@435 | 980 | return idx == MemNode::Address; |
duke@435 | 981 | } |
duke@435 | 982 | |
duke@435 | 983 | //--------------------------LoadBNode::Ideal-------------------------------------- |
duke@435 | 984 | // |
duke@435 | 985 | // If the previous store is to the same address as this load, |
duke@435 | 986 | // and the value stored was larger than a byte, replace this load |
duke@435 | 987 | // with the value stored truncated to a byte. If no truncation is |
duke@435 | 988 | // needed, the replacement is done in LoadNode::Identity(). |
duke@435 | 989 | // |
duke@435 | 990 | Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 991 | Node* mem = in(MemNode::Memory); |
duke@435 | 992 | Node* value = can_see_stored_value(mem,phase); |
duke@435 | 993 | if( value && !phase->type(value)->higher_equal( _type ) ) { |
duke@435 | 994 | Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); |
duke@435 | 995 | return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); |
duke@435 | 996 | } |
duke@435 | 997 | // Identity call will handle the case where truncation is not needed. |
duke@435 | 998 | return LoadNode::Ideal(phase, can_reshape); |
duke@435 | 999 | } |
duke@435 | 1000 | |
duke@435 | 1001 | //--------------------------LoadCNode::Ideal-------------------------------------- |
duke@435 | 1002 | // |
duke@435 | 1003 | // If the previous store is to the same address as this load, |
duke@435 | 1004 | // and the value stored was larger than a char, replace this load |
duke@435 | 1005 | // with the value stored truncated to a char. If no truncation is |
duke@435 | 1006 | // needed, the replacement is done in LoadNode::Identity(). |
duke@435 | 1007 | // |
duke@435 | 1008 | Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 1009 | Node* mem = in(MemNode::Memory); |
duke@435 | 1010 | Node* value = can_see_stored_value(mem,phase); |
duke@435 | 1011 | if( value && !phase->type(value)->higher_equal( _type ) ) |
duke@435 | 1012 | return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); |
duke@435 | 1013 | // Identity call will handle the case where truncation is not needed. |
duke@435 | 1014 | return LoadNode::Ideal(phase, can_reshape); |
duke@435 | 1015 | } |
duke@435 | 1016 | |
duke@435 | 1017 | //--------------------------LoadSNode::Ideal-------------------------------------- |
duke@435 | 1018 | // |
duke@435 | 1019 | // If the previous store is to the same address as this load, |
duke@435 | 1020 | // and the value stored was larger than a short, replace this load |
duke@435 | 1021 | // with the value stored truncated to a short. If no truncation is |
duke@435 | 1022 | // needed, the replacement is done in LoadNode::Identity(). |
duke@435 | 1023 | // |
duke@435 | 1024 | Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 1025 | Node* mem = in(MemNode::Memory); |
duke@435 | 1026 | Node* value = can_see_stored_value(mem,phase); |
duke@435 | 1027 | if( value && !phase->type(value)->higher_equal( _type ) ) { |
duke@435 | 1028 | Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); |
duke@435 | 1029 | return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); |
duke@435 | 1030 | } |
duke@435 | 1031 | // Identity call will handle the case where truncation is not needed. |
duke@435 | 1032 | return LoadNode::Ideal(phase, can_reshape); |
duke@435 | 1033 | } |
duke@435 | 1034 | |
duke@435 | 1035 | //============================================================================= |
duke@435 | 1036 | //------------------------------Value------------------------------------------ |
duke@435 | 1037 | const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { |
duke@435 | 1038 | // Either input is TOP ==> the result is TOP |
duke@435 | 1039 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
duke@435 | 1040 | if (t1 == Type::TOP) return Type::TOP; |
duke@435 | 1041 | Node *adr = in(MemNode::Address); |
duke@435 | 1042 | const Type *t2 = phase->type( adr ); |
duke@435 | 1043 | if (t2 == Type::TOP) return Type::TOP; |
duke@435 | 1044 | const TypePtr *tp = t2->is_ptr(); |
duke@435 | 1045 | if (TypePtr::above_centerline(tp->ptr()) || |
duke@435 | 1046 | tp->ptr() == TypePtr::Null) return Type::TOP; |
duke@435 | 1047 | |
duke@435 | 1048 | // Return a more precise klass, if possible |
duke@435 | 1049 | const TypeInstPtr *tinst = tp->isa_instptr(); |
duke@435 | 1050 | if (tinst != NULL) { |
duke@435 | 1051 | ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); |
duke@435 | 1052 | int offset = tinst->offset(); |
duke@435 | 1053 | if (ik == phase->C->env()->Class_klass() |
duke@435 | 1054 | && (offset == java_lang_Class::klass_offset_in_bytes() || |
duke@435 | 1055 | offset == java_lang_Class::array_klass_offset_in_bytes())) { |
duke@435 | 1056 | // We are loading a special hidden field from a Class mirror object, |
duke@435 | 1057 | // the field which points to the VM's Klass metaobject. |
duke@435 | 1058 | ciType* t = tinst->java_mirror_type(); |
duke@435 | 1059 | // java_mirror_type returns non-null for compile-time Class constants. |
duke@435 | 1060 | if (t != NULL) { |
duke@435 | 1061 | // constant oop => constant klass |
duke@435 | 1062 | if (offset == java_lang_Class::array_klass_offset_in_bytes()) { |
duke@435 | 1063 | return TypeKlassPtr::make(ciArrayKlass::make(t)); |
duke@435 | 1064 | } |
duke@435 | 1065 | if (!t->is_klass()) { |
duke@435 | 1066 | // a primitive Class (e.g., int.class) has NULL for a klass field |
duke@435 | 1067 | return TypePtr::NULL_PTR; |
duke@435 | 1068 | } |
duke@435 | 1069 | // (Folds up the 1st indirection in aClassConstant.getModifiers().) |
duke@435 | 1070 | return TypeKlassPtr::make(t->as_klass()); |
duke@435 | 1071 | } |
duke@435 | 1072 | // non-constant mirror, so we can't tell what's going on |
duke@435 | 1073 | } |
duke@435 | 1074 | if( !ik->is_loaded() ) |
duke@435 | 1075 | return _type; // Bail out if not loaded |
duke@435 | 1076 | if (offset == oopDesc::klass_offset_in_bytes()) { |
duke@435 | 1077 | if (tinst->klass_is_exact()) { |
duke@435 | 1078 | return TypeKlassPtr::make(ik); |
duke@435 | 1079 | } |
duke@435 | 1080 | // See if we can become precise: no subklasses and no interface |
duke@435 | 1081 | // (Note: We need to support verified interfaces.) |
duke@435 | 1082 | if (!ik->is_interface() && !ik->has_subklass()) { |
duke@435 | 1083 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
duke@435 | 1084 | // Add a dependence; if any subclass added we need to recompile |
duke@435 | 1085 | if (!ik->is_final()) { |
duke@435 | 1086 | // %%% should use stronger assert_unique_concrete_subtype instead |
duke@435 | 1087 | phase->C->dependencies()->assert_leaf_type(ik); |
duke@435 | 1088 | } |
duke@435 | 1089 | // Return precise klass |
duke@435 | 1090 | return TypeKlassPtr::make(ik); |
duke@435 | 1091 | } |
duke@435 | 1092 | |
duke@435 | 1093 | // Return root of possible klass |
duke@435 | 1094 | return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); |
duke@435 | 1095 | } |
duke@435 | 1096 | } |
duke@435 | 1097 | |
duke@435 | 1098 | // Check for loading klass from an array |
duke@435 | 1099 | const TypeAryPtr *tary = tp->isa_aryptr(); |
duke@435 | 1100 | if( tary != NULL ) { |
duke@435 | 1101 | ciKlass *tary_klass = tary->klass(); |
duke@435 | 1102 | if (tary_klass != NULL // can be NULL when at BOTTOM or TOP |
duke@435 | 1103 | && tary->offset() == oopDesc::klass_offset_in_bytes()) { |
duke@435 | 1104 | if (tary->klass_is_exact()) { |
duke@435 | 1105 | return TypeKlassPtr::make(tary_klass); |
duke@435 | 1106 | } |
duke@435 | 1107 | ciArrayKlass *ak = tary->klass()->as_array_klass(); |
duke@435 | 1108 | // If the klass is an object array, we defer the question to the |
duke@435 | 1109 | // array component klass. |
duke@435 | 1110 | if( ak->is_obj_array_klass() ) { |
duke@435 | 1111 | assert( ak->is_loaded(), "" ); |
duke@435 | 1112 | ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); |
duke@435 | 1113 | if( base_k->is_loaded() && base_k->is_instance_klass() ) { |
duke@435 | 1114 | ciInstanceKlass* ik = base_k->as_instance_klass(); |
duke@435 | 1115 | // See if we can become precise: no subklasses and no interface |
duke@435 | 1116 | if (!ik->is_interface() && !ik->has_subklass()) { |
duke@435 | 1117 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
duke@435 | 1118 | // Add a dependence; if any subclass added we need to recompile |
duke@435 | 1119 | if (!ik->is_final()) { |
duke@435 | 1120 | phase->C->dependencies()->assert_leaf_type(ik); |
duke@435 | 1121 | } |
duke@435 | 1122 | // Return precise array klass |
duke@435 | 1123 | return TypeKlassPtr::make(ak); |
duke@435 | 1124 | } |
duke@435 | 1125 | } |
duke@435 | 1126 | return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); |
duke@435 | 1127 | } else { // Found a type-array? |
duke@435 | 1128 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
duke@435 | 1129 | assert( ak->is_type_array_klass(), "" ); |
duke@435 | 1130 | return TypeKlassPtr::make(ak); // These are always precise |
duke@435 | 1131 | } |
duke@435 | 1132 | } |
duke@435 | 1133 | } |
duke@435 | 1134 | |
duke@435 | 1135 | // Check for loading klass from an array klass |
duke@435 | 1136 | const TypeKlassPtr *tkls = tp->isa_klassptr(); |
duke@435 | 1137 | if (tkls != NULL && !StressReflectiveCode) { |
duke@435 | 1138 | ciKlass* klass = tkls->klass(); |
duke@435 | 1139 | if( !klass->is_loaded() ) |
duke@435 | 1140 | return _type; // Bail out if not loaded |
duke@435 | 1141 | if( klass->is_obj_array_klass() && |
duke@435 | 1142 | (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { |
duke@435 | 1143 | ciKlass* elem = klass->as_obj_array_klass()->element_klass(); |
duke@435 | 1144 | // // Always returning precise element type is incorrect, |
duke@435 | 1145 | // // e.g., element type could be object and array may contain strings |
duke@435 | 1146 | // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); |
duke@435 | 1147 | |
duke@435 | 1148 | // The array's TypeKlassPtr was declared 'precise' or 'not precise' |
duke@435 | 1149 | // according to the element type's subclassing. |
duke@435 | 1150 | return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); |
duke@435 | 1151 | } |
duke@435 | 1152 | if( klass->is_instance_klass() && tkls->klass_is_exact() && |
duke@435 | 1153 | (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { |
duke@435 | 1154 | ciKlass* sup = klass->as_instance_klass()->super(); |
duke@435 | 1155 | // The field is Klass::_super. Return its (constant) value. |
duke@435 | 1156 | // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) |
duke@435 | 1157 | return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; |
duke@435 | 1158 | } |
duke@435 | 1159 | } |
duke@435 | 1160 | |
duke@435 | 1161 | // Bailout case |
duke@435 | 1162 | return LoadNode::Value(phase); |
duke@435 | 1163 | } |
duke@435 | 1164 | |
duke@435 | 1165 | //------------------------------Identity--------------------------------------- |
duke@435 | 1166 | // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. |
duke@435 | 1167 | // Also feed through the klass in Allocate(...klass...)._klass. |
duke@435 | 1168 | Node* LoadKlassNode::Identity( PhaseTransform *phase ) { |
duke@435 | 1169 | Node* x = LoadNode::Identity(phase); |
duke@435 | 1170 | if (x != this) return x; |
duke@435 | 1171 | |
duke@435 | 1172 | // Take apart the address into an oop and and offset. |
duke@435 | 1173 | // Return 'this' if we cannot. |
duke@435 | 1174 | Node* adr = in(MemNode::Address); |
duke@435 | 1175 | intptr_t offset = 0; |
duke@435 | 1176 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
duke@435 | 1177 | if (base == NULL) return this; |
duke@435 | 1178 | const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); |
duke@435 | 1179 | if (toop == NULL) return this; |
duke@435 | 1180 | |
duke@435 | 1181 | // We can fetch the klass directly through an AllocateNode. |
duke@435 | 1182 | // This works even if the klass is not constant (clone or newArray). |
duke@435 | 1183 | if (offset == oopDesc::klass_offset_in_bytes()) { |
duke@435 | 1184 | Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); |
duke@435 | 1185 | if (allocated_klass != NULL) { |
duke@435 | 1186 | return allocated_klass; |
duke@435 | 1187 | } |
duke@435 | 1188 | } |
duke@435 | 1189 | |
duke@435 | 1190 | // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. |
duke@435 | 1191 | // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. |
duke@435 | 1192 | // See inline_native_Class_query for occurrences of these patterns. |
duke@435 | 1193 | // Java Example: x.getClass().isAssignableFrom(y) |
duke@435 | 1194 | // Java Example: Array.newInstance(x.getClass().getComponentType(), n) |
duke@435 | 1195 | // |
duke@435 | 1196 | // This improves reflective code, often making the Class |
duke@435 | 1197 | // mirror go completely dead. (Current exception: Class |
duke@435 | 1198 | // mirrors may appear in debug info, but we could clean them out by |
duke@435 | 1199 | // introducing a new debug info operator for klassOop.java_mirror). |
duke@435 | 1200 | if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() |
duke@435 | 1201 | && (offset == java_lang_Class::klass_offset_in_bytes() || |
duke@435 | 1202 | offset == java_lang_Class::array_klass_offset_in_bytes())) { |
duke@435 | 1203 | // We are loading a special hidden field from a Class mirror, |
duke@435 | 1204 | // the field which points to its Klass or arrayKlass metaobject. |
duke@435 | 1205 | if (base->is_Load()) { |
duke@435 | 1206 | Node* adr2 = base->in(MemNode::Address); |
duke@435 | 1207 | const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); |
duke@435 | 1208 | if (tkls != NULL && !tkls->empty() |
duke@435 | 1209 | && (tkls->klass()->is_instance_klass() || |
duke@435 | 1210 | tkls->klass()->is_array_klass()) |
duke@435 | 1211 | && adr2->is_AddP() |
duke@435 | 1212 | ) { |
duke@435 | 1213 | int mirror_field = Klass::java_mirror_offset_in_bytes(); |
duke@435 | 1214 | if (offset == java_lang_Class::array_klass_offset_in_bytes()) { |
duke@435 | 1215 | mirror_field = in_bytes(arrayKlass::component_mirror_offset()); |
duke@435 | 1216 | } |
duke@435 | 1217 | if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { |
duke@435 | 1218 | return adr2->in(AddPNode::Base); |
duke@435 | 1219 | } |
duke@435 | 1220 | } |
duke@435 | 1221 | } |
duke@435 | 1222 | } |
duke@435 | 1223 | |
duke@435 | 1224 | return this; |
duke@435 | 1225 | } |
duke@435 | 1226 | |
duke@435 | 1227 | //------------------------------Value----------------------------------------- |
duke@435 | 1228 | const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { |
duke@435 | 1229 | // Either input is TOP ==> the result is TOP |
duke@435 | 1230 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
duke@435 | 1231 | if( t1 == Type::TOP ) return Type::TOP; |
duke@435 | 1232 | Node *adr = in(MemNode::Address); |
duke@435 | 1233 | const Type *t2 = phase->type( adr ); |
duke@435 | 1234 | if( t2 == Type::TOP ) return Type::TOP; |
duke@435 | 1235 | const TypePtr *tp = t2->is_ptr(); |
duke@435 | 1236 | if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; |
duke@435 | 1237 | const TypeAryPtr *tap = tp->isa_aryptr(); |
duke@435 | 1238 | if( !tap ) return _type; |
duke@435 | 1239 | return tap->size(); |
duke@435 | 1240 | } |
duke@435 | 1241 | |
duke@435 | 1242 | //------------------------------Identity--------------------------------------- |
duke@435 | 1243 | // Feed through the length in AllocateArray(...length...)._length. |
duke@435 | 1244 | Node* LoadRangeNode::Identity( PhaseTransform *phase ) { |
duke@435 | 1245 | Node* x = LoadINode::Identity(phase); |
duke@435 | 1246 | if (x != this) return x; |
duke@435 | 1247 | |
duke@435 | 1248 | // Take apart the address into an oop and and offset. |
duke@435 | 1249 | // Return 'this' if we cannot. |
duke@435 | 1250 | Node* adr = in(MemNode::Address); |
duke@435 | 1251 | intptr_t offset = 0; |
duke@435 | 1252 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
duke@435 | 1253 | if (base == NULL) return this; |
duke@435 | 1254 | const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); |
duke@435 | 1255 | if (tary == NULL) return this; |
duke@435 | 1256 | |
duke@435 | 1257 | // We can fetch the length directly through an AllocateArrayNode. |
duke@435 | 1258 | // This works even if the length is not constant (clone or newArray). |
duke@435 | 1259 | if (offset == arrayOopDesc::length_offset_in_bytes()) { |
duke@435 | 1260 | Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase); |
duke@435 | 1261 | if (allocated_length != NULL) { |
duke@435 | 1262 | return allocated_length; |
duke@435 | 1263 | } |
duke@435 | 1264 | } |
duke@435 | 1265 | |
duke@435 | 1266 | return this; |
duke@435 | 1267 | |
duke@435 | 1268 | } |
duke@435 | 1269 | //============================================================================= |
duke@435 | 1270 | //---------------------------StoreNode::make----------------------------------- |
duke@435 | 1271 | // Polymorphic factory method: |
duke@435 | 1272 | StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { |
duke@435 | 1273 | switch (bt) { |
duke@435 | 1274 | case T_BOOLEAN: |
duke@435 | 1275 | case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); |
duke@435 | 1276 | case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); |
duke@435 | 1277 | case T_CHAR: |
duke@435 | 1278 | case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); |
duke@435 | 1279 | case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); |
duke@435 | 1280 | case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); |
duke@435 | 1281 | case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); |
duke@435 | 1282 | case T_ADDRESS: |
duke@435 | 1283 | case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); |
duke@435 | 1284 | } |
duke@435 | 1285 | ShouldNotReachHere(); |
duke@435 | 1286 | return (StoreNode*)NULL; |
duke@435 | 1287 | } |
duke@435 | 1288 | |
duke@435 | 1289 | StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { |
duke@435 | 1290 | bool require_atomic = true; |
duke@435 | 1291 | return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); |
duke@435 | 1292 | } |
duke@435 | 1293 | |
duke@435 | 1294 | |
duke@435 | 1295 | //--------------------------bottom_type---------------------------------------- |
duke@435 | 1296 | const Type *StoreNode::bottom_type() const { |
duke@435 | 1297 | return Type::MEMORY; |
duke@435 | 1298 | } |
duke@435 | 1299 | |
duke@435 | 1300 | //------------------------------hash------------------------------------------- |
duke@435 | 1301 | uint StoreNode::hash() const { |
duke@435 | 1302 | // unroll addition of interesting fields |
duke@435 | 1303 | //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); |
duke@435 | 1304 | |
duke@435 | 1305 | // Since they are not commoned, do not hash them: |
duke@435 | 1306 | return NO_HASH; |
duke@435 | 1307 | } |
duke@435 | 1308 | |
duke@435 | 1309 | //------------------------------Ideal------------------------------------------ |
duke@435 | 1310 | // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). |
duke@435 | 1311 | // When a store immediately follows a relevant allocation/initialization, |
duke@435 | 1312 | // try to capture it into the initialization, or hoist it above. |
duke@435 | 1313 | Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 1314 | Node* p = MemNode::Ideal_common(phase, can_reshape); |
duke@435 | 1315 | if (p) return (p == NodeSentinel) ? NULL : p; |
duke@435 | 1316 | |
duke@435 | 1317 | Node* mem = in(MemNode::Memory); |
duke@435 | 1318 | Node* address = in(MemNode::Address); |
duke@435 | 1319 | |
duke@435 | 1320 | // Back-to-back stores to same address? Fold em up. |
duke@435 | 1321 | // Generally unsafe if I have intervening uses... |
duke@435 | 1322 | if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { |
duke@435 | 1323 | // Looking at a dead closed cycle of memory? |
duke@435 | 1324 | assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); |
duke@435 | 1325 | |
duke@435 | 1326 | assert(Opcode() == mem->Opcode() || |
duke@435 | 1327 | phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, |
duke@435 | 1328 | "no mismatched stores, except on raw memory"); |
duke@435 | 1329 | |
duke@435 | 1330 | if (mem->outcnt() == 1 && // check for intervening uses |
duke@435 | 1331 | mem->as_Store()->memory_size() <= this->memory_size()) { |
duke@435 | 1332 | // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. |
duke@435 | 1333 | // For example, 'mem' might be the final state at a conditional return. |
duke@435 | 1334 | // Or, 'mem' might be used by some node which is live at the same time |
duke@435 | 1335 | // 'this' is live, which might be unschedulable. So, require exactly |
duke@435 | 1336 | // ONE user, the 'this' store, until such time as we clone 'mem' for |
duke@435 | 1337 | // each of 'mem's uses (thus making the exactly-1-user-rule hold true). |
duke@435 | 1338 | if (can_reshape) { // (%%% is this an anachronism?) |
duke@435 | 1339 | set_req_X(MemNode::Memory, mem->in(MemNode::Memory), |
duke@435 | 1340 | phase->is_IterGVN()); |
duke@435 | 1341 | } else { |
duke@435 | 1342 | // It's OK to do this in the parser, since DU info is always accurate, |
duke@435 | 1343 | // and the parser always refers to nodes via SafePointNode maps. |
duke@435 | 1344 | set_req(MemNode::Memory, mem->in(MemNode::Memory)); |
duke@435 | 1345 | } |
duke@435 | 1346 | return this; |
duke@435 | 1347 | } |
duke@435 | 1348 | } |
duke@435 | 1349 | |
duke@435 | 1350 | // Capture an unaliased, unconditional, simple store into an initializer. |
duke@435 | 1351 | // Or, if it is independent of the allocation, hoist it above the allocation. |
duke@435 | 1352 | if (ReduceFieldZeroing && /*can_reshape &&*/ |
duke@435 | 1353 | mem->is_Proj() && mem->in(0)->is_Initialize()) { |
duke@435 | 1354 | InitializeNode* init = mem->in(0)->as_Initialize(); |
duke@435 | 1355 | intptr_t offset = init->can_capture_store(this, phase); |
duke@435 | 1356 | if (offset > 0) { |
duke@435 | 1357 | Node* moved = init->capture_store(this, offset, phase); |
duke@435 | 1358 | // If the InitializeNode captured me, it made a raw copy of me, |
duke@435 | 1359 | // and I need to disappear. |
duke@435 | 1360 | if (moved != NULL) { |
duke@435 | 1361 | // %%% hack to ensure that Ideal returns a new node: |
duke@435 | 1362 | mem = MergeMemNode::make(phase->C, mem); |
duke@435 | 1363 | return mem; // fold me away |
duke@435 | 1364 | } |
duke@435 | 1365 | } |
duke@435 | 1366 | } |
duke@435 | 1367 | |
duke@435 | 1368 | return NULL; // No further progress |
duke@435 | 1369 | } |
duke@435 | 1370 | |
duke@435 | 1371 | //------------------------------Value----------------------------------------- |
duke@435 | 1372 | const Type *StoreNode::Value( PhaseTransform *phase ) const { |
duke@435 | 1373 | // Either input is TOP ==> the result is TOP |
duke@435 | 1374 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
duke@435 | 1375 | if( t1 == Type::TOP ) return Type::TOP; |
duke@435 | 1376 | const Type *t2 = phase->type( in(MemNode::Address) ); |
duke@435 | 1377 | if( t2 == Type::TOP ) return Type::TOP; |
duke@435 | 1378 | const Type *t3 = phase->type( in(MemNode::ValueIn) ); |
duke@435 | 1379 | if( t3 == Type::TOP ) return Type::TOP; |
duke@435 | 1380 | return Type::MEMORY; |
duke@435 | 1381 | } |
duke@435 | 1382 | |
duke@435 | 1383 | //------------------------------Identity--------------------------------------- |
duke@435 | 1384 | // Remove redundant stores: |
duke@435 | 1385 | // Store(m, p, Load(m, p)) changes to m. |
duke@435 | 1386 | // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). |
duke@435 | 1387 | Node *StoreNode::Identity( PhaseTransform *phase ) { |
duke@435 | 1388 | Node* mem = in(MemNode::Memory); |
duke@435 | 1389 | Node* adr = in(MemNode::Address); |
duke@435 | 1390 | Node* val = in(MemNode::ValueIn); |
duke@435 | 1391 | |
duke@435 | 1392 | // Load then Store? Then the Store is useless |
duke@435 | 1393 | if (val->is_Load() && |
duke@435 | 1394 | phase->eqv_uncast( val->in(MemNode::Address), adr ) && |
duke@435 | 1395 | phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && |
duke@435 | 1396 | val->as_Load()->store_Opcode() == Opcode()) { |
duke@435 | 1397 | return mem; |
duke@435 | 1398 | } |
duke@435 | 1399 | |
duke@435 | 1400 | // Two stores in a row of the same value? |
duke@435 | 1401 | if (mem->is_Store() && |
duke@435 | 1402 | phase->eqv_uncast( mem->in(MemNode::Address), adr ) && |
duke@435 | 1403 | phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && |
duke@435 | 1404 | mem->Opcode() == Opcode()) { |
duke@435 | 1405 | return mem; |
duke@435 | 1406 | } |
duke@435 | 1407 | |
duke@435 | 1408 | // Store of zero anywhere into a freshly-allocated object? |
duke@435 | 1409 | // Then the store is useless. |
duke@435 | 1410 | // (It must already have been captured by the InitializeNode.) |
duke@435 | 1411 | if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { |
duke@435 | 1412 | // a newly allocated object is already all-zeroes everywhere |
duke@435 | 1413 | if (mem->is_Proj() && mem->in(0)->is_Allocate()) { |
duke@435 | 1414 | return mem; |
duke@435 | 1415 | } |
duke@435 | 1416 | |
duke@435 | 1417 | // the store may also apply to zero-bits in an earlier object |
duke@435 | 1418 | Node* prev_mem = find_previous_store(phase); |
duke@435 | 1419 | // Steps (a), (b): Walk past independent stores to find an exact match. |
duke@435 | 1420 | if (prev_mem != NULL) { |
duke@435 | 1421 | Node* prev_val = can_see_stored_value(prev_mem, phase); |
duke@435 | 1422 | if (prev_val != NULL && phase->eqv(prev_val, val)) { |
duke@435 | 1423 | // prev_val and val might differ by a cast; it would be good |
duke@435 | 1424 | // to keep the more informative of the two. |
duke@435 | 1425 | return mem; |
duke@435 | 1426 | } |
duke@435 | 1427 | } |
duke@435 | 1428 | } |
duke@435 | 1429 | |
duke@435 | 1430 | return this; |
duke@435 | 1431 | } |
duke@435 | 1432 | |
duke@435 | 1433 | //------------------------------match_edge------------------------------------- |
duke@435 | 1434 | // Do we Match on this edge index or not? Match only memory & value |
duke@435 | 1435 | uint StoreNode::match_edge(uint idx) const { |
duke@435 | 1436 | return idx == MemNode::Address || idx == MemNode::ValueIn; |
duke@435 | 1437 | } |
duke@435 | 1438 | |
duke@435 | 1439 | //------------------------------cmp-------------------------------------------- |
duke@435 | 1440 | // Do not common stores up together. They generally have to be split |
duke@435 | 1441 | // back up anyways, so do not bother. |
duke@435 | 1442 | uint StoreNode::cmp( const Node &n ) const { |
duke@435 | 1443 | return (&n == this); // Always fail except on self |
duke@435 | 1444 | } |
duke@435 | 1445 | |
duke@435 | 1446 | //------------------------------Ideal_masked_input----------------------------- |
duke@435 | 1447 | // Check for a useless mask before a partial-word store |
duke@435 | 1448 | // (StoreB ... (AndI valIn conIa) ) |
duke@435 | 1449 | // If (conIa & mask == mask) this simplifies to |
duke@435 | 1450 | // (StoreB ... (valIn) ) |
duke@435 | 1451 | Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { |
duke@435 | 1452 | Node *val = in(MemNode::ValueIn); |
duke@435 | 1453 | if( val->Opcode() == Op_AndI ) { |
duke@435 | 1454 | const TypeInt *t = phase->type( val->in(2) )->isa_int(); |
duke@435 | 1455 | if( t && t->is_con() && (t->get_con() & mask) == mask ) { |
duke@435 | 1456 | set_req(MemNode::ValueIn, val->in(1)); |
duke@435 | 1457 | return this; |
duke@435 | 1458 | } |
duke@435 | 1459 | } |
duke@435 | 1460 | return NULL; |
duke@435 | 1461 | } |
duke@435 | 1462 | |
duke@435 | 1463 | |
duke@435 | 1464 | //------------------------------Ideal_sign_extended_input---------------------- |
duke@435 | 1465 | // Check for useless sign-extension before a partial-word store |
duke@435 | 1466 | // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) |
duke@435 | 1467 | // If (conIL == conIR && conIR <= num_bits) this simplifies to |
duke@435 | 1468 | // (StoreB ... (valIn) ) |
duke@435 | 1469 | Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { |
duke@435 | 1470 | Node *val = in(MemNode::ValueIn); |
duke@435 | 1471 | if( val->Opcode() == Op_RShiftI ) { |
duke@435 | 1472 | const TypeInt *t = phase->type( val->in(2) )->isa_int(); |
duke@435 | 1473 | if( t && t->is_con() && (t->get_con() <= num_bits) ) { |
duke@435 | 1474 | Node *shl = val->in(1); |
duke@435 | 1475 | if( shl->Opcode() == Op_LShiftI ) { |
duke@435 | 1476 | const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); |
duke@435 | 1477 | if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { |
duke@435 | 1478 | set_req(MemNode::ValueIn, shl->in(1)); |
duke@435 | 1479 | return this; |
duke@435 | 1480 | } |
duke@435 | 1481 | } |
duke@435 | 1482 | } |
duke@435 | 1483 | } |
duke@435 | 1484 | return NULL; |
duke@435 | 1485 | } |
duke@435 | 1486 | |
duke@435 | 1487 | //------------------------------value_never_loaded----------------------------------- |
duke@435 | 1488 | // Determine whether there are any possible loads of the value stored. |
duke@435 | 1489 | // For simplicity, we actually check if there are any loads from the |
duke@435 | 1490 | // address stored to, not just for loads of the value stored by this node. |
duke@435 | 1491 | // |
duke@435 | 1492 | bool StoreNode::value_never_loaded( PhaseTransform *phase) const { |
duke@435 | 1493 | Node *adr = in(Address); |
duke@435 | 1494 | const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); |
duke@435 | 1495 | if (adr_oop == NULL) |
duke@435 | 1496 | return false; |
duke@435 | 1497 | if (!adr_oop->is_instance()) |
duke@435 | 1498 | return false; // if not a distinct instance, there may be aliases of the address |
duke@435 | 1499 | for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { |
duke@435 | 1500 | Node *use = adr->fast_out(i); |
duke@435 | 1501 | int opc = use->Opcode(); |
duke@435 | 1502 | if (use->is_Load() || use->is_LoadStore()) { |
duke@435 | 1503 | return false; |
duke@435 | 1504 | } |
duke@435 | 1505 | } |
duke@435 | 1506 | return true; |
duke@435 | 1507 | } |
duke@435 | 1508 | |
duke@435 | 1509 | //============================================================================= |
duke@435 | 1510 | //------------------------------Ideal------------------------------------------ |
duke@435 | 1511 | // If the store is from an AND mask that leaves the low bits untouched, then |
duke@435 | 1512 | // we can skip the AND operation. If the store is from a sign-extension |
duke@435 | 1513 | // (a left shift, then right shift) we can skip both. |
duke@435 | 1514 | Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
duke@435 | 1515 | Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); |
duke@435 | 1516 | if( progress != NULL ) return progress; |
duke@435 | 1517 | |
duke@435 | 1518 | progress = StoreNode::Ideal_sign_extended_input(phase, 24); |
duke@435 | 1519 | if( progress != NULL ) return progress; |
duke@435 | 1520 | |
duke@435 | 1521 | // Finally check the default case |
duke@435 | 1522 | return StoreNode::Ideal(phase, can_reshape); |
duke@435 | 1523 | } |
duke@435 | 1524 | |
duke@435 | 1525 | //============================================================================= |
duke@435 | 1526 | //------------------------------Ideal------------------------------------------ |
duke@435 | 1527 | // If the store is from an AND mask that leaves the low bits untouched, then |
duke@435 | 1528 | // we can skip the AND operation |
duke@435 | 1529 | Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
duke@435 | 1530 | Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); |
duke@435 | 1531 | if( progress != NULL ) return progress; |
duke@435 | 1532 | |
duke@435 | 1533 | progress = StoreNode::Ideal_sign_extended_input(phase, 16); |
duke@435 | 1534 | if( progress != NULL ) return progress; |
duke@435 | 1535 | |
duke@435 | 1536 | // Finally check the default case |
duke@435 | 1537 | return StoreNode::Ideal(phase, can_reshape); |
duke@435 | 1538 | } |
duke@435 | 1539 | |
duke@435 | 1540 | //============================================================================= |
duke@435 | 1541 | //------------------------------Identity--------------------------------------- |
duke@435 | 1542 | Node *StoreCMNode::Identity( PhaseTransform *phase ) { |
duke@435 | 1543 | // No need to card mark when storing a null ptr |
duke@435 | 1544 | Node* my_store = in(MemNode::OopStore); |
duke@435 | 1545 | if (my_store->is_Store()) { |
duke@435 | 1546 | const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); |
duke@435 | 1547 | if( t1 == TypePtr::NULL_PTR ) { |
duke@435 | 1548 | return in(MemNode::Memory); |
duke@435 | 1549 | } |
duke@435 | 1550 | } |
duke@435 | 1551 | return this; |
duke@435 | 1552 | } |
duke@435 | 1553 | |
duke@435 | 1554 | //------------------------------Value----------------------------------------- |
duke@435 | 1555 | const Type *StoreCMNode::Value( PhaseTransform *phase ) const { |
duke@435 | 1556 | // If extra input is TOP ==> the result is TOP |
duke@435 | 1557 | const Type *t1 = phase->type( in(MemNode::OopStore) ); |
duke@435 | 1558 | if( t1 == Type::TOP ) return Type::TOP; |
duke@435 | 1559 | |
duke@435 | 1560 | return StoreNode::Value( phase ); |
duke@435 | 1561 | } |
duke@435 | 1562 | |
duke@435 | 1563 | |
duke@435 | 1564 | //============================================================================= |
duke@435 | 1565 | //----------------------------------SCMemProjNode------------------------------ |
duke@435 | 1566 | const Type * SCMemProjNode::Value( PhaseTransform *phase ) const |
duke@435 | 1567 | { |
duke@435 | 1568 | return bottom_type(); |
duke@435 | 1569 | } |
duke@435 | 1570 | |
duke@435 | 1571 | //============================================================================= |
duke@435 | 1572 | LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { |
duke@435 | 1573 | init_req(MemNode::Control, c ); |
duke@435 | 1574 | init_req(MemNode::Memory , mem); |
duke@435 | 1575 | init_req(MemNode::Address, adr); |
duke@435 | 1576 | init_req(MemNode::ValueIn, val); |
duke@435 | 1577 | init_req( ExpectedIn, ex ); |
duke@435 | 1578 | init_class_id(Class_LoadStore); |
duke@435 | 1579 | |
duke@435 | 1580 | } |
duke@435 | 1581 | |
duke@435 | 1582 | //============================================================================= |
duke@435 | 1583 | //-------------------------------adr_type-------------------------------------- |
duke@435 | 1584 | // Do we Match on this edge index or not? Do not match memory |
duke@435 | 1585 | const TypePtr* ClearArrayNode::adr_type() const { |
duke@435 | 1586 | Node *adr = in(3); |
duke@435 | 1587 | return MemNode::calculate_adr_type(adr->bottom_type()); |
duke@435 | 1588 | } |
duke@435 | 1589 | |
duke@435 | 1590 | //------------------------------match_edge------------------------------------- |
duke@435 | 1591 | // Do we Match on this edge index or not? Do not match memory |
duke@435 | 1592 | uint ClearArrayNode::match_edge(uint idx) const { |
duke@435 | 1593 | return idx > 1; |
duke@435 | 1594 | } |
duke@435 | 1595 | |
duke@435 | 1596 | //------------------------------Identity--------------------------------------- |
duke@435 | 1597 | // Clearing a zero length array does nothing |
duke@435 | 1598 | Node *ClearArrayNode::Identity( PhaseTransform *phase ) { |
duke@435 | 1599 | return phase->type(in(2))->higher_equal(TypeInt::ZERO) ? in(1) : this; |
duke@435 | 1600 | } |
duke@435 | 1601 | |
duke@435 | 1602 | //------------------------------Idealize--------------------------------------- |
duke@435 | 1603 | // Clearing a short array is faster with stores |
duke@435 | 1604 | Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
duke@435 | 1605 | const int unit = BytesPerLong; |
duke@435 | 1606 | const TypeX* t = phase->type(in(2))->isa_intptr_t(); |
duke@435 | 1607 | if (!t) return NULL; |
duke@435 | 1608 | if (!t->is_con()) return NULL; |
duke@435 | 1609 | intptr_t raw_count = t->get_con(); |
duke@435 | 1610 | intptr_t size = raw_count; |
duke@435 | 1611 | if (!Matcher::init_array_count_is_in_bytes) size *= unit; |
duke@435 | 1612 | // Clearing nothing uses the Identity call. |
duke@435 | 1613 | // Negative clears are possible on dead ClearArrays |
duke@435 | 1614 | // (see jck test stmt114.stmt11402.val). |
duke@435 | 1615 | if (size <= 0 || size % unit != 0) return NULL; |
duke@435 | 1616 | intptr_t count = size / unit; |
duke@435 | 1617 | // Length too long; use fast hardware clear |
duke@435 | 1618 | if (size > Matcher::init_array_short_size) return NULL; |
duke@435 | 1619 | Node *mem = in(1); |
duke@435 | 1620 | if( phase->type(mem)==Type::TOP ) return NULL; |
duke@435 | 1621 | Node *adr = in(3); |
duke@435 | 1622 | const Type* at = phase->type(adr); |
duke@435 | 1623 | if( at==Type::TOP ) return NULL; |
duke@435 | 1624 | const TypePtr* atp = at->isa_ptr(); |
duke@435 | 1625 | // adjust atp to be the correct array element address type |
duke@435 | 1626 | if (atp == NULL) atp = TypePtr::BOTTOM; |
duke@435 | 1627 | else atp = atp->add_offset(Type::OffsetBot); |
duke@435 | 1628 | // Get base for derived pointer purposes |
duke@435 | 1629 | if( adr->Opcode() != Op_AddP ) Unimplemented(); |
duke@435 | 1630 | Node *base = adr->in(1); |
duke@435 | 1631 | |
duke@435 | 1632 | Node *zero = phase->makecon(TypeLong::ZERO); |
duke@435 | 1633 | Node *off = phase->MakeConX(BytesPerLong); |
duke@435 | 1634 | mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); |
duke@435 | 1635 | count--; |
duke@435 | 1636 | while( count-- ) { |
duke@435 | 1637 | mem = phase->transform(mem); |
duke@435 | 1638 | adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); |
duke@435 | 1639 | mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); |
duke@435 | 1640 | } |
duke@435 | 1641 | return mem; |
duke@435 | 1642 | } |
duke@435 | 1643 | |
duke@435 | 1644 | //----------------------------clear_memory------------------------------------- |
duke@435 | 1645 | // Generate code to initialize object storage to zero. |
duke@435 | 1646 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
duke@435 | 1647 | intptr_t start_offset, |
duke@435 | 1648 | Node* end_offset, |
duke@435 | 1649 | PhaseGVN* phase) { |
duke@435 | 1650 | Compile* C = phase->C; |
duke@435 | 1651 | intptr_t offset = start_offset; |
duke@435 | 1652 | |
duke@435 | 1653 | int unit = BytesPerLong; |
duke@435 | 1654 | if ((offset % unit) != 0) { |
duke@435 | 1655 | Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); |
duke@435 | 1656 | adr = phase->transform(adr); |
duke@435 | 1657 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
duke@435 | 1658 | mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); |
duke@435 | 1659 | mem = phase->transform(mem); |
duke@435 | 1660 | offset += BytesPerInt; |
duke@435 | 1661 | } |
duke@435 | 1662 | assert((offset % unit) == 0, ""); |
duke@435 | 1663 | |
duke@435 | 1664 | // Initialize the remaining stuff, if any, with a ClearArray. |
duke@435 | 1665 | return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); |
duke@435 | 1666 | } |
duke@435 | 1667 | |
duke@435 | 1668 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
duke@435 | 1669 | Node* start_offset, |
duke@435 | 1670 | Node* end_offset, |
duke@435 | 1671 | PhaseGVN* phase) { |
duke@435 | 1672 | Compile* C = phase->C; |
duke@435 | 1673 | int unit = BytesPerLong; |
duke@435 | 1674 | Node* zbase = start_offset; |
duke@435 | 1675 | Node* zend = end_offset; |
duke@435 | 1676 | |
duke@435 | 1677 | // Scale to the unit required by the CPU: |
duke@435 | 1678 | if (!Matcher::init_array_count_is_in_bytes) { |
duke@435 | 1679 | Node* shift = phase->intcon(exact_log2(unit)); |
duke@435 | 1680 | zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); |
duke@435 | 1681 | zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); |
duke@435 | 1682 | } |
duke@435 | 1683 | |
duke@435 | 1684 | Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); |
duke@435 | 1685 | Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); |
duke@435 | 1686 | |
duke@435 | 1687 | // Bulk clear double-words |
duke@435 | 1688 | Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); |
duke@435 | 1689 | mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); |
duke@435 | 1690 | return phase->transform(mem); |
duke@435 | 1691 | } |
duke@435 | 1692 | |
duke@435 | 1693 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
duke@435 | 1694 | intptr_t start_offset, |
duke@435 | 1695 | intptr_t end_offset, |
duke@435 | 1696 | PhaseGVN* phase) { |
duke@435 | 1697 | Compile* C = phase->C; |
duke@435 | 1698 | assert((end_offset % BytesPerInt) == 0, "odd end offset"); |
duke@435 | 1699 | intptr_t done_offset = end_offset; |
duke@435 | 1700 | if ((done_offset % BytesPerLong) != 0) { |
duke@435 | 1701 | done_offset -= BytesPerInt; |
duke@435 | 1702 | } |
duke@435 | 1703 | if (done_offset > start_offset) { |
duke@435 | 1704 | mem = clear_memory(ctl, mem, dest, |
duke@435 | 1705 | start_offset, phase->MakeConX(done_offset), phase); |
duke@435 | 1706 | } |
duke@435 | 1707 | if (done_offset < end_offset) { // emit the final 32-bit store |
duke@435 | 1708 | Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); |
duke@435 | 1709 | adr = phase->transform(adr); |
duke@435 | 1710 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
duke@435 | 1711 | mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); |
duke@435 | 1712 | mem = phase->transform(mem); |
duke@435 | 1713 | done_offset += BytesPerInt; |
duke@435 | 1714 | } |
duke@435 | 1715 | assert(done_offset == end_offset, ""); |
duke@435 | 1716 | return mem; |
duke@435 | 1717 | } |
duke@435 | 1718 | |
duke@435 | 1719 | //============================================================================= |
duke@435 | 1720 | // Do we match on this edge? No memory edges |
duke@435 | 1721 | uint StrCompNode::match_edge(uint idx) const { |
duke@435 | 1722 | return idx == 5 || idx == 6; |
duke@435 | 1723 | } |
duke@435 | 1724 | |
duke@435 | 1725 | //------------------------------Ideal------------------------------------------ |
duke@435 | 1726 | // Return a node which is more "ideal" than the current node. Strip out |
duke@435 | 1727 | // control copies |
duke@435 | 1728 | Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
duke@435 | 1729 | return remove_dead_region(phase, can_reshape) ? this : NULL; |
duke@435 | 1730 | } |
duke@435 | 1731 | |
duke@435 | 1732 | |
duke@435 | 1733 | //============================================================================= |
duke@435 | 1734 | MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) |
duke@435 | 1735 | : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), |
duke@435 | 1736 | _adr_type(C->get_adr_type(alias_idx)) |
duke@435 | 1737 | { |
duke@435 | 1738 | init_class_id(Class_MemBar); |
duke@435 | 1739 | Node* top = C->top(); |
duke@435 | 1740 | init_req(TypeFunc::I_O,top); |
duke@435 | 1741 | init_req(TypeFunc::FramePtr,top); |
duke@435 | 1742 | init_req(TypeFunc::ReturnAdr,top); |
duke@435 | 1743 | if (precedent != NULL) |
duke@435 | 1744 | init_req(TypeFunc::Parms, precedent); |
duke@435 | 1745 | } |
duke@435 | 1746 | |
duke@435 | 1747 | //------------------------------cmp-------------------------------------------- |
duke@435 | 1748 | uint MemBarNode::hash() const { return NO_HASH; } |
duke@435 | 1749 | uint MemBarNode::cmp( const Node &n ) const { |
duke@435 | 1750 | return (&n == this); // Always fail except on self |
duke@435 | 1751 | } |
duke@435 | 1752 | |
duke@435 | 1753 | //------------------------------make------------------------------------------- |
duke@435 | 1754 | MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { |
duke@435 | 1755 | int len = Precedent + (pn == NULL? 0: 1); |
duke@435 | 1756 | switch (opcode) { |
duke@435 | 1757 | case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); |
duke@435 | 1758 | case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); |
duke@435 | 1759 | case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); |
duke@435 | 1760 | case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); |
duke@435 | 1761 | case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); |
duke@435 | 1762 | default: ShouldNotReachHere(); return NULL; |
duke@435 | 1763 | } |
duke@435 | 1764 | } |
duke@435 | 1765 | |
duke@435 | 1766 | //------------------------------Ideal------------------------------------------ |
duke@435 | 1767 | // Return a node which is more "ideal" than the current node. Strip out |
duke@435 | 1768 | // control copies |
duke@435 | 1769 | Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 1770 | if (remove_dead_region(phase, can_reshape)) return this; |
duke@435 | 1771 | return NULL; |
duke@435 | 1772 | } |
duke@435 | 1773 | |
duke@435 | 1774 | //------------------------------Value------------------------------------------ |
duke@435 | 1775 | const Type *MemBarNode::Value( PhaseTransform *phase ) const { |
duke@435 | 1776 | if( !in(0) ) return Type::TOP; |
duke@435 | 1777 | if( phase->type(in(0)) == Type::TOP ) |
duke@435 | 1778 | return Type::TOP; |
duke@435 | 1779 | return TypeTuple::MEMBAR; |
duke@435 | 1780 | } |
duke@435 | 1781 | |
duke@435 | 1782 | //------------------------------match------------------------------------------ |
duke@435 | 1783 | // Construct projections for memory. |
duke@435 | 1784 | Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { |
duke@435 | 1785 | switch (proj->_con) { |
duke@435 | 1786 | case TypeFunc::Control: |
duke@435 | 1787 | case TypeFunc::Memory: |
duke@435 | 1788 | return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); |
duke@435 | 1789 | } |
duke@435 | 1790 | ShouldNotReachHere(); |
duke@435 | 1791 | return NULL; |
duke@435 | 1792 | } |
duke@435 | 1793 | |
duke@435 | 1794 | //===========================InitializeNode==================================== |
duke@435 | 1795 | // SUMMARY: |
duke@435 | 1796 | // This node acts as a memory barrier on raw memory, after some raw stores. |
duke@435 | 1797 | // The 'cooked' oop value feeds from the Initialize, not the Allocation. |
duke@435 | 1798 | // The Initialize can 'capture' suitably constrained stores as raw inits. |
duke@435 | 1799 | // It can coalesce related raw stores into larger units (called 'tiles'). |
duke@435 | 1800 | // It can avoid zeroing new storage for memory units which have raw inits. |
duke@435 | 1801 | // At macro-expansion, it is marked 'complete', and does not optimize further. |
duke@435 | 1802 | // |
duke@435 | 1803 | // EXAMPLE: |
duke@435 | 1804 | // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. |
duke@435 | 1805 | // ctl = incoming control; mem* = incoming memory |
duke@435 | 1806 | // (Note: A star * on a memory edge denotes I/O and other standard edges.) |
duke@435 | 1807 | // First allocate uninitialized memory and fill in the header: |
duke@435 | 1808 | // alloc = (Allocate ctl mem* 16 #short[].klass ...) |
duke@435 | 1809 | // ctl := alloc.Control; mem* := alloc.Memory* |
duke@435 | 1810 | // rawmem = alloc.Memory; rawoop = alloc.RawAddress |
duke@435 | 1811 | // Then initialize to zero the non-header parts of the raw memory block: |
duke@435 | 1812 | // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) |
duke@435 | 1813 | // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory |
duke@435 | 1814 | // After the initialize node executes, the object is ready for service: |
duke@435 | 1815 | // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) |
duke@435 | 1816 | // Suppose its body is immediately initialized as {1,2}: |
duke@435 | 1817 | // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) |
duke@435 | 1818 | // store2 = (StoreC init.Control store1 (+ oop 14) 2) |
duke@435 | 1819 | // mem.SLICE(#short[*]) := store2 |
duke@435 | 1820 | // |
duke@435 | 1821 | // DETAILS: |
duke@435 | 1822 | // An InitializeNode collects and isolates object initialization after |
duke@435 | 1823 | // an AllocateNode and before the next possible safepoint. As a |
duke@435 | 1824 | // memory barrier (MemBarNode), it keeps critical stores from drifting |
duke@435 | 1825 | // down past any safepoint or any publication of the allocation. |
duke@435 | 1826 | // Before this barrier, a newly-allocated object may have uninitialized bits. |
duke@435 | 1827 | // After this barrier, it may be treated as a real oop, and GC is allowed. |
duke@435 | 1828 | // |
duke@435 | 1829 | // The semantics of the InitializeNode include an implicit zeroing of |
duke@435 | 1830 | // the new object from object header to the end of the object. |
duke@435 | 1831 | // (The object header and end are determined by the AllocateNode.) |
duke@435 | 1832 | // |
duke@435 | 1833 | // Certain stores may be added as direct inputs to the InitializeNode. |
duke@435 | 1834 | // These stores must update raw memory, and they must be to addresses |
duke@435 | 1835 | // derived from the raw address produced by AllocateNode, and with |
duke@435 | 1836 | // a constant offset. They must be ordered by increasing offset. |
duke@435 | 1837 | // The first one is at in(RawStores), the last at in(req()-1). |
duke@435 | 1838 | // Unlike most memory operations, they are not linked in a chain, |
duke@435 | 1839 | // but are displayed in parallel as users of the rawmem output of |
duke@435 | 1840 | // the allocation. |
duke@435 | 1841 | // |
duke@435 | 1842 | // (See comments in InitializeNode::capture_store, which continue |
duke@435 | 1843 | // the example given above.) |
duke@435 | 1844 | // |
duke@435 | 1845 | // When the associated Allocate is macro-expanded, the InitializeNode |
duke@435 | 1846 | // may be rewritten to optimize collected stores. A ClearArrayNode |
duke@435 | 1847 | // may also be created at that point to represent any required zeroing. |
duke@435 | 1848 | // The InitializeNode is then marked 'complete', prohibiting further |
duke@435 | 1849 | // capturing of nearby memory operations. |
duke@435 | 1850 | // |
duke@435 | 1851 | // During macro-expansion, all captured initializations which store |
duke@435 | 1852 | // constant values of 32 bits or smaller are coalesced (if advantagous) |
duke@435 | 1853 | // into larger 'tiles' 32 or 64 bits. This allows an object to be |
duke@435 | 1854 | // initialized in fewer memory operations. Memory words which are |
duke@435 | 1855 | // covered by neither tiles nor non-constant stores are pre-zeroed |
duke@435 | 1856 | // by explicit stores of zero. (The code shape happens to do all |
duke@435 | 1857 | // zeroing first, then all other stores, with both sequences occurring |
duke@435 | 1858 | // in order of ascending offsets.) |
duke@435 | 1859 | // |
duke@435 | 1860 | // Alternatively, code may be inserted between an AllocateNode and its |
duke@435 | 1861 | // InitializeNode, to perform arbitrary initialization of the new object. |
duke@435 | 1862 | // E.g., the object copying intrinsics insert complex data transfers here. |
duke@435 | 1863 | // The initialization must then be marked as 'complete' disable the |
duke@435 | 1864 | // built-in zeroing semantics and the collection of initializing stores. |
duke@435 | 1865 | // |
duke@435 | 1866 | // While an InitializeNode is incomplete, reads from the memory state |
duke@435 | 1867 | // produced by it are optimizable if they match the control edge and |
duke@435 | 1868 | // new oop address associated with the allocation/initialization. |
duke@435 | 1869 | // They return a stored value (if the offset matches) or else zero. |
duke@435 | 1870 | // A write to the memory state, if it matches control and address, |
duke@435 | 1871 | // and if it is to a constant offset, may be 'captured' by the |
duke@435 | 1872 | // InitializeNode. It is cloned as a raw memory operation and rewired |
duke@435 | 1873 | // inside the initialization, to the raw oop produced by the allocation. |
duke@435 | 1874 | // Operations on addresses which are provably distinct (e.g., to |
duke@435 | 1875 | // other AllocateNodes) are allowed to bypass the initialization. |
duke@435 | 1876 | // |
duke@435 | 1877 | // The effect of all this is to consolidate object initialization |
duke@435 | 1878 | // (both arrays and non-arrays, both piecewise and bulk) into a |
duke@435 | 1879 | // single location, where it can be optimized as a unit. |
duke@435 | 1880 | // |
duke@435 | 1881 | // Only stores with an offset less than TrackedInitializationLimit words |
duke@435 | 1882 | // will be considered for capture by an InitializeNode. This puts a |
duke@435 | 1883 | // reasonable limit on the complexity of optimized initializations. |
duke@435 | 1884 | |
duke@435 | 1885 | //---------------------------InitializeNode------------------------------------ |
duke@435 | 1886 | InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) |
duke@435 | 1887 | : _is_complete(false), |
duke@435 | 1888 | MemBarNode(C, adr_type, rawoop) |
duke@435 | 1889 | { |
duke@435 | 1890 | init_class_id(Class_Initialize); |
duke@435 | 1891 | |
duke@435 | 1892 | assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); |
duke@435 | 1893 | assert(in(RawAddress) == rawoop, "proper init"); |
duke@435 | 1894 | // Note: allocation() can be NULL, for secondary initialization barriers |
duke@435 | 1895 | } |
duke@435 | 1896 | |
duke@435 | 1897 | // Since this node is not matched, it will be processed by the |
duke@435 | 1898 | // register allocator. Declare that there are no constraints |
duke@435 | 1899 | // on the allocation of the RawAddress edge. |
duke@435 | 1900 | const RegMask &InitializeNode::in_RegMask(uint idx) const { |
duke@435 | 1901 | // This edge should be set to top, by the set_complete. But be conservative. |
duke@435 | 1902 | if (idx == InitializeNode::RawAddress) |
duke@435 | 1903 | return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); |
duke@435 | 1904 | return RegMask::Empty; |
duke@435 | 1905 | } |
duke@435 | 1906 | |
duke@435 | 1907 | Node* InitializeNode::memory(uint alias_idx) { |
duke@435 | 1908 | Node* mem = in(Memory); |
duke@435 | 1909 | if (mem->is_MergeMem()) { |
duke@435 | 1910 | return mem->as_MergeMem()->memory_at(alias_idx); |
duke@435 | 1911 | } else { |
duke@435 | 1912 | // incoming raw memory is not split |
duke@435 | 1913 | return mem; |
duke@435 | 1914 | } |
duke@435 | 1915 | } |
duke@435 | 1916 | |
duke@435 | 1917 | bool InitializeNode::is_non_zero() { |
duke@435 | 1918 | if (is_complete()) return false; |
duke@435 | 1919 | remove_extra_zeroes(); |
duke@435 | 1920 | return (req() > RawStores); |
duke@435 | 1921 | } |
duke@435 | 1922 | |
duke@435 | 1923 | void InitializeNode::set_complete(PhaseGVN* phase) { |
duke@435 | 1924 | assert(!is_complete(), "caller responsibility"); |
duke@435 | 1925 | _is_complete = true; |
duke@435 | 1926 | |
duke@435 | 1927 | // After this node is complete, it contains a bunch of |
duke@435 | 1928 | // raw-memory initializations. There is no need for |
duke@435 | 1929 | // it to have anything to do with non-raw memory effects. |
duke@435 | 1930 | // Therefore, tell all non-raw users to re-optimize themselves, |
duke@435 | 1931 | // after skipping the memory effects of this initialization. |
duke@435 | 1932 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
duke@435 | 1933 | if (igvn) igvn->add_users_to_worklist(this); |
duke@435 | 1934 | } |
duke@435 | 1935 | |
duke@435 | 1936 | // convenience function |
duke@435 | 1937 | // return false if the init contains any stores already |
duke@435 | 1938 | bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { |
duke@435 | 1939 | InitializeNode* init = initialization(); |
duke@435 | 1940 | if (init == NULL || init->is_complete()) return false; |
duke@435 | 1941 | init->remove_extra_zeroes(); |
duke@435 | 1942 | // for now, if this allocation has already collected any inits, bail: |
duke@435 | 1943 | if (init->is_non_zero()) return false; |
duke@435 | 1944 | init->set_complete(phase); |
duke@435 | 1945 | return true; |
duke@435 | 1946 | } |
duke@435 | 1947 | |
duke@435 | 1948 | void InitializeNode::remove_extra_zeroes() { |
duke@435 | 1949 | if (req() == RawStores) return; |
duke@435 | 1950 | Node* zmem = zero_memory(); |
duke@435 | 1951 | uint fill = RawStores; |
duke@435 | 1952 | for (uint i = fill; i < req(); i++) { |
duke@435 | 1953 | Node* n = in(i); |
duke@435 | 1954 | if (n->is_top() || n == zmem) continue; // skip |
duke@435 | 1955 | if (fill < i) set_req(fill, n); // compact |
duke@435 | 1956 | ++fill; |
duke@435 | 1957 | } |
duke@435 | 1958 | // delete any empty spaces created: |
duke@435 | 1959 | while (fill < req()) { |
duke@435 | 1960 | del_req(fill); |
duke@435 | 1961 | } |
duke@435 | 1962 | } |
duke@435 | 1963 | |
duke@435 | 1964 | // Helper for remembering which stores go with which offsets. |
duke@435 | 1965 | intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { |
duke@435 | 1966 | if (!st->is_Store()) return -1; // can happen to dead code via subsume_node |
duke@435 | 1967 | intptr_t offset = -1; |
duke@435 | 1968 | Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), |
duke@435 | 1969 | phase, offset); |
duke@435 | 1970 | if (base == NULL) return -1; // something is dead, |
duke@435 | 1971 | if (offset < 0) return -1; // dead, dead |
duke@435 | 1972 | return offset; |
duke@435 | 1973 | } |
duke@435 | 1974 | |
duke@435 | 1975 | // Helper for proving that an initialization expression is |
duke@435 | 1976 | // "simple enough" to be folded into an object initialization. |
duke@435 | 1977 | // Attempts to prove that a store's initial value 'n' can be captured |
duke@435 | 1978 | // within the initialization without creating a vicious cycle, such as: |
duke@435 | 1979 | // { Foo p = new Foo(); p.next = p; } |
duke@435 | 1980 | // True for constants and parameters and small combinations thereof. |
duke@435 | 1981 | bool InitializeNode::detect_init_independence(Node* n, |
duke@435 | 1982 | bool st_is_pinned, |
duke@435 | 1983 | int& count) { |
duke@435 | 1984 | if (n == NULL) return true; // (can this really happen?) |
duke@435 | 1985 | if (n->is_Proj()) n = n->in(0); |
duke@435 | 1986 | if (n == this) return false; // found a cycle |
duke@435 | 1987 | if (n->is_Con()) return true; |
duke@435 | 1988 | if (n->is_Start()) return true; // params, etc., are OK |
duke@435 | 1989 | if (n->is_Root()) return true; // even better |
duke@435 | 1990 | |
duke@435 | 1991 | Node* ctl = n->in(0); |
duke@435 | 1992 | if (ctl != NULL && !ctl->is_top()) { |
duke@435 | 1993 | if (ctl->is_Proj()) ctl = ctl->in(0); |
duke@435 | 1994 | if (ctl == this) return false; |
duke@435 | 1995 | |
duke@435 | 1996 | // If we already know that the enclosing memory op is pinned right after |
duke@435 | 1997 | // the init, then any control flow that the store has picked up |
duke@435 | 1998 | // must have preceded the init, or else be equal to the init. |
duke@435 | 1999 | // Even after loop optimizations (which might change control edges) |
duke@435 | 2000 | // a store is never pinned *before* the availability of its inputs. |
duke@435 | 2001 | if (!MemNode::detect_dominating_control(ctl, this->in(0))) |
duke@435 | 2002 | return false; // failed to prove a good control |
duke@435 | 2003 | |
duke@435 | 2004 | } |
duke@435 | 2005 | |
duke@435 | 2006 | // Check data edges for possible dependencies on 'this'. |
duke@435 | 2007 | if ((count += 1) > 20) return false; // complexity limit |
duke@435 | 2008 | for (uint i = 1; i < n->req(); i++) { |
duke@435 | 2009 | Node* m = n->in(i); |
duke@435 | 2010 | if (m == NULL || m == n || m->is_top()) continue; |
duke@435 | 2011 | uint first_i = n->find_edge(m); |
duke@435 | 2012 | if (i != first_i) continue; // process duplicate edge just once |
duke@435 | 2013 | if (!detect_init_independence(m, st_is_pinned, count)) { |
duke@435 | 2014 | return false; |
duke@435 | 2015 | } |
duke@435 | 2016 | } |
duke@435 | 2017 | |
duke@435 | 2018 | return true; |
duke@435 | 2019 | } |
duke@435 | 2020 | |
duke@435 | 2021 | // Here are all the checks a Store must pass before it can be moved into |
duke@435 | 2022 | // an initialization. Returns zero if a check fails. |
duke@435 | 2023 | // On success, returns the (constant) offset to which the store applies, |
duke@435 | 2024 | // within the initialized memory. |
duke@435 | 2025 | intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { |
duke@435 | 2026 | const int FAIL = 0; |
duke@435 | 2027 | if (st->req() != MemNode::ValueIn + 1) |
duke@435 | 2028 | return FAIL; // an inscrutable StoreNode (card mark?) |
duke@435 | 2029 | Node* ctl = st->in(MemNode::Control); |
duke@435 | 2030 | if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) |
duke@435 | 2031 | return FAIL; // must be unconditional after the initialization |
duke@435 | 2032 | Node* mem = st->in(MemNode::Memory); |
duke@435 | 2033 | if (!(mem->is_Proj() && mem->in(0) == this)) |
duke@435 | 2034 | return FAIL; // must not be preceded by other stores |
duke@435 | 2035 | Node* adr = st->in(MemNode::Address); |
duke@435 | 2036 | intptr_t offset; |
duke@435 | 2037 | AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); |
duke@435 | 2038 | if (alloc == NULL) |
duke@435 | 2039 | return FAIL; // inscrutable address |
duke@435 | 2040 | if (alloc != allocation()) |
duke@435 | 2041 | return FAIL; // wrong allocation! (store needs to float up) |
duke@435 | 2042 | Node* val = st->in(MemNode::ValueIn); |
duke@435 | 2043 | int complexity_count = 0; |
duke@435 | 2044 | if (!detect_init_independence(val, true, complexity_count)) |
duke@435 | 2045 | return FAIL; // stored value must be 'simple enough' |
duke@435 | 2046 | |
duke@435 | 2047 | return offset; // success |
duke@435 | 2048 | } |
duke@435 | 2049 | |
duke@435 | 2050 | // Find the captured store in(i) which corresponds to the range |
duke@435 | 2051 | // [start..start+size) in the initialized object. |
duke@435 | 2052 | // If there is one, return its index i. If there isn't, return the |
duke@435 | 2053 | // negative of the index where it should be inserted. |
duke@435 | 2054 | // Return 0 if the queried range overlaps an initialization boundary |
duke@435 | 2055 | // or if dead code is encountered. |
duke@435 | 2056 | // If size_in_bytes is zero, do not bother with overlap checks. |
duke@435 | 2057 | int InitializeNode::captured_store_insertion_point(intptr_t start, |
duke@435 | 2058 | int size_in_bytes, |
duke@435 | 2059 | PhaseTransform* phase) { |
duke@435 | 2060 | const int FAIL = 0, MAX_STORE = BytesPerLong; |
duke@435 | 2061 | |
duke@435 | 2062 | if (is_complete()) |
duke@435 | 2063 | return FAIL; // arraycopy got here first; punt |
duke@435 | 2064 | |
duke@435 | 2065 | assert(allocation() != NULL, "must be present"); |
duke@435 | 2066 | |
duke@435 | 2067 | // no negatives, no header fields: |
duke@435 | 2068 | if (start < (intptr_t) sizeof(oopDesc)) return FAIL; |
duke@435 | 2069 | if (start < (intptr_t) sizeof(arrayOopDesc) && |
duke@435 | 2070 | start < (intptr_t) allocation()->minimum_header_size()) return FAIL; |
duke@435 | 2071 | |
duke@435 | 2072 | // after a certain size, we bail out on tracking all the stores: |
duke@435 | 2073 | intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); |
duke@435 | 2074 | if (start >= ti_limit) return FAIL; |
duke@435 | 2075 | |
duke@435 | 2076 | for (uint i = InitializeNode::RawStores, limit = req(); ; ) { |
duke@435 | 2077 | if (i >= limit) return -(int)i; // not found; here is where to put it |
duke@435 | 2078 | |
duke@435 | 2079 | Node* st = in(i); |
duke@435 | 2080 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 2081 | if (st_off < 0) { |
duke@435 | 2082 | if (st != zero_memory()) { |
duke@435 | 2083 | return FAIL; // bail out if there is dead garbage |
duke@435 | 2084 | } |
duke@435 | 2085 | } else if (st_off > start) { |
duke@435 | 2086 | // ...we are done, since stores are ordered |
duke@435 | 2087 | if (st_off < start + size_in_bytes) { |
duke@435 | 2088 | return FAIL; // the next store overlaps |
duke@435 | 2089 | } |
duke@435 | 2090 | return -(int)i; // not found; here is where to put it |
duke@435 | 2091 | } else if (st_off < start) { |
duke@435 | 2092 | if (size_in_bytes != 0 && |
duke@435 | 2093 | start < st_off + MAX_STORE && |
duke@435 | 2094 | start < st_off + st->as_Store()->memory_size()) { |
duke@435 | 2095 | return FAIL; // the previous store overlaps |
duke@435 | 2096 | } |
duke@435 | 2097 | } else { |
duke@435 | 2098 | if (size_in_bytes != 0 && |
duke@435 | 2099 | st->as_Store()->memory_size() != size_in_bytes) { |
duke@435 | 2100 | return FAIL; // mismatched store size |
duke@435 | 2101 | } |
duke@435 | 2102 | return i; |
duke@435 | 2103 | } |
duke@435 | 2104 | |
duke@435 | 2105 | ++i; |
duke@435 | 2106 | } |
duke@435 | 2107 | } |
duke@435 | 2108 | |
duke@435 | 2109 | // Look for a captured store which initializes at the offset 'start' |
duke@435 | 2110 | // with the given size. If there is no such store, and no other |
duke@435 | 2111 | // initialization interferes, then return zero_memory (the memory |
duke@435 | 2112 | // projection of the AllocateNode). |
duke@435 | 2113 | Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, |
duke@435 | 2114 | PhaseTransform* phase) { |
duke@435 | 2115 | assert(stores_are_sane(phase), ""); |
duke@435 | 2116 | int i = captured_store_insertion_point(start, size_in_bytes, phase); |
duke@435 | 2117 | if (i == 0) { |
duke@435 | 2118 | return NULL; // something is dead |
duke@435 | 2119 | } else if (i < 0) { |
duke@435 | 2120 | return zero_memory(); // just primordial zero bits here |
duke@435 | 2121 | } else { |
duke@435 | 2122 | Node* st = in(i); // here is the store at this position |
duke@435 | 2123 | assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); |
duke@435 | 2124 | return st; |
duke@435 | 2125 | } |
duke@435 | 2126 | } |
duke@435 | 2127 | |
duke@435 | 2128 | // Create, as a raw pointer, an address within my new object at 'offset'. |
duke@435 | 2129 | Node* InitializeNode::make_raw_address(intptr_t offset, |
duke@435 | 2130 | PhaseTransform* phase) { |
duke@435 | 2131 | Node* addr = in(RawAddress); |
duke@435 | 2132 | if (offset != 0) { |
duke@435 | 2133 | Compile* C = phase->C; |
duke@435 | 2134 | addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, |
duke@435 | 2135 | phase->MakeConX(offset)) ); |
duke@435 | 2136 | } |
duke@435 | 2137 | return addr; |
duke@435 | 2138 | } |
duke@435 | 2139 | |
duke@435 | 2140 | // Clone the given store, converting it into a raw store |
duke@435 | 2141 | // initializing a field or element of my new object. |
duke@435 | 2142 | // Caller is responsible for retiring the original store, |
duke@435 | 2143 | // with subsume_node or the like. |
duke@435 | 2144 | // |
duke@435 | 2145 | // From the example above InitializeNode::InitializeNode, |
duke@435 | 2146 | // here are the old stores to be captured: |
duke@435 | 2147 | // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) |
duke@435 | 2148 | // store2 = (StoreC init.Control store1 (+ oop 14) 2) |
duke@435 | 2149 | // |
duke@435 | 2150 | // Here is the changed code; note the extra edges on init: |
duke@435 | 2151 | // alloc = (Allocate ...) |
duke@435 | 2152 | // rawoop = alloc.RawAddress |
duke@435 | 2153 | // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) |
duke@435 | 2154 | // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) |
duke@435 | 2155 | // init = (Initialize alloc.Control alloc.Memory rawoop |
duke@435 | 2156 | // rawstore1 rawstore2) |
duke@435 | 2157 | // |
duke@435 | 2158 | Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, |
duke@435 | 2159 | PhaseTransform* phase) { |
duke@435 | 2160 | assert(stores_are_sane(phase), ""); |
duke@435 | 2161 | |
duke@435 | 2162 | if (start < 0) return NULL; |
duke@435 | 2163 | assert(can_capture_store(st, phase) == start, "sanity"); |
duke@435 | 2164 | |
duke@435 | 2165 | Compile* C = phase->C; |
duke@435 | 2166 | int size_in_bytes = st->memory_size(); |
duke@435 | 2167 | int i = captured_store_insertion_point(start, size_in_bytes, phase); |
duke@435 | 2168 | if (i == 0) return NULL; // bail out |
duke@435 | 2169 | Node* prev_mem = NULL; // raw memory for the captured store |
duke@435 | 2170 | if (i > 0) { |
duke@435 | 2171 | prev_mem = in(i); // there is a pre-existing store under this one |
duke@435 | 2172 | set_req(i, C->top()); // temporarily disconnect it |
duke@435 | 2173 | // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. |
duke@435 | 2174 | } else { |
duke@435 | 2175 | i = -i; // no pre-existing store |
duke@435 | 2176 | prev_mem = zero_memory(); // a slice of the newly allocated object |
duke@435 | 2177 | if (i > InitializeNode::RawStores && in(i-1) == prev_mem) |
duke@435 | 2178 | set_req(--i, C->top()); // reuse this edge; it has been folded away |
duke@435 | 2179 | else |
duke@435 | 2180 | ins_req(i, C->top()); // build a new edge |
duke@435 | 2181 | } |
duke@435 | 2182 | Node* new_st = st->clone(); |
duke@435 | 2183 | new_st->set_req(MemNode::Control, in(Control)); |
duke@435 | 2184 | new_st->set_req(MemNode::Memory, prev_mem); |
duke@435 | 2185 | new_st->set_req(MemNode::Address, make_raw_address(start, phase)); |
duke@435 | 2186 | new_st = phase->transform(new_st); |
duke@435 | 2187 | |
duke@435 | 2188 | // At this point, new_st might have swallowed a pre-existing store |
duke@435 | 2189 | // at the same offset, or perhaps new_st might have disappeared, |
duke@435 | 2190 | // if it redundantly stored the same value (or zero to fresh memory). |
duke@435 | 2191 | |
duke@435 | 2192 | // In any case, wire it in: |
duke@435 | 2193 | set_req(i, new_st); |
duke@435 | 2194 | |
duke@435 | 2195 | // The caller may now kill the old guy. |
duke@435 | 2196 | DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); |
duke@435 | 2197 | assert(check_st == new_st || check_st == NULL, "must be findable"); |
duke@435 | 2198 | assert(!is_complete(), ""); |
duke@435 | 2199 | return new_st; |
duke@435 | 2200 | } |
duke@435 | 2201 | |
duke@435 | 2202 | static bool store_constant(jlong* tiles, int num_tiles, |
duke@435 | 2203 | intptr_t st_off, int st_size, |
duke@435 | 2204 | jlong con) { |
duke@435 | 2205 | if ((st_off & (st_size-1)) != 0) |
duke@435 | 2206 | return false; // strange store offset (assume size==2**N) |
duke@435 | 2207 | address addr = (address)tiles + st_off; |
duke@435 | 2208 | assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); |
duke@435 | 2209 | switch (st_size) { |
duke@435 | 2210 | case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; |
duke@435 | 2211 | case sizeof(jchar): *(jchar*) addr = (jchar) con; break; |
duke@435 | 2212 | case sizeof(jint): *(jint*) addr = (jint) con; break; |
duke@435 | 2213 | case sizeof(jlong): *(jlong*) addr = (jlong) con; break; |
duke@435 | 2214 | default: return false; // strange store size (detect size!=2**N here) |
duke@435 | 2215 | } |
duke@435 | 2216 | return true; // return success to caller |
duke@435 | 2217 | } |
duke@435 | 2218 | |
duke@435 | 2219 | // Coalesce subword constants into int constants and possibly |
duke@435 | 2220 | // into long constants. The goal, if the CPU permits, |
duke@435 | 2221 | // is to initialize the object with a small number of 64-bit tiles. |
duke@435 | 2222 | // Also, convert floating-point constants to bit patterns. |
duke@435 | 2223 | // Non-constants are not relevant to this pass. |
duke@435 | 2224 | // |
duke@435 | 2225 | // In terms of the running example on InitializeNode::InitializeNode |
duke@435 | 2226 | // and InitializeNode::capture_store, here is the transformation |
duke@435 | 2227 | // of rawstore1 and rawstore2 into rawstore12: |
duke@435 | 2228 | // alloc = (Allocate ...) |
duke@435 | 2229 | // rawoop = alloc.RawAddress |
duke@435 | 2230 | // tile12 = 0x00010002 |
duke@435 | 2231 | // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) |
duke@435 | 2232 | // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) |
duke@435 | 2233 | // |
duke@435 | 2234 | void |
duke@435 | 2235 | InitializeNode::coalesce_subword_stores(intptr_t header_size, |
duke@435 | 2236 | Node* size_in_bytes, |
duke@435 | 2237 | PhaseGVN* phase) { |
duke@435 | 2238 | Compile* C = phase->C; |
duke@435 | 2239 | |
duke@435 | 2240 | assert(stores_are_sane(phase), ""); |
duke@435 | 2241 | // Note: After this pass, they are not completely sane, |
duke@435 | 2242 | // since there may be some overlaps. |
duke@435 | 2243 | |
duke@435 | 2244 | int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; |
duke@435 | 2245 | |
duke@435 | 2246 | intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); |
duke@435 | 2247 | intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); |
duke@435 | 2248 | size_limit = MIN2(size_limit, ti_limit); |
duke@435 | 2249 | size_limit = align_size_up(size_limit, BytesPerLong); |
duke@435 | 2250 | int num_tiles = size_limit / BytesPerLong; |
duke@435 | 2251 | |
duke@435 | 2252 | // allocate space for the tile map: |
duke@435 | 2253 | const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small |
duke@435 | 2254 | jlong tiles_buf[small_len]; |
duke@435 | 2255 | Node* nodes_buf[small_len]; |
duke@435 | 2256 | jlong inits_buf[small_len]; |
duke@435 | 2257 | jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] |
duke@435 | 2258 | : NEW_RESOURCE_ARRAY(jlong, num_tiles)); |
duke@435 | 2259 | Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] |
duke@435 | 2260 | : NEW_RESOURCE_ARRAY(Node*, num_tiles)); |
duke@435 | 2261 | jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] |
duke@435 | 2262 | : NEW_RESOURCE_ARRAY(jlong, num_tiles)); |
duke@435 | 2263 | // tiles: exact bitwise model of all primitive constants |
duke@435 | 2264 | // nodes: last constant-storing node subsumed into the tiles model |
duke@435 | 2265 | // inits: which bytes (in each tile) are touched by any initializations |
duke@435 | 2266 | |
duke@435 | 2267 | //// Pass A: Fill in the tile model with any relevant stores. |
duke@435 | 2268 | |
duke@435 | 2269 | Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); |
duke@435 | 2270 | Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); |
duke@435 | 2271 | Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); |
duke@435 | 2272 | Node* zmem = zero_memory(); // initially zero memory state |
duke@435 | 2273 | for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { |
duke@435 | 2274 | Node* st = in(i); |
duke@435 | 2275 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 2276 | |
duke@435 | 2277 | // Figure out the store's offset and constant value: |
duke@435 | 2278 | if (st_off < header_size) continue; //skip (ignore header) |
duke@435 | 2279 | if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) |
duke@435 | 2280 | int st_size = st->as_Store()->memory_size(); |
duke@435 | 2281 | if (st_off + st_size > size_limit) break; |
duke@435 | 2282 | |
duke@435 | 2283 | // Record which bytes are touched, whether by constant or not. |
duke@435 | 2284 | if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) |
duke@435 | 2285 | continue; // skip (strange store size) |
duke@435 | 2286 | |
duke@435 | 2287 | const Type* val = phase->type(st->in(MemNode::ValueIn)); |
duke@435 | 2288 | if (!val->singleton()) continue; //skip (non-con store) |
duke@435 | 2289 | BasicType type = val->basic_type(); |
duke@435 | 2290 | |
duke@435 | 2291 | jlong con = 0; |
duke@435 | 2292 | switch (type) { |
duke@435 | 2293 | case T_INT: con = val->is_int()->get_con(); break; |
duke@435 | 2294 | case T_LONG: con = val->is_long()->get_con(); break; |
duke@435 | 2295 | case T_FLOAT: con = jint_cast(val->getf()); break; |
duke@435 | 2296 | case T_DOUBLE: con = jlong_cast(val->getd()); break; |
duke@435 | 2297 | default: continue; //skip (odd store type) |
duke@435 | 2298 | } |
duke@435 | 2299 | |
duke@435 | 2300 | if (type == T_LONG && Matcher::isSimpleConstant64(con) && |
duke@435 | 2301 | st->Opcode() == Op_StoreL) { |
duke@435 | 2302 | continue; // This StoreL is already optimal. |
duke@435 | 2303 | } |
duke@435 | 2304 | |
duke@435 | 2305 | // Store down the constant. |
duke@435 | 2306 | store_constant(tiles, num_tiles, st_off, st_size, con); |
duke@435 | 2307 | |
duke@435 | 2308 | intptr_t j = st_off >> LogBytesPerLong; |
duke@435 | 2309 | |
duke@435 | 2310 | if (type == T_INT && st_size == BytesPerInt |
duke@435 | 2311 | && (st_off & BytesPerInt) == BytesPerInt) { |
duke@435 | 2312 | jlong lcon = tiles[j]; |
duke@435 | 2313 | if (!Matcher::isSimpleConstant64(lcon) && |
duke@435 | 2314 | st->Opcode() == Op_StoreI) { |
duke@435 | 2315 | // This StoreI is already optimal by itself. |
duke@435 | 2316 | jint* intcon = (jint*) &tiles[j]; |
duke@435 | 2317 | intcon[1] = 0; // undo the store_constant() |
duke@435 | 2318 | |
duke@435 | 2319 | // If the previous store is also optimal by itself, back up and |
duke@435 | 2320 | // undo the action of the previous loop iteration... if we can. |
duke@435 | 2321 | // But if we can't, just let the previous half take care of itself. |
duke@435 | 2322 | st = nodes[j]; |
duke@435 | 2323 | st_off -= BytesPerInt; |
duke@435 | 2324 | con = intcon[0]; |
duke@435 | 2325 | if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { |
duke@435 | 2326 | assert(st_off >= header_size, "still ignoring header"); |
duke@435 | 2327 | assert(get_store_offset(st, phase) == st_off, "must be"); |
duke@435 | 2328 | assert(in(i-1) == zmem, "must be"); |
duke@435 | 2329 | DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); |
duke@435 | 2330 | assert(con == tcon->is_int()->get_con(), "must be"); |
duke@435 | 2331 | // Undo the effects of the previous loop trip, which swallowed st: |
duke@435 | 2332 | intcon[0] = 0; // undo store_constant() |
duke@435 | 2333 | set_req(i-1, st); // undo set_req(i, zmem) |
duke@435 | 2334 | nodes[j] = NULL; // undo nodes[j] = st |
duke@435 | 2335 | --old_subword; // undo ++old_subword |
duke@435 | 2336 | } |
duke@435 | 2337 | continue; // This StoreI is already optimal. |
duke@435 | 2338 | } |
duke@435 | 2339 | } |
duke@435 | 2340 | |
duke@435 | 2341 | // This store is not needed. |
duke@435 | 2342 | set_req(i, zmem); |
duke@435 | 2343 | nodes[j] = st; // record for the moment |
duke@435 | 2344 | if (st_size < BytesPerLong) // something has changed |
duke@435 | 2345 | ++old_subword; // includes int/float, but who's counting... |
duke@435 | 2346 | else ++old_long; |
duke@435 | 2347 | } |
duke@435 | 2348 | |
duke@435 | 2349 | if ((old_subword + old_long) == 0) |
duke@435 | 2350 | return; // nothing more to do |
duke@435 | 2351 | |
duke@435 | 2352 | //// Pass B: Convert any non-zero tiles into optimal constant stores. |
duke@435 | 2353 | // Be sure to insert them before overlapping non-constant stores. |
duke@435 | 2354 | // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) |
duke@435 | 2355 | for (int j = 0; j < num_tiles; j++) { |
duke@435 | 2356 | jlong con = tiles[j]; |
duke@435 | 2357 | jlong init = inits[j]; |
duke@435 | 2358 | if (con == 0) continue; |
duke@435 | 2359 | jint con0, con1; // split the constant, address-wise |
duke@435 | 2360 | jint init0, init1; // split the init map, address-wise |
duke@435 | 2361 | { union { jlong con; jint intcon[2]; } u; |
duke@435 | 2362 | u.con = con; |
duke@435 | 2363 | con0 = u.intcon[0]; |
duke@435 | 2364 | con1 = u.intcon[1]; |
duke@435 | 2365 | u.con = init; |
duke@435 | 2366 | init0 = u.intcon[0]; |
duke@435 | 2367 | init1 = u.intcon[1]; |
duke@435 | 2368 | } |
duke@435 | 2369 | |
duke@435 | 2370 | Node* old = nodes[j]; |
duke@435 | 2371 | assert(old != NULL, "need the prior store"); |
duke@435 | 2372 | intptr_t offset = (j * BytesPerLong); |
duke@435 | 2373 | |
duke@435 | 2374 | bool split = !Matcher::isSimpleConstant64(con); |
duke@435 | 2375 | |
duke@435 | 2376 | if (offset < header_size) { |
duke@435 | 2377 | assert(offset + BytesPerInt >= header_size, "second int counts"); |
duke@435 | 2378 | assert(*(jint*)&tiles[j] == 0, "junk in header"); |
duke@435 | 2379 | split = true; // only the second word counts |
duke@435 | 2380 | // Example: int a[] = { 42 ... } |
duke@435 | 2381 | } else if (con0 == 0 && init0 == -1) { |
duke@435 | 2382 | split = true; // first word is covered by full inits |
duke@435 | 2383 | // Example: int a[] = { ... foo(), 42 ... } |
duke@435 | 2384 | } else if (con1 == 0 && init1 == -1) { |
duke@435 | 2385 | split = true; // second word is covered by full inits |
duke@435 | 2386 | // Example: int a[] = { ... 42, foo() ... } |
duke@435 | 2387 | } |
duke@435 | 2388 | |
duke@435 | 2389 | // Here's a case where init0 is neither 0 nor -1: |
duke@435 | 2390 | // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } |
duke@435 | 2391 | // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. |
duke@435 | 2392 | // In this case the tile is not split; it is (jlong)42. |
duke@435 | 2393 | // The big tile is stored down, and then the foo() value is inserted. |
duke@435 | 2394 | // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) |
duke@435 | 2395 | |
duke@435 | 2396 | Node* ctl = old->in(MemNode::Control); |
duke@435 | 2397 | Node* adr = make_raw_address(offset, phase); |
duke@435 | 2398 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
duke@435 | 2399 | |
duke@435 | 2400 | // One or two coalesced stores to plop down. |
duke@435 | 2401 | Node* st[2]; |
duke@435 | 2402 | intptr_t off[2]; |
duke@435 | 2403 | int nst = 0; |
duke@435 | 2404 | if (!split) { |
duke@435 | 2405 | ++new_long; |
duke@435 | 2406 | off[nst] = offset; |
duke@435 | 2407 | st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, |
duke@435 | 2408 | phase->longcon(con), T_LONG); |
duke@435 | 2409 | } else { |
duke@435 | 2410 | // Omit either if it is a zero. |
duke@435 | 2411 | if (con0 != 0) { |
duke@435 | 2412 | ++new_int; |
duke@435 | 2413 | off[nst] = offset; |
duke@435 | 2414 | st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, |
duke@435 | 2415 | phase->intcon(con0), T_INT); |
duke@435 | 2416 | } |
duke@435 | 2417 | if (con1 != 0) { |
duke@435 | 2418 | ++new_int; |
duke@435 | 2419 | offset += BytesPerInt; |
duke@435 | 2420 | adr = make_raw_address(offset, phase); |
duke@435 | 2421 | off[nst] = offset; |
duke@435 | 2422 | st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp, |
duke@435 | 2423 | phase->intcon(con1), T_INT); |
duke@435 | 2424 | } |
duke@435 | 2425 | } |
duke@435 | 2426 | |
duke@435 | 2427 | // Insert second store first, then the first before the second. |
duke@435 | 2428 | // Insert each one just before any overlapping non-constant stores. |
duke@435 | 2429 | while (nst > 0) { |
duke@435 | 2430 | Node* st1 = st[--nst]; |
duke@435 | 2431 | C->copy_node_notes_to(st1, old); |
duke@435 | 2432 | st1 = phase->transform(st1); |
duke@435 | 2433 | offset = off[nst]; |
duke@435 | 2434 | assert(offset >= header_size, "do not smash header"); |
duke@435 | 2435 | int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); |
duke@435 | 2436 | guarantee(ins_idx != 0, "must re-insert constant store"); |
duke@435 | 2437 | if (ins_idx < 0) ins_idx = -ins_idx; // never overlap |
duke@435 | 2438 | if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) |
duke@435 | 2439 | set_req(--ins_idx, st1); |
duke@435 | 2440 | else |
duke@435 | 2441 | ins_req(ins_idx, st1); |
duke@435 | 2442 | } |
duke@435 | 2443 | } |
duke@435 | 2444 | |
duke@435 | 2445 | if (PrintCompilation && WizardMode) |
duke@435 | 2446 | tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", |
duke@435 | 2447 | old_subword, old_long, new_int, new_long); |
duke@435 | 2448 | if (C->log() != NULL) |
duke@435 | 2449 | C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", |
duke@435 | 2450 | old_subword, old_long, new_int, new_long); |
duke@435 | 2451 | |
duke@435 | 2452 | // Clean up any remaining occurrences of zmem: |
duke@435 | 2453 | remove_extra_zeroes(); |
duke@435 | 2454 | } |
duke@435 | 2455 | |
duke@435 | 2456 | // Explore forward from in(start) to find the first fully initialized |
duke@435 | 2457 | // word, and return its offset. Skip groups of subword stores which |
duke@435 | 2458 | // together initialize full words. If in(start) is itself part of a |
duke@435 | 2459 | // fully initialized word, return the offset of in(start). If there |
duke@435 | 2460 | // are no following full-word stores, or if something is fishy, return |
duke@435 | 2461 | // a negative value. |
duke@435 | 2462 | intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { |
duke@435 | 2463 | int int_map = 0; |
duke@435 | 2464 | intptr_t int_map_off = 0; |
duke@435 | 2465 | const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for |
duke@435 | 2466 | |
duke@435 | 2467 | for (uint i = start, limit = req(); i < limit; i++) { |
duke@435 | 2468 | Node* st = in(i); |
duke@435 | 2469 | |
duke@435 | 2470 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 2471 | if (st_off < 0) break; // return conservative answer |
duke@435 | 2472 | |
duke@435 | 2473 | int st_size = st->as_Store()->memory_size(); |
duke@435 | 2474 | if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { |
duke@435 | 2475 | return st_off; // we found a complete word init |
duke@435 | 2476 | } |
duke@435 | 2477 | |
duke@435 | 2478 | // update the map: |
duke@435 | 2479 | |
duke@435 | 2480 | intptr_t this_int_off = align_size_down(st_off, BytesPerInt); |
duke@435 | 2481 | if (this_int_off != int_map_off) { |
duke@435 | 2482 | // reset the map: |
duke@435 | 2483 | int_map = 0; |
duke@435 | 2484 | int_map_off = this_int_off; |
duke@435 | 2485 | } |
duke@435 | 2486 | |
duke@435 | 2487 | int subword_off = st_off - this_int_off; |
duke@435 | 2488 | int_map |= right_n_bits(st_size) << subword_off; |
duke@435 | 2489 | if ((int_map & FULL_MAP) == FULL_MAP) { |
duke@435 | 2490 | return this_int_off; // we found a complete word init |
duke@435 | 2491 | } |
duke@435 | 2492 | |
duke@435 | 2493 | // Did this store hit or cross the word boundary? |
duke@435 | 2494 | intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); |
duke@435 | 2495 | if (next_int_off == this_int_off + BytesPerInt) { |
duke@435 | 2496 | // We passed the current int, without fully initializing it. |
duke@435 | 2497 | int_map_off = next_int_off; |
duke@435 | 2498 | int_map >>= BytesPerInt; |
duke@435 | 2499 | } else if (next_int_off > this_int_off + BytesPerInt) { |
duke@435 | 2500 | // We passed the current and next int. |
duke@435 | 2501 | return this_int_off + BytesPerInt; |
duke@435 | 2502 | } |
duke@435 | 2503 | } |
duke@435 | 2504 | |
duke@435 | 2505 | return -1; |
duke@435 | 2506 | } |
duke@435 | 2507 | |
duke@435 | 2508 | |
duke@435 | 2509 | // Called when the associated AllocateNode is expanded into CFG. |
duke@435 | 2510 | // At this point, we may perform additional optimizations. |
duke@435 | 2511 | // Linearize the stores by ascending offset, to make memory |
duke@435 | 2512 | // activity as coherent as possible. |
duke@435 | 2513 | Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, |
duke@435 | 2514 | intptr_t header_size, |
duke@435 | 2515 | Node* size_in_bytes, |
duke@435 | 2516 | PhaseGVN* phase) { |
duke@435 | 2517 | assert(!is_complete(), "not already complete"); |
duke@435 | 2518 | assert(stores_are_sane(phase), ""); |
duke@435 | 2519 | assert(allocation() != NULL, "must be present"); |
duke@435 | 2520 | |
duke@435 | 2521 | remove_extra_zeroes(); |
duke@435 | 2522 | |
duke@435 | 2523 | if (ReduceFieldZeroing || ReduceBulkZeroing) |
duke@435 | 2524 | // reduce instruction count for common initialization patterns |
duke@435 | 2525 | coalesce_subword_stores(header_size, size_in_bytes, phase); |
duke@435 | 2526 | |
duke@435 | 2527 | Node* zmem = zero_memory(); // initially zero memory state |
duke@435 | 2528 | Node* inits = zmem; // accumulating a linearized chain of inits |
duke@435 | 2529 | #ifdef ASSERT |
duke@435 | 2530 | intptr_t last_init_off = sizeof(oopDesc); // previous init offset |
duke@435 | 2531 | intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size |
duke@435 | 2532 | intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size |
duke@435 | 2533 | #endif |
duke@435 | 2534 | intptr_t zeroes_done = header_size; |
duke@435 | 2535 | |
duke@435 | 2536 | bool do_zeroing = true; // we might give up if inits are very sparse |
duke@435 | 2537 | int big_init_gaps = 0; // how many large gaps have we seen? |
duke@435 | 2538 | |
duke@435 | 2539 | if (ZeroTLAB) do_zeroing = false; |
duke@435 | 2540 | if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; |
duke@435 | 2541 | |
duke@435 | 2542 | for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { |
duke@435 | 2543 | Node* st = in(i); |
duke@435 | 2544 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 2545 | if (st_off < 0) |
duke@435 | 2546 | break; // unknown junk in the inits |
duke@435 | 2547 | if (st->in(MemNode::Memory) != zmem) |
duke@435 | 2548 | break; // complicated store chains somehow in list |
duke@435 | 2549 | |
duke@435 | 2550 | int st_size = st->as_Store()->memory_size(); |
duke@435 | 2551 | intptr_t next_init_off = st_off + st_size; |
duke@435 | 2552 | |
duke@435 | 2553 | if (do_zeroing && zeroes_done < next_init_off) { |
duke@435 | 2554 | // See if this store needs a zero before it or under it. |
duke@435 | 2555 | intptr_t zeroes_needed = st_off; |
duke@435 | 2556 | |
duke@435 | 2557 | if (st_size < BytesPerInt) { |
duke@435 | 2558 | // Look for subword stores which only partially initialize words. |
duke@435 | 2559 | // If we find some, we must lay down some word-level zeroes first, |
duke@435 | 2560 | // underneath the subword stores. |
duke@435 | 2561 | // |
duke@435 | 2562 | // Examples: |
duke@435 | 2563 | // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s |
duke@435 | 2564 | // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y |
duke@435 | 2565 | // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z |
duke@435 | 2566 | // |
duke@435 | 2567 | // Note: coalesce_subword_stores may have already done this, |
duke@435 | 2568 | // if it was prompted by constant non-zero subword initializers. |
duke@435 | 2569 | // But this case can still arise with non-constant stores. |
duke@435 | 2570 | |
duke@435 | 2571 | intptr_t next_full_store = find_next_fullword_store(i, phase); |
duke@435 | 2572 | |
duke@435 | 2573 | // In the examples above: |
duke@435 | 2574 | // in(i) p q r s x y z |
duke@435 | 2575 | // st_off 12 13 14 15 12 13 14 |
duke@435 | 2576 | // st_size 1 1 1 1 1 1 1 |
duke@435 | 2577 | // next_full_s. 12 16 16 16 16 16 16 |
duke@435 | 2578 | // z's_done 12 16 16 16 12 16 12 |
duke@435 | 2579 | // z's_needed 12 16 16 16 16 16 16 |
duke@435 | 2580 | // zsize 0 0 0 0 4 0 4 |
duke@435 | 2581 | if (next_full_store < 0) { |
duke@435 | 2582 | // Conservative tack: Zero to end of current word. |
duke@435 | 2583 | zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); |
duke@435 | 2584 | } else { |
duke@435 | 2585 | // Zero to beginning of next fully initialized word. |
duke@435 | 2586 | // Or, don't zero at all, if we are already in that word. |
duke@435 | 2587 | assert(next_full_store >= zeroes_needed, "must go forward"); |
duke@435 | 2588 | assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); |
duke@435 | 2589 | zeroes_needed = next_full_store; |
duke@435 | 2590 | } |
duke@435 | 2591 | } |
duke@435 | 2592 | |
duke@435 | 2593 | if (zeroes_needed > zeroes_done) { |
duke@435 | 2594 | intptr_t zsize = zeroes_needed - zeroes_done; |
duke@435 | 2595 | // Do some incremental zeroing on rawmem, in parallel with inits. |
duke@435 | 2596 | zeroes_done = align_size_down(zeroes_done, BytesPerInt); |
duke@435 | 2597 | rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, |
duke@435 | 2598 | zeroes_done, zeroes_needed, |
duke@435 | 2599 | phase); |
duke@435 | 2600 | zeroes_done = zeroes_needed; |
duke@435 | 2601 | if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) |
duke@435 | 2602 | do_zeroing = false; // leave the hole, next time |
duke@435 | 2603 | } |
duke@435 | 2604 | } |
duke@435 | 2605 | |
duke@435 | 2606 | // Collect the store and move on: |
duke@435 | 2607 | st->set_req(MemNode::Memory, inits); |
duke@435 | 2608 | inits = st; // put it on the linearized chain |
duke@435 | 2609 | set_req(i, zmem); // unhook from previous position |
duke@435 | 2610 | |
duke@435 | 2611 | if (zeroes_done == st_off) |
duke@435 | 2612 | zeroes_done = next_init_off; |
duke@435 | 2613 | |
duke@435 | 2614 | assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); |
duke@435 | 2615 | |
duke@435 | 2616 | #ifdef ASSERT |
duke@435 | 2617 | // Various order invariants. Weaker than stores_are_sane because |
duke@435 | 2618 | // a large constant tile can be filled in by smaller non-constant stores. |
duke@435 | 2619 | assert(st_off >= last_init_off, "inits do not reverse"); |
duke@435 | 2620 | last_init_off = st_off; |
duke@435 | 2621 | const Type* val = NULL; |
duke@435 | 2622 | if (st_size >= BytesPerInt && |
duke@435 | 2623 | (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && |
duke@435 | 2624 | (int)val->basic_type() < (int)T_OBJECT) { |
duke@435 | 2625 | assert(st_off >= last_tile_end, "tiles do not overlap"); |
duke@435 | 2626 | assert(st_off >= last_init_end, "tiles do not overwrite inits"); |
duke@435 | 2627 | last_tile_end = MAX2(last_tile_end, next_init_off); |
duke@435 | 2628 | } else { |
duke@435 | 2629 | intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); |
duke@435 | 2630 | assert(st_tile_end >= last_tile_end, "inits stay with tiles"); |
duke@435 | 2631 | assert(st_off >= last_init_end, "inits do not overlap"); |
duke@435 | 2632 | last_init_end = next_init_off; // it's a non-tile |
duke@435 | 2633 | } |
duke@435 | 2634 | #endif //ASSERT |
duke@435 | 2635 | } |
duke@435 | 2636 | |
duke@435 | 2637 | remove_extra_zeroes(); // clear out all the zmems left over |
duke@435 | 2638 | add_req(inits); |
duke@435 | 2639 | |
duke@435 | 2640 | if (!ZeroTLAB) { |
duke@435 | 2641 | // If anything remains to be zeroed, zero it all now. |
duke@435 | 2642 | zeroes_done = align_size_down(zeroes_done, BytesPerInt); |
duke@435 | 2643 | // if it is the last unused 4 bytes of an instance, forget about it |
duke@435 | 2644 | intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); |
duke@435 | 2645 | if (zeroes_done + BytesPerLong >= size_limit) { |
duke@435 | 2646 | assert(allocation() != NULL, ""); |
duke@435 | 2647 | Node* klass_node = allocation()->in(AllocateNode::KlassNode); |
duke@435 | 2648 | ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); |
duke@435 | 2649 | if (zeroes_done == k->layout_helper()) |
duke@435 | 2650 | zeroes_done = size_limit; |
duke@435 | 2651 | } |
duke@435 | 2652 | if (zeroes_done < size_limit) { |
duke@435 | 2653 | rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, |
duke@435 | 2654 | zeroes_done, size_in_bytes, phase); |
duke@435 | 2655 | } |
duke@435 | 2656 | } |
duke@435 | 2657 | |
duke@435 | 2658 | set_complete(phase); |
duke@435 | 2659 | return rawmem; |
duke@435 | 2660 | } |
duke@435 | 2661 | |
duke@435 | 2662 | |
duke@435 | 2663 | #ifdef ASSERT |
duke@435 | 2664 | bool InitializeNode::stores_are_sane(PhaseTransform* phase) { |
duke@435 | 2665 | if (is_complete()) |
duke@435 | 2666 | return true; // stores could be anything at this point |
duke@435 | 2667 | intptr_t last_off = sizeof(oopDesc); |
duke@435 | 2668 | for (uint i = InitializeNode::RawStores; i < req(); i++) { |
duke@435 | 2669 | Node* st = in(i); |
duke@435 | 2670 | intptr_t st_off = get_store_offset(st, phase); |
duke@435 | 2671 | if (st_off < 0) continue; // ignore dead garbage |
duke@435 | 2672 | if (last_off > st_off) { |
duke@435 | 2673 | tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); |
duke@435 | 2674 | this->dump(2); |
duke@435 | 2675 | assert(false, "ascending store offsets"); |
duke@435 | 2676 | return false; |
duke@435 | 2677 | } |
duke@435 | 2678 | last_off = st_off + st->as_Store()->memory_size(); |
duke@435 | 2679 | } |
duke@435 | 2680 | return true; |
duke@435 | 2681 | } |
duke@435 | 2682 | #endif //ASSERT |
duke@435 | 2683 | |
duke@435 | 2684 | |
duke@435 | 2685 | |
duke@435 | 2686 | |
duke@435 | 2687 | //============================MergeMemNode===================================== |
duke@435 | 2688 | // |
duke@435 | 2689 | // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several |
duke@435 | 2690 | // contributing store or call operations. Each contributor provides the memory |
duke@435 | 2691 | // state for a particular "alias type" (see Compile::alias_type). For example, |
duke@435 | 2692 | // if a MergeMem has an input X for alias category #6, then any memory reference |
duke@435 | 2693 | // to alias category #6 may use X as its memory state input, as an exact equivalent |
duke@435 | 2694 | // to using the MergeMem as a whole. |
duke@435 | 2695 | // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) |
duke@435 | 2696 | // |
duke@435 | 2697 | // (Here, the <N> notation gives the index of the relevant adr_type.) |
duke@435 | 2698 | // |
duke@435 | 2699 | // In one special case (and more cases in the future), alias categories overlap. |
duke@435 | 2700 | // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory |
duke@435 | 2701 | // states. Therefore, if a MergeMem has only one contributing input W for Bot, |
duke@435 | 2702 | // it is exactly equivalent to that state W: |
duke@435 | 2703 | // MergeMem(<Bot>: W) <==> W |
duke@435 | 2704 | // |
duke@435 | 2705 | // Usually, the merge has more than one input. In that case, where inputs |
duke@435 | 2706 | // overlap (i.e., one is Bot), the narrower alias type determines the memory |
duke@435 | 2707 | // state for that type, and the wider alias type (Bot) fills in everywhere else: |
duke@435 | 2708 | // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) |
duke@435 | 2709 | // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) |
duke@435 | 2710 | // |
duke@435 | 2711 | // A merge can take a "wide" memory state as one of its narrow inputs. |
duke@435 | 2712 | // This simply means that the merge observes out only the relevant parts of |
duke@435 | 2713 | // the wide input. That is, wide memory states arriving at narrow merge inputs |
duke@435 | 2714 | // are implicitly "filtered" or "sliced" as necessary. (This is rare.) |
duke@435 | 2715 | // |
duke@435 | 2716 | // These rules imply that MergeMem nodes may cascade (via their <Bot> links), |
duke@435 | 2717 | // and that memory slices "leak through": |
duke@435 | 2718 | // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) |
duke@435 | 2719 | // |
duke@435 | 2720 | // But, in such a cascade, repeated memory slices can "block the leak": |
duke@435 | 2721 | // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') |
duke@435 | 2722 | // |
duke@435 | 2723 | // In the last example, Y is not part of the combined memory state of the |
duke@435 | 2724 | // outermost MergeMem. The system must, of course, prevent unschedulable |
duke@435 | 2725 | // memory states from arising, so you can be sure that the state Y is somehow |
duke@435 | 2726 | // a precursor to state Y'. |
duke@435 | 2727 | // |
duke@435 | 2728 | // |
duke@435 | 2729 | // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array |
duke@435 | 2730 | // of each MergeMemNode array are exactly the numerical alias indexes, including |
duke@435 | 2731 | // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions |
duke@435 | 2732 | // Compile::alias_type (and kin) produce and manage these indexes. |
duke@435 | 2733 | // |
duke@435 | 2734 | // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. |
duke@435 | 2735 | // (Note that this provides quick access to the top node inside MergeMem methods, |
duke@435 | 2736 | // without the need to reach out via TLS to Compile::current.) |
duke@435 | 2737 | // |
duke@435 | 2738 | // As a consequence of what was just described, a MergeMem that represents a full |
duke@435 | 2739 | // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, |
duke@435 | 2740 | // containing all alias categories. |
duke@435 | 2741 | // |
duke@435 | 2742 | // MergeMem nodes never (?) have control inputs, so in(0) is NULL. |
duke@435 | 2743 | // |
duke@435 | 2744 | // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either |
duke@435 | 2745 | // a memory state for the alias type <N>, or else the top node, meaning that |
duke@435 | 2746 | // there is no particular input for that alias type. Note that the length of |
duke@435 | 2747 | // a MergeMem is variable, and may be extended at any time to accommodate new |
duke@435 | 2748 | // memory states at larger alias indexes. When merges grow, they are of course |
duke@435 | 2749 | // filled with "top" in the unused in() positions. |
duke@435 | 2750 | // |
duke@435 | 2751 | // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. |
duke@435 | 2752 | // (Top was chosen because it works smoothly with passes like GCM.) |
duke@435 | 2753 | // |
duke@435 | 2754 | // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is |
duke@435 | 2755 | // the type of random VM bits like TLS references.) Since it is always the |
duke@435 | 2756 | // first non-Bot memory slice, some low-level loops use it to initialize an |
duke@435 | 2757 | // index variable: for (i = AliasIdxRaw; i < req(); i++). |
duke@435 | 2758 | // |
duke@435 | 2759 | // |
duke@435 | 2760 | // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns |
duke@435 | 2761 | // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns |
duke@435 | 2762 | // the memory state for alias type <N>, or (if there is no particular slice at <N>, |
duke@435 | 2763 | // it returns the base memory. To prevent bugs, memory_at does not accept <Top> |
duke@435 | 2764 | // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over |
duke@435 | 2765 | // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. |
duke@435 | 2766 | // |
duke@435 | 2767 | // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't |
duke@435 | 2768 | // really that different from the other memory inputs. An abbreviation called |
duke@435 | 2769 | // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. |
duke@435 | 2770 | // |
duke@435 | 2771 | // |
duke@435 | 2772 | // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent |
duke@435 | 2773 | // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi |
duke@435 | 2774 | // that "emerges though" the base memory will be marked as excluding the alias types |
duke@435 | 2775 | // of the other (narrow-memory) copies which "emerged through" the narrow edges: |
duke@435 | 2776 | // |
duke@435 | 2777 | // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) |
duke@435 | 2778 | // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) |
duke@435 | 2779 | // |
duke@435 | 2780 | // This strange "subtraction" effect is necessary to ensure IGVN convergence. |
duke@435 | 2781 | // (It is currently unimplemented.) As you can see, the resulting merge is |
duke@435 | 2782 | // actually a disjoint union of memory states, rather than an overlay. |
duke@435 | 2783 | // |
duke@435 | 2784 | |
duke@435 | 2785 | //------------------------------MergeMemNode----------------------------------- |
duke@435 | 2786 | Node* MergeMemNode::make_empty_memory() { |
duke@435 | 2787 | Node* empty_memory = (Node*) Compile::current()->top(); |
duke@435 | 2788 | assert(empty_memory->is_top(), "correct sentinel identity"); |
duke@435 | 2789 | return empty_memory; |
duke@435 | 2790 | } |
duke@435 | 2791 | |
duke@435 | 2792 | MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { |
duke@435 | 2793 | init_class_id(Class_MergeMem); |
duke@435 | 2794 | // all inputs are nullified in Node::Node(int) |
duke@435 | 2795 | // set_input(0, NULL); // no control input |
duke@435 | 2796 | |
duke@435 | 2797 | // Initialize the edges uniformly to top, for starters. |
duke@435 | 2798 | Node* empty_mem = make_empty_memory(); |
duke@435 | 2799 | for (uint i = Compile::AliasIdxTop; i < req(); i++) { |
duke@435 | 2800 | init_req(i,empty_mem); |
duke@435 | 2801 | } |
duke@435 | 2802 | assert(empty_memory() == empty_mem, ""); |
duke@435 | 2803 | |
duke@435 | 2804 | if( new_base != NULL && new_base->is_MergeMem() ) { |
duke@435 | 2805 | MergeMemNode* mdef = new_base->as_MergeMem(); |
duke@435 | 2806 | assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); |
duke@435 | 2807 | for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { |
duke@435 | 2808 | mms.set_memory(mms.memory2()); |
duke@435 | 2809 | } |
duke@435 | 2810 | assert(base_memory() == mdef->base_memory(), ""); |
duke@435 | 2811 | } else { |
duke@435 | 2812 | set_base_memory(new_base); |
duke@435 | 2813 | } |
duke@435 | 2814 | } |
duke@435 | 2815 | |
duke@435 | 2816 | // Make a new, untransformed MergeMem with the same base as 'mem'. |
duke@435 | 2817 | // If mem is itself a MergeMem, populate the result with the same edges. |
duke@435 | 2818 | MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { |
duke@435 | 2819 | return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); |
duke@435 | 2820 | } |
duke@435 | 2821 | |
duke@435 | 2822 | //------------------------------cmp-------------------------------------------- |
duke@435 | 2823 | uint MergeMemNode::hash() const { return NO_HASH; } |
duke@435 | 2824 | uint MergeMemNode::cmp( const Node &n ) const { |
duke@435 | 2825 | return (&n == this); // Always fail except on self |
duke@435 | 2826 | } |
duke@435 | 2827 | |
duke@435 | 2828 | //------------------------------Identity--------------------------------------- |
duke@435 | 2829 | Node* MergeMemNode::Identity(PhaseTransform *phase) { |
duke@435 | 2830 | // Identity if this merge point does not record any interesting memory |
duke@435 | 2831 | // disambiguations. |
duke@435 | 2832 | Node* base_mem = base_memory(); |
duke@435 | 2833 | Node* empty_mem = empty_memory(); |
duke@435 | 2834 | if (base_mem != empty_mem) { // Memory path is not dead? |
duke@435 | 2835 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 2836 | Node* mem = in(i); |
duke@435 | 2837 | if (mem != empty_mem && mem != base_mem) { |
duke@435 | 2838 | return this; // Many memory splits; no change |
duke@435 | 2839 | } |
duke@435 | 2840 | } |
duke@435 | 2841 | } |
duke@435 | 2842 | return base_mem; // No memory splits; ID on the one true input |
duke@435 | 2843 | } |
duke@435 | 2844 | |
duke@435 | 2845 | //------------------------------Ideal------------------------------------------ |
duke@435 | 2846 | // This method is invoked recursively on chains of MergeMem nodes |
duke@435 | 2847 | Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
duke@435 | 2848 | // Remove chain'd MergeMems |
duke@435 | 2849 | // |
duke@435 | 2850 | // This is delicate, because the each "in(i)" (i >= Raw) is interpreted |
duke@435 | 2851 | // relative to the "in(Bot)". Since we are patching both at the same time, |
duke@435 | 2852 | // we have to be careful to read each "in(i)" relative to the old "in(Bot)", |
duke@435 | 2853 | // but rewrite each "in(i)" relative to the new "in(Bot)". |
duke@435 | 2854 | Node *progress = NULL; |
duke@435 | 2855 | |
duke@435 | 2856 | |
duke@435 | 2857 | Node* old_base = base_memory(); |
duke@435 | 2858 | Node* empty_mem = empty_memory(); |
duke@435 | 2859 | if (old_base == empty_mem) |
duke@435 | 2860 | return NULL; // Dead memory path. |
duke@435 | 2861 | |
duke@435 | 2862 | MergeMemNode* old_mbase; |
duke@435 | 2863 | if (old_base != NULL && old_base->is_MergeMem()) |
duke@435 | 2864 | old_mbase = old_base->as_MergeMem(); |
duke@435 | 2865 | else |
duke@435 | 2866 | old_mbase = NULL; |
duke@435 | 2867 | Node* new_base = old_base; |
duke@435 | 2868 | |
duke@435 | 2869 | // simplify stacked MergeMems in base memory |
duke@435 | 2870 | if (old_mbase) new_base = old_mbase->base_memory(); |
duke@435 | 2871 | |
duke@435 | 2872 | // the base memory might contribute new slices beyond my req() |
duke@435 | 2873 | if (old_mbase) grow_to_match(old_mbase); |
duke@435 | 2874 | |
duke@435 | 2875 | // Look carefully at the base node if it is a phi. |
duke@435 | 2876 | PhiNode* phi_base; |
duke@435 | 2877 | if (new_base != NULL && new_base->is_Phi()) |
duke@435 | 2878 | phi_base = new_base->as_Phi(); |
duke@435 | 2879 | else |
duke@435 | 2880 | phi_base = NULL; |
duke@435 | 2881 | |
duke@435 | 2882 | Node* phi_reg = NULL; |
duke@435 | 2883 | uint phi_len = (uint)-1; |
duke@435 | 2884 | if (phi_base != NULL && !phi_base->is_copy()) { |
duke@435 | 2885 | // do not examine phi if degraded to a copy |
duke@435 | 2886 | phi_reg = phi_base->region(); |
duke@435 | 2887 | phi_len = phi_base->req(); |
duke@435 | 2888 | // see if the phi is unfinished |
duke@435 | 2889 | for (uint i = 1; i < phi_len; i++) { |
duke@435 | 2890 | if (phi_base->in(i) == NULL) { |
duke@435 | 2891 | // incomplete phi; do not look at it yet! |
duke@435 | 2892 | phi_reg = NULL; |
duke@435 | 2893 | phi_len = (uint)-1; |
duke@435 | 2894 | break; |
duke@435 | 2895 | } |
duke@435 | 2896 | } |
duke@435 | 2897 | } |
duke@435 | 2898 | |
duke@435 | 2899 | // Note: We do not call verify_sparse on entry, because inputs |
duke@435 | 2900 | // can normalize to the base_memory via subsume_node or similar |
duke@435 | 2901 | // mechanisms. This method repairs that damage. |
duke@435 | 2902 | |
duke@435 | 2903 | assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); |
duke@435 | 2904 | |
duke@435 | 2905 | // Look at each slice. |
duke@435 | 2906 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 2907 | Node* old_in = in(i); |
duke@435 | 2908 | // calculate the old memory value |
duke@435 | 2909 | Node* old_mem = old_in; |
duke@435 | 2910 | if (old_mem == empty_mem) old_mem = old_base; |
duke@435 | 2911 | assert(old_mem == memory_at(i), ""); |
duke@435 | 2912 | |
duke@435 | 2913 | // maybe update (reslice) the old memory value |
duke@435 | 2914 | |
duke@435 | 2915 | // simplify stacked MergeMems |
duke@435 | 2916 | Node* new_mem = old_mem; |
duke@435 | 2917 | MergeMemNode* old_mmem; |
duke@435 | 2918 | if (old_mem != NULL && old_mem->is_MergeMem()) |
duke@435 | 2919 | old_mmem = old_mem->as_MergeMem(); |
duke@435 | 2920 | else |
duke@435 | 2921 | old_mmem = NULL; |
duke@435 | 2922 | if (old_mmem == this) { |
duke@435 | 2923 | // This can happen if loops break up and safepoints disappear. |
duke@435 | 2924 | // A merge of BotPtr (default) with a RawPtr memory derived from a |
duke@435 | 2925 | // safepoint can be rewritten to a merge of the same BotPtr with |
duke@435 | 2926 | // the BotPtr phi coming into the loop. If that phi disappears |
duke@435 | 2927 | // also, we can end up with a self-loop of the mergemem. |
duke@435 | 2928 | // In general, if loops degenerate and memory effects disappear, |
duke@435 | 2929 | // a mergemem can be left looking at itself. This simply means |
duke@435 | 2930 | // that the mergemem's default should be used, since there is |
duke@435 | 2931 | // no longer any apparent effect on this slice. |
duke@435 | 2932 | // Note: If a memory slice is a MergeMem cycle, it is unreachable |
duke@435 | 2933 | // from start. Update the input to TOP. |
duke@435 | 2934 | new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; |
duke@435 | 2935 | } |
duke@435 | 2936 | else if (old_mmem != NULL) { |
duke@435 | 2937 | new_mem = old_mmem->memory_at(i); |
duke@435 | 2938 | } |
duke@435 | 2939 | // else preceeding memory was not a MergeMem |
duke@435 | 2940 | |
duke@435 | 2941 | // replace equivalent phis (unfortunately, they do not GVN together) |
duke@435 | 2942 | if (new_mem != NULL && new_mem != new_base && |
duke@435 | 2943 | new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { |
duke@435 | 2944 | if (new_mem->is_Phi()) { |
duke@435 | 2945 | PhiNode* phi_mem = new_mem->as_Phi(); |
duke@435 | 2946 | for (uint i = 1; i < phi_len; i++) { |
duke@435 | 2947 | if (phi_base->in(i) != phi_mem->in(i)) { |
duke@435 | 2948 | phi_mem = NULL; |
duke@435 | 2949 | break; |
duke@435 | 2950 | } |
duke@435 | 2951 | } |
duke@435 | 2952 | if (phi_mem != NULL) { |
duke@435 | 2953 | // equivalent phi nodes; revert to the def |
duke@435 | 2954 | new_mem = new_base; |
duke@435 | 2955 | } |
duke@435 | 2956 | } |
duke@435 | 2957 | } |
duke@435 | 2958 | |
duke@435 | 2959 | // maybe store down a new value |
duke@435 | 2960 | Node* new_in = new_mem; |
duke@435 | 2961 | if (new_in == new_base) new_in = empty_mem; |
duke@435 | 2962 | |
duke@435 | 2963 | if (new_in != old_in) { |
duke@435 | 2964 | // Warning: Do not combine this "if" with the previous "if" |
duke@435 | 2965 | // A memory slice might have be be rewritten even if it is semantically |
duke@435 | 2966 | // unchanged, if the base_memory value has changed. |
duke@435 | 2967 | set_req(i, new_in); |
duke@435 | 2968 | progress = this; // Report progress |
duke@435 | 2969 | } |
duke@435 | 2970 | } |
duke@435 | 2971 | |
duke@435 | 2972 | if (new_base != old_base) { |
duke@435 | 2973 | set_req(Compile::AliasIdxBot, new_base); |
duke@435 | 2974 | // Don't use set_base_memory(new_base), because we need to update du. |
duke@435 | 2975 | assert(base_memory() == new_base, ""); |
duke@435 | 2976 | progress = this; |
duke@435 | 2977 | } |
duke@435 | 2978 | |
duke@435 | 2979 | if( base_memory() == this ) { |
duke@435 | 2980 | // a self cycle indicates this memory path is dead |
duke@435 | 2981 | set_req(Compile::AliasIdxBot, empty_mem); |
duke@435 | 2982 | } |
duke@435 | 2983 | |
duke@435 | 2984 | // Resolve external cycles by calling Ideal on a MergeMem base_memory |
duke@435 | 2985 | // Recursion must occur after the self cycle check above |
duke@435 | 2986 | if( base_memory()->is_MergeMem() ) { |
duke@435 | 2987 | MergeMemNode *new_mbase = base_memory()->as_MergeMem(); |
duke@435 | 2988 | Node *m = phase->transform(new_mbase); // Rollup any cycles |
duke@435 | 2989 | if( m != NULL && (m->is_top() || |
duke@435 | 2990 | m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { |
duke@435 | 2991 | // propagate rollup of dead cycle to self |
duke@435 | 2992 | set_req(Compile::AliasIdxBot, empty_mem); |
duke@435 | 2993 | } |
duke@435 | 2994 | } |
duke@435 | 2995 | |
duke@435 | 2996 | if( base_memory() == empty_mem ) { |
duke@435 | 2997 | progress = this; |
duke@435 | 2998 | // Cut inputs during Parse phase only. |
duke@435 | 2999 | // During Optimize phase a dead MergeMem node will be subsumed by Top. |
duke@435 | 3000 | if( !can_reshape ) { |
duke@435 | 3001 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 3002 | if( in(i) != empty_mem ) { set_req(i, empty_mem); } |
duke@435 | 3003 | } |
duke@435 | 3004 | } |
duke@435 | 3005 | } |
duke@435 | 3006 | |
duke@435 | 3007 | if( !progress && base_memory()->is_Phi() && can_reshape ) { |
duke@435 | 3008 | // Check if PhiNode::Ideal's "Split phis through memory merges" |
duke@435 | 3009 | // transform should be attempted. Look for this->phi->this cycle. |
duke@435 | 3010 | uint merge_width = req(); |
duke@435 | 3011 | if (merge_width > Compile::AliasIdxRaw) { |
duke@435 | 3012 | PhiNode* phi = base_memory()->as_Phi(); |
duke@435 | 3013 | for( uint i = 1; i < phi->req(); ++i ) {// For all paths in |
duke@435 | 3014 | if (phi->in(i) == this) { |
duke@435 | 3015 | phase->is_IterGVN()->_worklist.push(phi); |
duke@435 | 3016 | break; |
duke@435 | 3017 | } |
duke@435 | 3018 | } |
duke@435 | 3019 | } |
duke@435 | 3020 | } |
duke@435 | 3021 | |
duke@435 | 3022 | assert(verify_sparse(), "please, no dups of base"); |
duke@435 | 3023 | return progress; |
duke@435 | 3024 | } |
duke@435 | 3025 | |
duke@435 | 3026 | //-------------------------set_base_memory------------------------------------- |
duke@435 | 3027 | void MergeMemNode::set_base_memory(Node *new_base) { |
duke@435 | 3028 | Node* empty_mem = empty_memory(); |
duke@435 | 3029 | set_req(Compile::AliasIdxBot, new_base); |
duke@435 | 3030 | assert(memory_at(req()) == new_base, "must set default memory"); |
duke@435 | 3031 | // Clear out other occurrences of new_base: |
duke@435 | 3032 | if (new_base != empty_mem) { |
duke@435 | 3033 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 3034 | if (in(i) == new_base) set_req(i, empty_mem); |
duke@435 | 3035 | } |
duke@435 | 3036 | } |
duke@435 | 3037 | } |
duke@435 | 3038 | |
duke@435 | 3039 | //------------------------------out_RegMask------------------------------------ |
duke@435 | 3040 | const RegMask &MergeMemNode::out_RegMask() const { |
duke@435 | 3041 | return RegMask::Empty; |
duke@435 | 3042 | } |
duke@435 | 3043 | |
duke@435 | 3044 | //------------------------------dump_spec-------------------------------------- |
duke@435 | 3045 | #ifndef PRODUCT |
duke@435 | 3046 | void MergeMemNode::dump_spec(outputStream *st) const { |
duke@435 | 3047 | st->print(" {"); |
duke@435 | 3048 | Node* base_mem = base_memory(); |
duke@435 | 3049 | for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { |
duke@435 | 3050 | Node* mem = memory_at(i); |
duke@435 | 3051 | if (mem == base_mem) { st->print(" -"); continue; } |
duke@435 | 3052 | st->print( " N%d:", mem->_idx ); |
duke@435 | 3053 | Compile::current()->get_adr_type(i)->dump_on(st); |
duke@435 | 3054 | } |
duke@435 | 3055 | st->print(" }"); |
duke@435 | 3056 | } |
duke@435 | 3057 | #endif // !PRODUCT |
duke@435 | 3058 | |
duke@435 | 3059 | |
duke@435 | 3060 | #ifdef ASSERT |
duke@435 | 3061 | static bool might_be_same(Node* a, Node* b) { |
duke@435 | 3062 | if (a == b) return true; |
duke@435 | 3063 | if (!(a->is_Phi() || b->is_Phi())) return false; |
duke@435 | 3064 | // phis shift around during optimization |
duke@435 | 3065 | return true; // pretty stupid... |
duke@435 | 3066 | } |
duke@435 | 3067 | |
duke@435 | 3068 | // verify a narrow slice (either incoming or outgoing) |
duke@435 | 3069 | static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { |
duke@435 | 3070 | if (!VerifyAliases) return; // don't bother to verify unless requested |
duke@435 | 3071 | if (is_error_reported()) return; // muzzle asserts when debugging an error |
duke@435 | 3072 | if (Node::in_dump()) return; // muzzle asserts when printing |
duke@435 | 3073 | assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); |
duke@435 | 3074 | assert(n != NULL, ""); |
duke@435 | 3075 | // Elide intervening MergeMem's |
duke@435 | 3076 | while (n->is_MergeMem()) { |
duke@435 | 3077 | n = n->as_MergeMem()->memory_at(alias_idx); |
duke@435 | 3078 | } |
duke@435 | 3079 | Compile* C = Compile::current(); |
duke@435 | 3080 | const TypePtr* n_adr_type = n->adr_type(); |
duke@435 | 3081 | if (n == m->empty_memory()) { |
duke@435 | 3082 | // Implicit copy of base_memory() |
duke@435 | 3083 | } else if (n_adr_type != TypePtr::BOTTOM) { |
duke@435 | 3084 | assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); |
duke@435 | 3085 | assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); |
duke@435 | 3086 | } else { |
duke@435 | 3087 | // A few places like make_runtime_call "know" that VM calls are narrow, |
duke@435 | 3088 | // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. |
duke@435 | 3089 | bool expected_wide_mem = false; |
duke@435 | 3090 | if (n == m->base_memory()) { |
duke@435 | 3091 | expected_wide_mem = true; |
duke@435 | 3092 | } else if (alias_idx == Compile::AliasIdxRaw || |
duke@435 | 3093 | n == m->memory_at(Compile::AliasIdxRaw)) { |
duke@435 | 3094 | expected_wide_mem = true; |
duke@435 | 3095 | } else if (!C->alias_type(alias_idx)->is_rewritable()) { |
duke@435 | 3096 | // memory can "leak through" calls on channels that |
duke@435 | 3097 | // are write-once. Allow this also. |
duke@435 | 3098 | expected_wide_mem = true; |
duke@435 | 3099 | } |
duke@435 | 3100 | assert(expected_wide_mem, "expected narrow slice replacement"); |
duke@435 | 3101 | } |
duke@435 | 3102 | } |
duke@435 | 3103 | #else // !ASSERT |
duke@435 | 3104 | #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op |
duke@435 | 3105 | #endif |
duke@435 | 3106 | |
duke@435 | 3107 | |
duke@435 | 3108 | //-----------------------------memory_at--------------------------------------- |
duke@435 | 3109 | Node* MergeMemNode::memory_at(uint alias_idx) const { |
duke@435 | 3110 | assert(alias_idx >= Compile::AliasIdxRaw || |
duke@435 | 3111 | alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, |
duke@435 | 3112 | "must avoid base_memory and AliasIdxTop"); |
duke@435 | 3113 | |
duke@435 | 3114 | // Otherwise, it is a narrow slice. |
duke@435 | 3115 | Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); |
duke@435 | 3116 | Compile *C = Compile::current(); |
duke@435 | 3117 | if (is_empty_memory(n)) { |
duke@435 | 3118 | // the array is sparse; empty slots are the "top" node |
duke@435 | 3119 | n = base_memory(); |
duke@435 | 3120 | assert(Node::in_dump() |
duke@435 | 3121 | || n == NULL || n->bottom_type() == Type::TOP |
duke@435 | 3122 | || n->adr_type() == TypePtr::BOTTOM |
duke@435 | 3123 | || n->adr_type() == TypeRawPtr::BOTTOM |
duke@435 | 3124 | || Compile::current()->AliasLevel() == 0, |
duke@435 | 3125 | "must be a wide memory"); |
duke@435 | 3126 | // AliasLevel == 0 if we are organizing the memory states manually. |
duke@435 | 3127 | // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. |
duke@435 | 3128 | } else { |
duke@435 | 3129 | // make sure the stored slice is sane |
duke@435 | 3130 | #ifdef ASSERT |
duke@435 | 3131 | if (is_error_reported() || Node::in_dump()) { |
duke@435 | 3132 | } else if (might_be_same(n, base_memory())) { |
duke@435 | 3133 | // Give it a pass: It is a mostly harmless repetition of the base. |
duke@435 | 3134 | // This can arise normally from node subsumption during optimization. |
duke@435 | 3135 | } else { |
duke@435 | 3136 | verify_memory_slice(this, alias_idx, n); |
duke@435 | 3137 | } |
duke@435 | 3138 | #endif |
duke@435 | 3139 | } |
duke@435 | 3140 | return n; |
duke@435 | 3141 | } |
duke@435 | 3142 | |
duke@435 | 3143 | //---------------------------set_memory_at------------------------------------- |
duke@435 | 3144 | void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { |
duke@435 | 3145 | verify_memory_slice(this, alias_idx, n); |
duke@435 | 3146 | Node* empty_mem = empty_memory(); |
duke@435 | 3147 | if (n == base_memory()) n = empty_mem; // collapse default |
duke@435 | 3148 | uint need_req = alias_idx+1; |
duke@435 | 3149 | if (req() < need_req) { |
duke@435 | 3150 | if (n == empty_mem) return; // already the default, so do not grow me |
duke@435 | 3151 | // grow the sparse array |
duke@435 | 3152 | do { |
duke@435 | 3153 | add_req(empty_mem); |
duke@435 | 3154 | } while (req() < need_req); |
duke@435 | 3155 | } |
duke@435 | 3156 | set_req( alias_idx, n ); |
duke@435 | 3157 | } |
duke@435 | 3158 | |
duke@435 | 3159 | |
duke@435 | 3160 | |
duke@435 | 3161 | //--------------------------iteration_setup------------------------------------ |
duke@435 | 3162 | void MergeMemNode::iteration_setup(const MergeMemNode* other) { |
duke@435 | 3163 | if (other != NULL) { |
duke@435 | 3164 | grow_to_match(other); |
duke@435 | 3165 | // invariant: the finite support of mm2 is within mm->req() |
duke@435 | 3166 | #ifdef ASSERT |
duke@435 | 3167 | for (uint i = req(); i < other->req(); i++) { |
duke@435 | 3168 | assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); |
duke@435 | 3169 | } |
duke@435 | 3170 | #endif |
duke@435 | 3171 | } |
duke@435 | 3172 | // Replace spurious copies of base_memory by top. |
duke@435 | 3173 | Node* base_mem = base_memory(); |
duke@435 | 3174 | if (base_mem != NULL && !base_mem->is_top()) { |
duke@435 | 3175 | for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { |
duke@435 | 3176 | if (in(i) == base_mem) |
duke@435 | 3177 | set_req(i, empty_memory()); |
duke@435 | 3178 | } |
duke@435 | 3179 | } |
duke@435 | 3180 | } |
duke@435 | 3181 | |
duke@435 | 3182 | //---------------------------grow_to_match------------------------------------- |
duke@435 | 3183 | void MergeMemNode::grow_to_match(const MergeMemNode* other) { |
duke@435 | 3184 | Node* empty_mem = empty_memory(); |
duke@435 | 3185 | assert(other->is_empty_memory(empty_mem), "consistent sentinels"); |
duke@435 | 3186 | // look for the finite support of the other memory |
duke@435 | 3187 | for (uint i = other->req(); --i >= req(); ) { |
duke@435 | 3188 | if (other->in(i) != empty_mem) { |
duke@435 | 3189 | uint new_len = i+1; |
duke@435 | 3190 | while (req() < new_len) add_req(empty_mem); |
duke@435 | 3191 | break; |
duke@435 | 3192 | } |
duke@435 | 3193 | } |
duke@435 | 3194 | } |
duke@435 | 3195 | |
duke@435 | 3196 | //---------------------------verify_sparse------------------------------------- |
duke@435 | 3197 | #ifndef PRODUCT |
duke@435 | 3198 | bool MergeMemNode::verify_sparse() const { |
duke@435 | 3199 | assert(is_empty_memory(make_empty_memory()), "sane sentinel"); |
duke@435 | 3200 | Node* base_mem = base_memory(); |
duke@435 | 3201 | // The following can happen in degenerate cases, since empty==top. |
duke@435 | 3202 | if (is_empty_memory(base_mem)) return true; |
duke@435 | 3203 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
duke@435 | 3204 | assert(in(i) != NULL, "sane slice"); |
duke@435 | 3205 | if (in(i) == base_mem) return false; // should have been the sentinel value! |
duke@435 | 3206 | } |
duke@435 | 3207 | return true; |
duke@435 | 3208 | } |
duke@435 | 3209 | |
duke@435 | 3210 | bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { |
duke@435 | 3211 | Node* n; |
duke@435 | 3212 | n = mm->in(idx); |
duke@435 | 3213 | if (mem == n) return true; // might be empty_memory() |
duke@435 | 3214 | n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); |
duke@435 | 3215 | if (mem == n) return true; |
duke@435 | 3216 | while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { |
duke@435 | 3217 | if (mem == n) return true; |
duke@435 | 3218 | if (n == NULL) break; |
duke@435 | 3219 | } |
duke@435 | 3220 | return false; |
duke@435 | 3221 | } |
duke@435 | 3222 | #endif // !PRODUCT |