J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1 | /* |
Vladimir Kozlov | d1191bb | 2012-06-15 01:25:19 -0700 | [diff] [blame] | 2 | * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 | * |
| 5 | * This code is free software; you can redistribute it and/or modify it |
| 6 | * under the terms of the GNU General Public License version 2 only, as |
| 7 | * published by the Free Software Foundation. |
| 8 | * |
| 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
| 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 12 | * version 2 for more details (a copy is included in the LICENSE file that |
| 13 | * accompanied this code). |
| 14 | * |
| 15 | * You should have received a copy of the GNU General Public License version |
| 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
| 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
| 18 | * |
Erik Trimble | ba7c173 | 2010-05-27 19:08:38 -0700 | [diff] [blame] | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
| 20 | * or visit www.oracle.com if you need additional information or have any |
| 21 | * questions. |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 22 | * |
| 23 | */ |
| 24 | |
Stefan Karlsson | 8006fe8 | 2010-11-23 13:22:55 -0800 | [diff] [blame] | 25 | #include "precompiled.hpp" |
| 26 | #include "classfile/systemDictionary.hpp" |
| 27 | #include "compiler/compileLog.hpp" |
| 28 | #include "memory/allocation.inline.hpp" |
| 29 | #include "oops/objArrayKlass.hpp" |
| 30 | #include "opto/addnode.hpp" |
| 31 | #include "opto/cfgnode.hpp" |
| 32 | #include "opto/compile.hpp" |
| 33 | #include "opto/connode.hpp" |
| 34 | #include "opto/loopnode.hpp" |
| 35 | #include "opto/machnode.hpp" |
| 36 | #include "opto/matcher.hpp" |
| 37 | #include "opto/memnode.hpp" |
| 38 | #include "opto/mulnode.hpp" |
| 39 | #include "opto/phaseX.hpp" |
| 40 | #include "opto/regmask.hpp" |
| 41 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 42 | // Portions of code courtesy of Clifford Click |
| 43 | |
| 44 | // Optimization - Graph Style |
| 45 | |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 46 | static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); |
| 47 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 48 | //============================================================================= |
| 49 | uint MemNode::size_of() const { return sizeof(*this); } |
| 50 | |
| 51 | const TypePtr *MemNode::adr_type() const { |
| 52 | Node* adr = in(Address); |
| 53 | const TypePtr* cross_check = NULL; |
| 54 | DEBUG_ONLY(cross_check = _adr_type); |
| 55 | return calculate_adr_type(adr->bottom_type(), cross_check); |
| 56 | } |
| 57 | |
| 58 | #ifndef PRODUCT |
| 59 | void MemNode::dump_spec(outputStream *st) const { |
| 60 | if (in(Address) == NULL) return; // node is dead |
| 61 | #ifndef ASSERT |
| 62 | // fake the missing field |
| 63 | const TypePtr* _adr_type = NULL; |
| 64 | if (in(Address) != NULL) |
| 65 | _adr_type = in(Address)->bottom_type()->isa_ptr(); |
| 66 | #endif |
| 67 | dump_adr_type(this, _adr_type, st); |
| 68 | |
| 69 | Compile* C = Compile::current(); |
| 70 | if( C->alias_type(_adr_type)->is_volatile() ) |
| 71 | st->print(" Volatile!"); |
| 72 | } |
| 73 | |
| 74 | void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { |
| 75 | st->print(" @"); |
| 76 | if (adr_type == NULL) { |
| 77 | st->print("NULL"); |
| 78 | } else { |
| 79 | adr_type->dump_on(st); |
| 80 | Compile* C = Compile::current(); |
| 81 | Compile::AliasType* atp = NULL; |
| 82 | if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); |
| 83 | if (atp == NULL) |
| 84 | st->print(", idx=?\?;"); |
| 85 | else if (atp->index() == Compile::AliasIdxBot) |
| 86 | st->print(", idx=Bot;"); |
| 87 | else if (atp->index() == Compile::AliasIdxTop) |
| 88 | st->print(", idx=Top;"); |
| 89 | else if (atp->index() == Compile::AliasIdxRaw) |
| 90 | st->print(", idx=Raw;"); |
| 91 | else { |
| 92 | ciField* field = atp->field(); |
| 93 | if (field) { |
| 94 | st->print(", name="); |
| 95 | field->print_name_on(st); |
| 96 | } |
| 97 | st->print(", idx=%d;", atp->index()); |
| 98 | } |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | extern void print_alias_types(); |
| 103 | |
| 104 | #endif |
| 105 | |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 106 | Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { |
| 107 | assert((t_oop != NULL), "sanity"); |
| 108 | bool is_instance = t_oop->is_known_instance_field(); |
| 109 | bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && |
| 110 | (load != NULL) && load->is_Load() && |
| 111 | (phase->is_IterGVN() != NULL); |
| 112 | if (!(is_instance || is_boxed_value_load)) |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 113 | return mchain; // don't try to optimize non-instance types |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 114 | uint instance_id = t_oop->instance_id(); |
Vladimir Kozlov | 50c4a23 | 2008-07-28 17:12:52 -0700 | [diff] [blame] | 115 | Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 116 | Node *prev = NULL; |
| 117 | Node *result = mchain; |
| 118 | while (prev != result) { |
| 119 | prev = result; |
Vladimir Kozlov | 50c4a23 | 2008-07-28 17:12:52 -0700 | [diff] [blame] | 120 | if (result == start_mem) |
Christian Thalinger | 05d1de7 | 2009-02-27 13:27:09 -0800 | [diff] [blame] | 121 | break; // hit one of our sentinels |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 122 | // skip over a call which does not affect this memory slice |
| 123 | if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { |
| 124 | Node *proj_in = result->in(0); |
Vladimir Kozlov | 50c4a23 | 2008-07-28 17:12:52 -0700 | [diff] [blame] | 125 | if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { |
Christian Thalinger | 05d1de7 | 2009-02-27 13:27:09 -0800 | [diff] [blame] | 126 | break; // hit one of our sentinels |
Vladimir Kozlov | 50c4a23 | 2008-07-28 17:12:52 -0700 | [diff] [blame] | 127 | } else if (proj_in->is_Call()) { |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 128 | CallNode *call = proj_in->as_Call(); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 129 | if (!call->may_modify(t_oop, phase)) { // returns false for instances |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 130 | result = call->in(TypeFunc::Memory); |
| 131 | } |
| 132 | } else if (proj_in->is_Initialize()) { |
| 133 | AllocateNode* alloc = proj_in->as_Initialize()->allocation(); |
| 134 | // Stop if this is the initialization for the object instance which |
| 135 | // which contains this memory slice, otherwise skip over it. |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 136 | if ((alloc == NULL) || (alloc->_idx == instance_id)) { |
| 137 | break; |
| 138 | } |
| 139 | if (is_instance) { |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 140 | result = proj_in->in(TypeFunc::Memory); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 141 | } else if (is_boxed_value_load) { |
| 142 | Node* klass = alloc->in(AllocateNode::KlassNode); |
| 143 | const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); |
| 144 | if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { |
| 145 | result = proj_in->in(TypeFunc::Memory); // not related allocation |
| 146 | } |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 147 | } |
| 148 | } else if (proj_in->is_MemBar()) { |
| 149 | result = proj_in->in(TypeFunc::Memory); |
Vladimir Kozlov | 50c4a23 | 2008-07-28 17:12:52 -0700 | [diff] [blame] | 150 | } else { |
| 151 | assert(false, "unexpected projection"); |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 152 | } |
Vladimir Kozlov | 9f5ca02 | 2009-12-09 16:40:45 -0800 | [diff] [blame] | 153 | } else if (result->is_ClearArray()) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 154 | if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { |
Vladimir Kozlov | 9f5ca02 | 2009-12-09 16:40:45 -0800 | [diff] [blame] | 155 | // Can not bypass initialization of the instance |
| 156 | // we are looking for. |
| 157 | break; |
| 158 | } |
| 159 | // Otherwise skip it (the call updated 'result' value). |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 160 | } else if (result->is_MergeMem()) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 161 | result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 162 | } |
| 163 | } |
| 164 | return result; |
| 165 | } |
| 166 | |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 167 | Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { |
| 168 | const TypeOopPtr* t_oop = t_adr->isa_oopptr(); |
| 169 | if (t_oop == NULL) |
| 170 | return mchain; // don't try to optimize non-oop types |
| 171 | Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); |
| 172 | bool is_instance = t_oop->is_known_instance_field(); |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 173 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 174 | if (is_instance && igvn != NULL && result->is_Phi()) { |
| 175 | PhiNode *mphi = result->as_Phi(); |
| 176 | assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); |
| 177 | const TypePtr *t = mphi->adr_type(); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 178 | if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || |
Vladimir Kozlov | 4213e62 | 2008-06-26 13:34:00 -0700 | [diff] [blame] | 179 | t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && |
Vladimir Kozlov | fae3906 | 2008-07-16 16:04:39 -0700 | [diff] [blame] | 180 | t->is_oopptr()->cast_to_exactness(true) |
| 181 | ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) |
| 182 | ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 183 | // clone the Phi with our address type |
| 184 | result = mphi->split_out_instance(t_adr, igvn); |
| 185 | } else { |
| 186 | assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); |
| 187 | } |
| 188 | } |
| 189 | return result; |
| 190 | } |
| 191 | |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 192 | static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { |
| 193 | uint alias_idx = phase->C->get_alias_index(tp); |
| 194 | Node *mem = mmem; |
| 195 | #ifdef ASSERT |
| 196 | { |
| 197 | // Check that current type is consistent with the alias index used during graph construction |
| 198 | assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); |
| 199 | bool consistent = adr_check == NULL || adr_check->empty() || |
| 200 | phase->C->must_alias(adr_check, alias_idx ); |
| 201 | // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] |
| 202 | if( !consistent && adr_check != NULL && !adr_check->empty() && |
Chuck Rasbold | 2e672ac | 2008-05-29 16:22:09 -0700 | [diff] [blame] | 203 | tp->isa_aryptr() && tp->offset() == Type::OffsetBot && |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 204 | adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && |
| 205 | ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || |
| 206 | adr_check->offset() == oopDesc::klass_offset_in_bytes() || |
| 207 | adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { |
| 208 | // don't assert if it is dead code. |
| 209 | consistent = true; |
| 210 | } |
| 211 | if( !consistent ) { |
| 212 | st->print("alias_idx==%d, adr_check==", alias_idx); |
| 213 | if( adr_check == NULL ) { |
| 214 | st->print("NULL"); |
| 215 | } else { |
| 216 | adr_check->dump(); |
| 217 | } |
| 218 | st->cr(); |
| 219 | print_alias_types(); |
| 220 | assert(consistent, "adr_check must match alias idx"); |
| 221 | } |
| 222 | } |
| 223 | #endif |
Tom Rodriguez | f4b4eae | 2010-09-22 13:01:12 -0700 | [diff] [blame] | 224 | // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 225 | // means an array I have not precisely typed yet. Do not do any |
| 226 | // alias stuff with it any time soon. |
Tom Rodriguez | f4b4eae | 2010-09-22 13:01:12 -0700 | [diff] [blame] | 227 | const TypeOopPtr *toop = tp->isa_oopptr(); |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 228 | if( tp->base() != Type::AnyPtr && |
Tom Rodriguez | f4b4eae | 2010-09-22 13:01:12 -0700 | [diff] [blame] | 229 | !(toop && |
| 230 | toop->klass() != NULL && |
| 231 | toop->klass()->is_java_lang_Object() && |
| 232 | toop->offset() == Type::OffsetBot) ) { |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 233 | // compress paths and change unreachable cycles to TOP |
| 234 | // If not, we can update the input infinitely along a MergeMem cycle |
| 235 | // Equivalent code in PhiNode::Ideal |
| 236 | Node* m = phase->transform(mmem); |
Christian Thalinger | 05d1de7 | 2009-02-27 13:27:09 -0800 | [diff] [blame] | 237 | // If transformed to a MergeMem, get the desired slice |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 238 | // Otherwise the returned node represents memory for every slice |
| 239 | mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; |
| 240 | // Update input if it is progress over what we have now |
| 241 | } |
| 242 | return mem; |
| 243 | } |
| 244 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 245 | //--------------------------Ideal_common--------------------------------------- |
| 246 | // Look for degenerate control and memory inputs. Bypass MergeMem inputs. |
| 247 | // Unhook non-raw memories from complete (macro-expanded) initializations. |
| 248 | Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { |
| 249 | // If our control input is a dead region, kill all below the region |
| 250 | Node *ctl = in(MemNode::Control); |
| 251 | if (ctl && remove_dead_region(phase, can_reshape)) |
| 252 | return this; |
Vladimir Kozlov | 3730631 | 2008-08-27 09:15:46 -0700 | [diff] [blame] | 253 | ctl = in(MemNode::Control); |
| 254 | // Don't bother trying to transform a dead node |
Vladimir Kozlov | 010f640 | 2013-03-06 12:25:57 -0800 | [diff] [blame] | 255 | if (ctl && ctl->is_top()) return NodeSentinel; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 256 | |
Vladimir Kozlov | ca3a350 | 2009-04-07 19:04:24 -0700 | [diff] [blame] | 257 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
| 258 | // Wait if control on the worklist. |
| 259 | if (ctl && can_reshape && igvn != NULL) { |
| 260 | Node* bol = NULL; |
| 261 | Node* cmp = NULL; |
| 262 | if (ctl->in(0)->is_If()) { |
| 263 | assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); |
| 264 | bol = ctl->in(0)->in(1); |
| 265 | if (bol->is_Bool()) |
| 266 | cmp = ctl->in(0)->in(1)->in(1); |
| 267 | } |
| 268 | if (igvn->_worklist.member(ctl) || |
| 269 | (bol != NULL && igvn->_worklist.member(bol)) || |
| 270 | (cmp != NULL && igvn->_worklist.member(cmp)) ) { |
| 271 | // This control path may be dead. |
| 272 | // Delay this memory node transformation until the control is processed. |
| 273 | phase->is_IterGVN()->_worklist.push(this); |
| 274 | return NodeSentinel; // caller will return NULL |
| 275 | } |
| 276 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 277 | // Ignore if memory is dead, or self-loop |
| 278 | Node *mem = in(MemNode::Memory); |
Vladimir Kozlov | 010f640 | 2013-03-06 12:25:57 -0800 | [diff] [blame] | 279 | if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL |
| 280 | assert(mem != this, "dead loop in MemNode::Ideal"); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 281 | |
Vladimir Kozlov | 78034a3 | 2011-12-02 21:37:19 -0800 | [diff] [blame] | 282 | if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { |
| 283 | // This memory slice may be dead. |
| 284 | // Delay this mem node transformation until the memory is processed. |
| 285 | phase->is_IterGVN()->_worklist.push(this); |
| 286 | return NodeSentinel; // caller will return NULL |
| 287 | } |
| 288 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 289 | Node *address = in(MemNode::Address); |
Vladimir Kozlov | 010f640 | 2013-03-06 12:25:57 -0800 | [diff] [blame] | 290 | const Type *t_adr = phase->type(address); |
| 291 | if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 292 | |
Vladimir Kozlov | 010f640 | 2013-03-06 12:25:57 -0800 | [diff] [blame] | 293 | if (can_reshape && igvn != NULL && |
Vladimir Kozlov | 7cc5573 | 2010-09-30 18:31:45 -0700 | [diff] [blame] | 294 | (igvn->_worklist.member(address) || |
Vladimir Kozlov | 010f640 | 2013-03-06 12:25:57 -0800 | [diff] [blame] | 295 | igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { |
Vladimir Kozlov | 7aae40a | 2008-11-07 09:29:38 -0800 | [diff] [blame] | 296 | // The address's base and type may change when the address is processed. |
| 297 | // Delay this mem node transformation until the address is processed. |
| 298 | phase->is_IterGVN()->_worklist.push(this); |
| 299 | return NodeSentinel; // caller will return NULL |
| 300 | } |
| 301 | |
Vladimir Kozlov | c5744bd | 2009-11-04 14:43:50 -0800 | [diff] [blame] | 302 | // Do NOT remove or optimize the next lines: ensure a new alias index |
| 303 | // is allocated for an oop pointer type before Escape Analysis. |
| 304 | // Note: C++ will not remove it since the call has side effect. |
Vladimir Kozlov | 010f640 | 2013-03-06 12:25:57 -0800 | [diff] [blame] | 305 | if (t_adr->isa_oopptr()) { |
Vladimir Kozlov | c5744bd | 2009-11-04 14:43:50 -0800 | [diff] [blame] | 306 | int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); |
| 307 | } |
| 308 | |
Vladimir Kozlov | ca3a350 | 2009-04-07 19:04:24 -0700 | [diff] [blame] | 309 | #ifdef ASSERT |
| 310 | Node* base = NULL; |
| 311 | if (address->is_AddP()) |
| 312 | base = address->in(AddPNode::Base); |
Vladimir Kozlov | 010f640 | 2013-03-06 12:25:57 -0800 | [diff] [blame] | 313 | if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && |
| 314 | !t_adr->isa_rawptr()) { |
| 315 | // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. |
| 316 | Compile* C = phase->C; |
| 317 | tty->cr(); |
| 318 | tty->print_cr("===== NULL+offs not RAW address ====="); |
| 319 | if (C->is_dead_node(this->_idx)) tty->print_cr("'this' is dead"); |
| 320 | if ((ctl != NULL) && C->is_dead_node(ctl->_idx)) tty->print_cr("'ctl' is dead"); |
| 321 | if (C->is_dead_node(mem->_idx)) tty->print_cr("'mem' is dead"); |
| 322 | if (C->is_dead_node(address->_idx)) tty->print_cr("'address' is dead"); |
| 323 | if (C->is_dead_node(base->_idx)) tty->print_cr("'base' is dead"); |
| 324 | tty->cr(); |
| 325 | base->dump(1); |
| 326 | tty->cr(); |
| 327 | this->dump(2); |
| 328 | tty->print("this->adr_type(): "); adr_type()->dump(); tty->cr(); |
| 329 | tty->print("phase->type(address): "); t_adr->dump(); tty->cr(); |
| 330 | tty->print("phase->type(base): "); phase->type(address)->dump(); tty->cr(); |
| 331 | tty->cr(); |
| 332 | } |
Vladimir Kozlov | ca3a350 | 2009-04-07 19:04:24 -0700 | [diff] [blame] | 333 | assert(base == NULL || t_adr->isa_rawptr() || |
| 334 | !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); |
| 335 | #endif |
| 336 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 337 | // Avoid independent memory operations |
| 338 | Node* old_mem = mem; |
| 339 | |
Vladimir Kozlov | dc6ad19 | 2008-02-29 19:57:41 -0800 | [diff] [blame] | 340 | // The code which unhooks non-raw memories from complete (macro-expanded) |
| 341 | // initializations was removed. After macro-expansion all stores catched |
| 342 | // by Initialize node became raw stores and there is no information |
| 343 | // which memory slices they modify. So it is unsafe to move any memory |
| 344 | // operation above these stores. Also in most cases hooked non-raw memories |
| 345 | // were already unhooked by using information from detect_ptr_independence() |
| 346 | // and find_previous_store(). |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 347 | |
| 348 | if (mem->is_MergeMem()) { |
| 349 | MergeMemNode* mmem = mem->as_MergeMem(); |
| 350 | const TypePtr *tp = t_adr->is_ptr(); |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 351 | |
| 352 | mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 353 | } |
| 354 | |
| 355 | if (mem != old_mem) { |
| 356 | set_req(MemNode::Memory, mem); |
Roland Westrelin | fe92862 | 2013-02-25 14:13:04 +0100 | [diff] [blame] | 357 | if (can_reshape && old_mem->outcnt() == 0) { |
| 358 | igvn->_worklist.push(old_mem); |
| 359 | } |
Vladimir Kozlov | 3730631 | 2008-08-27 09:15:46 -0700 | [diff] [blame] | 360 | if (phase->type( mem ) == Type::TOP) return NodeSentinel; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 361 | return this; |
| 362 | } |
| 363 | |
| 364 | // let the subclass continue analyzing... |
| 365 | return NULL; |
| 366 | } |
| 367 | |
| 368 | // Helper function for proving some simple control dominations. |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 369 | // Attempt to prove that all control inputs of 'dom' dominate 'sub'. |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 370 | // Already assumes that 'dom' is available at 'sub', and that 'sub' |
| 371 | // is not a constant (dominated by the method's StartNode). |
| 372 | // Used by MemNode::find_previous_store to prove that the |
| 373 | // control input of a memory operation predates (dominates) |
| 374 | // an allocation it wants to look past. |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 375 | bool MemNode::all_controls_dominate(Node* dom, Node* sub) { |
| 376 | if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) |
| 377 | return false; // Conservative answer for dead code |
| 378 | |
Vladimir Kozlov | 6978df8 | 2008-06-13 15:08:56 -0700 | [diff] [blame] | 379 | // Check 'dom'. Skip Proj and CatchProj nodes. |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 380 | dom = dom->find_exact_control(dom); |
| 381 | if (dom == NULL || dom->is_top()) |
| 382 | return false; // Conservative answer for dead code |
| 383 | |
Vladimir Kozlov | 6978df8 | 2008-06-13 15:08:56 -0700 | [diff] [blame] | 384 | if (dom == sub) { |
| 385 | // For the case when, for example, 'sub' is Initialize and the original |
| 386 | // 'dom' is Proj node of the 'sub'. |
| 387 | return false; |
| 388 | } |
| 389 | |
Vladimir Kozlov | 60c74ba | 2008-05-15 22:43:11 -0700 | [diff] [blame] | 390 | if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 391 | return true; |
| 392 | |
| 393 | // 'dom' dominates 'sub' if its control edge and control edges |
| 394 | // of all its inputs dominate or equal to sub's control edge. |
| 395 | |
| 396 | // Currently 'sub' is either Allocate, Initialize or Start nodes. |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 397 | // Or Region for the check in LoadNode::Ideal(); |
| 398 | // 'sub' should have sub->in(0) != NULL. |
| 399 | assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 400 | sub->is_Region() || sub->is_Call(), "expecting only these nodes"); |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 401 | |
| 402 | // Get control edge of 'sub'. |
Vladimir Kozlov | 6978df8 | 2008-06-13 15:08:56 -0700 | [diff] [blame] | 403 | Node* orig_sub = sub; |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 404 | sub = sub->find_exact_control(sub->in(0)); |
| 405 | if (sub == NULL || sub->is_top()) |
| 406 | return false; // Conservative answer for dead code |
| 407 | |
| 408 | assert(sub->is_CFG(), "expecting control"); |
| 409 | |
| 410 | if (sub == dom) |
| 411 | return true; |
| 412 | |
| 413 | if (sub->is_Start() || sub->is_Root()) |
| 414 | return false; |
| 415 | |
| 416 | { |
| 417 | // Check all control edges of 'dom'. |
| 418 | |
| 419 | ResourceMark rm; |
| 420 | Arena* arena = Thread::current()->resource_area(); |
| 421 | Node_List nlist(arena); |
| 422 | Unique_Node_List dom_list(arena); |
| 423 | |
| 424 | dom_list.push(dom); |
| 425 | bool only_dominating_controls = false; |
| 426 | |
| 427 | for (uint next = 0; next < dom_list.size(); next++) { |
| 428 | Node* n = dom_list.at(next); |
Vladimir Kozlov | 6978df8 | 2008-06-13 15:08:56 -0700 | [diff] [blame] | 429 | if (n == orig_sub) |
| 430 | return false; // One of dom's inputs dominated by sub. |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 431 | if (!n->is_CFG() && n->pinned()) { |
| 432 | // Check only own control edge for pinned non-control nodes. |
| 433 | n = n->find_exact_control(n->in(0)); |
| 434 | if (n == NULL || n->is_top()) |
| 435 | return false; // Conservative answer for dead code |
| 436 | assert(n->is_CFG(), "expecting control"); |
Vladimir Kozlov | 6978df8 | 2008-06-13 15:08:56 -0700 | [diff] [blame] | 437 | dom_list.push(n); |
| 438 | } else if (n->is_Con() || n->is_Start() || n->is_Root()) { |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 439 | only_dominating_controls = true; |
| 440 | } else if (n->is_CFG()) { |
| 441 | if (n->dominates(sub, nlist)) |
| 442 | only_dominating_controls = true; |
| 443 | else |
| 444 | return false; |
| 445 | } else { |
| 446 | // First, own control edge. |
| 447 | Node* m = n->find_exact_control(n->in(0)); |
Vladimir Kozlov | 60c74ba | 2008-05-15 22:43:11 -0700 | [diff] [blame] | 448 | if (m != NULL) { |
| 449 | if (m->is_top()) |
| 450 | return false; // Conservative answer for dead code |
| 451 | dom_list.push(m); |
| 452 | } |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 453 | // Now, the rest of edges. |
| 454 | uint cnt = n->req(); |
| 455 | for (uint i = 1; i < cnt; i++) { |
| 456 | m = n->find_exact_control(n->in(i)); |
| 457 | if (m == NULL || m->is_top()) |
| 458 | continue; |
| 459 | dom_list.push(m); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 460 | } |
| 461 | } |
| 462 | } |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 463 | return only_dominating_controls; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 464 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 465 | } |
| 466 | |
| 467 | //---------------------detect_ptr_independence--------------------------------- |
| 468 | // Used by MemNode::find_previous_store to prove that two base |
| 469 | // pointers are never equal. |
| 470 | // The pointers are accompanied by their associated allocations, |
| 471 | // if any, which have been previously discovered by the caller. |
| 472 | bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, |
| 473 | Node* p2, AllocateNode* a2, |
| 474 | PhaseTransform* phase) { |
| 475 | // Attempt to prove that these two pointers cannot be aliased. |
| 476 | // They may both manifestly be allocations, and they should differ. |
| 477 | // Or, if they are not both allocations, they can be distinct constants. |
| 478 | // Otherwise, one is an allocation and the other a pre-existing value. |
| 479 | if (a1 == NULL && a2 == NULL) { // neither an allocation |
| 480 | return (p1 != p2) && p1->is_Con() && p2->is_Con(); |
| 481 | } else if (a1 != NULL && a2 != NULL) { // both allocations |
| 482 | return (a1 != a2); |
| 483 | } else if (a1 != NULL) { // one allocation a1 |
| 484 | // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 485 | return all_controls_dominate(p2, a1); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 486 | } else { //(a2 != NULL) // one allocation a2 |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 487 | return all_controls_dominate(p1, a2); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 488 | } |
| 489 | return false; |
| 490 | } |
| 491 | |
| 492 | |
| 493 | // The logic for reordering loads and stores uses four steps: |
| 494 | // (a) Walk carefully past stores and initializations which we |
| 495 | // can prove are independent of this load. |
| 496 | // (b) Observe that the next memory state makes an exact match |
| 497 | // with self (load or store), and locate the relevant store. |
| 498 | // (c) Ensure that, if we were to wire self directly to the store, |
| 499 | // the optimizer would fold it up somehow. |
| 500 | // (d) Do the rewiring, and return, depending on some other part of |
| 501 | // the optimizer to fold up the load. |
| 502 | // This routine handles steps (a) and (b). Steps (c) and (d) are |
| 503 | // specific to loads and stores, so they are handled by the callers. |
| 504 | // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) |
| 505 | // |
| 506 | Node* MemNode::find_previous_store(PhaseTransform* phase) { |
| 507 | Node* ctrl = in(MemNode::Control); |
| 508 | Node* adr = in(MemNode::Address); |
| 509 | intptr_t offset = 0; |
| 510 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
| 511 | AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); |
| 512 | |
| 513 | if (offset == Type::OffsetBot) |
| 514 | return NULL; // cannot unalias unless there are precise offsets |
| 515 | |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 516 | const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); |
| 517 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 518 | intptr_t size_in_bytes = memory_size(); |
| 519 | |
| 520 | Node* mem = in(MemNode::Memory); // start searching here... |
| 521 | |
| 522 | int cnt = 50; // Cycle limiter |
| 523 | for (;;) { // While we can dance past unrelated stores... |
| 524 | if (--cnt < 0) break; // Caught in cycle or a complicated dance? |
| 525 | |
| 526 | if (mem->is_Store()) { |
| 527 | Node* st_adr = mem->in(MemNode::Address); |
| 528 | intptr_t st_offset = 0; |
| 529 | Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); |
| 530 | if (st_base == NULL) |
| 531 | break; // inscrutable pointer |
| 532 | if (st_offset != offset && st_offset != Type::OffsetBot) { |
| 533 | const int MAX_STORE = BytesPerLong; |
| 534 | if (st_offset >= offset + size_in_bytes || |
| 535 | st_offset <= offset - MAX_STORE || |
| 536 | st_offset <= offset - mem->as_Store()->memory_size()) { |
| 537 | // Success: The offsets are provably independent. |
| 538 | // (You may ask, why not just test st_offset != offset and be done? |
| 539 | // The answer is that stores of different sizes can co-exist |
| 540 | // in the same sequence of RawMem effects. We sometimes initialize |
| 541 | // a whole 'tile' of array elements with a single jint or jlong.) |
| 542 | mem = mem->in(MemNode::Memory); |
| 543 | continue; // (a) advance through independent store memory |
| 544 | } |
| 545 | } |
| 546 | if (st_base != base && |
| 547 | detect_ptr_independence(base, alloc, |
| 548 | st_base, |
| 549 | AllocateNode::Ideal_allocation(st_base, phase), |
| 550 | phase)) { |
| 551 | // Success: The bases are provably independent. |
| 552 | mem = mem->in(MemNode::Memory); |
| 553 | continue; // (a) advance through independent store memory |
| 554 | } |
| 555 | |
| 556 | // (b) At this point, if the bases or offsets do not agree, we lose, |
| 557 | // since we have not managed to prove 'this' and 'mem' independent. |
| 558 | if (st_base == base && st_offset == offset) { |
| 559 | return mem; // let caller handle steps (c), (d) |
| 560 | } |
| 561 | |
| 562 | } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { |
| 563 | InitializeNode* st_init = mem->in(0)->as_Initialize(); |
| 564 | AllocateNode* st_alloc = st_init->allocation(); |
| 565 | if (st_alloc == NULL) |
| 566 | break; // something degenerated |
| 567 | bool known_identical = false; |
| 568 | bool known_independent = false; |
| 569 | if (alloc == st_alloc) |
| 570 | known_identical = true; |
| 571 | else if (alloc != NULL) |
| 572 | known_independent = true; |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 573 | else if (all_controls_dominate(this, st_alloc)) |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 574 | known_independent = true; |
| 575 | |
| 576 | if (known_independent) { |
| 577 | // The bases are provably independent: Either they are |
| 578 | // manifestly distinct allocations, or else the control |
| 579 | // of this load dominates the store's allocation. |
| 580 | int alias_idx = phase->C->get_alias_index(adr_type()); |
| 581 | if (alias_idx == Compile::AliasIdxRaw) { |
| 582 | mem = st_alloc->in(TypeFunc::Memory); |
| 583 | } else { |
| 584 | mem = st_init->memory(alias_idx); |
| 585 | } |
| 586 | continue; // (a) advance through independent store memory |
| 587 | } |
| 588 | |
| 589 | // (b) at this point, if we are not looking at a store initializing |
| 590 | // the same allocation we are loading from, we lose. |
| 591 | if (known_identical) { |
| 592 | // From caller, can_see_stored_value will consult find_captured_store. |
| 593 | return mem; // let caller handle steps (c), (d) |
| 594 | } |
| 595 | |
Vladimir Kozlov | 4213e62 | 2008-06-26 13:34:00 -0700 | [diff] [blame] | 596 | } else if (addr_t != NULL && addr_t->is_known_instance_field()) { |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 597 | // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. |
| 598 | if (mem->is_Proj() && mem->in(0)->is_Call()) { |
| 599 | CallNode *call = mem->in(0)->as_Call(); |
| 600 | if (!call->may_modify(addr_t, phase)) { |
| 601 | mem = call->in(TypeFunc::Memory); |
| 602 | continue; // (a) advance through independent call memory |
| 603 | } |
| 604 | } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { |
| 605 | mem = mem->in(0)->in(TypeFunc::Memory); |
| 606 | continue; // (a) advance through independent MemBar memory |
Vladimir Kozlov | 9f5ca02 | 2009-12-09 16:40:45 -0800 | [diff] [blame] | 607 | } else if (mem->is_ClearArray()) { |
| 608 | if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { |
| 609 | // (the call updated 'mem' value) |
| 610 | continue; // (a) advance through independent allocation memory |
| 611 | } else { |
| 612 | // Can not bypass initialization of the instance |
| 613 | // we are looking for. |
| 614 | return mem; |
| 615 | } |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 616 | } else if (mem->is_MergeMem()) { |
| 617 | int alias_idx = phase->C->get_alias_index(adr_type()); |
| 618 | mem = mem->as_MergeMem()->memory_at(alias_idx); |
| 619 | continue; // (a) advance through independent MergeMem memory |
| 620 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 621 | } |
| 622 | |
| 623 | // Unless there is an explicit 'continue', we must bail out here, |
| 624 | // because 'mem' is an inscrutable memory state (e.g., a call). |
| 625 | break; |
| 626 | } |
| 627 | |
| 628 | return NULL; // bail out |
| 629 | } |
| 630 | |
| 631 | //----------------------calculate_adr_type------------------------------------- |
| 632 | // Helper function. Notices when the given type of address hits top or bottom. |
| 633 | // Also, asserts a cross-check of the type against the expected address type. |
| 634 | const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { |
| 635 | if (t == Type::TOP) return NULL; // does not touch memory any more? |
| 636 | #ifdef PRODUCT |
| 637 | cross_check = NULL; |
| 638 | #else |
| 639 | if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; |
| 640 | #endif |
| 641 | const TypePtr* tp = t->isa_ptr(); |
| 642 | if (tp == NULL) { |
| 643 | assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); |
| 644 | return TypePtr::BOTTOM; // touches lots of memory |
| 645 | } else { |
| 646 | #ifdef ASSERT |
| 647 | // %%%% [phh] We don't check the alias index if cross_check is |
| 648 | // TypeRawPtr::BOTTOM. Needs to be investigated. |
| 649 | if (cross_check != NULL && |
| 650 | cross_check != TypePtr::BOTTOM && |
| 651 | cross_check != TypeRawPtr::BOTTOM) { |
| 652 | // Recheck the alias index, to see if it has changed (due to a bug). |
| 653 | Compile* C = Compile::current(); |
| 654 | assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), |
| 655 | "must stay in the original alias category"); |
| 656 | // The type of the address must be contained in the adr_type, |
| 657 | // disregarding "null"-ness. |
| 658 | // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) |
| 659 | const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); |
| 660 | assert(cross_check->meet(tp_notnull) == cross_check, |
| 661 | "real address must not escape from expected memory type"); |
| 662 | } |
| 663 | #endif |
| 664 | return tp; |
| 665 | } |
| 666 | } |
| 667 | |
| 668 | //------------------------adr_phi_is_loop_invariant---------------------------- |
| 669 | // A helper function for Ideal_DU_postCCP to check if a Phi in a counted |
| 670 | // loop is loop invariant. Make a quick traversal of Phi and associated |
| 671 | // CastPP nodes, looking to see if they are a closed group within the loop. |
| 672 | bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { |
| 673 | // The idea is that the phi-nest must boil down to only CastPP nodes |
| 674 | // with the same data. This implies that any path into the loop already |
| 675 | // includes such a CastPP, and so the original cast, whatever its input, |
| 676 | // must be covered by an equivalent cast, with an earlier control input. |
| 677 | ResourceMark rm; |
| 678 | |
| 679 | // The loop entry input of the phi should be the unique dominating |
| 680 | // node for every Phi/CastPP in the loop. |
| 681 | Unique_Node_List closure; |
| 682 | closure.push(adr_phi->in(LoopNode::EntryControl)); |
| 683 | |
| 684 | // Add the phi node and the cast to the worklist. |
| 685 | Unique_Node_List worklist; |
| 686 | worklist.push(adr_phi); |
| 687 | if( cast != NULL ){ |
| 688 | if( !cast->is_ConstraintCast() ) return false; |
| 689 | worklist.push(cast); |
| 690 | } |
| 691 | |
| 692 | // Begin recursive walk of phi nodes. |
| 693 | while( worklist.size() ){ |
| 694 | // Take a node off the worklist |
| 695 | Node *n = worklist.pop(); |
| 696 | if( !closure.member(n) ){ |
| 697 | // Add it to the closure. |
| 698 | closure.push(n); |
| 699 | // Make a sanity check to ensure we don't waste too much time here. |
| 700 | if( closure.size() > 20) return false; |
| 701 | // This node is OK if: |
| 702 | // - it is a cast of an identical value |
| 703 | // - or it is a phi node (then we add its inputs to the worklist) |
| 704 | // Otherwise, the node is not OK, and we presume the cast is not invariant |
| 705 | if( n->is_ConstraintCast() ){ |
| 706 | worklist.push(n->in(1)); |
| 707 | } else if( n->is_Phi() ) { |
| 708 | for( uint i = 1; i < n->req(); i++ ) { |
| 709 | worklist.push(n->in(i)); |
| 710 | } |
| 711 | } else { |
| 712 | return false; |
| 713 | } |
| 714 | } |
| 715 | } |
| 716 | |
| 717 | // Quit when the worklist is empty, and we've found no offending nodes. |
| 718 | return true; |
| 719 | } |
| 720 | |
| 721 | //------------------------------Ideal_DU_postCCP------------------------------- |
| 722 | // Find any cast-away of null-ness and keep its control. Null cast-aways are |
| 723 | // going away in this pass and we need to make this memory op depend on the |
| 724 | // gating null check. |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 725 | Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { |
| 726 | return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); |
| 727 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 728 | |
| 729 | // I tried to leave the CastPP's in. This makes the graph more accurate in |
| 730 | // some sense; we get to keep around the knowledge that an oop is not-null |
| 731 | // after some test. Alas, the CastPP's interfere with GVN (some values are |
| 732 | // the regular oop, some are the CastPP of the oop, all merge at Phi's which |
| 733 | // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed |
| 734 | // some of the more trivial cases in the optimizer. Removing more useless |
| 735 | // Phi's started allowing Loads to illegally float above null checks. I gave |
| 736 | // up on this approach. CNC 10/20/2000 |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 737 | // This static method may be called not from MemNode (EncodePNode calls it). |
| 738 | // Only the control edge of the node 'n' might be updated. |
| 739 | Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 740 | Node *skipped_cast = NULL; |
| 741 | // Need a null check? Regular static accesses do not because they are |
| 742 | // from constant addresses. Array ops are gated by the range check (which |
| 743 | // always includes a NULL check). Just check field ops. |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 744 | if( n->in(MemNode::Control) == NULL ) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 745 | // Scan upwards for the highest location we can place this memory op. |
| 746 | while( true ) { |
| 747 | switch( adr->Opcode() ) { |
| 748 | |
| 749 | case Op_AddP: // No change to NULL-ness, so peek thru AddP's |
| 750 | adr = adr->in(AddPNode::Base); |
| 751 | continue; |
| 752 | |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 753 | case Op_DecodeN: // No change to NULL-ness, so peek thru |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 754 | case Op_DecodeNKlass: |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 755 | adr = adr->in(1); |
| 756 | continue; |
| 757 | |
Vladimir Kozlov | ae6d0d3 | 2012-06-11 14:58:51 -0700 | [diff] [blame] | 758 | case Op_EncodeP: |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 759 | case Op_EncodePKlass: |
Vladimir Kozlov | ae6d0d3 | 2012-06-11 14:58:51 -0700 | [diff] [blame] | 760 | // EncodeP node's control edge could be set by this method |
| 761 | // when EncodeP node depends on CastPP node. |
| 762 | // |
| 763 | // Use its control edge for memory op because EncodeP may go away |
| 764 | // later when it is folded with following or preceding DecodeN node. |
| 765 | if (adr->in(0) == NULL) { |
| 766 | // Keep looking for cast nodes. |
| 767 | adr = adr->in(1); |
| 768 | continue; |
| 769 | } |
| 770 | ccp->hash_delete(n); |
| 771 | n->set_req(MemNode::Control, adr->in(0)); |
| 772 | ccp->hash_insert(n); |
| 773 | return n; |
| 774 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 775 | case Op_CastPP: |
| 776 | // If the CastPP is useless, just peek on through it. |
| 777 | if( ccp->type(adr) == ccp->type(adr->in(1)) ) { |
| 778 | // Remember the cast that we've peeked though. If we peek |
| 779 | // through more than one, then we end up remembering the highest |
| 780 | // one, that is, if in a loop, the one closest to the top. |
| 781 | skipped_cast = adr; |
| 782 | adr = adr->in(1); |
| 783 | continue; |
| 784 | } |
| 785 | // CastPP is going away in this pass! We need this memory op to be |
| 786 | // control-dependent on the test that is guarding the CastPP. |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 787 | ccp->hash_delete(n); |
| 788 | n->set_req(MemNode::Control, adr->in(0)); |
| 789 | ccp->hash_insert(n); |
| 790 | return n; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 791 | |
| 792 | case Op_Phi: |
| 793 | // Attempt to float above a Phi to some dominating point. |
| 794 | if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { |
| 795 | // If we've already peeked through a Cast (which could have set the |
| 796 | // control), we can't float above a Phi, because the skipped Cast |
| 797 | // may not be loop invariant. |
| 798 | if (adr_phi_is_loop_invariant(adr, skipped_cast)) { |
| 799 | adr = adr->in(1); |
| 800 | continue; |
| 801 | } |
| 802 | } |
| 803 | |
| 804 | // Intentional fallthrough! |
| 805 | |
| 806 | // No obvious dominating point. The mem op is pinned below the Phi |
| 807 | // by the Phi itself. If the Phi goes away (no true value is merged) |
| 808 | // then the mem op can float, but not indefinitely. It must be pinned |
| 809 | // behind the controls leading to the Phi. |
| 810 | case Op_CheckCastPP: |
| 811 | // These usually stick around to change address type, however a |
| 812 | // useless one can be elided and we still need to pick up a control edge |
| 813 | if (adr->in(0) == NULL) { |
| 814 | // This CheckCastPP node has NO control and is likely useless. But we |
| 815 | // need check further up the ancestor chain for a control input to keep |
| 816 | // the node in place. 4959717. |
| 817 | skipped_cast = adr; |
| 818 | adr = adr->in(1); |
| 819 | continue; |
| 820 | } |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 821 | ccp->hash_delete(n); |
| 822 | n->set_req(MemNode::Control, adr->in(0)); |
| 823 | ccp->hash_insert(n); |
| 824 | return n; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 825 | |
| 826 | // List of "safe" opcodes; those that implicitly block the memory |
| 827 | // op below any null check. |
| 828 | case Op_CastX2P: // no null checks on native pointers |
| 829 | case Op_Parm: // 'this' pointer is not null |
| 830 | case Op_LoadP: // Loading from within a klass |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 831 | case Op_LoadN: // Loading from within a klass |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 832 | case Op_LoadKlass: // Loading from within a klass |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 833 | case Op_LoadNKlass: // Loading from within a klass |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 834 | case Op_ConP: // Loading from a klass |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 835 | case Op_ConN: // Loading from a klass |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 836 | case Op_ConNKlass: // Loading from a klass |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 837 | case Op_CreateEx: // Sucking up the guts of an exception oop |
| 838 | case Op_Con: // Reading from TLS |
| 839 | case Op_CMoveP: // CMoveP is pinned |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 840 | case Op_CMoveN: // CMoveN is pinned |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 841 | break; // No progress |
| 842 | |
| 843 | case Op_Proj: // Direct call to an allocation routine |
| 844 | case Op_SCMemProj: // Memory state from store conditional ops |
| 845 | #ifdef ASSERT |
| 846 | { |
| 847 | assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); |
| 848 | const Node* call = adr->in(0); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 849 | if (call->is_CallJava()) { |
| 850 | const CallJavaNode* call_java = call->as_CallJava(); |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 851 | const TypeTuple *r = call_java->tf()->range(); |
| 852 | assert(r->cnt() > TypeFunc::Parms, "must return value"); |
| 853 | const Type* ret_type = r->field_at(TypeFunc::Parms); |
| 854 | assert(ret_type && ret_type->isa_ptr(), "must return pointer"); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 855 | // We further presume that this is one of |
| 856 | // new_instance_Java, new_array_Java, or |
| 857 | // the like, but do not assert for this. |
| 858 | } else if (call->is_Allocate()) { |
| 859 | // similar case to new_instance_Java, etc. |
| 860 | } else if (!call->is_CallLeaf()) { |
| 861 | // Projections from fetch_oop (OSR) are allowed as well. |
| 862 | ShouldNotReachHere(); |
| 863 | } |
| 864 | } |
| 865 | #endif |
| 866 | break; |
| 867 | default: |
| 868 | ShouldNotReachHere(); |
| 869 | } |
| 870 | break; |
| 871 | } |
| 872 | } |
| 873 | |
| 874 | return NULL; // No progress |
| 875 | } |
| 876 | |
| 877 | |
| 878 | //============================================================================= |
| 879 | uint LoadNode::size_of() const { return sizeof(*this); } |
| 880 | uint LoadNode::cmp( const Node &n ) const |
| 881 | { return !Type::cmp( _type, ((LoadNode&)n)._type ); } |
| 882 | const Type *LoadNode::bottom_type() const { return _type; } |
| 883 | uint LoadNode::ideal_reg() const { |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 884 | return _type->ideal_reg(); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 885 | } |
| 886 | |
| 887 | #ifndef PRODUCT |
| 888 | void LoadNode::dump_spec(outputStream *st) const { |
| 889 | MemNode::dump_spec(st); |
| 890 | if( !Verbose && !WizardMode ) { |
| 891 | // standard dump does this in Verbose and WizardMode |
| 892 | st->print(" #"); _type->dump_on(st); |
| 893 | } |
| 894 | } |
| 895 | #endif |
| 896 | |
Vladimir Kozlov | 21f481e | 2010-06-15 18:07:27 -0700 | [diff] [blame] | 897 | #ifdef ASSERT |
| 898 | //----------------------------is_immutable_value------------------------------- |
| 899 | // Helper function to allow a raw load without control edge for some cases |
| 900 | bool LoadNode::is_immutable_value(Node* adr) { |
| 901 | return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && |
| 902 | adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && |
| 903 | (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == |
| 904 | in_bytes(JavaThread::osthread_offset()))); |
| 905 | } |
| 906 | #endif |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 907 | |
| 908 | //----------------------------LoadNode::make----------------------------------- |
| 909 | // Polymorphic factory method: |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 910 | Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { |
| 911 | Compile* C = gvn.C; |
| 912 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 913 | // sanity check the alias category against the created node type |
| 914 | assert(!(adr_type->isa_oopptr() && |
| 915 | adr_type->offset() == oopDesc::klass_offset_in_bytes()), |
| 916 | "use LoadKlassNode instead"); |
| 917 | assert(!(adr_type->isa_aryptr() && |
| 918 | adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), |
| 919 | "use LoadRangeNode instead"); |
Vladimir Kozlov | 21f481e | 2010-06-15 18:07:27 -0700 | [diff] [blame] | 920 | // Check control edge of raw loads |
| 921 | assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || |
| 922 | // oop will be recorded in oop map if load crosses safepoint |
| 923 | rt->isa_oopptr() || is_immutable_value(adr), |
| 924 | "raw memory operations should have control edge"); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 925 | switch (bt) { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 926 | case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); |
| 927 | case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); |
| 928 | case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); |
| 929 | case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); |
| 930 | case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); |
| 931 | case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); |
| 932 | case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt ); |
| 933 | case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt ); |
| 934 | case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 935 | case T_OBJECT: |
| 936 | #ifdef _LP64 |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 937 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 938 | Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); |
| 939 | return new (C) DecodeNNode(load, load->bottom_type()->make_ptr()); |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 940 | } else |
| 941 | #endif |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 942 | { |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 943 | assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 944 | return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 945 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 946 | } |
| 947 | ShouldNotReachHere(); |
| 948 | return (LoadNode*)NULL; |
| 949 | } |
| 950 | |
| 951 | LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { |
| 952 | bool require_atomic = true; |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 953 | return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 954 | } |
| 955 | |
| 956 | |
| 957 | |
| 958 | |
| 959 | //------------------------------hash------------------------------------------- |
| 960 | uint LoadNode::hash() const { |
| 961 | // unroll addition of interesting fields |
| 962 | return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); |
| 963 | } |
| 964 | |
| 965 | //---------------------------can_see_stored_value------------------------------ |
| 966 | // This routine exists to make sure this set of tests is done the same |
| 967 | // everywhere. We need to make a coordinated change: first LoadNode::Ideal |
| 968 | // will change the graph shape in a way which makes memory alive twice at the |
| 969 | // same time (uses the Oracle model of aliasing), then some |
| 970 | // LoadXNode::Identity will fold things back to the equivalence-class model |
| 971 | // of aliasing. |
| 972 | Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { |
| 973 | Node* ld_adr = in(MemNode::Address); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 974 | intptr_t ld_off = 0; |
| 975 | AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 976 | const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 977 | Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; |
| 978 | // This is more general than load from boxing objects. |
| 979 | if (phase->C->eliminate_boxing() && (atp != NULL) && |
| 980 | (atp->index() >= Compile::AliasIdxRaw) && |
| 981 | (atp->field() != NULL) && !atp->field()->is_volatile()) { |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 982 | uint alias_idx = atp->index(); |
| 983 | bool final = atp->field()->is_final(); |
| 984 | Node* result = NULL; |
| 985 | Node* current = st; |
| 986 | // Skip through chains of MemBarNodes checking the MergeMems for |
| 987 | // new states for the slice of this load. Stop once any other |
| 988 | // kind of node is encountered. Loads from final memory can skip |
| 989 | // through any kind of MemBar but normal loads shouldn't skip |
| 990 | // through MemBarAcquire since the could allow them to move out of |
| 991 | // a synchronized region. |
| 992 | while (current->is_Proj()) { |
| 993 | int opc = current->in(0)->Opcode(); |
Roland Westrelin | b543a07 | 2011-08-02 18:36:40 +0200 | [diff] [blame] | 994 | if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) || |
| 995 | opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder || |
| 996 | opc == Op_MemBarReleaseLock) { |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 997 | Node* mem = current->in(0)->in(TypeFunc::Memory); |
| 998 | if (mem->is_MergeMem()) { |
| 999 | MergeMemNode* merge = mem->as_MergeMem(); |
| 1000 | Node* new_st = merge->memory_at(alias_idx); |
| 1001 | if (new_st == merge->base_memory()) { |
| 1002 | // Keep searching |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1003 | current = new_st; |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1004 | continue; |
| 1005 | } |
| 1006 | // Save the new memory state for the slice and fall through |
| 1007 | // to exit. |
| 1008 | result = new_st; |
| 1009 | } |
| 1010 | } |
| 1011 | break; |
| 1012 | } |
| 1013 | if (result != NULL) { |
| 1014 | st = result; |
| 1015 | } |
| 1016 | } |
| 1017 | |
| 1018 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1019 | // Loop around twice in the case Load -> Initialize -> Store. |
| 1020 | // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) |
| 1021 | for (int trip = 0; trip <= 1; trip++) { |
| 1022 | |
| 1023 | if (st->is_Store()) { |
| 1024 | Node* st_adr = st->in(MemNode::Address); |
| 1025 | if (!phase->eqv(st_adr, ld_adr)) { |
| 1026 | // Try harder before giving up... Match raw and non-raw pointers. |
| 1027 | intptr_t st_off = 0; |
| 1028 | AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); |
| 1029 | if (alloc == NULL) return NULL; |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1030 | if (alloc != ld_alloc) return NULL; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1031 | if (ld_off != st_off) return NULL; |
| 1032 | // At this point we have proven something like this setup: |
| 1033 | // A = Allocate(...) |
| 1034 | // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) |
| 1035 | // S = StoreQ(, AddP(, A.Parm , #Off), V) |
| 1036 | // (Actually, we haven't yet proven the Q's are the same.) |
| 1037 | // In other words, we are loading from a casted version of |
| 1038 | // the same pointer-and-offset that we stored to. |
| 1039 | // Thus, we are able to replace L by V. |
| 1040 | } |
| 1041 | // Now prove that we have a LoadQ matched to a StoreQ, for some Q. |
| 1042 | if (store_Opcode() != st->Opcode()) |
| 1043 | return NULL; |
| 1044 | return st->in(MemNode::ValueIn); |
| 1045 | } |
| 1046 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1047 | // A load from a freshly-created object always returns zero. |
| 1048 | // (This can happen after LoadNode::Ideal resets the load's memory input |
| 1049 | // to find_captured_store, which returned InitializeNode::zero_memory.) |
| 1050 | if (st->is_Proj() && st->in(0)->is_Allocate() && |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1051 | (st->in(0) == ld_alloc) && |
| 1052 | (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1053 | // return a zero value for the load's basic type |
| 1054 | // (This is one of the few places where a generic PhaseTransform |
| 1055 | // can create new nodes. Think of it as lazily manifesting |
| 1056 | // virtually pre-existing constants.) |
| 1057 | return phase->zerocon(memory_type()); |
| 1058 | } |
| 1059 | |
| 1060 | // A load from an initialization barrier can match a captured store. |
| 1061 | if (st->is_Proj() && st->in(0)->is_Initialize()) { |
| 1062 | InitializeNode* init = st->in(0)->as_Initialize(); |
| 1063 | AllocateNode* alloc = init->allocation(); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1064 | if ((alloc != NULL) && (alloc == ld_alloc)) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1065 | // examine a captured store value |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1066 | st = init->find_captured_store(ld_off, memory_size(), phase); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1067 | if (st != NULL) |
| 1068 | continue; // take one more trip around |
| 1069 | } |
| 1070 | } |
| 1071 | |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1072 | // Load boxed value from result of valueOf() call is input parameter. |
| 1073 | if (this->is_Load() && ld_adr->is_AddP() && |
| 1074 | (tp != NULL) && tp->is_ptr_to_boxed_value()) { |
| 1075 | intptr_t ignore = 0; |
| 1076 | Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); |
| 1077 | if (base != NULL && base->is_Proj() && |
| 1078 | base->as_Proj()->_con == TypeFunc::Parms && |
| 1079 | base->in(0)->is_CallStaticJava() && |
| 1080 | base->in(0)->as_CallStaticJava()->is_boxing_method()) { |
| 1081 | return base->in(0)->in(TypeFunc::Parms); |
| 1082 | } |
| 1083 | } |
| 1084 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1085 | break; |
| 1086 | } |
| 1087 | |
| 1088 | return NULL; |
| 1089 | } |
| 1090 | |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 1091 | //----------------------is_instance_field_load_with_local_phi------------------ |
| 1092 | bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1093 | if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && |
| 1094 | in(Address)->is_AddP() ) { |
| 1095 | const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); |
| 1096 | // Only instances and boxed values. |
| 1097 | if( t_oop != NULL && |
| 1098 | (t_oop->is_ptr_to_boxed_value() || |
| 1099 | t_oop->is_known_instance_field()) && |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 1100 | t_oop->offset() != Type::OffsetBot && |
| 1101 | t_oop->offset() != Type::OffsetTop) { |
| 1102 | return true; |
| 1103 | } |
| 1104 | } |
| 1105 | return false; |
| 1106 | } |
| 1107 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1108 | //------------------------------Identity--------------------------------------- |
| 1109 | // Loads are identity if previous store is to same address |
| 1110 | Node *LoadNode::Identity( PhaseTransform *phase ) { |
| 1111 | // If the previous store-maker is the right kind of Store, and the store is |
| 1112 | // to the same address, then we are equal to the value stored. |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1113 | Node* mem = in(Memory); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1114 | Node* value = can_see_stored_value(mem, phase); |
| 1115 | if( value ) { |
| 1116 | // byte, short & char stores truncate naturally. |
| 1117 | // A load has to load the truncated value which requires |
| 1118 | // some sort of masking operation and that requires an |
| 1119 | // Ideal call instead of an Identity call. |
| 1120 | if (memory_size() < BytesPerInt) { |
| 1121 | // If the input to the store does not fit with the load's result type, |
| 1122 | // it must be truncated via an Ideal call. |
| 1123 | if (!phase->type(value)->higher_equal(phase->type(this))) |
| 1124 | return this; |
| 1125 | } |
| 1126 | // (This works even when value is a Con, but LoadNode::Value |
| 1127 | // usually runs first, producing the singleton type of the Con.) |
| 1128 | return value; |
| 1129 | } |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 1130 | |
| 1131 | // Search for an existing data phi which was generated before for the same |
Christian Thalinger | 05d1de7 | 2009-02-27 13:27:09 -0800 | [diff] [blame] | 1132 | // instance's field to avoid infinite generation of phis in a loop. |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 1133 | Node *region = mem->in(0); |
| 1134 | if (is_instance_field_load_with_local_phi(region)) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1135 | const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 1136 | int this_index = phase->C->get_alias_index(addr_t); |
| 1137 | int this_offset = addr_t->offset(); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1138 | int this_iid = addr_t->instance_id(); |
| 1139 | if (!addr_t->is_known_instance() && |
| 1140 | addr_t->is_ptr_to_boxed_value()) { |
| 1141 | // Use _idx of address base (could be Phi node) for boxed values. |
| 1142 | intptr_t ignore = 0; |
| 1143 | Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); |
| 1144 | this_iid = base->_idx; |
| 1145 | } |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 1146 | const Type* this_type = bottom_type(); |
| 1147 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
| 1148 | Node* phi = region->fast_out(i); |
| 1149 | if (phi->is_Phi() && phi != mem && |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1150 | phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) { |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 1151 | return phi; |
| 1152 | } |
| 1153 | } |
| 1154 | } |
| 1155 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1156 | return this; |
| 1157 | } |
| 1158 | |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1159 | // We're loading from an object which has autobox behaviour. |
| 1160 | // If this object is result of a valueOf call we'll have a phi |
| 1161 | // merging a newly allocated object and a load from the cache. |
| 1162 | // We want to replace this load with the original incoming |
| 1163 | // argument to the valueOf call. |
| 1164 | Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1165 | assert(phase->C->eliminate_boxing(), "sanity"); |
| 1166 | intptr_t ignore = 0; |
| 1167 | Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); |
| 1168 | if ((base == NULL) || base->is_Phi()) { |
| 1169 | // Push the loads from the phi that comes from valueOf up |
| 1170 | // through it to allow elimination of the loads and the recovery |
| 1171 | // of the original value. It is done in split_through_phi(). |
| 1172 | return NULL; |
Vladimir Kozlov | 73e8e58 | 2009-02-18 13:53:42 -0800 | [diff] [blame] | 1173 | } else if (base->is_Load() || |
| 1174 | base->is_DecodeN() && base->in(1)->is_Load()) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1175 | // Eliminate the load of boxed value for integer types from the cache |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1176 | // array by deriving the value from the index into the array. |
| 1177 | // Capture the offset of the load and then reverse the computation. |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1178 | |
| 1179 | // Get LoadN node which loads a boxing object from 'cache' array. |
| 1180 | if (base->is_DecodeN()) { |
| 1181 | base = base->in(1); |
Vladimir Kozlov | 73e8e58 | 2009-02-18 13:53:42 -0800 | [diff] [blame] | 1182 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1183 | if (!base->in(Address)->is_AddP()) { |
| 1184 | return NULL; // Complex address |
| 1185 | } |
| 1186 | AddPNode* address = base->in(Address)->as_AddP(); |
| 1187 | Node* cache_base = address->in(AddPNode::Base); |
| 1188 | if ((cache_base != NULL) && cache_base->is_DecodeN()) { |
| 1189 | // Get ConP node which is static 'cache' field. |
| 1190 | cache_base = cache_base->in(1); |
| 1191 | } |
| 1192 | if ((cache_base != NULL) && cache_base->is_Con()) { |
| 1193 | const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); |
| 1194 | if ((base_type != NULL) && base_type->is_autobox_cache()) { |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1195 | Node* elements[4]; |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1196 | int shift = exact_log2(type2aelembytes(T_OBJECT)); |
| 1197 | int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); |
| 1198 | if ((count > 0) && elements[0]->is_Con() && |
| 1199 | ((count == 1) || |
| 1200 | (count == 2) && elements[1]->Opcode() == Op_LShiftX && |
| 1201 | elements[1]->in(2) == phase->intcon(shift))) { |
| 1202 | ciObjArray* array = base_type->const_oop()->as_obj_array(); |
| 1203 | // Fetch the box object cache[0] at the base of the array and get its value |
| 1204 | ciInstance* box = array->obj_at(0)->as_instance(); |
| 1205 | ciInstanceKlass* ik = box->klass()->as_instance_klass(); |
| 1206 | assert(ik->is_box_klass(), "sanity"); |
| 1207 | assert(ik->nof_nonstatic_fields() == 1, "change following code"); |
| 1208 | if (ik->nof_nonstatic_fields() == 1) { |
| 1209 | // This should be true nonstatic_field_at requires calling |
| 1210 | // nof_nonstatic_fields so check it anyway |
| 1211 | ciConstant c = box->field_value(ik->nonstatic_field_at(0)); |
| 1212 | BasicType bt = c.basic_type(); |
| 1213 | // Only integer types have boxing cache. |
| 1214 | assert(bt == T_BOOLEAN || bt == T_CHAR || |
| 1215 | bt == T_BYTE || bt == T_SHORT || |
| 1216 | bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt))); |
| 1217 | jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); |
| 1218 | if (cache_low != (int)cache_low) { |
| 1219 | return NULL; // should not happen since cache is array indexed by value |
| 1220 | } |
| 1221 | jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); |
| 1222 | if (offset != (int)offset) { |
| 1223 | return NULL; // should not happen since cache is array indexed by value |
| 1224 | } |
| 1225 | // Add up all the offsets making of the address of the load |
| 1226 | Node* result = elements[0]; |
| 1227 | for (int i = 1; i < count; i++) { |
| 1228 | result = phase->transform(new (phase->C) AddXNode(result, elements[i])); |
| 1229 | } |
| 1230 | // Remove the constant offset from the address and then |
| 1231 | result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset))); |
| 1232 | // remove the scaling of the offset to recover the original index. |
| 1233 | if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { |
| 1234 | // Peel the shift off directly but wrap it in a dummy node |
| 1235 | // since Ideal can't return existing nodes |
| 1236 | result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0)); |
| 1237 | } else if (result->is_Add() && result->in(2)->is_Con() && |
| 1238 | result->in(1)->Opcode() == Op_LShiftX && |
| 1239 | result->in(1)->in(2) == phase->intcon(shift)) { |
| 1240 | // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) |
| 1241 | // but for boxing cache access we know that X<<Z will not overflow |
| 1242 | // (there is range check) so we do this optimizatrion by hand here. |
| 1243 | Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift)); |
| 1244 | result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con)); |
| 1245 | } else { |
| 1246 | result = new (phase->C) RShiftXNode(result, phase->intcon(shift)); |
| 1247 | } |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1248 | #ifdef _LP64 |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1249 | if (bt != T_LONG) { |
| 1250 | result = new (phase->C) ConvL2INode(phase->transform(result)); |
| 1251 | } |
| 1252 | #else |
| 1253 | if (bt == T_LONG) { |
| 1254 | result = new (phase->C) ConvI2LNode(phase->transform(result)); |
| 1255 | } |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1256 | #endif |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1257 | return result; |
| 1258 | } |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1259 | } |
| 1260 | } |
| 1261 | } |
| 1262 | } |
| 1263 | return NULL; |
| 1264 | } |
| 1265 | |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1266 | static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { |
| 1267 | Node* region = phi->in(0); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1268 | if (region == NULL) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1269 | return false; // Wait stable graph |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1270 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1271 | uint cnt = phi->req(); |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1272 | for (uint i = 1; i < cnt; i++) { |
| 1273 | Node* rc = region->in(i); |
| 1274 | if (rc == NULL || phase->type(rc) == Type::TOP) |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1275 | return false; // Wait stable graph |
| 1276 | Node* in = phi->in(i); |
| 1277 | if (in == NULL || phase->type(in) == Type::TOP) |
| 1278 | return false; // Wait stable graph |
| 1279 | } |
| 1280 | return true; |
| 1281 | } |
| 1282 | //------------------------------split_through_phi------------------------------ |
| 1283 | // Split instance or boxed field load through Phi. |
| 1284 | Node *LoadNode::split_through_phi(PhaseGVN *phase) { |
| 1285 | Node* mem = in(Memory); |
| 1286 | Node* address = in(Address); |
| 1287 | const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); |
| 1288 | |
| 1289 | assert((t_oop != NULL) && |
| 1290 | (t_oop->is_known_instance_field() || |
| 1291 | t_oop->is_ptr_to_boxed_value()), "invalide conditions"); |
| 1292 | |
| 1293 | Compile* C = phase->C; |
| 1294 | intptr_t ignore = 0; |
| 1295 | Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); |
| 1296 | bool base_is_phi = (base != NULL) && base->is_Phi(); |
| 1297 | bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && |
| 1298 | (base != NULL) && (base == address->in(AddPNode::Base)) && |
| 1299 | phase->type(base)->higher_equal(TypePtr::NOTNULL); |
| 1300 | |
| 1301 | if (!((mem->is_Phi() || base_is_phi) && |
| 1302 | (load_boxed_values || t_oop->is_known_instance_field()))) { |
| 1303 | return NULL; // memory is not Phi |
| 1304 | } |
| 1305 | |
| 1306 | if (mem->is_Phi()) { |
| 1307 | if (!stable_phi(mem->as_Phi(), phase)) { |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1308 | return NULL; // Wait stable graph |
| 1309 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1310 | uint cnt = mem->req(); |
| 1311 | // Check for loop invariant memory. |
| 1312 | if (cnt == 3) { |
| 1313 | for (uint i = 1; i < cnt; i++) { |
| 1314 | Node* in = mem->in(i); |
| 1315 | Node* m = optimize_memory_chain(in, t_oop, this, phase); |
| 1316 | if (m == mem) { |
| 1317 | set_req(Memory, mem->in(cnt - i)); |
| 1318 | return this; // made change |
| 1319 | } |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1320 | } |
| 1321 | } |
| 1322 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1323 | if (base_is_phi) { |
| 1324 | if (!stable_phi(base->as_Phi(), phase)) { |
| 1325 | return NULL; // Wait stable graph |
| 1326 | } |
| 1327 | uint cnt = base->req(); |
| 1328 | // Check for loop invariant memory. |
| 1329 | if (cnt == 3) { |
| 1330 | for (uint i = 1; i < cnt; i++) { |
| 1331 | if (base->in(i) == base) { |
| 1332 | return NULL; // Wait stable graph |
| 1333 | } |
| 1334 | } |
| 1335 | } |
| 1336 | } |
| 1337 | |
| 1338 | bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); |
| 1339 | |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1340 | // Split through Phi (see original code in loopopts.cpp). |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1341 | assert(C->have_alias_type(t_oop), "instance should have alias type"); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1342 | |
| 1343 | // Do nothing here if Identity will find a value |
| 1344 | // (to avoid infinite chain of value phis generation). |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1345 | if (!phase->eqv(this, this->Identity(phase))) |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1346 | return NULL; |
| 1347 | |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1348 | // Select Region to split through. |
| 1349 | Node* region; |
| 1350 | if (!base_is_phi) { |
| 1351 | assert(mem->is_Phi(), "sanity"); |
| 1352 | region = mem->in(0); |
| 1353 | // Skip if the region dominates some control edge of the address. |
| 1354 | if (!MemNode::all_controls_dominate(address, region)) |
| 1355 | return NULL; |
| 1356 | } else if (!mem->is_Phi()) { |
| 1357 | assert(base_is_phi, "sanity"); |
| 1358 | region = base->in(0); |
| 1359 | // Skip if the region dominates some control edge of the memory. |
| 1360 | if (!MemNode::all_controls_dominate(mem, region)) |
| 1361 | return NULL; |
| 1362 | } else if (base->in(0) != mem->in(0)) { |
| 1363 | assert(base_is_phi && mem->is_Phi(), "sanity"); |
| 1364 | if (MemNode::all_controls_dominate(mem, base->in(0))) { |
| 1365 | region = base->in(0); |
| 1366 | } else if (MemNode::all_controls_dominate(address, mem->in(0))) { |
| 1367 | region = mem->in(0); |
| 1368 | } else { |
| 1369 | return NULL; // complex graph |
| 1370 | } |
| 1371 | } else { |
| 1372 | assert(base->in(0) == mem->in(0), "sanity"); |
| 1373 | region = mem->in(0); |
| 1374 | } |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1375 | |
| 1376 | const Type* this_type = this->bottom_type(); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1377 | int this_index = C->get_alias_index(t_oop); |
| 1378 | int this_offset = t_oop->offset(); |
| 1379 | int this_iid = t_oop->instance_id(); |
| 1380 | if (!t_oop->is_known_instance() && load_boxed_values) { |
| 1381 | // Use _idx of address base for boxed values. |
| 1382 | this_iid = base->_idx; |
| 1383 | } |
| 1384 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
| 1385 | Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1386 | for (uint i = 1; i < region->req(); i++) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1387 | Node* x; |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1388 | Node* the_clone = NULL; |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1389 | if (region->in(i) == C->top()) { |
| 1390 | x = C->top(); // Dead path? Use a dead data op |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1391 | } else { |
| 1392 | x = this->clone(); // Else clone up the data op |
| 1393 | the_clone = x; // Remember for possible deletion. |
| 1394 | // Alter data node to use pre-phi inputs |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1395 | if (this->in(0) == region) { |
| 1396 | x->set_req(0, region->in(i)); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1397 | } else { |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1398 | x->set_req(0, NULL); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1399 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1400 | if (mem->is_Phi() && (mem->in(0) == region)) { |
| 1401 | x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. |
| 1402 | } |
| 1403 | if (address->is_Phi() && address->in(0) == region) { |
| 1404 | x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone |
| 1405 | } |
| 1406 | if (base_is_phi && (base->in(0) == region)) { |
| 1407 | Node* base_x = base->in(i); // Clone address for loads from boxed objects. |
| 1408 | Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset))); |
| 1409 | x->set_req(Address, adr_x); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1410 | } |
| 1411 | } |
| 1412 | // Check for a 'win' on some paths |
| 1413 | const Type *t = x->Value(igvn); |
| 1414 | |
| 1415 | bool singleton = t->singleton(); |
| 1416 | |
| 1417 | // See comments in PhaseIdealLoop::split_thru_phi(). |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1418 | if (singleton && t == Type::TOP) { |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1419 | singleton &= region->is_Loop() && (i != LoopNode::EntryControl); |
| 1420 | } |
| 1421 | |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1422 | if (singleton) { |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1423 | x = igvn->makecon(t); |
| 1424 | } else { |
| 1425 | // We now call Identity to try to simplify the cloned node. |
| 1426 | // Note that some Identity methods call phase->type(this). |
| 1427 | // Make sure that the type array is big enough for |
| 1428 | // our new node, even though we may throw the node away. |
| 1429 | // (This tweaking with igvn only works because x is a new node.) |
| 1430 | igvn->set_type(x, t); |
Vladimir Kozlov | 897278f | 2008-08-01 10:06:45 -0700 | [diff] [blame] | 1431 | // If x is a TypeNode, capture any more-precise type permanently into Node |
Christian Thalinger | 05d1de7 | 2009-02-27 13:27:09 -0800 | [diff] [blame] | 1432 | // otherwise it will be not updated during igvn->transform since |
Vladimir Kozlov | 897278f | 2008-08-01 10:06:45 -0700 | [diff] [blame] | 1433 | // igvn->type(x) is set to x->Value() already. |
| 1434 | x->raise_bottom_type(t); |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1435 | Node *y = x->Identity(igvn); |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1436 | if (y != x) { |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1437 | x = y; |
| 1438 | } else { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1439 | y = igvn->hash_find_insert(x); |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1440 | if (y) { |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1441 | x = y; |
| 1442 | } else { |
| 1443 | // Else x is a new node we are keeping |
| 1444 | // We do not need register_new_node_with_optimizer |
| 1445 | // because set_type has already been called. |
| 1446 | igvn->_worklist.push(x); |
| 1447 | } |
| 1448 | } |
| 1449 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1450 | if (x != the_clone && the_clone != NULL) { |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1451 | igvn->remove_dead_node(the_clone); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1452 | } |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1453 | phi->set_req(i, x); |
| 1454 | } |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1455 | // Record Phi |
| 1456 | igvn->register_new_node_with_optimizer(phi); |
| 1457 | return phi; |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1458 | } |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1459 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1460 | //------------------------------Ideal------------------------------------------ |
| 1461 | // If the load is from Field memory and the pointer is non-null, we can |
| 1462 | // zero out the control input. |
| 1463 | // If the offset is constant and the base is an object allocation, |
| 1464 | // try to hook me up to the exact initializing store. |
| 1465 | Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
| 1466 | Node* p = MemNode::Ideal_common(phase, can_reshape); |
| 1467 | if (p) return (p == NodeSentinel) ? NULL : p; |
| 1468 | |
| 1469 | Node* ctrl = in(MemNode::Control); |
| 1470 | Node* address = in(MemNode::Address); |
| 1471 | |
| 1472 | // Skip up past a SafePoint control. Cannot do this for Stores because |
| 1473 | // pointer stores & cardmarks must stay on the same side of a SafePoint. |
| 1474 | if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && |
| 1475 | phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { |
| 1476 | ctrl = ctrl->in(0); |
| 1477 | set_req(MemNode::Control,ctrl); |
| 1478 | } |
| 1479 | |
Vladimir Kozlov | ca3a350 | 2009-04-07 19:04:24 -0700 | [diff] [blame] | 1480 | intptr_t ignore = 0; |
| 1481 | Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); |
| 1482 | if (base != NULL |
| 1483 | && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { |
| 1484 | // Check for useless control edge in some common special cases |
| 1485 | if (in(MemNode::Control) != NULL |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1486 | && phase->type(base)->higher_equal(TypePtr::NOTNULL) |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 1487 | && all_controls_dominate(base, phase->C->start())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1488 | // A method-invariant, non-null address (constant or 'this' argument). |
| 1489 | set_req(MemNode::Control, NULL); |
| 1490 | } |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1491 | } |
| 1492 | |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 1493 | Node* mem = in(MemNode::Memory); |
| 1494 | const TypePtr *addr_t = phase->type(address)->isa_ptr(); |
| 1495 | |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1496 | if (can_reshape && (addr_t != NULL)) { |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 1497 | // try to optimize our memory input |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1498 | Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 1499 | if (opt_mem != mem) { |
| 1500 | set_req(MemNode::Memory, opt_mem); |
Vladimir Kozlov | 3730631 | 2008-08-27 09:15:46 -0700 | [diff] [blame] | 1501 | if (phase->type( opt_mem ) == Type::TOP) return NULL; |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 1502 | return this; |
| 1503 | } |
| 1504 | const TypeOopPtr *t_oop = addr_t->isa_oopptr(); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1505 | if ((t_oop != NULL) && |
| 1506 | (t_oop->is_known_instance_field() || |
| 1507 | t_oop->is_ptr_to_boxed_value())) { |
Vladimir Kozlov | 09b7902 | 2011-11-09 07:25:51 -0800 | [diff] [blame] | 1508 | PhaseIterGVN *igvn = phase->is_IterGVN(); |
| 1509 | if (igvn != NULL && igvn->_worklist.member(opt_mem)) { |
| 1510 | // Delay this transformation until memory Phi is processed. |
| 1511 | phase->is_IterGVN()->_worklist.push(this); |
| 1512 | return NULL; |
| 1513 | } |
Vladimir Kozlov | 757229d | 2008-05-21 10:45:07 -0700 | [diff] [blame] | 1514 | // Split instance field load through Phi. |
| 1515 | Node* result = split_through_phi(phase); |
| 1516 | if (result != NULL) return result; |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1517 | |
| 1518 | if (t_oop->is_ptr_to_boxed_value()) { |
| 1519 | Node* result = eliminate_autobox(phase); |
| 1520 | if (result != NULL) return result; |
| 1521 | } |
Vladimir Kozlov | cdd2796 | 2008-03-20 15:11:44 -0700 | [diff] [blame] | 1522 | } |
| 1523 | } |
| 1524 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1525 | // Check for prior store with a different base or offset; make Load |
| 1526 | // independent. Skip through any number of them. Bail out if the stores |
| 1527 | // are in an endless dead cycle and report no progress. This is a key |
| 1528 | // transform for Reflection. However, if after skipping through the Stores |
| 1529 | // we can't then fold up against a prior store do NOT do the transform as |
| 1530 | // this amounts to using the 'Oracle' model of aliasing. It leaves the same |
| 1531 | // array memory alive twice: once for the hoisted Load and again after the |
| 1532 | // bypassed Store. This situation only works if EVERYBODY who does |
| 1533 | // anti-dependence work knows how to bypass. I.e. we need all |
| 1534 | // anti-dependence checks to ask the same Oracle. Right now, that Oracle is |
| 1535 | // the alias index stuff. So instead, peek through Stores and IFF we can |
| 1536 | // fold up, do so. |
| 1537 | Node* prev_mem = find_previous_store(phase); |
| 1538 | // Steps (a), (b): Walk past independent stores to find an exact match. |
| 1539 | if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { |
| 1540 | // (c) See if we can fold up on the spot, but don't fold up here. |
Christian Thalinger | 3b8452d | 2009-01-26 16:22:12 +0100 | [diff] [blame] | 1541 | // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1542 | // just return a prior value, which is done by Identity calls. |
| 1543 | if (can_see_stored_value(prev_mem, phase)) { |
| 1544 | // Make ready for step (d): |
| 1545 | set_req(MemNode::Memory, prev_mem); |
| 1546 | return this; |
| 1547 | } |
| 1548 | } |
| 1549 | |
| 1550 | return NULL; // No further progress |
| 1551 | } |
| 1552 | |
| 1553 | // Helper to recognize certain Klass fields which are invariant across |
| 1554 | // some group of array types (e.g., int[] or all T[] where T < Object). |
| 1555 | const Type* |
| 1556 | LoadNode::load_array_final_field(const TypeKlassPtr *tkls, |
| 1557 | ciKlass* klass) const { |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 1558 | if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1559 | // The field is Klass::_modifier_flags. Return its (constant) value. |
| 1560 | // (Folds up the 2nd indirection in aClassConstant.getModifiers().) |
| 1561 | assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); |
| 1562 | return TypeInt::make(klass->modifier_flags()); |
| 1563 | } |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 1564 | if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1565 | // The field is Klass::_access_flags. Return its (constant) value. |
| 1566 | // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) |
| 1567 | assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); |
| 1568 | return TypeInt::make(klass->access_flags()); |
| 1569 | } |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 1570 | if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1571 | // The field is Klass::_layout_helper. Return its constant value if known. |
| 1572 | assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); |
| 1573 | return TypeInt::make(klass->layout_helper()); |
| 1574 | } |
| 1575 | |
| 1576 | // No match. |
| 1577 | return NULL; |
| 1578 | } |
| 1579 | |
| 1580 | //------------------------------Value----------------------------------------- |
| 1581 | const Type *LoadNode::Value( PhaseTransform *phase ) const { |
| 1582 | // Either input is TOP ==> the result is TOP |
| 1583 | Node* mem = in(MemNode::Memory); |
| 1584 | const Type *t1 = phase->type(mem); |
| 1585 | if (t1 == Type::TOP) return Type::TOP; |
| 1586 | Node* adr = in(MemNode::Address); |
| 1587 | const TypePtr* tp = phase->type(adr)->isa_ptr(); |
| 1588 | if (tp == NULL || tp->empty()) return Type::TOP; |
| 1589 | int off = tp->offset(); |
| 1590 | assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); |
Christian Thalinger | 1038fed | 2011-09-02 04:28:59 -0700 | [diff] [blame] | 1591 | Compile* C = phase->C; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1592 | |
| 1593 | // Try to guess loaded type from pointer type |
| 1594 | if (tp->base() == Type::AryPtr) { |
| 1595 | const Type *t = tp->is_aryptr()->elem(); |
| 1596 | // Don't do this for integer types. There is only potential profit if |
| 1597 | // the element type t is lower than _type; that is, for int types, if _type is |
| 1598 | // more restrictive than t. This only happens here if one is short and the other |
| 1599 | // char (both 16 bits), and in those cases we've made an intentional decision |
| 1600 | // to use one kind of load over the other. See AndINode::Ideal and 4965907. |
| 1601 | // Also, do not try to narrow the type for a LoadKlass, regardless of offset. |
| 1602 | // |
| 1603 | // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) |
| 1604 | // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier |
| 1605 | // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been |
| 1606 | // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, |
| 1607 | // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. |
| 1608 | // In fact, that could have been the original type of p1, and p1 could have |
| 1609 | // had an original form like p1:(AddP x x (LShiftL quux 3)), where the |
| 1610 | // expression (LShiftL quux 3) independently optimized to the constant 8. |
| 1611 | if ((t->isa_int() == NULL) && (t->isa_long() == NULL) |
Vladimir Kozlov | d1191bb | 2012-06-15 01:25:19 -0700 | [diff] [blame] | 1612 | && (_type->isa_vect() == NULL) |
Vladimir Kozlov | 897278f | 2008-08-01 10:06:45 -0700 | [diff] [blame] | 1613 | && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1614 | // t might actually be lower than _type, if _type is a unique |
| 1615 | // concrete subclass of abstract class t. |
| 1616 | // Make sure the reference is not into the header, by comparing |
| 1617 | // the offset against the offset of the start of the array's data. |
| 1618 | // Different array types begin at slightly different offsets (12 vs. 16). |
| 1619 | // We choose T_BYTE as an example base type that is least restrictive |
| 1620 | // as to alignment, which will therefore produce the smallest |
| 1621 | // possible base offset. |
| 1622 | const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
| 1623 | if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? |
| 1624 | const Type* jt = t->join(_type); |
| 1625 | // In any case, do not allow the join, per se, to empty out the type. |
| 1626 | if (jt->empty() && !t->empty()) { |
| 1627 | // This can happen if a interface-typed array narrows to a class type. |
| 1628 | jt = _type; |
| 1629 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1630 | #ifdef ASSERT |
| 1631 | if (phase->C->eliminate_boxing() && adr->is_AddP()) { |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1632 | // The pointers in the autobox arrays are always non-null |
Vladimir Kozlov | ca3a350 | 2009-04-07 19:04:24 -0700 | [diff] [blame] | 1633 | Node* base = adr->in(AddPNode::Base); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1634 | if ((base != NULL) && base->is_DecodeN()) { |
| 1635 | // Get LoadN node which loads IntegerCache.cache field |
| 1636 | base = base->in(1); |
| 1637 | } |
| 1638 | if ((base != NULL) && base->is_Con()) { |
| 1639 | const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); |
| 1640 | if ((base_type != NULL) && base_type->is_autobox_cache()) { |
| 1641 | // It could be narrow oop |
| 1642 | assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); |
Tom Rodriguez | 10c473e | 2007-12-05 09:01:00 -0800 | [diff] [blame] | 1643 | } |
| 1644 | } |
| 1645 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1646 | #endif |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1647 | return jt; |
| 1648 | } |
| 1649 | } |
| 1650 | } else if (tp->base() == Type::InstPtr) { |
Christian Thalinger | 1038fed | 2011-09-02 04:28:59 -0700 | [diff] [blame] | 1651 | ciEnv* env = C->env(); |
Tom Rodriguez | 9db2092 | 2009-11-12 09:24:21 -0800 | [diff] [blame] | 1652 | const TypeInstPtr* tinst = tp->is_instptr(); |
| 1653 | ciKlass* klass = tinst->klass(); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1654 | assert( off != Type::OffsetBot || |
| 1655 | // arrays can be cast to Objects |
| 1656 | tp->is_oopptr()->klass()->is_java_lang_Object() || |
| 1657 | // unsafe field access may not have a constant offset |
Christian Thalinger | 1038fed | 2011-09-02 04:28:59 -0700 | [diff] [blame] | 1658 | C->has_unsafe_access(), |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1659 | "Field accesses must be precise" ); |
| 1660 | // For oop loads, we expect the _type to be precise |
Christian Thalinger | 1038fed | 2011-09-02 04:28:59 -0700 | [diff] [blame] | 1661 | if (klass == env->String_klass() && |
Tom Rodriguez | 9db2092 | 2009-11-12 09:24:21 -0800 | [diff] [blame] | 1662 | adr->is_AddP() && off != Type::OffsetBot) { |
Vladimir Kozlov | a74bc73 | 2011-02-26 12:10:54 -0800 | [diff] [blame] | 1663 | // For constant Strings treat the final fields as compile time constants. |
Tom Rodriguez | 9db2092 | 2009-11-12 09:24:21 -0800 | [diff] [blame] | 1664 | Node* base = adr->in(AddPNode::Base); |
Tom Rodriguez | 84aedee | 2010-08-30 17:27:35 -0700 | [diff] [blame] | 1665 | const TypeOopPtr* t = phase->type(base)->isa_oopptr(); |
| 1666 | if (t != NULL && t->singleton()) { |
Christian Thalinger | 1038fed | 2011-09-02 04:28:59 -0700 | [diff] [blame] | 1667 | ciField* field = env->String_klass()->get_field_by_offset(off, false); |
Vladimir Kozlov | a74bc73 | 2011-02-26 12:10:54 -0800 | [diff] [blame] | 1668 | if (field != NULL && field->is_final()) { |
| 1669 | ciObject* string = t->const_oop(); |
| 1670 | ciConstant constant = string->as_instance()->field_value(field); |
| 1671 | if (constant.basic_type() == T_INT) { |
| 1672 | return TypeInt::make(constant.as_int()); |
| 1673 | } else if (constant.basic_type() == T_ARRAY) { |
| 1674 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
John Coomes | fe899cc | 2011-03-22 13:36:33 -0700 | [diff] [blame] | 1675 | return TypeNarrowOop::make_from_constant(constant.as_object(), true); |
Vladimir Kozlov | a74bc73 | 2011-02-26 12:10:54 -0800 | [diff] [blame] | 1676 | } else { |
John Coomes | fe899cc | 2011-03-22 13:36:33 -0700 | [diff] [blame] | 1677 | return TypeOopPtr::make_from_constant(constant.as_object(), true); |
Vladimir Kozlov | a74bc73 | 2011-02-26 12:10:54 -0800 | [diff] [blame] | 1678 | } |
Tom Rodriguez | 9db2092 | 2009-11-12 09:24:21 -0800 | [diff] [blame] | 1679 | } |
| 1680 | } |
| 1681 | } |
| 1682 | } |
Christian Thalinger | 1038fed | 2011-09-02 04:28:59 -0700 | [diff] [blame] | 1683 | // Optimizations for constant objects |
| 1684 | ciObject* const_oop = tinst->const_oop(); |
| 1685 | if (const_oop != NULL) { |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1686 | // For constant Boxed value treat the target field as a compile time constant. |
| 1687 | if (tinst->is_ptr_to_boxed_value()) { |
| 1688 | return tinst->get_const_boxed_value(); |
| 1689 | } else |
Christian Thalinger | 1038fed | 2011-09-02 04:28:59 -0700 | [diff] [blame] | 1690 | // For constant CallSites treat the target field as a compile time constant. |
| 1691 | if (const_oop->is_call_site()) { |
| 1692 | ciCallSite* call_site = const_oop->as_call_site(); |
| 1693 | ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); |
| 1694 | if (field != NULL && field->is_call_site_target()) { |
| 1695 | ciMethodHandle* target = call_site->get_target(); |
| 1696 | if (target != NULL) { // just in case |
| 1697 | ciConstant constant(T_OBJECT, target); |
| 1698 | const Type* t; |
| 1699 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
| 1700 | t = TypeNarrowOop::make_from_constant(constant.as_object(), true); |
| 1701 | } else { |
| 1702 | t = TypeOopPtr::make_from_constant(constant.as_object(), true); |
| 1703 | } |
| 1704 | // Add a dependence for invalidation of the optimization. |
| 1705 | if (!call_site->is_constant_call_site()) { |
| 1706 | C->dependencies()->assert_call_site_target_value(call_site, target); |
| 1707 | } |
| 1708 | return t; |
| 1709 | } |
| 1710 | } |
| 1711 | } |
| 1712 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1713 | } else if (tp->base() == Type::KlassPtr) { |
| 1714 | assert( off != Type::OffsetBot || |
| 1715 | // arrays can be cast to Objects |
| 1716 | tp->is_klassptr()->klass()->is_java_lang_Object() || |
| 1717 | // also allow array-loading from the primary supertype |
| 1718 | // array during subtype checks |
| 1719 | Opcode() == Op_LoadKlass, |
| 1720 | "Field accesses must be precise" ); |
| 1721 | // For klass/static loads, we expect the _type to be precise |
| 1722 | } |
| 1723 | |
| 1724 | const TypeKlassPtr *tkls = tp->isa_klassptr(); |
| 1725 | if (tkls != NULL && !StressReflectiveCode) { |
| 1726 | ciKlass* klass = tkls->klass(); |
| 1727 | if (klass->is_loaded() && tkls->klass_is_exact()) { |
| 1728 | // We are loading a field from a Klass metaobject whose identity |
| 1729 | // is known at compile time (the type is "exact" or "precise"). |
| 1730 | // Check for fields we know are maintained as constants by the VM. |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 1731 | if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1732 | // The field is Klass::_super_check_offset. Return its (constant) value. |
| 1733 | // (Folds up type checking code.) |
| 1734 | assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); |
| 1735 | return TypeInt::make(klass->super_check_offset()); |
| 1736 | } |
| 1737 | // Compute index into primary_supers array |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 1738 | juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1739 | // Check for overflowing; use unsigned compare to handle the negative case. |
| 1740 | if( depth < ciKlass::primary_super_limit() ) { |
| 1741 | // The field is an element of Klass::_primary_supers. Return its (constant) value. |
| 1742 | // (Folds up type checking code.) |
| 1743 | assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); |
| 1744 | ciKlass *ss = klass->super_of_depth(depth); |
| 1745 | return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; |
| 1746 | } |
| 1747 | const Type* aift = load_array_final_field(tkls, klass); |
| 1748 | if (aift != NULL) return aift; |
Coleen Phillimore | 28047d6 | 2012-09-29 06:40:00 -0400 | [diff] [blame] | 1749 | if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset()) |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1750 | && klass->is_array_klass()) { |
Coleen Phillimore | 28047d6 | 2012-09-29 06:40:00 -0400 | [diff] [blame] | 1751 | // The field is ArrayKlass::_component_mirror. Return its (constant) value. |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1752 | // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) |
| 1753 | assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); |
| 1754 | return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); |
| 1755 | } |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 1756 | if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1757 | // The field is Klass::_java_mirror. Return its (constant) value. |
| 1758 | // (Folds up the 2nd indirection in anObjConstant.getClass().) |
| 1759 | assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); |
| 1760 | return TypeInstPtr::make(klass->java_mirror()); |
| 1761 | } |
| 1762 | } |
| 1763 | |
| 1764 | // We can still check if we are loading from the primary_supers array at a |
| 1765 | // shallow enough depth. Even though the klass is not exact, entries less |
| 1766 | // than or equal to its super depth are correct. |
| 1767 | if (klass->is_loaded() ) { |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 1768 | ciType *inner = klass; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1769 | while( inner->is_obj_array_klass() ) |
| 1770 | inner = inner->as_obj_array_klass()->base_element_type(); |
| 1771 | if( inner->is_instance_klass() && |
| 1772 | !inner->as_instance_klass()->flags().is_interface() ) { |
| 1773 | // Compute index into primary_supers array |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 1774 | juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1775 | // Check for overflowing; use unsigned compare to handle the negative case. |
| 1776 | if( depth < ciKlass::primary_super_limit() && |
| 1777 | depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case |
| 1778 | // The field is an element of Klass::_primary_supers. Return its (constant) value. |
| 1779 | // (Folds up type checking code.) |
| 1780 | assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); |
| 1781 | ciKlass *ss = klass->super_of_depth(depth); |
| 1782 | return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; |
| 1783 | } |
| 1784 | } |
| 1785 | } |
| 1786 | |
| 1787 | // If the type is enough to determine that the thing is not an array, |
| 1788 | // we can give the layout_helper a positive interval type. |
| 1789 | // This will help short-circuit some reflective code. |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 1790 | if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1791 | && !klass->is_array_klass() // not directly typed as an array |
| 1792 | && !klass->is_interface() // specifically not Serializable & Cloneable |
| 1793 | && !klass->is_java_lang_Object() // not the supertype of all T[] |
| 1794 | ) { |
| 1795 | // Note: When interfaces are reliable, we can narrow the interface |
| 1796 | // test to (klass != Serializable && klass != Cloneable). |
| 1797 | assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); |
| 1798 | jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); |
| 1799 | // The key property of this type is that it folds up tests |
| 1800 | // for array-ness, since it proves that the layout_helper is positive. |
| 1801 | // Thus, a generic value like the basic object layout helper works fine. |
| 1802 | return TypeInt::make(min_size, max_jint, Type::WidenMin); |
| 1803 | } |
| 1804 | } |
| 1805 | |
| 1806 | // If we are loading from a freshly-allocated object, produce a zero, |
| 1807 | // if the load is provably beyond the header of the object. |
| 1808 | // (Also allow a variable load from a fresh array to produce zero.) |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1809 | const TypeOopPtr *tinst = tp->isa_oopptr(); |
| 1810 | bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 1811 | bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); |
| 1812 | if (ReduceFieldZeroing || is_instance || is_boxed_value) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1813 | Node* value = can_see_stored_value(mem,phase); |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1814 | if (value != NULL && value->is_Con()) { |
| 1815 | assert(value->bottom_type()->higher_equal(_type),"sanity"); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1816 | return value->bottom_type(); |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1817 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1818 | } |
| 1819 | |
Vladimir Kozlov | abc5f94 | 2011-04-20 18:29:35 -0700 | [diff] [blame] | 1820 | if (is_instance) { |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 1821 | // If we have an instance type and our memory input is the |
| 1822 | // programs's initial memory state, there is no matching store, |
| 1823 | // so just return a zero of the appropriate type |
| 1824 | Node *mem = in(MemNode::Memory); |
| 1825 | if (mem->is_Parm() && mem->in(0)->is_Start()) { |
| 1826 | assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); |
| 1827 | return Type::get_zero_type(_type->basic_type()); |
| 1828 | } |
| 1829 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1830 | return _type; |
| 1831 | } |
| 1832 | |
| 1833 | //------------------------------match_edge------------------------------------- |
| 1834 | // Do we Match on this edge index or not? Match only the address. |
| 1835 | uint LoadNode::match_edge(uint idx) const { |
| 1836 | return idx == MemNode::Address; |
| 1837 | } |
| 1838 | |
| 1839 | //--------------------------LoadBNode::Ideal-------------------------------------- |
| 1840 | // |
| 1841 | // If the previous store is to the same address as this load, |
| 1842 | // and the value stored was larger than a byte, replace this load |
| 1843 | // with the value stored truncated to a byte. If no truncation is |
| 1844 | // needed, the replacement is done in LoadNode::Identity(). |
| 1845 | // |
| 1846 | Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
| 1847 | Node* mem = in(MemNode::Memory); |
| 1848 | Node* value = can_see_stored_value(mem,phase); |
| 1849 | if( value && !phase->type(value)->higher_equal( _type ) ) { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 1850 | Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) ); |
| 1851 | return new (phase->C) RShiftINode(result, phase->intcon(24)); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1852 | } |
| 1853 | // Identity call will handle the case where truncation is not needed. |
| 1854 | return LoadNode::Ideal(phase, can_reshape); |
| 1855 | } |
| 1856 | |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1857 | const Type* LoadBNode::Value(PhaseTransform *phase) const { |
| 1858 | Node* mem = in(MemNode::Memory); |
| 1859 | Node* value = can_see_stored_value(mem,phase); |
Vladimir Kozlov | 3770ece | 2012-01-24 17:04:51 -0800 | [diff] [blame] | 1860 | if (value != NULL && value->is_Con() && |
| 1861 | !value->bottom_type()->higher_equal(_type)) { |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1862 | // If the input to the store does not fit with the load's result type, |
| 1863 | // it must be truncated. We can't delay until Ideal call since |
| 1864 | // a singleton Value is needed for split_thru_phi optimization. |
| 1865 | int con = value->get_int(); |
| 1866 | return TypeInt::make((con << 24) >> 24); |
| 1867 | } |
| 1868 | return LoadNode::Value(phase); |
| 1869 | } |
| 1870 | |
Christian Thalinger | 89cea91 | 2009-03-09 03:17:11 -0700 | [diff] [blame] | 1871 | //--------------------------LoadUBNode::Ideal------------------------------------- |
| 1872 | // |
| 1873 | // If the previous store is to the same address as this load, |
| 1874 | // and the value stored was larger than a byte, replace this load |
| 1875 | // with the value stored truncated to a byte. If no truncation is |
| 1876 | // needed, the replacement is done in LoadNode::Identity(). |
| 1877 | // |
| 1878 | Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { |
| 1879 | Node* mem = in(MemNode::Memory); |
| 1880 | Node* value = can_see_stored_value(mem, phase); |
| 1881 | if (value && !phase->type(value)->higher_equal(_type)) |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 1882 | return new (phase->C) AndINode(value, phase->intcon(0xFF)); |
Christian Thalinger | 89cea91 | 2009-03-09 03:17:11 -0700 | [diff] [blame] | 1883 | // Identity call will handle the case where truncation is not needed. |
| 1884 | return LoadNode::Ideal(phase, can_reshape); |
| 1885 | } |
| 1886 | |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1887 | const Type* LoadUBNode::Value(PhaseTransform *phase) const { |
| 1888 | Node* mem = in(MemNode::Memory); |
| 1889 | Node* value = can_see_stored_value(mem,phase); |
Vladimir Kozlov | 3770ece | 2012-01-24 17:04:51 -0800 | [diff] [blame] | 1890 | if (value != NULL && value->is_Con() && |
| 1891 | !value->bottom_type()->higher_equal(_type)) { |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1892 | // If the input to the store does not fit with the load's result type, |
| 1893 | // it must be truncated. We can't delay until Ideal call since |
| 1894 | // a singleton Value is needed for split_thru_phi optimization. |
| 1895 | int con = value->get_int(); |
| 1896 | return TypeInt::make(con & 0xFF); |
| 1897 | } |
| 1898 | return LoadNode::Value(phase); |
| 1899 | } |
| 1900 | |
Christian Thalinger | 3b8452d | 2009-01-26 16:22:12 +0100 | [diff] [blame] | 1901 | //--------------------------LoadUSNode::Ideal------------------------------------- |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1902 | // |
| 1903 | // If the previous store is to the same address as this load, |
| 1904 | // and the value stored was larger than a char, replace this load |
| 1905 | // with the value stored truncated to a char. If no truncation is |
| 1906 | // needed, the replacement is done in LoadNode::Identity(). |
| 1907 | // |
Christian Thalinger | 3b8452d | 2009-01-26 16:22:12 +0100 | [diff] [blame] | 1908 | Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1909 | Node* mem = in(MemNode::Memory); |
| 1910 | Node* value = can_see_stored_value(mem,phase); |
| 1911 | if( value && !phase->type(value)->higher_equal( _type ) ) |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 1912 | return new (phase->C) AndINode(value,phase->intcon(0xFFFF)); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1913 | // Identity call will handle the case where truncation is not needed. |
| 1914 | return LoadNode::Ideal(phase, can_reshape); |
| 1915 | } |
| 1916 | |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1917 | const Type* LoadUSNode::Value(PhaseTransform *phase) const { |
| 1918 | Node* mem = in(MemNode::Memory); |
| 1919 | Node* value = can_see_stored_value(mem,phase); |
Vladimir Kozlov | 3770ece | 2012-01-24 17:04:51 -0800 | [diff] [blame] | 1920 | if (value != NULL && value->is_Con() && |
| 1921 | !value->bottom_type()->higher_equal(_type)) { |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1922 | // If the input to the store does not fit with the load's result type, |
| 1923 | // it must be truncated. We can't delay until Ideal call since |
| 1924 | // a singleton Value is needed for split_thru_phi optimization. |
| 1925 | int con = value->get_int(); |
| 1926 | return TypeInt::make(con & 0xFFFF); |
| 1927 | } |
| 1928 | return LoadNode::Value(phase); |
| 1929 | } |
| 1930 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1931 | //--------------------------LoadSNode::Ideal-------------------------------------- |
| 1932 | // |
| 1933 | // If the previous store is to the same address as this load, |
| 1934 | // and the value stored was larger than a short, replace this load |
| 1935 | // with the value stored truncated to a short. If no truncation is |
| 1936 | // needed, the replacement is done in LoadNode::Identity(). |
| 1937 | // |
| 1938 | Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
| 1939 | Node* mem = in(MemNode::Memory); |
| 1940 | Node* value = can_see_stored_value(mem,phase); |
| 1941 | if( value && !phase->type(value)->higher_equal( _type ) ) { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 1942 | Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) ); |
| 1943 | return new (phase->C) RShiftINode(result, phase->intcon(16)); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1944 | } |
| 1945 | // Identity call will handle the case where truncation is not needed. |
| 1946 | return LoadNode::Ideal(phase, can_reshape); |
| 1947 | } |
| 1948 | |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1949 | const Type* LoadSNode::Value(PhaseTransform *phase) const { |
| 1950 | Node* mem = in(MemNode::Memory); |
| 1951 | Node* value = can_see_stored_value(mem,phase); |
Vladimir Kozlov | 3770ece | 2012-01-24 17:04:51 -0800 | [diff] [blame] | 1952 | if (value != NULL && value->is_Con() && |
| 1953 | !value->bottom_type()->higher_equal(_type)) { |
Vladimir Kozlov | 5b3f2ef | 2012-01-20 09:43:06 -0800 | [diff] [blame] | 1954 | // If the input to the store does not fit with the load's result type, |
| 1955 | // it must be truncated. We can't delay until Ideal call since |
| 1956 | // a singleton Value is needed for split_thru_phi optimization. |
| 1957 | int con = value->get_int(); |
| 1958 | return TypeInt::make((con << 16) >> 16); |
| 1959 | } |
| 1960 | return LoadNode::Value(phase); |
| 1961 | } |
| 1962 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1963 | //============================================================================= |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 1964 | //----------------------------LoadKlassNode::make------------------------------ |
| 1965 | // Polymorphic factory method: |
| 1966 | Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { |
| 1967 | Compile* C = gvn.C; |
| 1968 | Node *ctl = NULL; |
| 1969 | // sanity check the alias category against the created node type |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 1970 | const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); |
| 1971 | assert(adr_type != NULL, "expecting TypeKlassPtr"); |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 1972 | #ifdef _LP64 |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 1973 | if (adr_type->is_ptr_to_narrowklass()) { |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 1974 | assert(UseCompressedKlassPointers, "no compressed klasses"); |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 1975 | Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass())); |
| 1976 | return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 1977 | } |
Vladimir Kozlov | a1f5099 | 2008-05-29 12:04:14 -0700 | [diff] [blame] | 1978 | #endif |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 1979 | assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 1980 | return new (C) LoadKlassNode(ctl, mem, adr, at, tk); |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 1981 | } |
| 1982 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1983 | //------------------------------Value------------------------------------------ |
| 1984 | const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 1985 | return klass_value_common(phase); |
| 1986 | } |
| 1987 | |
| 1988 | const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 1989 | // Either input is TOP ==> the result is TOP |
| 1990 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
| 1991 | if (t1 == Type::TOP) return Type::TOP; |
| 1992 | Node *adr = in(MemNode::Address); |
| 1993 | const Type *t2 = phase->type( adr ); |
| 1994 | if (t2 == Type::TOP) return Type::TOP; |
| 1995 | const TypePtr *tp = t2->is_ptr(); |
| 1996 | if (TypePtr::above_centerline(tp->ptr()) || |
| 1997 | tp->ptr() == TypePtr::Null) return Type::TOP; |
| 1998 | |
| 1999 | // Return a more precise klass, if possible |
| 2000 | const TypeInstPtr *tinst = tp->isa_instptr(); |
| 2001 | if (tinst != NULL) { |
| 2002 | ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); |
| 2003 | int offset = tinst->offset(); |
| 2004 | if (ik == phase->C->env()->Class_klass() |
| 2005 | && (offset == java_lang_Class::klass_offset_in_bytes() || |
| 2006 | offset == java_lang_Class::array_klass_offset_in_bytes())) { |
| 2007 | // We are loading a special hidden field from a Class mirror object, |
| 2008 | // the field which points to the VM's Klass metaobject. |
| 2009 | ciType* t = tinst->java_mirror_type(); |
| 2010 | // java_mirror_type returns non-null for compile-time Class constants. |
| 2011 | if (t != NULL) { |
| 2012 | // constant oop => constant klass |
| 2013 | if (offset == java_lang_Class::array_klass_offset_in_bytes()) { |
| 2014 | return TypeKlassPtr::make(ciArrayKlass::make(t)); |
| 2015 | } |
| 2016 | if (!t->is_klass()) { |
| 2017 | // a primitive Class (e.g., int.class) has NULL for a klass field |
| 2018 | return TypePtr::NULL_PTR; |
| 2019 | } |
| 2020 | // (Folds up the 1st indirection in aClassConstant.getModifiers().) |
| 2021 | return TypeKlassPtr::make(t->as_klass()); |
| 2022 | } |
| 2023 | // non-constant mirror, so we can't tell what's going on |
| 2024 | } |
| 2025 | if( !ik->is_loaded() ) |
| 2026 | return _type; // Bail out if not loaded |
| 2027 | if (offset == oopDesc::klass_offset_in_bytes()) { |
| 2028 | if (tinst->klass_is_exact()) { |
| 2029 | return TypeKlassPtr::make(ik); |
| 2030 | } |
| 2031 | // See if we can become precise: no subklasses and no interface |
| 2032 | // (Note: We need to support verified interfaces.) |
| 2033 | if (!ik->is_interface() && !ik->has_subklass()) { |
| 2034 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
| 2035 | // Add a dependence; if any subclass added we need to recompile |
| 2036 | if (!ik->is_final()) { |
| 2037 | // %%% should use stronger assert_unique_concrete_subtype instead |
| 2038 | phase->C->dependencies()->assert_leaf_type(ik); |
| 2039 | } |
| 2040 | // Return precise klass |
| 2041 | return TypeKlassPtr::make(ik); |
| 2042 | } |
| 2043 | |
| 2044 | // Return root of possible klass |
| 2045 | return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); |
| 2046 | } |
| 2047 | } |
| 2048 | |
| 2049 | // Check for loading klass from an array |
| 2050 | const TypeAryPtr *tary = tp->isa_aryptr(); |
| 2051 | if( tary != NULL ) { |
| 2052 | ciKlass *tary_klass = tary->klass(); |
| 2053 | if (tary_klass != NULL // can be NULL when at BOTTOM or TOP |
| 2054 | && tary->offset() == oopDesc::klass_offset_in_bytes()) { |
| 2055 | if (tary->klass_is_exact()) { |
| 2056 | return TypeKlassPtr::make(tary_klass); |
| 2057 | } |
| 2058 | ciArrayKlass *ak = tary->klass()->as_array_klass(); |
| 2059 | // If the klass is an object array, we defer the question to the |
| 2060 | // array component klass. |
| 2061 | if( ak->is_obj_array_klass() ) { |
| 2062 | assert( ak->is_loaded(), "" ); |
| 2063 | ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); |
| 2064 | if( base_k->is_loaded() && base_k->is_instance_klass() ) { |
| 2065 | ciInstanceKlass* ik = base_k->as_instance_klass(); |
| 2066 | // See if we can become precise: no subklasses and no interface |
| 2067 | if (!ik->is_interface() && !ik->has_subklass()) { |
| 2068 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
| 2069 | // Add a dependence; if any subclass added we need to recompile |
| 2070 | if (!ik->is_final()) { |
| 2071 | phase->C->dependencies()->assert_leaf_type(ik); |
| 2072 | } |
| 2073 | // Return precise array klass |
| 2074 | return TypeKlassPtr::make(ak); |
| 2075 | } |
| 2076 | } |
| 2077 | return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); |
| 2078 | } else { // Found a type-array? |
| 2079 | //assert(!UseExactTypes, "this code should be useless with exact types"); |
| 2080 | assert( ak->is_type_array_klass(), "" ); |
| 2081 | return TypeKlassPtr::make(ak); // These are always precise |
| 2082 | } |
| 2083 | } |
| 2084 | } |
| 2085 | |
| 2086 | // Check for loading klass from an array klass |
| 2087 | const TypeKlassPtr *tkls = tp->isa_klassptr(); |
| 2088 | if (tkls != NULL && !StressReflectiveCode) { |
| 2089 | ciKlass* klass = tkls->klass(); |
| 2090 | if( !klass->is_loaded() ) |
| 2091 | return _type; // Bail out if not loaded |
| 2092 | if( klass->is_obj_array_klass() && |
Coleen Phillimore | 28047d6 | 2012-09-29 06:40:00 -0400 | [diff] [blame] | 2093 | tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2094 | ciKlass* elem = klass->as_obj_array_klass()->element_klass(); |
| 2095 | // // Always returning precise element type is incorrect, |
| 2096 | // // e.g., element type could be object and array may contain strings |
| 2097 | // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); |
| 2098 | |
| 2099 | // The array's TypeKlassPtr was declared 'precise' or 'not precise' |
| 2100 | // according to the element type's subclassing. |
| 2101 | return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); |
| 2102 | } |
| 2103 | if( klass->is_instance_klass() && tkls->klass_is_exact() && |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 2104 | tkls->offset() == in_bytes(Klass::super_offset())) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2105 | ciKlass* sup = klass->as_instance_klass()->super(); |
| 2106 | // The field is Klass::_super. Return its (constant) value. |
| 2107 | // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) |
| 2108 | return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; |
| 2109 | } |
| 2110 | } |
| 2111 | |
| 2112 | // Bailout case |
| 2113 | return LoadNode::Value(phase); |
| 2114 | } |
| 2115 | |
| 2116 | //------------------------------Identity--------------------------------------- |
| 2117 | // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. |
| 2118 | // Also feed through the klass in Allocate(...klass...)._klass. |
| 2119 | Node* LoadKlassNode::Identity( PhaseTransform *phase ) { |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 2120 | return klass_identity_common(phase); |
| 2121 | } |
| 2122 | |
| 2123 | Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2124 | Node* x = LoadNode::Identity(phase); |
| 2125 | if (x != this) return x; |
| 2126 | |
| 2127 | // Take apart the address into an oop and and offset. |
| 2128 | // Return 'this' if we cannot. |
| 2129 | Node* adr = in(MemNode::Address); |
| 2130 | intptr_t offset = 0; |
| 2131 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
| 2132 | if (base == NULL) return this; |
| 2133 | const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); |
| 2134 | if (toop == NULL) return this; |
| 2135 | |
| 2136 | // We can fetch the klass directly through an AllocateNode. |
| 2137 | // This works even if the klass is not constant (clone or newArray). |
| 2138 | if (offset == oopDesc::klass_offset_in_bytes()) { |
| 2139 | Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); |
| 2140 | if (allocated_klass != NULL) { |
| 2141 | return allocated_klass; |
| 2142 | } |
| 2143 | } |
| 2144 | |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 2145 | // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. |
Coleen Phillimore | 28047d6 | 2012-09-29 06:40:00 -0400 | [diff] [blame] | 2146 | // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass. |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2147 | // See inline_native_Class_query for occurrences of these patterns. |
| 2148 | // Java Example: x.getClass().isAssignableFrom(y) |
| 2149 | // Java Example: Array.newInstance(x.getClass().getComponentType(), n) |
| 2150 | // |
| 2151 | // This improves reflective code, often making the Class |
| 2152 | // mirror go completely dead. (Current exception: Class |
| 2153 | // mirrors may appear in debug info, but we could clean them out by |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 2154 | // introducing a new debug info operator for Klass*.java_mirror). |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2155 | if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() |
| 2156 | && (offset == java_lang_Class::klass_offset_in_bytes() || |
| 2157 | offset == java_lang_Class::array_klass_offset_in_bytes())) { |
| 2158 | // We are loading a special hidden field from a Class mirror, |
Coleen Phillimore | 28047d6 | 2012-09-29 06:40:00 -0400 | [diff] [blame] | 2159 | // the field which points to its Klass or ArrayKlass metaobject. |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2160 | if (base->is_Load()) { |
| 2161 | Node* adr2 = base->in(MemNode::Address); |
| 2162 | const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); |
| 2163 | if (tkls != NULL && !tkls->empty() |
| 2164 | && (tkls->klass()->is_instance_klass() || |
| 2165 | tkls->klass()->is_array_klass()) |
| 2166 | && adr2->is_AddP() |
| 2167 | ) { |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 2168 | int mirror_field = in_bytes(Klass::java_mirror_offset()); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2169 | if (offset == java_lang_Class::array_klass_offset_in_bytes()) { |
Coleen Phillimore | 28047d6 | 2012-09-29 06:40:00 -0400 | [diff] [blame] | 2170 | mirror_field = in_bytes(ArrayKlass::component_mirror_offset()); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2171 | } |
Stefan Karlsson | e057d60 | 2011-12-07 11:35:03 +0100 | [diff] [blame] | 2172 | if (tkls->offset() == mirror_field) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2173 | return adr2->in(AddPNode::Base); |
| 2174 | } |
| 2175 | } |
| 2176 | } |
| 2177 | } |
| 2178 | |
| 2179 | return this; |
| 2180 | } |
| 2181 | |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 2182 | |
| 2183 | //------------------------------Value------------------------------------------ |
| 2184 | const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { |
| 2185 | const Type *t = klass_value_common(phase); |
Vladimir Kozlov | a671e7c | 2008-06-24 10:43:29 -0700 | [diff] [blame] | 2186 | if (t == Type::TOP) |
| 2187 | return t; |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 2188 | |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 2189 | return t->make_narrowklass(); |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 2190 | } |
| 2191 | |
| 2192 | //------------------------------Identity--------------------------------------- |
| 2193 | // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. |
| 2194 | // Also feed through the klass in Allocate(...klass...)._klass. |
| 2195 | Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { |
| 2196 | Node *x = klass_identity_common(phase); |
| 2197 | |
| 2198 | const Type *t = phase->type( x ); |
| 2199 | if( t == Type::TOP ) return x; |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 2200 | if( t->isa_narrowklass()) return x; |
| 2201 | assert (!t->isa_narrowoop(), "no narrow oop here"); |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 2202 | |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 2203 | return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass())); |
Vladimir Kozlov | 7603542 | 2008-05-21 13:46:23 -0700 | [diff] [blame] | 2204 | } |
| 2205 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2206 | //------------------------------Value----------------------------------------- |
| 2207 | const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { |
| 2208 | // Either input is TOP ==> the result is TOP |
| 2209 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
| 2210 | if( t1 == Type::TOP ) return Type::TOP; |
| 2211 | Node *adr = in(MemNode::Address); |
| 2212 | const Type *t2 = phase->type( adr ); |
| 2213 | if( t2 == Type::TOP ) return Type::TOP; |
| 2214 | const TypePtr *tp = t2->is_ptr(); |
| 2215 | if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; |
| 2216 | const TypeAryPtr *tap = tp->isa_aryptr(); |
| 2217 | if( !tap ) return _type; |
| 2218 | return tap->size(); |
| 2219 | } |
| 2220 | |
Chuck Rasbold | eee15b1 | 2008-09-17 08:29:17 -0700 | [diff] [blame] | 2221 | //-------------------------------Ideal--------------------------------------- |
| 2222 | // Feed through the length in AllocateArray(...length...)._length. |
| 2223 | Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
| 2224 | Node* p = MemNode::Ideal_common(phase, can_reshape); |
| 2225 | if (p) return (p == NodeSentinel) ? NULL : p; |
| 2226 | |
| 2227 | // Take apart the address into an oop and and offset. |
| 2228 | // Return 'this' if we cannot. |
| 2229 | Node* adr = in(MemNode::Address); |
| 2230 | intptr_t offset = 0; |
| 2231 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
| 2232 | if (base == NULL) return NULL; |
| 2233 | const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); |
| 2234 | if (tary == NULL) return NULL; |
| 2235 | |
| 2236 | // We can fetch the length directly through an AllocateArrayNode. |
| 2237 | // This works even if the length is not constant (clone or newArray). |
| 2238 | if (offset == arrayOopDesc::length_offset_in_bytes()) { |
| 2239 | AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); |
| 2240 | if (alloc != NULL) { |
| 2241 | Node* allocated_length = alloc->Ideal_length(); |
| 2242 | Node* len = alloc->make_ideal_length(tary, phase); |
| 2243 | if (allocated_length != len) { |
| 2244 | // New CastII improves on this. |
| 2245 | return len; |
| 2246 | } |
| 2247 | } |
| 2248 | } |
| 2249 | |
| 2250 | return NULL; |
| 2251 | } |
| 2252 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2253 | //------------------------------Identity--------------------------------------- |
| 2254 | // Feed through the length in AllocateArray(...length...)._length. |
| 2255 | Node* LoadRangeNode::Identity( PhaseTransform *phase ) { |
| 2256 | Node* x = LoadINode::Identity(phase); |
| 2257 | if (x != this) return x; |
| 2258 | |
| 2259 | // Take apart the address into an oop and and offset. |
| 2260 | // Return 'this' if we cannot. |
| 2261 | Node* adr = in(MemNode::Address); |
| 2262 | intptr_t offset = 0; |
| 2263 | Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
| 2264 | if (base == NULL) return this; |
| 2265 | const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); |
| 2266 | if (tary == NULL) return this; |
| 2267 | |
| 2268 | // We can fetch the length directly through an AllocateArrayNode. |
| 2269 | // This works even if the length is not constant (clone or newArray). |
| 2270 | if (offset == arrayOopDesc::length_offset_in_bytes()) { |
Chuck Rasbold | eee15b1 | 2008-09-17 08:29:17 -0700 | [diff] [blame] | 2271 | AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); |
| 2272 | if (alloc != NULL) { |
| 2273 | Node* allocated_length = alloc->Ideal_length(); |
| 2274 | // Do not allow make_ideal_length to allocate a CastII node. |
| 2275 | Node* len = alloc->make_ideal_length(tary, phase, false); |
| 2276 | if (allocated_length == len) { |
| 2277 | // Return allocated_length only if it would not be improved by a CastII. |
| 2278 | return allocated_length; |
| 2279 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2280 | } |
| 2281 | } |
| 2282 | |
| 2283 | return this; |
| 2284 | |
| 2285 | } |
Chuck Rasbold | eee15b1 | 2008-09-17 08:29:17 -0700 | [diff] [blame] | 2286 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2287 | //============================================================================= |
| 2288 | //---------------------------StoreNode::make----------------------------------- |
| 2289 | // Polymorphic factory method: |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 2290 | StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { |
| 2291 | Compile* C = gvn.C; |
Vladimir Kozlov | 21f481e | 2010-06-15 18:07:27 -0700 | [diff] [blame] | 2292 | assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw || |
| 2293 | ctl != NULL, "raw memory operations should have control edge"); |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 2294 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2295 | switch (bt) { |
| 2296 | case T_BOOLEAN: |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2297 | case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val); |
| 2298 | case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2299 | case T_CHAR: |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2300 | case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val); |
| 2301 | case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val); |
| 2302 | case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val); |
| 2303 | case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val); |
Jon Masamitsu | 5c58d27 | 2012-09-01 13:25:18 -0400 | [diff] [blame] | 2304 | case T_METADATA: |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2305 | case T_ADDRESS: |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 2306 | case T_OBJECT: |
| 2307 | #ifdef _LP64 |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 2308 | if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2309 | val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop())); |
| 2310 | return new (C) StoreNNode(ctl, mem, adr, adr_type, val); |
Roland Westrelin | 61eb5a0 | 2012-10-09 10:11:38 +0200 | [diff] [blame] | 2311 | } else if (adr->bottom_type()->is_ptr_to_narrowklass() || |
| 2312 | (UseCompressedKlassPointers && val->bottom_type()->isa_klassptr() && |
| 2313 | adr->bottom_type()->isa_rawptr())) { |
| 2314 | val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); |
| 2315 | return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val); |
| 2316 | } |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 2317 | #endif |
Vladimir Kozlov | a671e7c | 2008-06-24 10:43:29 -0700 | [diff] [blame] | 2318 | { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2319 | return new (C) StorePNode(ctl, mem, adr, adr_type, val); |
Vladimir Kozlov | a671e7c | 2008-06-24 10:43:29 -0700 | [diff] [blame] | 2320 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2321 | } |
| 2322 | ShouldNotReachHere(); |
| 2323 | return (StoreNode*)NULL; |
| 2324 | } |
| 2325 | |
| 2326 | StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { |
| 2327 | bool require_atomic = true; |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2328 | return new (C) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2329 | } |
| 2330 | |
| 2331 | |
| 2332 | //--------------------------bottom_type---------------------------------------- |
| 2333 | const Type *StoreNode::bottom_type() const { |
| 2334 | return Type::MEMORY; |
| 2335 | } |
| 2336 | |
| 2337 | //------------------------------hash------------------------------------------- |
| 2338 | uint StoreNode::hash() const { |
| 2339 | // unroll addition of interesting fields |
| 2340 | //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); |
| 2341 | |
| 2342 | // Since they are not commoned, do not hash them: |
| 2343 | return NO_HASH; |
| 2344 | } |
| 2345 | |
| 2346 | //------------------------------Ideal------------------------------------------ |
| 2347 | // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). |
| 2348 | // When a store immediately follows a relevant allocation/initialization, |
| 2349 | // try to capture it into the initialization, or hoist it above. |
| 2350 | Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
| 2351 | Node* p = MemNode::Ideal_common(phase, can_reshape); |
| 2352 | if (p) return (p == NodeSentinel) ? NULL : p; |
| 2353 | |
| 2354 | Node* mem = in(MemNode::Memory); |
| 2355 | Node* address = in(MemNode::Address); |
| 2356 | |
Tom Rodriguez | aac4647 | 2011-04-05 19:14:03 -0700 | [diff] [blame] | 2357 | // Back-to-back stores to same address? Fold em up. Generally |
| 2358 | // unsafe if I have intervening uses... Also disallowed for StoreCM |
| 2359 | // since they must follow each StoreP operation. Redundant StoreCMs |
| 2360 | // are eliminated just before matching in final_graph_reshape. |
Vladimir Kozlov | 791afc4 | 2012-01-10 18:05:38 -0800 | [diff] [blame] | 2361 | if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && |
Tom Rodriguez | aac4647 | 2011-04-05 19:14:03 -0700 | [diff] [blame] | 2362 | mem->Opcode() != Op_StoreCM) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2363 | // Looking at a dead closed cycle of memory? |
| 2364 | assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); |
| 2365 | |
| 2366 | assert(Opcode() == mem->Opcode() || |
| 2367 | phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, |
| 2368 | "no mismatched stores, except on raw memory"); |
| 2369 | |
| 2370 | if (mem->outcnt() == 1 && // check for intervening uses |
| 2371 | mem->as_Store()->memory_size() <= this->memory_size()) { |
| 2372 | // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. |
| 2373 | // For example, 'mem' might be the final state at a conditional return. |
| 2374 | // Or, 'mem' might be used by some node which is live at the same time |
| 2375 | // 'this' is live, which might be unschedulable. So, require exactly |
| 2376 | // ONE user, the 'this' store, until such time as we clone 'mem' for |
| 2377 | // each of 'mem's uses (thus making the exactly-1-user-rule hold true). |
| 2378 | if (can_reshape) { // (%%% is this an anachronism?) |
| 2379 | set_req_X(MemNode::Memory, mem->in(MemNode::Memory), |
| 2380 | phase->is_IterGVN()); |
| 2381 | } else { |
| 2382 | // It's OK to do this in the parser, since DU info is always accurate, |
| 2383 | // and the parser always refers to nodes via SafePointNode maps. |
| 2384 | set_req(MemNode::Memory, mem->in(MemNode::Memory)); |
| 2385 | } |
| 2386 | return this; |
| 2387 | } |
| 2388 | } |
| 2389 | |
| 2390 | // Capture an unaliased, unconditional, simple store into an initializer. |
| 2391 | // Or, if it is independent of the allocation, hoist it above the allocation. |
| 2392 | if (ReduceFieldZeroing && /*can_reshape &&*/ |
| 2393 | mem->is_Proj() && mem->in(0)->is_Initialize()) { |
| 2394 | InitializeNode* init = mem->in(0)->as_Initialize(); |
Roland Westrelin | fe92862 | 2013-02-25 14:13:04 +0100 | [diff] [blame] | 2395 | intptr_t offset = init->can_capture_store(this, phase, can_reshape); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2396 | if (offset > 0) { |
Roland Westrelin | fe92862 | 2013-02-25 14:13:04 +0100 | [diff] [blame] | 2397 | Node* moved = init->capture_store(this, offset, phase, can_reshape); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2398 | // If the InitializeNode captured me, it made a raw copy of me, |
| 2399 | // and I need to disappear. |
| 2400 | if (moved != NULL) { |
| 2401 | // %%% hack to ensure that Ideal returns a new node: |
| 2402 | mem = MergeMemNode::make(phase->C, mem); |
| 2403 | return mem; // fold me away |
| 2404 | } |
| 2405 | } |
| 2406 | } |
| 2407 | |
| 2408 | return NULL; // No further progress |
| 2409 | } |
| 2410 | |
| 2411 | //------------------------------Value----------------------------------------- |
| 2412 | const Type *StoreNode::Value( PhaseTransform *phase ) const { |
| 2413 | // Either input is TOP ==> the result is TOP |
| 2414 | const Type *t1 = phase->type( in(MemNode::Memory) ); |
| 2415 | if( t1 == Type::TOP ) return Type::TOP; |
| 2416 | const Type *t2 = phase->type( in(MemNode::Address) ); |
| 2417 | if( t2 == Type::TOP ) return Type::TOP; |
| 2418 | const Type *t3 = phase->type( in(MemNode::ValueIn) ); |
| 2419 | if( t3 == Type::TOP ) return Type::TOP; |
| 2420 | return Type::MEMORY; |
| 2421 | } |
| 2422 | |
| 2423 | //------------------------------Identity--------------------------------------- |
| 2424 | // Remove redundant stores: |
| 2425 | // Store(m, p, Load(m, p)) changes to m. |
| 2426 | // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). |
| 2427 | Node *StoreNode::Identity( PhaseTransform *phase ) { |
| 2428 | Node* mem = in(MemNode::Memory); |
| 2429 | Node* adr = in(MemNode::Address); |
| 2430 | Node* val = in(MemNode::ValueIn); |
| 2431 | |
| 2432 | // Load then Store? Then the Store is useless |
| 2433 | if (val->is_Load() && |
Vladimir Kozlov | 791afc4 | 2012-01-10 18:05:38 -0800 | [diff] [blame] | 2434 | val->in(MemNode::Address)->eqv_uncast(adr) && |
| 2435 | val->in(MemNode::Memory )->eqv_uncast(mem) && |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2436 | val->as_Load()->store_Opcode() == Opcode()) { |
| 2437 | return mem; |
| 2438 | } |
| 2439 | |
| 2440 | // Two stores in a row of the same value? |
| 2441 | if (mem->is_Store() && |
Vladimir Kozlov | 791afc4 | 2012-01-10 18:05:38 -0800 | [diff] [blame] | 2442 | mem->in(MemNode::Address)->eqv_uncast(adr) && |
| 2443 | mem->in(MemNode::ValueIn)->eqv_uncast(val) && |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2444 | mem->Opcode() == Opcode()) { |
| 2445 | return mem; |
| 2446 | } |
| 2447 | |
| 2448 | // Store of zero anywhere into a freshly-allocated object? |
| 2449 | // Then the store is useless. |
| 2450 | // (It must already have been captured by the InitializeNode.) |
| 2451 | if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { |
| 2452 | // a newly allocated object is already all-zeroes everywhere |
| 2453 | if (mem->is_Proj() && mem->in(0)->is_Allocate()) { |
| 2454 | return mem; |
| 2455 | } |
| 2456 | |
| 2457 | // the store may also apply to zero-bits in an earlier object |
| 2458 | Node* prev_mem = find_previous_store(phase); |
| 2459 | // Steps (a), (b): Walk past independent stores to find an exact match. |
| 2460 | if (prev_mem != NULL) { |
| 2461 | Node* prev_val = can_see_stored_value(prev_mem, phase); |
| 2462 | if (prev_val != NULL && phase->eqv(prev_val, val)) { |
| 2463 | // prev_val and val might differ by a cast; it would be good |
| 2464 | // to keep the more informative of the two. |
| 2465 | return mem; |
| 2466 | } |
| 2467 | } |
| 2468 | } |
| 2469 | |
| 2470 | return this; |
| 2471 | } |
| 2472 | |
| 2473 | //------------------------------match_edge------------------------------------- |
| 2474 | // Do we Match on this edge index or not? Match only memory & value |
| 2475 | uint StoreNode::match_edge(uint idx) const { |
| 2476 | return idx == MemNode::Address || idx == MemNode::ValueIn; |
| 2477 | } |
| 2478 | |
| 2479 | //------------------------------cmp-------------------------------------------- |
| 2480 | // Do not common stores up together. They generally have to be split |
| 2481 | // back up anyways, so do not bother. |
| 2482 | uint StoreNode::cmp( const Node &n ) const { |
| 2483 | return (&n == this); // Always fail except on self |
| 2484 | } |
| 2485 | |
| 2486 | //------------------------------Ideal_masked_input----------------------------- |
| 2487 | // Check for a useless mask before a partial-word store |
| 2488 | // (StoreB ... (AndI valIn conIa) ) |
| 2489 | // If (conIa & mask == mask) this simplifies to |
| 2490 | // (StoreB ... (valIn) ) |
| 2491 | Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { |
| 2492 | Node *val = in(MemNode::ValueIn); |
| 2493 | if( val->Opcode() == Op_AndI ) { |
| 2494 | const TypeInt *t = phase->type( val->in(2) )->isa_int(); |
| 2495 | if( t && t->is_con() && (t->get_con() & mask) == mask ) { |
| 2496 | set_req(MemNode::ValueIn, val->in(1)); |
| 2497 | return this; |
| 2498 | } |
| 2499 | } |
| 2500 | return NULL; |
| 2501 | } |
| 2502 | |
| 2503 | |
| 2504 | //------------------------------Ideal_sign_extended_input---------------------- |
| 2505 | // Check for useless sign-extension before a partial-word store |
| 2506 | // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) |
| 2507 | // If (conIL == conIR && conIR <= num_bits) this simplifies to |
| 2508 | // (StoreB ... (valIn) ) |
| 2509 | Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { |
| 2510 | Node *val = in(MemNode::ValueIn); |
| 2511 | if( val->Opcode() == Op_RShiftI ) { |
| 2512 | const TypeInt *t = phase->type( val->in(2) )->isa_int(); |
| 2513 | if( t && t->is_con() && (t->get_con() <= num_bits) ) { |
| 2514 | Node *shl = val->in(1); |
| 2515 | if( shl->Opcode() == Op_LShiftI ) { |
| 2516 | const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); |
| 2517 | if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { |
| 2518 | set_req(MemNode::ValueIn, shl->in(1)); |
| 2519 | return this; |
| 2520 | } |
| 2521 | } |
| 2522 | } |
| 2523 | } |
| 2524 | return NULL; |
| 2525 | } |
| 2526 | |
| 2527 | //------------------------------value_never_loaded----------------------------------- |
| 2528 | // Determine whether there are any possible loads of the value stored. |
| 2529 | // For simplicity, we actually check if there are any loads from the |
| 2530 | // address stored to, not just for loads of the value stored by this node. |
| 2531 | // |
| 2532 | bool StoreNode::value_never_loaded( PhaseTransform *phase) const { |
| 2533 | Node *adr = in(Address); |
| 2534 | const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); |
| 2535 | if (adr_oop == NULL) |
| 2536 | return false; |
Vladimir Kozlov | 4213e62 | 2008-06-26 13:34:00 -0700 | [diff] [blame] | 2537 | if (!adr_oop->is_known_instance_field()) |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2538 | return false; // if not a distinct instance, there may be aliases of the address |
| 2539 | for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { |
| 2540 | Node *use = adr->fast_out(i); |
| 2541 | int opc = use->Opcode(); |
| 2542 | if (use->is_Load() || use->is_LoadStore()) { |
| 2543 | return false; |
| 2544 | } |
| 2545 | } |
| 2546 | return true; |
| 2547 | } |
| 2548 | |
| 2549 | //============================================================================= |
| 2550 | //------------------------------Ideal------------------------------------------ |
| 2551 | // If the store is from an AND mask that leaves the low bits untouched, then |
| 2552 | // we can skip the AND operation. If the store is from a sign-extension |
| 2553 | // (a left shift, then right shift) we can skip both. |
| 2554 | Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
| 2555 | Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); |
| 2556 | if( progress != NULL ) return progress; |
| 2557 | |
| 2558 | progress = StoreNode::Ideal_sign_extended_input(phase, 24); |
| 2559 | if( progress != NULL ) return progress; |
| 2560 | |
| 2561 | // Finally check the default case |
| 2562 | return StoreNode::Ideal(phase, can_reshape); |
| 2563 | } |
| 2564 | |
| 2565 | //============================================================================= |
| 2566 | //------------------------------Ideal------------------------------------------ |
| 2567 | // If the store is from an AND mask that leaves the low bits untouched, then |
| 2568 | // we can skip the AND operation |
| 2569 | Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
| 2570 | Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); |
| 2571 | if( progress != NULL ) return progress; |
| 2572 | |
| 2573 | progress = StoreNode::Ideal_sign_extended_input(phase, 16); |
| 2574 | if( progress != NULL ) return progress; |
| 2575 | |
| 2576 | // Finally check the default case |
| 2577 | return StoreNode::Ideal(phase, can_reshape); |
| 2578 | } |
| 2579 | |
| 2580 | //============================================================================= |
| 2581 | //------------------------------Identity--------------------------------------- |
| 2582 | Node *StoreCMNode::Identity( PhaseTransform *phase ) { |
| 2583 | // No need to card mark when storing a null ptr |
| 2584 | Node* my_store = in(MemNode::OopStore); |
| 2585 | if (my_store->is_Store()) { |
| 2586 | const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); |
| 2587 | if( t1 == TypePtr::NULL_PTR ) { |
| 2588 | return in(MemNode::Memory); |
| 2589 | } |
| 2590 | } |
| 2591 | return this; |
| 2592 | } |
| 2593 | |
Changpeng Fang | c492f4c | 2009-09-14 09:49:54 -0700 | [diff] [blame] | 2594 | //============================================================================= |
| 2595 | //------------------------------Ideal--------------------------------------- |
| 2596 | Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
| 2597 | Node* progress = StoreNode::Ideal(phase, can_reshape); |
| 2598 | if (progress != NULL) return progress; |
| 2599 | |
| 2600 | Node* my_store = in(MemNode::OopStore); |
| 2601 | if (my_store->is_MergeMem()) { |
| 2602 | Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); |
| 2603 | set_req(MemNode::OopStore, mem); |
| 2604 | return this; |
| 2605 | } |
| 2606 | |
| 2607 | return NULL; |
| 2608 | } |
| 2609 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2610 | //------------------------------Value----------------------------------------- |
| 2611 | const Type *StoreCMNode::Value( PhaseTransform *phase ) const { |
Vladimir Kozlov | 5070812 | 2008-03-11 11:04:40 -0700 | [diff] [blame] | 2612 | // Either input is TOP ==> the result is TOP |
| 2613 | const Type *t = phase->type( in(MemNode::Memory) ); |
| 2614 | if( t == Type::TOP ) return Type::TOP; |
| 2615 | t = phase->type( in(MemNode::Address) ); |
| 2616 | if( t == Type::TOP ) return Type::TOP; |
| 2617 | t = phase->type( in(MemNode::ValueIn) ); |
| 2618 | if( t == Type::TOP ) return Type::TOP; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2619 | // If extra input is TOP ==> the result is TOP |
Vladimir Kozlov | 5070812 | 2008-03-11 11:04:40 -0700 | [diff] [blame] | 2620 | t = phase->type( in(MemNode::OopStore) ); |
| 2621 | if( t == Type::TOP ) return Type::TOP; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2622 | |
| 2623 | return StoreNode::Value( phase ); |
| 2624 | } |
| 2625 | |
| 2626 | |
| 2627 | //============================================================================= |
| 2628 | //----------------------------------SCMemProjNode------------------------------ |
| 2629 | const Type * SCMemProjNode::Value( PhaseTransform *phase ) const |
| 2630 | { |
| 2631 | return bottom_type(); |
| 2632 | } |
| 2633 | |
| 2634 | //============================================================================= |
Roland Westrelin | 3025406 | 2012-09-20 16:49:17 +0200 | [diff] [blame] | 2635 | //----------------------------------LoadStoreNode------------------------------ |
| 2636 | LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) |
| 2637 | : Node(required), |
| 2638 | _type(rt), |
| 2639 | _adr_type(at) |
| 2640 | { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2641 | init_req(MemNode::Control, c ); |
| 2642 | init_req(MemNode::Memory , mem); |
| 2643 | init_req(MemNode::Address, adr); |
| 2644 | init_req(MemNode::ValueIn, val); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2645 | init_class_id(Class_LoadStore); |
Roland Westrelin | 3025406 | 2012-09-20 16:49:17 +0200 | [diff] [blame] | 2646 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2647 | |
Roland Westrelin | 3025406 | 2012-09-20 16:49:17 +0200 | [diff] [blame] | 2648 | uint LoadStoreNode::ideal_reg() const { |
| 2649 | return _type->ideal_reg(); |
| 2650 | } |
| 2651 | |
| 2652 | bool LoadStoreNode::result_not_used() const { |
| 2653 | for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { |
| 2654 | Node *x = fast_out(i); |
| 2655 | if (x->Opcode() == Op_SCMemProj) continue; |
| 2656 | return false; |
| 2657 | } |
| 2658 | return true; |
| 2659 | } |
| 2660 | |
| 2661 | uint LoadStoreNode::size_of() const { return sizeof(*this); } |
| 2662 | |
| 2663 | //============================================================================= |
| 2664 | //----------------------------------LoadStoreConditionalNode-------------------- |
| 2665 | LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { |
| 2666 | init_req(ExpectedIn, ex ); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2667 | } |
| 2668 | |
| 2669 | //============================================================================= |
| 2670 | //-------------------------------adr_type-------------------------------------- |
| 2671 | // Do we Match on this edge index or not? Do not match memory |
| 2672 | const TypePtr* ClearArrayNode::adr_type() const { |
| 2673 | Node *adr = in(3); |
| 2674 | return MemNode::calculate_adr_type(adr->bottom_type()); |
| 2675 | } |
| 2676 | |
| 2677 | //------------------------------match_edge------------------------------------- |
| 2678 | // Do we Match on this edge index or not? Do not match memory |
| 2679 | uint ClearArrayNode::match_edge(uint idx) const { |
| 2680 | return idx > 1; |
| 2681 | } |
| 2682 | |
| 2683 | //------------------------------Identity--------------------------------------- |
| 2684 | // Clearing a zero length array does nothing |
| 2685 | Node *ClearArrayNode::Identity( PhaseTransform *phase ) { |
Tom Rodriguez | 8e1624d | 2008-03-18 23:44:46 -0700 | [diff] [blame] | 2686 | return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2687 | } |
| 2688 | |
| 2689 | //------------------------------Idealize--------------------------------------- |
| 2690 | // Clearing a short array is faster with stores |
| 2691 | Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
| 2692 | const int unit = BytesPerLong; |
| 2693 | const TypeX* t = phase->type(in(2))->isa_intptr_t(); |
| 2694 | if (!t) return NULL; |
| 2695 | if (!t->is_con()) return NULL; |
| 2696 | intptr_t raw_count = t->get_con(); |
| 2697 | intptr_t size = raw_count; |
| 2698 | if (!Matcher::init_array_count_is_in_bytes) size *= unit; |
| 2699 | // Clearing nothing uses the Identity call. |
| 2700 | // Negative clears are possible on dead ClearArrays |
| 2701 | // (see jck test stmt114.stmt11402.val). |
| 2702 | if (size <= 0 || size % unit != 0) return NULL; |
| 2703 | intptr_t count = size / unit; |
| 2704 | // Length too long; use fast hardware clear |
| 2705 | if (size > Matcher::init_array_short_size) return NULL; |
| 2706 | Node *mem = in(1); |
| 2707 | if( phase->type(mem)==Type::TOP ) return NULL; |
| 2708 | Node *adr = in(3); |
| 2709 | const Type* at = phase->type(adr); |
| 2710 | if( at==Type::TOP ) return NULL; |
| 2711 | const TypePtr* atp = at->isa_ptr(); |
| 2712 | // adjust atp to be the correct array element address type |
| 2713 | if (atp == NULL) atp = TypePtr::BOTTOM; |
| 2714 | else atp = atp->add_offset(Type::OffsetBot); |
| 2715 | // Get base for derived pointer purposes |
| 2716 | if( adr->Opcode() != Op_AddP ) Unimplemented(); |
| 2717 | Node *base = adr->in(1); |
| 2718 | |
| 2719 | Node *zero = phase->makecon(TypeLong::ZERO); |
| 2720 | Node *off = phase->MakeConX(BytesPerLong); |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2721 | mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2722 | count--; |
| 2723 | while( count-- ) { |
| 2724 | mem = phase->transform(mem); |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2725 | adr = phase->transform(new (phase->C) AddPNode(base,adr,off)); |
| 2726 | mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2727 | } |
| 2728 | return mem; |
| 2729 | } |
| 2730 | |
Vladimir Kozlov | 9f5ca02 | 2009-12-09 16:40:45 -0800 | [diff] [blame] | 2731 | //----------------------------step_through---------------------------------- |
| 2732 | // Return allocation input memory edge if it is different instance |
| 2733 | // or itself if it is the one we are looking for. |
| 2734 | bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { |
| 2735 | Node* n = *np; |
| 2736 | assert(n->is_ClearArray(), "sanity"); |
| 2737 | intptr_t offset; |
| 2738 | AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); |
| 2739 | // This method is called only before Allocate nodes are expanded during |
| 2740 | // macro nodes expansion. Before that ClearArray nodes are only generated |
| 2741 | // in LibraryCallKit::generate_arraycopy() which follows allocations. |
| 2742 | assert(alloc != NULL, "should have allocation"); |
| 2743 | if (alloc->_idx == instance_id) { |
| 2744 | // Can not bypass initialization of the instance we are looking for. |
| 2745 | return false; |
| 2746 | } |
| 2747 | // Otherwise skip it. |
| 2748 | InitializeNode* init = alloc->initialization(); |
| 2749 | if (init != NULL) |
| 2750 | *np = init->in(TypeFunc::Memory); |
| 2751 | else |
| 2752 | *np = alloc->in(TypeFunc::Memory); |
| 2753 | return true; |
| 2754 | } |
| 2755 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2756 | //----------------------------clear_memory------------------------------------- |
| 2757 | // Generate code to initialize object storage to zero. |
| 2758 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
| 2759 | intptr_t start_offset, |
| 2760 | Node* end_offset, |
| 2761 | PhaseGVN* phase) { |
| 2762 | Compile* C = phase->C; |
| 2763 | intptr_t offset = start_offset; |
| 2764 | |
| 2765 | int unit = BytesPerLong; |
| 2766 | if ((offset % unit) != 0) { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2767 | Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset)); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2768 | adr = phase->transform(adr); |
| 2769 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 2770 | mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2771 | mem = phase->transform(mem); |
| 2772 | offset += BytesPerInt; |
| 2773 | } |
| 2774 | assert((offset % unit) == 0, ""); |
| 2775 | |
| 2776 | // Initialize the remaining stuff, if any, with a ClearArray. |
| 2777 | return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); |
| 2778 | } |
| 2779 | |
| 2780 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
| 2781 | Node* start_offset, |
| 2782 | Node* end_offset, |
| 2783 | PhaseGVN* phase) { |
Tom Rodriguez | 8e1624d | 2008-03-18 23:44:46 -0700 | [diff] [blame] | 2784 | if (start_offset == end_offset) { |
| 2785 | // nothing to do |
| 2786 | return mem; |
| 2787 | } |
| 2788 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2789 | Compile* C = phase->C; |
| 2790 | int unit = BytesPerLong; |
| 2791 | Node* zbase = start_offset; |
| 2792 | Node* zend = end_offset; |
| 2793 | |
| 2794 | // Scale to the unit required by the CPU: |
| 2795 | if (!Matcher::init_array_count_is_in_bytes) { |
| 2796 | Node* shift = phase->intcon(exact_log2(unit)); |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2797 | zbase = phase->transform( new(C) URShiftXNode(zbase, shift) ); |
| 2798 | zend = phase->transform( new(C) URShiftXNode(zend, shift) ); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2799 | } |
| 2800 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2801 | // Bulk clear double-words |
Vladimir Kozlov | cfcd28f | 2013-01-03 15:09:55 -0800 | [diff] [blame] | 2802 | Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) ); |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2803 | Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) ); |
| 2804 | mem = new (C) ClearArrayNode(ctl, mem, zsize, adr); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2805 | return phase->transform(mem); |
| 2806 | } |
| 2807 | |
| 2808 | Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, |
| 2809 | intptr_t start_offset, |
| 2810 | intptr_t end_offset, |
| 2811 | PhaseGVN* phase) { |
Tom Rodriguez | 8e1624d | 2008-03-18 23:44:46 -0700 | [diff] [blame] | 2812 | if (start_offset == end_offset) { |
| 2813 | // nothing to do |
| 2814 | return mem; |
| 2815 | } |
| 2816 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2817 | Compile* C = phase->C; |
| 2818 | assert((end_offset % BytesPerInt) == 0, "odd end offset"); |
| 2819 | intptr_t done_offset = end_offset; |
| 2820 | if ((done_offset % BytesPerLong) != 0) { |
| 2821 | done_offset -= BytesPerInt; |
| 2822 | } |
| 2823 | if (done_offset > start_offset) { |
| 2824 | mem = clear_memory(ctl, mem, dest, |
| 2825 | start_offset, phase->MakeConX(done_offset), phase); |
| 2826 | } |
| 2827 | if (done_offset < end_offset) { // emit the final 32-bit store |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2828 | Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset)); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2829 | adr = phase->transform(adr); |
| 2830 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 2831 | mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2832 | mem = phase->transform(mem); |
| 2833 | done_offset += BytesPerInt; |
| 2834 | } |
| 2835 | assert(done_offset == end_offset, ""); |
| 2836 | return mem; |
| 2837 | } |
| 2838 | |
| 2839 | //============================================================================= |
Vladimir Kozlov | 3ecb9f5 | 2011-03-30 12:08:49 -0700 | [diff] [blame] | 2840 | // Do not match memory edge. |
| 2841 | uint StrIntrinsicNode::match_edge(uint idx) const { |
| 2842 | return idx == 2 || idx == 3; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2843 | } |
| 2844 | |
| 2845 | //------------------------------Ideal------------------------------------------ |
| 2846 | // Return a node which is more "ideal" than the current node. Strip out |
| 2847 | // control copies |
Vladimir Kozlov | 3ecb9f5 | 2011-03-30 12:08:49 -0700 | [diff] [blame] | 2848 | Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
| 2849 | if (remove_dead_region(phase, can_reshape)) return this; |
Vladimir Kozlov | 8d2ee23 | 2011-11-16 09:13:57 -0800 | [diff] [blame] | 2850 | // Don't bother trying to transform a dead node |
| 2851 | if (in(0) && in(0)->is_top()) return NULL; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2852 | |
Vladimir Kozlov | 0493606 | 2011-03-31 13:22:34 -0700 | [diff] [blame] | 2853 | if (can_reshape) { |
| 2854 | Node* mem = phase->transform(in(MemNode::Memory)); |
| 2855 | // If transformed to a MergeMem, get the desired slice |
| 2856 | uint alias_idx = phase->C->get_alias_index(adr_type()); |
| 2857 | mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; |
| 2858 | if (mem != in(MemNode::Memory)) { |
| 2859 | set_req(MemNode::Memory, mem); |
| 2860 | return this; |
| 2861 | } |
| 2862 | } |
Vladimir Kozlov | 3ecb9f5 | 2011-03-30 12:08:49 -0700 | [diff] [blame] | 2863 | return NULL; |
Chuck Rasbold | 2e672ac | 2008-05-29 16:22:09 -0700 | [diff] [blame] | 2864 | } |
| 2865 | |
Vladimir Kozlov | 8d2ee23 | 2011-11-16 09:13:57 -0800 | [diff] [blame] | 2866 | //------------------------------Value------------------------------------------ |
| 2867 | const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { |
| 2868 | if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; |
| 2869 | return bottom_type(); |
| 2870 | } |
| 2871 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2872 | //============================================================================= |
Vladimir Kozlov | 9e0c61f | 2013-01-22 15:34:16 -0800 | [diff] [blame] | 2873 | //------------------------------match_edge------------------------------------- |
| 2874 | // Do not match memory edge |
| 2875 | uint EncodeISOArrayNode::match_edge(uint idx) const { |
| 2876 | return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len) |
| 2877 | } |
| 2878 | |
| 2879 | //------------------------------Ideal------------------------------------------ |
| 2880 | // Return a node which is more "ideal" than the current node. Strip out |
| 2881 | // control copies |
| 2882 | Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
| 2883 | return remove_dead_region(phase, can_reshape) ? this : NULL; |
| 2884 | } |
| 2885 | |
| 2886 | //------------------------------Value------------------------------------------ |
| 2887 | const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const { |
| 2888 | if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; |
| 2889 | return bottom_type(); |
| 2890 | } |
| 2891 | |
| 2892 | //============================================================================= |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2893 | MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) |
| 2894 | : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), |
| 2895 | _adr_type(C->get_adr_type(alias_idx)) |
| 2896 | { |
| 2897 | init_class_id(Class_MemBar); |
| 2898 | Node* top = C->top(); |
| 2899 | init_req(TypeFunc::I_O,top); |
| 2900 | init_req(TypeFunc::FramePtr,top); |
| 2901 | init_req(TypeFunc::ReturnAdr,top); |
| 2902 | if (precedent != NULL) |
| 2903 | init_req(TypeFunc::Parms, precedent); |
| 2904 | } |
| 2905 | |
| 2906 | //------------------------------cmp-------------------------------------------- |
| 2907 | uint MemBarNode::hash() const { return NO_HASH; } |
| 2908 | uint MemBarNode::cmp( const Node &n ) const { |
| 2909 | return (&n == this); // Always fail except on self |
| 2910 | } |
| 2911 | |
| 2912 | //------------------------------make------------------------------------------- |
| 2913 | MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2914 | switch (opcode) { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2915 | case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn); |
| 2916 | case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn); |
| 2917 | case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn); |
| 2918 | case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn); |
| 2919 | case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn); |
| 2920 | case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn); |
| 2921 | case Op_Initialize: return new(C) InitializeNode(C, atp, pn); |
| 2922 | case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2923 | default: ShouldNotReachHere(); return NULL; |
| 2924 | } |
| 2925 | } |
| 2926 | |
| 2927 | //------------------------------Ideal------------------------------------------ |
| 2928 | // Return a node which is more "ideal" than the current node. Strip out |
| 2929 | // control copies |
| 2930 | Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
Vladimir Kozlov | 9f5ca02 | 2009-12-09 16:40:45 -0800 | [diff] [blame] | 2931 | if (remove_dead_region(phase, can_reshape)) return this; |
Vladimir Kozlov | 8d2ee23 | 2011-11-16 09:13:57 -0800 | [diff] [blame] | 2932 | // Don't bother trying to transform a dead node |
| 2933 | if (in(0) && in(0)->is_top()) return NULL; |
Vladimir Kozlov | 9f5ca02 | 2009-12-09 16:40:45 -0800 | [diff] [blame] | 2934 | |
| 2935 | // Eliminate volatile MemBars for scalar replaced objects. |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 2936 | if (can_reshape && req() == (Precedent+1)) { |
| 2937 | bool eliminate = false; |
| 2938 | int opc = Opcode(); |
| 2939 | if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { |
| 2940 | // Volatile field loads and stores. |
| 2941 | Node* my_mem = in(MemBarNode::Precedent); |
| 2942 | if (my_mem != NULL && my_mem->is_Mem()) { |
| 2943 | const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); |
| 2944 | // Check for scalar replaced object reference. |
| 2945 | if( t_oop != NULL && t_oop->is_known_instance_field() && |
| 2946 | t_oop->offset() != Type::OffsetBot && |
| 2947 | t_oop->offset() != Type::OffsetTop) { |
| 2948 | eliminate = true; |
| 2949 | } |
Vladimir Kozlov | 9f5ca02 | 2009-12-09 16:40:45 -0800 | [diff] [blame] | 2950 | } |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 2951 | } else if (opc == Op_MemBarRelease) { |
| 2952 | // Final field stores. |
| 2953 | Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); |
| 2954 | if ((alloc != NULL) && alloc->is_Allocate() && |
| 2955 | alloc->as_Allocate()->_is_non_escaping) { |
| 2956 | // The allocated object does not escape. |
| 2957 | eliminate = true; |
| 2958 | } |
| 2959 | } |
| 2960 | if (eliminate) { |
| 2961 | // Replace MemBar projections by its inputs. |
| 2962 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
| 2963 | igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); |
| 2964 | igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); |
| 2965 | // Must return either the original node (now dead) or a new node |
| 2966 | // (Do not return a top here, since that would break the uniqueness of top.) |
| 2967 | return new (phase->C) ConINode(TypeInt::ZERO); |
Vladimir Kozlov | 9f5ca02 | 2009-12-09 16:40:45 -0800 | [diff] [blame] | 2968 | } |
| 2969 | } |
| 2970 | return NULL; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2971 | } |
| 2972 | |
| 2973 | //------------------------------Value------------------------------------------ |
| 2974 | const Type *MemBarNode::Value( PhaseTransform *phase ) const { |
| 2975 | if( !in(0) ) return Type::TOP; |
| 2976 | if( phase->type(in(0)) == Type::TOP ) |
| 2977 | return Type::TOP; |
| 2978 | return TypeTuple::MEMBAR; |
| 2979 | } |
| 2980 | |
| 2981 | //------------------------------match------------------------------------------ |
| 2982 | // Construct projections for memory. |
| 2983 | Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { |
| 2984 | switch (proj->_con) { |
| 2985 | case TypeFunc::Control: |
| 2986 | case TypeFunc::Memory: |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 2987 | return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 2988 | } |
| 2989 | ShouldNotReachHere(); |
| 2990 | return NULL; |
| 2991 | } |
| 2992 | |
| 2993 | //===========================InitializeNode==================================== |
| 2994 | // SUMMARY: |
| 2995 | // This node acts as a memory barrier on raw memory, after some raw stores. |
| 2996 | // The 'cooked' oop value feeds from the Initialize, not the Allocation. |
| 2997 | // The Initialize can 'capture' suitably constrained stores as raw inits. |
| 2998 | // It can coalesce related raw stores into larger units (called 'tiles'). |
| 2999 | // It can avoid zeroing new storage for memory units which have raw inits. |
| 3000 | // At macro-expansion, it is marked 'complete', and does not optimize further. |
| 3001 | // |
| 3002 | // EXAMPLE: |
| 3003 | // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. |
| 3004 | // ctl = incoming control; mem* = incoming memory |
| 3005 | // (Note: A star * on a memory edge denotes I/O and other standard edges.) |
| 3006 | // First allocate uninitialized memory and fill in the header: |
| 3007 | // alloc = (Allocate ctl mem* 16 #short[].klass ...) |
| 3008 | // ctl := alloc.Control; mem* := alloc.Memory* |
| 3009 | // rawmem = alloc.Memory; rawoop = alloc.RawAddress |
| 3010 | // Then initialize to zero the non-header parts of the raw memory block: |
| 3011 | // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) |
| 3012 | // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory |
| 3013 | // After the initialize node executes, the object is ready for service: |
| 3014 | // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) |
| 3015 | // Suppose its body is immediately initialized as {1,2}: |
| 3016 | // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) |
| 3017 | // store2 = (StoreC init.Control store1 (+ oop 14) 2) |
| 3018 | // mem.SLICE(#short[*]) := store2 |
| 3019 | // |
| 3020 | // DETAILS: |
| 3021 | // An InitializeNode collects and isolates object initialization after |
| 3022 | // an AllocateNode and before the next possible safepoint. As a |
| 3023 | // memory barrier (MemBarNode), it keeps critical stores from drifting |
| 3024 | // down past any safepoint or any publication of the allocation. |
| 3025 | // Before this barrier, a newly-allocated object may have uninitialized bits. |
| 3026 | // After this barrier, it may be treated as a real oop, and GC is allowed. |
| 3027 | // |
| 3028 | // The semantics of the InitializeNode include an implicit zeroing of |
| 3029 | // the new object from object header to the end of the object. |
| 3030 | // (The object header and end are determined by the AllocateNode.) |
| 3031 | // |
| 3032 | // Certain stores may be added as direct inputs to the InitializeNode. |
| 3033 | // These stores must update raw memory, and they must be to addresses |
| 3034 | // derived from the raw address produced by AllocateNode, and with |
| 3035 | // a constant offset. They must be ordered by increasing offset. |
| 3036 | // The first one is at in(RawStores), the last at in(req()-1). |
| 3037 | // Unlike most memory operations, they are not linked in a chain, |
| 3038 | // but are displayed in parallel as users of the rawmem output of |
| 3039 | // the allocation. |
| 3040 | // |
| 3041 | // (See comments in InitializeNode::capture_store, which continue |
| 3042 | // the example given above.) |
| 3043 | // |
| 3044 | // When the associated Allocate is macro-expanded, the InitializeNode |
| 3045 | // may be rewritten to optimize collected stores. A ClearArrayNode |
| 3046 | // may also be created at that point to represent any required zeroing. |
| 3047 | // The InitializeNode is then marked 'complete', prohibiting further |
| 3048 | // capturing of nearby memory operations. |
| 3049 | // |
| 3050 | // During macro-expansion, all captured initializations which store |
Christian Thalinger | 05d1de7 | 2009-02-27 13:27:09 -0800 | [diff] [blame] | 3051 | // constant values of 32 bits or smaller are coalesced (if advantageous) |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3052 | // into larger 'tiles' 32 or 64 bits. This allows an object to be |
| 3053 | // initialized in fewer memory operations. Memory words which are |
| 3054 | // covered by neither tiles nor non-constant stores are pre-zeroed |
| 3055 | // by explicit stores of zero. (The code shape happens to do all |
| 3056 | // zeroing first, then all other stores, with both sequences occurring |
| 3057 | // in order of ascending offsets.) |
| 3058 | // |
| 3059 | // Alternatively, code may be inserted between an AllocateNode and its |
| 3060 | // InitializeNode, to perform arbitrary initialization of the new object. |
| 3061 | // E.g., the object copying intrinsics insert complex data transfers here. |
| 3062 | // The initialization must then be marked as 'complete' disable the |
| 3063 | // built-in zeroing semantics and the collection of initializing stores. |
| 3064 | // |
| 3065 | // While an InitializeNode is incomplete, reads from the memory state |
| 3066 | // produced by it are optimizable if they match the control edge and |
| 3067 | // new oop address associated with the allocation/initialization. |
| 3068 | // They return a stored value (if the offset matches) or else zero. |
| 3069 | // A write to the memory state, if it matches control and address, |
| 3070 | // and if it is to a constant offset, may be 'captured' by the |
| 3071 | // InitializeNode. It is cloned as a raw memory operation and rewired |
| 3072 | // inside the initialization, to the raw oop produced by the allocation. |
| 3073 | // Operations on addresses which are provably distinct (e.g., to |
| 3074 | // other AllocateNodes) are allowed to bypass the initialization. |
| 3075 | // |
| 3076 | // The effect of all this is to consolidate object initialization |
| 3077 | // (both arrays and non-arrays, both piecewise and bulk) into a |
| 3078 | // single location, where it can be optimized as a unit. |
| 3079 | // |
| 3080 | // Only stores with an offset less than TrackedInitializationLimit words |
| 3081 | // will be considered for capture by an InitializeNode. This puts a |
| 3082 | // reasonable limit on the complexity of optimized initializations. |
| 3083 | |
| 3084 | //---------------------------InitializeNode------------------------------------ |
| 3085 | InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) |
Roland Westrelin | 97439fb | 2011-12-20 16:56:50 +0100 | [diff] [blame] | 3086 | : _is_complete(Incomplete), _does_not_escape(false), |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3087 | MemBarNode(C, adr_type, rawoop) |
| 3088 | { |
| 3089 | init_class_id(Class_Initialize); |
| 3090 | |
| 3091 | assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); |
| 3092 | assert(in(RawAddress) == rawoop, "proper init"); |
| 3093 | // Note: allocation() can be NULL, for secondary initialization barriers |
| 3094 | } |
| 3095 | |
| 3096 | // Since this node is not matched, it will be processed by the |
| 3097 | // register allocator. Declare that there are no constraints |
| 3098 | // on the allocation of the RawAddress edge. |
| 3099 | const RegMask &InitializeNode::in_RegMask(uint idx) const { |
| 3100 | // This edge should be set to top, by the set_complete. But be conservative. |
| 3101 | if (idx == InitializeNode::RawAddress) |
| 3102 | return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); |
| 3103 | return RegMask::Empty; |
| 3104 | } |
| 3105 | |
| 3106 | Node* InitializeNode::memory(uint alias_idx) { |
| 3107 | Node* mem = in(Memory); |
| 3108 | if (mem->is_MergeMem()) { |
| 3109 | return mem->as_MergeMem()->memory_at(alias_idx); |
| 3110 | } else { |
| 3111 | // incoming raw memory is not split |
| 3112 | return mem; |
| 3113 | } |
| 3114 | } |
| 3115 | |
| 3116 | bool InitializeNode::is_non_zero() { |
| 3117 | if (is_complete()) return false; |
| 3118 | remove_extra_zeroes(); |
| 3119 | return (req() > RawStores); |
| 3120 | } |
| 3121 | |
| 3122 | void InitializeNode::set_complete(PhaseGVN* phase) { |
| 3123 | assert(!is_complete(), "caller responsibility"); |
Vladimir Kozlov | 47e357e | 2011-09-26 10:24:05 -0700 | [diff] [blame] | 3124 | _is_complete = Complete; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3125 | |
| 3126 | // After this node is complete, it contains a bunch of |
| 3127 | // raw-memory initializations. There is no need for |
| 3128 | // it to have anything to do with non-raw memory effects. |
| 3129 | // Therefore, tell all non-raw users to re-optimize themselves, |
| 3130 | // after skipping the memory effects of this initialization. |
| 3131 | PhaseIterGVN* igvn = phase->is_IterGVN(); |
| 3132 | if (igvn) igvn->add_users_to_worklist(this); |
| 3133 | } |
| 3134 | |
| 3135 | // convenience function |
| 3136 | // return false if the init contains any stores already |
| 3137 | bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { |
| 3138 | InitializeNode* init = initialization(); |
| 3139 | if (init == NULL || init->is_complete()) return false; |
| 3140 | init->remove_extra_zeroes(); |
| 3141 | // for now, if this allocation has already collected any inits, bail: |
| 3142 | if (init->is_non_zero()) return false; |
| 3143 | init->set_complete(phase); |
| 3144 | return true; |
| 3145 | } |
| 3146 | |
| 3147 | void InitializeNode::remove_extra_zeroes() { |
| 3148 | if (req() == RawStores) return; |
| 3149 | Node* zmem = zero_memory(); |
| 3150 | uint fill = RawStores; |
| 3151 | for (uint i = fill; i < req(); i++) { |
| 3152 | Node* n = in(i); |
| 3153 | if (n->is_top() || n == zmem) continue; // skip |
| 3154 | if (fill < i) set_req(fill, n); // compact |
| 3155 | ++fill; |
| 3156 | } |
| 3157 | // delete any empty spaces created: |
| 3158 | while (fill < req()) { |
| 3159 | del_req(fill); |
| 3160 | } |
| 3161 | } |
| 3162 | |
| 3163 | // Helper for remembering which stores go with which offsets. |
| 3164 | intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { |
| 3165 | if (!st->is_Store()) return -1; // can happen to dead code via subsume_node |
| 3166 | intptr_t offset = -1; |
| 3167 | Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), |
| 3168 | phase, offset); |
| 3169 | if (base == NULL) return -1; // something is dead, |
| 3170 | if (offset < 0) return -1; // dead, dead |
| 3171 | return offset; |
| 3172 | } |
| 3173 | |
| 3174 | // Helper for proving that an initialization expression is |
| 3175 | // "simple enough" to be folded into an object initialization. |
| 3176 | // Attempts to prove that a store's initial value 'n' can be captured |
| 3177 | // within the initialization without creating a vicious cycle, such as: |
| 3178 | // { Foo p = new Foo(); p.next = p; } |
| 3179 | // True for constants and parameters and small combinations thereof. |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 3180 | bool InitializeNode::detect_init_independence(Node* n, int& count) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3181 | if (n == NULL) return true; // (can this really happen?) |
| 3182 | if (n->is_Proj()) n = n->in(0); |
| 3183 | if (n == this) return false; // found a cycle |
| 3184 | if (n->is_Con()) return true; |
| 3185 | if (n->is_Start()) return true; // params, etc., are OK |
| 3186 | if (n->is_Root()) return true; // even better |
| 3187 | |
| 3188 | Node* ctl = n->in(0); |
| 3189 | if (ctl != NULL && !ctl->is_top()) { |
| 3190 | if (ctl->is_Proj()) ctl = ctl->in(0); |
| 3191 | if (ctl == this) return false; |
| 3192 | |
| 3193 | // If we already know that the enclosing memory op is pinned right after |
| 3194 | // the init, then any control flow that the store has picked up |
| 3195 | // must have preceded the init, or else be equal to the init. |
| 3196 | // Even after loop optimizations (which might change control edges) |
| 3197 | // a store is never pinned *before* the availability of its inputs. |
Vladimir Kozlov | df8fc19 | 2008-04-16 19:19:48 -0700 | [diff] [blame] | 3198 | if (!MemNode::all_controls_dominate(n, this)) |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3199 | return false; // failed to prove a good control |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3200 | } |
| 3201 | |
| 3202 | // Check data edges for possible dependencies on 'this'. |
| 3203 | if ((count += 1) > 20) return false; // complexity limit |
| 3204 | for (uint i = 1; i < n->req(); i++) { |
| 3205 | Node* m = n->in(i); |
| 3206 | if (m == NULL || m == n || m->is_top()) continue; |
| 3207 | uint first_i = n->find_edge(m); |
| 3208 | if (i != first_i) continue; // process duplicate edge just once |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 3209 | if (!detect_init_independence(m, count)) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3210 | return false; |
| 3211 | } |
| 3212 | } |
| 3213 | |
| 3214 | return true; |
| 3215 | } |
| 3216 | |
| 3217 | // Here are all the checks a Store must pass before it can be moved into |
| 3218 | // an initialization. Returns zero if a check fails. |
| 3219 | // On success, returns the (constant) offset to which the store applies, |
| 3220 | // within the initialized memory. |
Roland Westrelin | fe92862 | 2013-02-25 14:13:04 +0100 | [diff] [blame] | 3221 | intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3222 | const int FAIL = 0; |
| 3223 | if (st->req() != MemNode::ValueIn + 1) |
| 3224 | return FAIL; // an inscrutable StoreNode (card mark?) |
| 3225 | Node* ctl = st->in(MemNode::Control); |
| 3226 | if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) |
| 3227 | return FAIL; // must be unconditional after the initialization |
| 3228 | Node* mem = st->in(MemNode::Memory); |
| 3229 | if (!(mem->is_Proj() && mem->in(0) == this)) |
| 3230 | return FAIL; // must not be preceded by other stores |
| 3231 | Node* adr = st->in(MemNode::Address); |
| 3232 | intptr_t offset; |
| 3233 | AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); |
| 3234 | if (alloc == NULL) |
| 3235 | return FAIL; // inscrutable address |
| 3236 | if (alloc != allocation()) |
| 3237 | return FAIL; // wrong allocation! (store needs to float up) |
| 3238 | Node* val = st->in(MemNode::ValueIn); |
| 3239 | int complexity_count = 0; |
Vladimir Kozlov | b4977e8 | 2013-05-08 15:08:01 -0700 | [diff] [blame^] | 3240 | if (!detect_init_independence(val, complexity_count)) |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3241 | return FAIL; // stored value must be 'simple enough' |
| 3242 | |
Roland Westrelin | fe92862 | 2013-02-25 14:13:04 +0100 | [diff] [blame] | 3243 | // The Store can be captured only if nothing after the allocation |
| 3244 | // and before the Store is using the memory location that the store |
| 3245 | // overwrites. |
| 3246 | bool failed = false; |
| 3247 | // If is_complete_with_arraycopy() is true the shape of the graph is |
| 3248 | // well defined and is safe so no need for extra checks. |
| 3249 | if (!is_complete_with_arraycopy()) { |
| 3250 | // We are going to look at each use of the memory state following |
| 3251 | // the allocation to make sure nothing reads the memory that the |
| 3252 | // Store writes. |
| 3253 | const TypePtr* t_adr = phase->type(adr)->isa_ptr(); |
| 3254 | int alias_idx = phase->C->get_alias_index(t_adr); |
| 3255 | ResourceMark rm; |
| 3256 | Unique_Node_List mems; |
| 3257 | mems.push(mem); |
| 3258 | Node* unique_merge = NULL; |
| 3259 | for (uint next = 0; next < mems.size(); ++next) { |
| 3260 | Node *m = mems.at(next); |
| 3261 | for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { |
| 3262 | Node *n = m->fast_out(j); |
| 3263 | if (n->outcnt() == 0) { |
| 3264 | continue; |
| 3265 | } |
| 3266 | if (n == st) { |
| 3267 | continue; |
| 3268 | } else if (n->in(0) != NULL && n->in(0) != ctl) { |
| 3269 | // If the control of this use is different from the control |
| 3270 | // of the Store which is right after the InitializeNode then |
| 3271 | // this node cannot be between the InitializeNode and the |
| 3272 | // Store. |
| 3273 | continue; |
| 3274 | } else if (n->is_MergeMem()) { |
| 3275 | if (n->as_MergeMem()->memory_at(alias_idx) == m) { |
| 3276 | // We can hit a MergeMemNode (that will likely go away |
| 3277 | // later) that is a direct use of the memory state |
| 3278 | // following the InitializeNode on the same slice as the |
| 3279 | // store node that we'd like to capture. We need to check |
| 3280 | // the uses of the MergeMemNode. |
| 3281 | mems.push(n); |
| 3282 | } |
| 3283 | } else if (n->is_Mem()) { |
| 3284 | Node* other_adr = n->in(MemNode::Address); |
| 3285 | if (other_adr == adr) { |
| 3286 | failed = true; |
| 3287 | break; |
| 3288 | } else { |
| 3289 | const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); |
| 3290 | if (other_t_adr != NULL) { |
| 3291 | int other_alias_idx = phase->C->get_alias_index(other_t_adr); |
| 3292 | if (other_alias_idx == alias_idx) { |
| 3293 | // A load from the same memory slice as the store right |
| 3294 | // after the InitializeNode. We check the control of the |
| 3295 | // object/array that is loaded from. If it's the same as |
| 3296 | // the store control then we cannot capture the store. |
| 3297 | assert(!n->is_Store(), "2 stores to same slice on same control?"); |
| 3298 | Node* base = other_adr; |
| 3299 | assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); |
| 3300 | base = base->in(AddPNode::Base); |
| 3301 | if (base != NULL) { |
| 3302 | base = base->uncast(); |
| 3303 | if (base->is_Proj() && base->in(0) == alloc) { |
| 3304 | failed = true; |
| 3305 | break; |
| 3306 | } |
| 3307 | } |
| 3308 | } |
| 3309 | } |
| 3310 | } |
| 3311 | } else { |
| 3312 | failed = true; |
| 3313 | break; |
| 3314 | } |
| 3315 | } |
| 3316 | } |
| 3317 | } |
| 3318 | if (failed) { |
| 3319 | if (!can_reshape) { |
| 3320 | // We decided we couldn't capture the store during parsing. We |
| 3321 | // should try again during the next IGVN once the graph is |
| 3322 | // cleaner. |
| 3323 | phase->C->record_for_igvn(st); |
| 3324 | } |
| 3325 | return FAIL; |
| 3326 | } |
| 3327 | |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3328 | return offset; // success |
| 3329 | } |
| 3330 | |
| 3331 | // Find the captured store in(i) which corresponds to the range |
| 3332 | // [start..start+size) in the initialized object. |
| 3333 | // If there is one, return its index i. If there isn't, return the |
| 3334 | // negative of the index where it should be inserted. |
| 3335 | // Return 0 if the queried range overlaps an initialization boundary |
| 3336 | // or if dead code is encountered. |
| 3337 | // If size_in_bytes is zero, do not bother with overlap checks. |
| 3338 | int InitializeNode::captured_store_insertion_point(intptr_t start, |
| 3339 | int size_in_bytes, |
| 3340 | PhaseTransform* phase) { |
| 3341 | const int FAIL = 0, MAX_STORE = BytesPerLong; |
| 3342 | |
| 3343 | if (is_complete()) |
| 3344 | return FAIL; // arraycopy got here first; punt |
| 3345 | |
| 3346 | assert(allocation() != NULL, "must be present"); |
| 3347 | |
| 3348 | // no negatives, no header fields: |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 3349 | if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3350 | |
| 3351 | // after a certain size, we bail out on tracking all the stores: |
| 3352 | intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); |
| 3353 | if (start >= ti_limit) return FAIL; |
| 3354 | |
| 3355 | for (uint i = InitializeNode::RawStores, limit = req(); ; ) { |
| 3356 | if (i >= limit) return -(int)i; // not found; here is where to put it |
| 3357 | |
| 3358 | Node* st = in(i); |
| 3359 | intptr_t st_off = get_store_offset(st, phase); |
| 3360 | if (st_off < 0) { |
| 3361 | if (st != zero_memory()) { |
| 3362 | return FAIL; // bail out if there is dead garbage |
| 3363 | } |
| 3364 | } else if (st_off > start) { |
| 3365 | // ...we are done, since stores are ordered |
| 3366 | if (st_off < start + size_in_bytes) { |
| 3367 | return FAIL; // the next store overlaps |
| 3368 | } |
| 3369 | return -(int)i; // not found; here is where to put it |
| 3370 | } else if (st_off < start) { |
| 3371 | if (size_in_bytes != 0 && |
| 3372 | start < st_off + MAX_STORE && |
| 3373 | start < st_off + st->as_Store()->memory_size()) { |
| 3374 | return FAIL; // the previous store overlaps |
| 3375 | } |
| 3376 | } else { |
| 3377 | if (size_in_bytes != 0 && |
| 3378 | st->as_Store()->memory_size() != size_in_bytes) { |
| 3379 | return FAIL; // mismatched store size |
| 3380 | } |
| 3381 | return i; |
| 3382 | } |
| 3383 | |
| 3384 | ++i; |
| 3385 | } |
| 3386 | } |
| 3387 | |
| 3388 | // Look for a captured store which initializes at the offset 'start' |
| 3389 | // with the given size. If there is no such store, and no other |
| 3390 | // initialization interferes, then return zero_memory (the memory |
| 3391 | // projection of the AllocateNode). |
| 3392 | Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, |
| 3393 | PhaseTransform* phase) { |
| 3394 | assert(stores_are_sane(phase), ""); |
| 3395 | int i = captured_store_insertion_point(start, size_in_bytes, phase); |
| 3396 | if (i == 0) { |
| 3397 | return NULL; // something is dead |
| 3398 | } else if (i < 0) { |
| 3399 | return zero_memory(); // just primordial zero bits here |
| 3400 | } else { |
| 3401 | Node* st = in(i); // here is the store at this position |
| 3402 | assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); |
| 3403 | return st; |
| 3404 | } |
| 3405 | } |
| 3406 | |
| 3407 | // Create, as a raw pointer, an address within my new object at 'offset'. |
| 3408 | Node* InitializeNode::make_raw_address(intptr_t offset, |
| 3409 | PhaseTransform* phase) { |
| 3410 | Node* addr = in(RawAddress); |
| 3411 | if (offset != 0) { |
| 3412 | Compile* C = phase->C; |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 3413 | addr = phase->transform( new (C) AddPNode(C->top(), addr, |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3414 | phase->MakeConX(offset)) ); |
| 3415 | } |
| 3416 | return addr; |
| 3417 | } |
| 3418 | |
| 3419 | // Clone the given store, converting it into a raw store |
| 3420 | // initializing a field or element of my new object. |
| 3421 | // Caller is responsible for retiring the original store, |
| 3422 | // with subsume_node or the like. |
| 3423 | // |
| 3424 | // From the example above InitializeNode::InitializeNode, |
| 3425 | // here are the old stores to be captured: |
| 3426 | // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) |
| 3427 | // store2 = (StoreC init.Control store1 (+ oop 14) 2) |
| 3428 | // |
| 3429 | // Here is the changed code; note the extra edges on init: |
| 3430 | // alloc = (Allocate ...) |
| 3431 | // rawoop = alloc.RawAddress |
| 3432 | // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) |
| 3433 | // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) |
| 3434 | // init = (Initialize alloc.Control alloc.Memory rawoop |
| 3435 | // rawstore1 rawstore2) |
| 3436 | // |
| 3437 | Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, |
Roland Westrelin | fe92862 | 2013-02-25 14:13:04 +0100 | [diff] [blame] | 3438 | PhaseTransform* phase, bool can_reshape) { |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3439 | assert(stores_are_sane(phase), ""); |
| 3440 | |
| 3441 | if (start < 0) return NULL; |
Roland Westrelin | fe92862 | 2013-02-25 14:13:04 +0100 | [diff] [blame] | 3442 | assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3443 | |
| 3444 | Compile* C = phase->C; |
| 3445 | int size_in_bytes = st->memory_size(); |
| 3446 | int i = captured_store_insertion_point(start, size_in_bytes, phase); |
| 3447 | if (i == 0) return NULL; // bail out |
| 3448 | Node* prev_mem = NULL; // raw memory for the captured store |
| 3449 | if (i > 0) { |
| 3450 | prev_mem = in(i); // there is a pre-existing store under this one |
| 3451 | set_req(i, C->top()); // temporarily disconnect it |
| 3452 | // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. |
| 3453 | } else { |
| 3454 | i = -i; // no pre-existing store |
| 3455 | prev_mem = zero_memory(); // a slice of the newly allocated object |
| 3456 | if (i > InitializeNode::RawStores && in(i-1) == prev_mem) |
| 3457 | set_req(--i, C->top()); // reuse this edge; it has been folded away |
| 3458 | else |
| 3459 | ins_req(i, C->top()); // build a new edge |
| 3460 | } |
| 3461 | Node* new_st = st->clone(); |
| 3462 | new_st->set_req(MemNode::Control, in(Control)); |
| 3463 | new_st->set_req(MemNode::Memory, prev_mem); |
| 3464 | new_st->set_req(MemNode::Address, make_raw_address(start, phase)); |
| 3465 | new_st = phase->transform(new_st); |
| 3466 | |
| 3467 | // At this point, new_st might have swallowed a pre-existing store |
| 3468 | // at the same offset, or perhaps new_st might have disappeared, |
| 3469 | // if it redundantly stored the same value (or zero to fresh memory). |
| 3470 | |
| 3471 | // In any case, wire it in: |
| 3472 | set_req(i, new_st); |
| 3473 | |
| 3474 | // The caller may now kill the old guy. |
| 3475 | DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); |
| 3476 | assert(check_st == new_st || check_st == NULL, "must be findable"); |
| 3477 | assert(!is_complete(), ""); |
| 3478 | return new_st; |
| 3479 | } |
| 3480 | |
| 3481 | static bool store_constant(jlong* tiles, int num_tiles, |
| 3482 | intptr_t st_off, int st_size, |
| 3483 | jlong con) { |
| 3484 | if ((st_off & (st_size-1)) != 0) |
| 3485 | return false; // strange store offset (assume size==2**N) |
| 3486 | address addr = (address)tiles + st_off; |
| 3487 | assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); |
| 3488 | switch (st_size) { |
| 3489 | case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; |
| 3490 | case sizeof(jchar): *(jchar*) addr = (jchar) con; break; |
| 3491 | case sizeof(jint): *(jint*) addr = (jint) con; break; |
| 3492 | case sizeof(jlong): *(jlong*) addr = (jlong) con; break; |
| 3493 | default: return false; // strange store size (detect size!=2**N here) |
| 3494 | } |
| 3495 | return true; // return success to caller |
| 3496 | } |
| 3497 | |
| 3498 | // Coalesce subword constants into int constants and possibly |
| 3499 | // into long constants. The goal, if the CPU permits, |
| 3500 | // is to initialize the object with a small number of 64-bit tiles. |
| 3501 | // Also, convert floating-point constants to bit patterns. |
| 3502 | // Non-constants are not relevant to this pass. |
| 3503 | // |
| 3504 | // In terms of the running example on InitializeNode::InitializeNode |
| 3505 | // and InitializeNode::capture_store, here is the transformation |
| 3506 | // of rawstore1 and rawstore2 into rawstore12: |
| 3507 | // alloc = (Allocate ...) |
| 3508 | // rawoop = alloc.RawAddress |
| 3509 | // tile12 = 0x00010002 |
| 3510 | // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) |
| 3511 | // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) |
| 3512 | // |
| 3513 | void |
| 3514 | InitializeNode::coalesce_subword_stores(intptr_t header_size, |
| 3515 | Node* size_in_bytes, |
| 3516 | PhaseGVN* phase) { |
| 3517 | Compile* C = phase->C; |
| 3518 | |
| 3519 | assert(stores_are_sane(phase), ""); |
| 3520 | // Note: After this pass, they are not completely sane, |
| 3521 | // since there may be some overlaps. |
| 3522 | |
| 3523 | int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; |
| 3524 | |
| 3525 | intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); |
| 3526 | intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); |
| 3527 | size_limit = MIN2(size_limit, ti_limit); |
| 3528 | size_limit = align_size_up(size_limit, BytesPerLong); |
| 3529 | int num_tiles = size_limit / BytesPerLong; |
| 3530 | |
| 3531 | // allocate space for the tile map: |
| 3532 | const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small |
| 3533 | jlong tiles_buf[small_len]; |
| 3534 | Node* nodes_buf[small_len]; |
| 3535 | jlong inits_buf[small_len]; |
| 3536 | jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] |
| 3537 | : NEW_RESOURCE_ARRAY(jlong, num_tiles)); |
| 3538 | Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] |
| 3539 | : NEW_RESOURCE_ARRAY(Node*, num_tiles)); |
| 3540 | jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] |
| 3541 | : NEW_RESOURCE_ARRAY(jlong, num_tiles)); |
| 3542 | // tiles: exact bitwise model of all primitive constants |
| 3543 | // nodes: last constant-storing node subsumed into the tiles model |
| 3544 | // inits: which bytes (in each tile) are touched by any initializations |
| 3545 | |
| 3546 | //// Pass A: Fill in the tile model with any relevant stores. |
| 3547 | |
| 3548 | Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); |
| 3549 | Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); |
| 3550 | Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); |
| 3551 | Node* zmem = zero_memory(); // initially zero memory state |
| 3552 | for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { |
| 3553 | Node* st = in(i); |
| 3554 | intptr_t st_off = get_store_offset(st, phase); |
| 3555 | |
| 3556 | // Figure out the store's offset and constant value: |
| 3557 | if (st_off < header_size) continue; //skip (ignore header) |
| 3558 | if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) |
| 3559 | int st_size = st->as_Store()->memory_size(); |
| 3560 | if (st_off + st_size > size_limit) break; |
| 3561 | |
| 3562 | // Record which bytes are touched, whether by constant or not. |
| 3563 | if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) |
| 3564 | continue; // skip (strange store size) |
| 3565 | |
| 3566 | const Type* val = phase->type(st->in(MemNode::ValueIn)); |
| 3567 | if (!val->singleton()) continue; //skip (non-con store) |
| 3568 | BasicType type = val->basic_type(); |
| 3569 | |
| 3570 | jlong con = 0; |
| 3571 | switch (type) { |
| 3572 | case T_INT: con = val->is_int()->get_con(); break; |
| 3573 | case T_LONG: con = val->is_long()->get_con(); break; |
| 3574 | case T_FLOAT: con = jint_cast(val->getf()); break; |
| 3575 | case T_DOUBLE: con = jlong_cast(val->getd()); break; |
| 3576 | default: continue; //skip (odd store type) |
| 3577 | } |
| 3578 | |
| 3579 | if (type == T_LONG && Matcher::isSimpleConstant64(con) && |
| 3580 | st->Opcode() == Op_StoreL) { |
| 3581 | continue; // This StoreL is already optimal. |
| 3582 | } |
| 3583 | |
| 3584 | // Store down the constant. |
| 3585 | store_constant(tiles, num_tiles, st_off, st_size, con); |
| 3586 | |
| 3587 | intptr_t j = st_off >> LogBytesPerLong; |
| 3588 | |
| 3589 | if (type == T_INT && st_size == BytesPerInt |
| 3590 | && (st_off & BytesPerInt) == BytesPerInt) { |
| 3591 | jlong lcon = tiles[j]; |
| 3592 | if (!Matcher::isSimpleConstant64(lcon) && |
| 3593 | st->Opcode() == Op_StoreI) { |
| 3594 | // This StoreI is already optimal by itself. |
| 3595 | jint* intcon = (jint*) &tiles[j]; |
| 3596 | intcon[1] = 0; // undo the store_constant() |
| 3597 | |
| 3598 | // If the previous store is also optimal by itself, back up and |
| 3599 | // undo the action of the previous loop iteration... if we can. |
| 3600 | // But if we can't, just let the previous half take care of itself. |
| 3601 | st = nodes[j]; |
| 3602 | st_off -= BytesPerInt; |
| 3603 | con = intcon[0]; |
| 3604 | if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { |
| 3605 | assert(st_off >= header_size, "still ignoring header"); |
| 3606 | assert(get_store_offset(st, phase) == st_off, "must be"); |
| 3607 | assert(in(i-1) == zmem, "must be"); |
| 3608 | DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); |
| 3609 | assert(con == tcon->is_int()->get_con(), "must be"); |
| 3610 | // Undo the effects of the previous loop trip, which swallowed st: |
| 3611 | intcon[0] = 0; // undo store_constant() |
| 3612 | set_req(i-1, st); // undo set_req(i, zmem) |
| 3613 | nodes[j] = NULL; // undo nodes[j] = st |
| 3614 | --old_subword; // undo ++old_subword |
| 3615 | } |
| 3616 | continue; // This StoreI is already optimal. |
| 3617 | } |
| 3618 | } |
| 3619 | |
| 3620 | // This store is not needed. |
| 3621 | set_req(i, zmem); |
| 3622 | nodes[j] = st; // record for the moment |
| 3623 | if (st_size < BytesPerLong) // something has changed |
| 3624 | ++old_subword; // includes int/float, but who's counting... |
| 3625 | else ++old_long; |
| 3626 | } |
| 3627 | |
| 3628 | if ((old_subword + old_long) == 0) |
| 3629 | return; // nothing more to do |
| 3630 | |
| 3631 | //// Pass B: Convert any non-zero tiles into optimal constant stores. |
| 3632 | // Be sure to insert them before overlapping non-constant stores. |
| 3633 | // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) |
| 3634 | for (int j = 0; j < num_tiles; j++) { |
| 3635 | jlong con = tiles[j]; |
| 3636 | jlong init = inits[j]; |
| 3637 | if (con == 0) continue; |
| 3638 | jint con0, con1; // split the constant, address-wise |
| 3639 | jint init0, init1; // split the init map, address-wise |
| 3640 | { union { jlong con; jint intcon[2]; } u; |
| 3641 | u.con = con; |
| 3642 | con0 = u.intcon[0]; |
| 3643 | con1 = u.intcon[1]; |
| 3644 | u.con = init; |
| 3645 | init0 = u.intcon[0]; |
| 3646 | init1 = u.intcon[1]; |
| 3647 | } |
| 3648 | |
| 3649 | Node* old = nodes[j]; |
| 3650 | assert(old != NULL, "need the prior store"); |
| 3651 | intptr_t offset = (j * BytesPerLong); |
| 3652 | |
| 3653 | bool split = !Matcher::isSimpleConstant64(con); |
| 3654 | |
| 3655 | if (offset < header_size) { |
| 3656 | assert(offset + BytesPerInt >= header_size, "second int counts"); |
| 3657 | assert(*(jint*)&tiles[j] == 0, "junk in header"); |
| 3658 | split = true; // only the second word counts |
| 3659 | // Example: int a[] = { 42 ... } |
| 3660 | } else if (con0 == 0 && init0 == -1) { |
| 3661 | split = true; // first word is covered by full inits |
| 3662 | // Example: int a[] = { ... foo(), 42 ... } |
| 3663 | } else if (con1 == 0 && init1 == -1) { |
| 3664 | split = true; // second word is covered by full inits |
| 3665 | // Example: int a[] = { ... 42, foo() ... } |
| 3666 | } |
| 3667 | |
| 3668 | // Here's a case where init0 is neither 0 nor -1: |
| 3669 | // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } |
| 3670 | // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. |
| 3671 | // In this case the tile is not split; it is (jlong)42. |
| 3672 | // The big tile is stored down, and then the foo() value is inserted. |
| 3673 | // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) |
| 3674 | |
| 3675 | Node* ctl = old->in(MemNode::Control); |
| 3676 | Node* adr = make_raw_address(offset, phase); |
| 3677 | const TypePtr* atp = TypeRawPtr::BOTTOM; |
| 3678 | |
| 3679 | // One or two coalesced stores to plop down. |
| 3680 | Node* st[2]; |
| 3681 | intptr_t off[2]; |
| 3682 | int nst = 0; |
| 3683 | if (!split) { |
| 3684 | ++new_long; |
| 3685 | off[nst] = offset; |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 3686 | st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3687 | phase->longcon(con), T_LONG); |
| 3688 | } else { |
| 3689 | // Omit either if it is a zero. |
| 3690 | if (con0 != 0) { |
| 3691 | ++new_int; |
| 3692 | off[nst] = offset; |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 3693 | st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3694 | phase->intcon(con0), T_INT); |
| 3695 | } |
| 3696 | if (con1 != 0) { |
| 3697 | ++new_int; |
| 3698 | offset += BytesPerInt; |
| 3699 | adr = make_raw_address(offset, phase); |
| 3700 | off[nst] = offset; |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 3701 | st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3702 | phase->intcon(con1), T_INT); |
| 3703 | } |
| 3704 | } |
| 3705 | |
| 3706 | // Insert second store first, then the first before the second. |
| 3707 | // Insert each one just before any overlapping non-constant stores. |
| 3708 | while (nst > 0) { |
| 3709 | Node* st1 = st[--nst]; |
| 3710 | C->copy_node_notes_to(st1, old); |
| 3711 | st1 = phase->transform(st1); |
| 3712 | offset = off[nst]; |
| 3713 | assert(offset >= header_size, "do not smash header"); |
| 3714 | int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); |
| 3715 | guarantee(ins_idx != 0, "must re-insert constant store"); |
| 3716 | if (ins_idx < 0) ins_idx = -ins_idx; // never overlap |
| 3717 | if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) |
| 3718 | set_req(--ins_idx, st1); |
| 3719 | else |
| 3720 | ins_req(ins_idx, st1); |
| 3721 | } |
| 3722 | } |
| 3723 | |
| 3724 | if (PrintCompilation && WizardMode) |
| 3725 | tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", |
| 3726 | old_subword, old_long, new_int, new_long); |
| 3727 | if (C->log() != NULL) |
| 3728 | C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", |
| 3729 | old_subword, old_long, new_int, new_long); |
| 3730 | |
| 3731 | // Clean up any remaining occurrences of zmem: |
| 3732 | remove_extra_zeroes(); |
| 3733 | } |
| 3734 | |
| 3735 | // Explore forward from in(start) to find the first fully initialized |
| 3736 | // word, and return its offset. Skip groups of subword stores which |
| 3737 | // together initialize full words. If in(start) is itself part of a |
| 3738 | // fully initialized word, return the offset of in(start). If there |
| 3739 | // are no following full-word stores, or if something is fishy, return |
| 3740 | // a negative value. |
| 3741 | intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { |
| 3742 | int int_map = 0; |
| 3743 | intptr_t int_map_off = 0; |
| 3744 | const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for |
| 3745 | |
| 3746 | for (uint i = start, limit = req(); i < limit; i++) { |
| 3747 | Node* st = in(i); |
| 3748 | |
| 3749 | intptr_t st_off = get_store_offset(st, phase); |
| 3750 | if (st_off < 0) break; // return conservative answer |
| 3751 | |
| 3752 | int st_size = st->as_Store()->memory_size(); |
| 3753 | if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { |
| 3754 | return st_off; // we found a complete word init |
| 3755 | } |
| 3756 | |
| 3757 | // update the map: |
| 3758 | |
| 3759 | intptr_t this_int_off = align_size_down(st_off, BytesPerInt); |
| 3760 | if (this_int_off != int_map_off) { |
| 3761 | // reset the map: |
| 3762 | int_map = 0; |
| 3763 | int_map_off = this_int_off; |
| 3764 | } |
| 3765 | |
| 3766 | int subword_off = st_off - this_int_off; |
| 3767 | int_map |= right_n_bits(st_size) << subword_off; |
| 3768 | if ((int_map & FULL_MAP) == FULL_MAP) { |
| 3769 | return this_int_off; // we found a complete word init |
| 3770 | } |
| 3771 | |
| 3772 | // Did this store hit or cross the word boundary? |
| 3773 | intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); |
| 3774 | if (next_int_off == this_int_off + BytesPerInt) { |
| 3775 | // We passed the current int, without fully initializing it. |
| 3776 | int_map_off = next_int_off; |
| 3777 | int_map >>= BytesPerInt; |
| 3778 | } else if (next_int_off > this_int_off + BytesPerInt) { |
| 3779 | // We passed the current and next int. |
| 3780 | return this_int_off + BytesPerInt; |
| 3781 | } |
| 3782 | } |
| 3783 | |
| 3784 | return -1; |
| 3785 | } |
| 3786 | |
| 3787 | |
| 3788 | // Called when the associated AllocateNode is expanded into CFG. |
| 3789 | // At this point, we may perform additional optimizations. |
| 3790 | // Linearize the stores by ascending offset, to make memory |
| 3791 | // activity as coherent as possible. |
| 3792 | Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, |
| 3793 | intptr_t header_size, |
| 3794 | Node* size_in_bytes, |
| 3795 | PhaseGVN* phase) { |
| 3796 | assert(!is_complete(), "not already complete"); |
| 3797 | assert(stores_are_sane(phase), ""); |
| 3798 | assert(allocation() != NULL, "must be present"); |
| 3799 | |
| 3800 | remove_extra_zeroes(); |
| 3801 | |
| 3802 | if (ReduceFieldZeroing || ReduceBulkZeroing) |
| 3803 | // reduce instruction count for common initialization patterns |
| 3804 | coalesce_subword_stores(header_size, size_in_bytes, phase); |
| 3805 | |
| 3806 | Node* zmem = zero_memory(); // initially zero memory state |
| 3807 | Node* inits = zmem; // accumulating a linearized chain of inits |
| 3808 | #ifdef ASSERT |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 3809 | intptr_t first_offset = allocation()->minimum_header_size(); |
| 3810 | intptr_t last_init_off = first_offset; // previous init offset |
| 3811 | intptr_t last_init_end = first_offset; // previous init offset+size |
| 3812 | intptr_t last_tile_end = first_offset; // previous tile offset+size |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3813 | #endif |
| 3814 | intptr_t zeroes_done = header_size; |
| 3815 | |
| 3816 | bool do_zeroing = true; // we might give up if inits are very sparse |
| 3817 | int big_init_gaps = 0; // how many large gaps have we seen? |
| 3818 | |
| 3819 | if (ZeroTLAB) do_zeroing = false; |
| 3820 | if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; |
| 3821 | |
| 3822 | for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { |
| 3823 | Node* st = in(i); |
| 3824 | intptr_t st_off = get_store_offset(st, phase); |
| 3825 | if (st_off < 0) |
| 3826 | break; // unknown junk in the inits |
| 3827 | if (st->in(MemNode::Memory) != zmem) |
| 3828 | break; // complicated store chains somehow in list |
| 3829 | |
| 3830 | int st_size = st->as_Store()->memory_size(); |
| 3831 | intptr_t next_init_off = st_off + st_size; |
| 3832 | |
| 3833 | if (do_zeroing && zeroes_done < next_init_off) { |
| 3834 | // See if this store needs a zero before it or under it. |
| 3835 | intptr_t zeroes_needed = st_off; |
| 3836 | |
| 3837 | if (st_size < BytesPerInt) { |
| 3838 | // Look for subword stores which only partially initialize words. |
| 3839 | // If we find some, we must lay down some word-level zeroes first, |
| 3840 | // underneath the subword stores. |
| 3841 | // |
| 3842 | // Examples: |
| 3843 | // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s |
| 3844 | // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y |
| 3845 | // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z |
| 3846 | // |
| 3847 | // Note: coalesce_subword_stores may have already done this, |
| 3848 | // if it was prompted by constant non-zero subword initializers. |
| 3849 | // But this case can still arise with non-constant stores. |
| 3850 | |
| 3851 | intptr_t next_full_store = find_next_fullword_store(i, phase); |
| 3852 | |
| 3853 | // In the examples above: |
| 3854 | // in(i) p q r s x y z |
| 3855 | // st_off 12 13 14 15 12 13 14 |
| 3856 | // st_size 1 1 1 1 1 1 1 |
| 3857 | // next_full_s. 12 16 16 16 16 16 16 |
| 3858 | // z's_done 12 16 16 16 12 16 12 |
| 3859 | // z's_needed 12 16 16 16 16 16 16 |
| 3860 | // zsize 0 0 0 0 4 0 4 |
| 3861 | if (next_full_store < 0) { |
| 3862 | // Conservative tack: Zero to end of current word. |
| 3863 | zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); |
| 3864 | } else { |
| 3865 | // Zero to beginning of next fully initialized word. |
| 3866 | // Or, don't zero at all, if we are already in that word. |
| 3867 | assert(next_full_store >= zeroes_needed, "must go forward"); |
| 3868 | assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); |
| 3869 | zeroes_needed = next_full_store; |
| 3870 | } |
| 3871 | } |
| 3872 | |
| 3873 | if (zeroes_needed > zeroes_done) { |
| 3874 | intptr_t zsize = zeroes_needed - zeroes_done; |
| 3875 | // Do some incremental zeroing on rawmem, in parallel with inits. |
| 3876 | zeroes_done = align_size_down(zeroes_done, BytesPerInt); |
| 3877 | rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, |
| 3878 | zeroes_done, zeroes_needed, |
| 3879 | phase); |
| 3880 | zeroes_done = zeroes_needed; |
| 3881 | if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) |
| 3882 | do_zeroing = false; // leave the hole, next time |
| 3883 | } |
| 3884 | } |
| 3885 | |
| 3886 | // Collect the store and move on: |
| 3887 | st->set_req(MemNode::Memory, inits); |
| 3888 | inits = st; // put it on the linearized chain |
| 3889 | set_req(i, zmem); // unhook from previous position |
| 3890 | |
| 3891 | if (zeroes_done == st_off) |
| 3892 | zeroes_done = next_init_off; |
| 3893 | |
| 3894 | assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); |
| 3895 | |
| 3896 | #ifdef ASSERT |
| 3897 | // Various order invariants. Weaker than stores_are_sane because |
| 3898 | // a large constant tile can be filled in by smaller non-constant stores. |
| 3899 | assert(st_off >= last_init_off, "inits do not reverse"); |
| 3900 | last_init_off = st_off; |
| 3901 | const Type* val = NULL; |
| 3902 | if (st_size >= BytesPerInt && |
| 3903 | (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && |
| 3904 | (int)val->basic_type() < (int)T_OBJECT) { |
| 3905 | assert(st_off >= last_tile_end, "tiles do not overlap"); |
| 3906 | assert(st_off >= last_init_end, "tiles do not overwrite inits"); |
| 3907 | last_tile_end = MAX2(last_tile_end, next_init_off); |
| 3908 | } else { |
| 3909 | intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); |
| 3910 | assert(st_tile_end >= last_tile_end, "inits stay with tiles"); |
| 3911 | assert(st_off >= last_init_end, "inits do not overlap"); |
| 3912 | last_init_end = next_init_off; // it's a non-tile |
| 3913 | } |
| 3914 | #endif //ASSERT |
| 3915 | } |
| 3916 | |
| 3917 | remove_extra_zeroes(); // clear out all the zmems left over |
| 3918 | add_req(inits); |
| 3919 | |
| 3920 | if (!ZeroTLAB) { |
| 3921 | // If anything remains to be zeroed, zero it all now. |
| 3922 | zeroes_done = align_size_down(zeroes_done, BytesPerInt); |
| 3923 | // if it is the last unused 4 bytes of an instance, forget about it |
| 3924 | intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); |
| 3925 | if (zeroes_done + BytesPerLong >= size_limit) { |
| 3926 | assert(allocation() != NULL, ""); |
Tom Rodriguez | 532dec3 | 2010-12-01 10:16:31 -0800 | [diff] [blame] | 3927 | if (allocation()->Opcode() == Op_Allocate) { |
| 3928 | Node* klass_node = allocation()->in(AllocateNode::KlassNode); |
| 3929 | ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); |
| 3930 | if (zeroes_done == k->layout_helper()) |
| 3931 | zeroes_done = size_limit; |
| 3932 | } |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3933 | } |
| 3934 | if (zeroes_done < size_limit) { |
| 3935 | rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, |
| 3936 | zeroes_done, size_in_bytes, phase); |
| 3937 | } |
| 3938 | } |
| 3939 | |
| 3940 | set_complete(phase); |
| 3941 | return rawmem; |
| 3942 | } |
| 3943 | |
| 3944 | |
| 3945 | #ifdef ASSERT |
| 3946 | bool InitializeNode::stores_are_sane(PhaseTransform* phase) { |
| 3947 | if (is_complete()) |
| 3948 | return true; // stores could be anything at this point |
Coleen Phillimore | 4a831d4 | 2008-04-13 17:43:42 -0400 | [diff] [blame] | 3949 | assert(allocation() != NULL, "must be present"); |
| 3950 | intptr_t last_off = allocation()->minimum_header_size(); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 3951 | for (uint i = InitializeNode::RawStores; i < req(); i++) { |
| 3952 | Node* st = in(i); |
| 3953 | intptr_t st_off = get_store_offset(st, phase); |
| 3954 | if (st_off < 0) continue; // ignore dead garbage |
| 3955 | if (last_off > st_off) { |
| 3956 | tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); |
| 3957 | this->dump(2); |
| 3958 | assert(false, "ascending store offsets"); |
| 3959 | return false; |
| 3960 | } |
| 3961 | last_off = st_off + st->as_Store()->memory_size(); |
| 3962 | } |
| 3963 | return true; |
| 3964 | } |
| 3965 | #endif //ASSERT |
| 3966 | |
| 3967 | |
| 3968 | |
| 3969 | |
| 3970 | //============================MergeMemNode===================================== |
| 3971 | // |
| 3972 | // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several |
| 3973 | // contributing store or call operations. Each contributor provides the memory |
| 3974 | // state for a particular "alias type" (see Compile::alias_type). For example, |
| 3975 | // if a MergeMem has an input X for alias category #6, then any memory reference |
| 3976 | // to alias category #6 may use X as its memory state input, as an exact equivalent |
| 3977 | // to using the MergeMem as a whole. |
| 3978 | // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) |
| 3979 | // |
| 3980 | // (Here, the <N> notation gives the index of the relevant adr_type.) |
| 3981 | // |
| 3982 | // In one special case (and more cases in the future), alias categories overlap. |
| 3983 | // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory |
| 3984 | // states. Therefore, if a MergeMem has only one contributing input W for Bot, |
| 3985 | // it is exactly equivalent to that state W: |
| 3986 | // MergeMem(<Bot>: W) <==> W |
| 3987 | // |
| 3988 | // Usually, the merge has more than one input. In that case, where inputs |
| 3989 | // overlap (i.e., one is Bot), the narrower alias type determines the memory |
| 3990 | // state for that type, and the wider alias type (Bot) fills in everywhere else: |
| 3991 | // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) |
| 3992 | // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) |
| 3993 | // |
| 3994 | // A merge can take a "wide" memory state as one of its narrow inputs. |
| 3995 | // This simply means that the merge observes out only the relevant parts of |
| 3996 | // the wide input. That is, wide memory states arriving at narrow merge inputs |
| 3997 | // are implicitly "filtered" or "sliced" as necessary. (This is rare.) |
| 3998 | // |
| 3999 | // These rules imply that MergeMem nodes may cascade (via their <Bot> links), |
| 4000 | // and that memory slices "leak through": |
| 4001 | // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) |
| 4002 | // |
| 4003 | // But, in such a cascade, repeated memory slices can "block the leak": |
| 4004 | // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') |
| 4005 | // |
| 4006 | // In the last example, Y is not part of the combined memory state of the |
| 4007 | // outermost MergeMem. The system must, of course, prevent unschedulable |
| 4008 | // memory states from arising, so you can be sure that the state Y is somehow |
| 4009 | // a precursor to state Y'. |
| 4010 | // |
| 4011 | // |
| 4012 | // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array |
| 4013 | // of each MergeMemNode array are exactly the numerical alias indexes, including |
| 4014 | // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions |
| 4015 | // Compile::alias_type (and kin) produce and manage these indexes. |
| 4016 | // |
| 4017 | // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. |
| 4018 | // (Note that this provides quick access to the top node inside MergeMem methods, |
| 4019 | // without the need to reach out via TLS to Compile::current.) |
| 4020 | // |
| 4021 | // As a consequence of what was just described, a MergeMem that represents a full |
| 4022 | // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, |
| 4023 | // containing all alias categories. |
| 4024 | // |
| 4025 | // MergeMem nodes never (?) have control inputs, so in(0) is NULL. |
| 4026 | // |
| 4027 | // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either |
| 4028 | // a memory state for the alias type <N>, or else the top node, meaning that |
| 4029 | // there is no particular input for that alias type. Note that the length of |
| 4030 | // a MergeMem is variable, and may be extended at any time to accommodate new |
| 4031 | // memory states at larger alias indexes. When merges grow, they are of course |
| 4032 | // filled with "top" in the unused in() positions. |
| 4033 | // |
| 4034 | // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. |
| 4035 | // (Top was chosen because it works smoothly with passes like GCM.) |
| 4036 | // |
| 4037 | // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is |
| 4038 | // the type of random VM bits like TLS references.) Since it is always the |
| 4039 | // first non-Bot memory slice, some low-level loops use it to initialize an |
| 4040 | // index variable: for (i = AliasIdxRaw; i < req(); i++). |
| 4041 | // |
| 4042 | // |
| 4043 | // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns |
| 4044 | // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns |
| 4045 | // the memory state for alias type <N>, or (if there is no particular slice at <N>, |
| 4046 | // it returns the base memory. To prevent bugs, memory_at does not accept <Top> |
| 4047 | // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over |
| 4048 | // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. |
| 4049 | // |
| 4050 | // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't |
| 4051 | // really that different from the other memory inputs. An abbreviation called |
| 4052 | // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. |
| 4053 | // |
| 4054 | // |
| 4055 | // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent |
| 4056 | // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi |
| 4057 | // that "emerges though" the base memory will be marked as excluding the alias types |
| 4058 | // of the other (narrow-memory) copies which "emerged through" the narrow edges: |
| 4059 | // |
| 4060 | // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) |
| 4061 | // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) |
| 4062 | // |
| 4063 | // This strange "subtraction" effect is necessary to ensure IGVN convergence. |
| 4064 | // (It is currently unimplemented.) As you can see, the resulting merge is |
| 4065 | // actually a disjoint union of memory states, rather than an overlay. |
| 4066 | // |
| 4067 | |
| 4068 | //------------------------------MergeMemNode----------------------------------- |
| 4069 | Node* MergeMemNode::make_empty_memory() { |
| 4070 | Node* empty_memory = (Node*) Compile::current()->top(); |
| 4071 | assert(empty_memory->is_top(), "correct sentinel identity"); |
| 4072 | return empty_memory; |
| 4073 | } |
| 4074 | |
| 4075 | MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { |
| 4076 | init_class_id(Class_MergeMem); |
| 4077 | // all inputs are nullified in Node::Node(int) |
| 4078 | // set_input(0, NULL); // no control input |
| 4079 | |
| 4080 | // Initialize the edges uniformly to top, for starters. |
| 4081 | Node* empty_mem = make_empty_memory(); |
| 4082 | for (uint i = Compile::AliasIdxTop; i < req(); i++) { |
| 4083 | init_req(i,empty_mem); |
| 4084 | } |
| 4085 | assert(empty_memory() == empty_mem, ""); |
| 4086 | |
| 4087 | if( new_base != NULL && new_base->is_MergeMem() ) { |
| 4088 | MergeMemNode* mdef = new_base->as_MergeMem(); |
| 4089 | assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); |
| 4090 | for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { |
| 4091 | mms.set_memory(mms.memory2()); |
| 4092 | } |
| 4093 | assert(base_memory() == mdef->base_memory(), ""); |
| 4094 | } else { |
| 4095 | set_base_memory(new_base); |
| 4096 | } |
| 4097 | } |
| 4098 | |
| 4099 | // Make a new, untransformed MergeMem with the same base as 'mem'. |
| 4100 | // If mem is itself a MergeMem, populate the result with the same edges. |
| 4101 | MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { |
Bharadwaj Yadavalli | d5d2e78 | 2012-09-27 09:38:42 -0700 | [diff] [blame] | 4102 | return new(C) MergeMemNode(mem); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 4103 | } |
| 4104 | |
| 4105 | //------------------------------cmp-------------------------------------------- |
| 4106 | uint MergeMemNode::hash() const { return NO_HASH; } |
| 4107 | uint MergeMemNode::cmp( const Node &n ) const { |
| 4108 | return (&n == this); // Always fail except on self |
| 4109 | } |
| 4110 | |
| 4111 | //------------------------------Identity--------------------------------------- |
| 4112 | Node* MergeMemNode::Identity(PhaseTransform *phase) { |
| 4113 | // Identity if this merge point does not record any interesting memory |
| 4114 | // disambiguations. |
| 4115 | Node* base_mem = base_memory(); |
| 4116 | Node* empty_mem = empty_memory(); |
| 4117 | if (base_mem != empty_mem) { // Memory path is not dead? |
| 4118 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
| 4119 | Node* mem = in(i); |
| 4120 | if (mem != empty_mem && mem != base_mem) { |
| 4121 | return this; // Many memory splits; no change |
| 4122 | } |
| 4123 | } |
| 4124 | } |
| 4125 | return base_mem; // No memory splits; ID on the one true input |
| 4126 | } |
| 4127 | |
| 4128 | //------------------------------Ideal------------------------------------------ |
| 4129 | // This method is invoked recursively on chains of MergeMem nodes |
| 4130 | Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
| 4131 | // Remove chain'd MergeMems |
| 4132 | // |
| 4133 | // This is delicate, because the each "in(i)" (i >= Raw) is interpreted |
| 4134 | // relative to the "in(Bot)". Since we are patching both at the same time, |
| 4135 | // we have to be careful to read each "in(i)" relative to the old "in(Bot)", |
| 4136 | // but rewrite each "in(i)" relative to the new "in(Bot)". |
| 4137 | Node *progress = NULL; |
| 4138 | |
| 4139 | |
| 4140 | Node* old_base = base_memory(); |
| 4141 | Node* empty_mem = empty_memory(); |
| 4142 | if (old_base == empty_mem) |
| 4143 | return NULL; // Dead memory path. |
| 4144 | |
| 4145 | MergeMemNode* old_mbase; |
| 4146 | if (old_base != NULL && old_base->is_MergeMem()) |
| 4147 | old_mbase = old_base->as_MergeMem(); |
| 4148 | else |
| 4149 | old_mbase = NULL; |
| 4150 | Node* new_base = old_base; |
| 4151 | |
| 4152 | // simplify stacked MergeMems in base memory |
| 4153 | if (old_mbase) new_base = old_mbase->base_memory(); |
| 4154 | |
| 4155 | // the base memory might contribute new slices beyond my req() |
| 4156 | if (old_mbase) grow_to_match(old_mbase); |
| 4157 | |
| 4158 | // Look carefully at the base node if it is a phi. |
| 4159 | PhiNode* phi_base; |
| 4160 | if (new_base != NULL && new_base->is_Phi()) |
| 4161 | phi_base = new_base->as_Phi(); |
| 4162 | else |
| 4163 | phi_base = NULL; |
| 4164 | |
| 4165 | Node* phi_reg = NULL; |
| 4166 | uint phi_len = (uint)-1; |
| 4167 | if (phi_base != NULL && !phi_base->is_copy()) { |
| 4168 | // do not examine phi if degraded to a copy |
| 4169 | phi_reg = phi_base->region(); |
| 4170 | phi_len = phi_base->req(); |
| 4171 | // see if the phi is unfinished |
| 4172 | for (uint i = 1; i < phi_len; i++) { |
| 4173 | if (phi_base->in(i) == NULL) { |
| 4174 | // incomplete phi; do not look at it yet! |
| 4175 | phi_reg = NULL; |
| 4176 | phi_len = (uint)-1; |
| 4177 | break; |
| 4178 | } |
| 4179 | } |
| 4180 | } |
| 4181 | |
| 4182 | // Note: We do not call verify_sparse on entry, because inputs |
| 4183 | // can normalize to the base_memory via subsume_node or similar |
| 4184 | // mechanisms. This method repairs that damage. |
| 4185 | |
| 4186 | assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); |
| 4187 | |
| 4188 | // Look at each slice. |
| 4189 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
| 4190 | Node* old_in = in(i); |
| 4191 | // calculate the old memory value |
| 4192 | Node* old_mem = old_in; |
| 4193 | if (old_mem == empty_mem) old_mem = old_base; |
| 4194 | assert(old_mem == memory_at(i), ""); |
| 4195 | |
| 4196 | // maybe update (reslice) the old memory value |
| 4197 | |
| 4198 | // simplify stacked MergeMems |
| 4199 | Node* new_mem = old_mem; |
| 4200 | MergeMemNode* old_mmem; |
| 4201 | if (old_mem != NULL && old_mem->is_MergeMem()) |
| 4202 | old_mmem = old_mem->as_MergeMem(); |
| 4203 | else |
| 4204 | old_mmem = NULL; |
| 4205 | if (old_mmem == this) { |
| 4206 | // This can happen if loops break up and safepoints disappear. |
| 4207 | // A merge of BotPtr (default) with a RawPtr memory derived from a |
| 4208 | // safepoint can be rewritten to a merge of the same BotPtr with |
| 4209 | // the BotPtr phi coming into the loop. If that phi disappears |
| 4210 | // also, we can end up with a self-loop of the mergemem. |
| 4211 | // In general, if loops degenerate and memory effects disappear, |
| 4212 | // a mergemem can be left looking at itself. This simply means |
| 4213 | // that the mergemem's default should be used, since there is |
| 4214 | // no longer any apparent effect on this slice. |
| 4215 | // Note: If a memory slice is a MergeMem cycle, it is unreachable |
| 4216 | // from start. Update the input to TOP. |
| 4217 | new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; |
| 4218 | } |
| 4219 | else if (old_mmem != NULL) { |
| 4220 | new_mem = old_mmem->memory_at(i); |
| 4221 | } |
Christian Thalinger | 05d1de7 | 2009-02-27 13:27:09 -0800 | [diff] [blame] | 4222 | // else preceding memory was not a MergeMem |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 4223 | |
| 4224 | // replace equivalent phis (unfortunately, they do not GVN together) |
| 4225 | if (new_mem != NULL && new_mem != new_base && |
| 4226 | new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { |
| 4227 | if (new_mem->is_Phi()) { |
| 4228 | PhiNode* phi_mem = new_mem->as_Phi(); |
| 4229 | for (uint i = 1; i < phi_len; i++) { |
| 4230 | if (phi_base->in(i) != phi_mem->in(i)) { |
| 4231 | phi_mem = NULL; |
| 4232 | break; |
| 4233 | } |
| 4234 | } |
| 4235 | if (phi_mem != NULL) { |
| 4236 | // equivalent phi nodes; revert to the def |
| 4237 | new_mem = new_base; |
| 4238 | } |
| 4239 | } |
| 4240 | } |
| 4241 | |
| 4242 | // maybe store down a new value |
| 4243 | Node* new_in = new_mem; |
| 4244 | if (new_in == new_base) new_in = empty_mem; |
| 4245 | |
| 4246 | if (new_in != old_in) { |
| 4247 | // Warning: Do not combine this "if" with the previous "if" |
| 4248 | // A memory slice might have be be rewritten even if it is semantically |
| 4249 | // unchanged, if the base_memory value has changed. |
| 4250 | set_req(i, new_in); |
| 4251 | progress = this; // Report progress |
| 4252 | } |
| 4253 | } |
| 4254 | |
| 4255 | if (new_base != old_base) { |
| 4256 | set_req(Compile::AliasIdxBot, new_base); |
| 4257 | // Don't use set_base_memory(new_base), because we need to update du. |
| 4258 | assert(base_memory() == new_base, ""); |
| 4259 | progress = this; |
| 4260 | } |
| 4261 | |
| 4262 | if( base_memory() == this ) { |
| 4263 | // a self cycle indicates this memory path is dead |
| 4264 | set_req(Compile::AliasIdxBot, empty_mem); |
| 4265 | } |
| 4266 | |
| 4267 | // Resolve external cycles by calling Ideal on a MergeMem base_memory |
| 4268 | // Recursion must occur after the self cycle check above |
| 4269 | if( base_memory()->is_MergeMem() ) { |
| 4270 | MergeMemNode *new_mbase = base_memory()->as_MergeMem(); |
| 4271 | Node *m = phase->transform(new_mbase); // Rollup any cycles |
| 4272 | if( m != NULL && (m->is_top() || |
| 4273 | m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { |
| 4274 | // propagate rollup of dead cycle to self |
| 4275 | set_req(Compile::AliasIdxBot, empty_mem); |
| 4276 | } |
| 4277 | } |
| 4278 | |
| 4279 | if( base_memory() == empty_mem ) { |
| 4280 | progress = this; |
| 4281 | // Cut inputs during Parse phase only. |
| 4282 | // During Optimize phase a dead MergeMem node will be subsumed by Top. |
| 4283 | if( !can_reshape ) { |
| 4284 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
| 4285 | if( in(i) != empty_mem ) { set_req(i, empty_mem); } |
| 4286 | } |
| 4287 | } |
| 4288 | } |
| 4289 | |
| 4290 | if( !progress && base_memory()->is_Phi() && can_reshape ) { |
| 4291 | // Check if PhiNode::Ideal's "Split phis through memory merges" |
| 4292 | // transform should be attempted. Look for this->phi->this cycle. |
| 4293 | uint merge_width = req(); |
| 4294 | if (merge_width > Compile::AliasIdxRaw) { |
| 4295 | PhiNode* phi = base_memory()->as_Phi(); |
| 4296 | for( uint i = 1; i < phi->req(); ++i ) {// For all paths in |
| 4297 | if (phi->in(i) == this) { |
| 4298 | phase->is_IterGVN()->_worklist.push(phi); |
| 4299 | break; |
| 4300 | } |
| 4301 | } |
| 4302 | } |
| 4303 | } |
| 4304 | |
Vladimir Kozlov | 30dc0ed | 2008-03-13 16:31:32 -0700 | [diff] [blame] | 4305 | assert(progress || verify_sparse(), "please, no dups of base"); |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 4306 | return progress; |
| 4307 | } |
| 4308 | |
| 4309 | //-------------------------set_base_memory------------------------------------- |
| 4310 | void MergeMemNode::set_base_memory(Node *new_base) { |
| 4311 | Node* empty_mem = empty_memory(); |
| 4312 | set_req(Compile::AliasIdxBot, new_base); |
| 4313 | assert(memory_at(req()) == new_base, "must set default memory"); |
| 4314 | // Clear out other occurrences of new_base: |
| 4315 | if (new_base != empty_mem) { |
| 4316 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
| 4317 | if (in(i) == new_base) set_req(i, empty_mem); |
| 4318 | } |
| 4319 | } |
| 4320 | } |
| 4321 | |
| 4322 | //------------------------------out_RegMask------------------------------------ |
| 4323 | const RegMask &MergeMemNode::out_RegMask() const { |
| 4324 | return RegMask::Empty; |
| 4325 | } |
| 4326 | |
| 4327 | //------------------------------dump_spec-------------------------------------- |
| 4328 | #ifndef PRODUCT |
| 4329 | void MergeMemNode::dump_spec(outputStream *st) const { |
| 4330 | st->print(" {"); |
| 4331 | Node* base_mem = base_memory(); |
| 4332 | for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { |
| 4333 | Node* mem = memory_at(i); |
| 4334 | if (mem == base_mem) { st->print(" -"); continue; } |
| 4335 | st->print( " N%d:", mem->_idx ); |
| 4336 | Compile::current()->get_adr_type(i)->dump_on(st); |
| 4337 | } |
| 4338 | st->print(" }"); |
| 4339 | } |
| 4340 | #endif // !PRODUCT |
| 4341 | |
| 4342 | |
| 4343 | #ifdef ASSERT |
| 4344 | static bool might_be_same(Node* a, Node* b) { |
| 4345 | if (a == b) return true; |
| 4346 | if (!(a->is_Phi() || b->is_Phi())) return false; |
| 4347 | // phis shift around during optimization |
| 4348 | return true; // pretty stupid... |
| 4349 | } |
| 4350 | |
| 4351 | // verify a narrow slice (either incoming or outgoing) |
| 4352 | static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { |
| 4353 | if (!VerifyAliases) return; // don't bother to verify unless requested |
| 4354 | if (is_error_reported()) return; // muzzle asserts when debugging an error |
| 4355 | if (Node::in_dump()) return; // muzzle asserts when printing |
| 4356 | assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); |
| 4357 | assert(n != NULL, ""); |
| 4358 | // Elide intervening MergeMem's |
| 4359 | while (n->is_MergeMem()) { |
| 4360 | n = n->as_MergeMem()->memory_at(alias_idx); |
| 4361 | } |
| 4362 | Compile* C = Compile::current(); |
| 4363 | const TypePtr* n_adr_type = n->adr_type(); |
| 4364 | if (n == m->empty_memory()) { |
| 4365 | // Implicit copy of base_memory() |
| 4366 | } else if (n_adr_type != TypePtr::BOTTOM) { |
| 4367 | assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); |
| 4368 | assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); |
| 4369 | } else { |
| 4370 | // A few places like make_runtime_call "know" that VM calls are narrow, |
| 4371 | // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. |
| 4372 | bool expected_wide_mem = false; |
| 4373 | if (n == m->base_memory()) { |
| 4374 | expected_wide_mem = true; |
| 4375 | } else if (alias_idx == Compile::AliasIdxRaw || |
| 4376 | n == m->memory_at(Compile::AliasIdxRaw)) { |
| 4377 | expected_wide_mem = true; |
| 4378 | } else if (!C->alias_type(alias_idx)->is_rewritable()) { |
| 4379 | // memory can "leak through" calls on channels that |
| 4380 | // are write-once. Allow this also. |
| 4381 | expected_wide_mem = true; |
| 4382 | } |
| 4383 | assert(expected_wide_mem, "expected narrow slice replacement"); |
| 4384 | } |
| 4385 | } |
| 4386 | #else // !ASSERT |
| 4387 | #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op |
| 4388 | #endif |
| 4389 | |
| 4390 | |
| 4391 | //-----------------------------memory_at--------------------------------------- |
| 4392 | Node* MergeMemNode::memory_at(uint alias_idx) const { |
| 4393 | assert(alias_idx >= Compile::AliasIdxRaw || |
| 4394 | alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, |
| 4395 | "must avoid base_memory and AliasIdxTop"); |
| 4396 | |
| 4397 | // Otherwise, it is a narrow slice. |
| 4398 | Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); |
| 4399 | Compile *C = Compile::current(); |
| 4400 | if (is_empty_memory(n)) { |
| 4401 | // the array is sparse; empty slots are the "top" node |
| 4402 | n = base_memory(); |
| 4403 | assert(Node::in_dump() |
| 4404 | || n == NULL || n->bottom_type() == Type::TOP |
Vladimir Kozlov | afa5d11 | 2011-02-24 10:28:20 -0800 | [diff] [blame] | 4405 | || n->adr_type() == NULL // address is TOP |
J. Duke | 8153779 | 2007-12-01 00:00:00 +0000 | [diff] [blame] | 4406 | || n->adr_type() == TypePtr::BOTTOM |
| 4407 | || n->adr_type() == TypeRawPtr::BOTTOM |
| 4408 | || Compile::current()->AliasLevel() == 0, |
| 4409 | "must be a wide memory"); |
| 4410 | // AliasLevel == 0 if we are organizing the memory states manually. |
| 4411 | // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. |
| 4412 | } else { |
| 4413 | // make sure the stored slice is sane |
| 4414 | #ifdef ASSERT |
| 4415 | if (is_error_reported() || Node::in_dump()) { |
| 4416 | } else if (might_be_same(n, base_memory())) { |
| 4417 | // Give it a pass: It is a mostly harmless repetition of the base. |
| 4418 | // This can arise normally from node subsumption during optimization. |
| 4419 | } else { |
| 4420 | verify_memory_slice(this, alias_idx, n); |
| 4421 | } |
| 4422 | #endif |
| 4423 | } |
| 4424 | return n; |
| 4425 | } |
| 4426 | |
| 4427 | //---------------------------set_memory_at------------------------------------- |
| 4428 | void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { |
| 4429 | verify_memory_slice(this, alias_idx, n); |
| 4430 | Node* empty_mem = empty_memory(); |
| 4431 | if (n == base_memory()) n = empty_mem; // collapse default |
| 4432 | uint need_req = alias_idx+1; |
| 4433 | if (req() < need_req) { |
| 4434 | if (n == empty_mem) return; // already the default, so do not grow me |
| 4435 | // grow the sparse array |
| 4436 | do { |
| 4437 | add_req(empty_mem); |
| 4438 | } while (req() < need_req); |
| 4439 | } |
| 4440 | set_req( alias_idx, n ); |
| 4441 | } |
| 4442 | |
| 4443 | |
| 4444 | |
| 4445 | //--------------------------iteration_setup------------------------------------ |
| 4446 | void MergeMemNode::iteration_setup(const MergeMemNode* other) { |
| 4447 | if (other != NULL) { |
| 4448 | grow_to_match(other); |
| 4449 | // invariant: the finite support of mm2 is within mm->req() |
| 4450 | #ifdef ASSERT |
| 4451 | for (uint i = req(); i < other->req(); i++) { |
| 4452 | assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); |
| 4453 | } |
| 4454 | #endif |
| 4455 | } |
| 4456 | // Replace spurious copies of base_memory by top. |
| 4457 | Node* base_mem = base_memory(); |
| 4458 | if (base_mem != NULL && !base_mem->is_top()) { |
| 4459 | for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { |
| 4460 | if (in(i) == base_mem) |
| 4461 | set_req(i, empty_memory()); |
| 4462 | } |
| 4463 | } |
| 4464 | } |
| 4465 | |
| 4466 | //---------------------------grow_to_match------------------------------------- |
| 4467 | void MergeMemNode::grow_to_match(const MergeMemNode* other) { |
| 4468 | Node* empty_mem = empty_memory(); |
| 4469 | assert(other->is_empty_memory(empty_mem), "consistent sentinels"); |
| 4470 | // look for the finite support of the other memory |
| 4471 | for (uint i = other->req(); --i >= req(); ) { |
| 4472 | if (other->in(i) != empty_mem) { |
| 4473 | uint new_len = i+1; |
| 4474 | while (req() < new_len) add_req(empty_mem); |
| 4475 | break; |
| 4476 | } |
| 4477 | } |
| 4478 | } |
| 4479 | |
| 4480 | //---------------------------verify_sparse------------------------------------- |
| 4481 | #ifndef PRODUCT |
| 4482 | bool MergeMemNode::verify_sparse() const { |
| 4483 | assert(is_empty_memory(make_empty_memory()), "sane sentinel"); |
| 4484 | Node* base_mem = base_memory(); |
| 4485 | // The following can happen in degenerate cases, since empty==top. |
| 4486 | if (is_empty_memory(base_mem)) return true; |
| 4487 | for (uint i = Compile::AliasIdxRaw; i < req(); i++) { |
| 4488 | assert(in(i) != NULL, "sane slice"); |
| 4489 | if (in(i) == base_mem) return false; // should have been the sentinel value! |
| 4490 | } |
| 4491 | return true; |
| 4492 | } |
| 4493 | |
| 4494 | bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { |
| 4495 | Node* n; |
| 4496 | n = mm->in(idx); |
| 4497 | if (mem == n) return true; // might be empty_memory() |
| 4498 | n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); |
| 4499 | if (mem == n) return true; |
| 4500 | while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { |
| 4501 | if (mem == n) return true; |
| 4502 | if (n == NULL) break; |
| 4503 | } |
| 4504 | return false; |
| 4505 | } |
| 4506 | #endif // !PRODUCT |