| 1 | // Copyright (c) 2013, the Dart project authors. Please see the AUTHORS file |
| 2 | // for details. All rights reserved. Use of this source code is governed by a |
| 3 | // BSD-style license that can be found in the LICENSE file. |
| 4 | #include "vm/compiler/jit/jit_call_specializer.h" |
| 5 | |
| 6 | #include "vm/bit_vector.h" |
| 7 | #include "vm/compiler/backend/branch_optimizer.h" |
| 8 | #include "vm/compiler/backend/flow_graph_compiler.h" |
| 9 | #include "vm/compiler/backend/il.h" |
| 10 | #include "vm/compiler/backend/il_printer.h" |
| 11 | #include "vm/compiler/backend/inliner.h" |
| 12 | #include "vm/compiler/backend/range_analysis.h" |
| 13 | #include "vm/compiler/cha.h" |
| 14 | #include "vm/compiler/frontend/flow_graph_builder.h" |
| 15 | #include "vm/compiler/jit/compiler.h" |
| 16 | #include "vm/cpu.h" |
| 17 | #include "vm/dart_entry.h" |
| 18 | #include "vm/exceptions.h" |
| 19 | #include "vm/hash_map.h" |
| 20 | #include "vm/object_store.h" |
| 21 | #include "vm/parser.h" |
| 22 | #include "vm/resolver.h" |
| 23 | #include "vm/scopes.h" |
| 24 | #include "vm/stack_frame.h" |
| 25 | #include "vm/symbols.h" |
| 26 | |
| 27 | namespace dart { |
| 28 | |
| 29 | // Quick access to the current isolate and zone. |
| 30 | #define I (isolate()) |
| 31 | #define Z (zone()) |
| 32 | |
| 33 | JitCallSpecializer::JitCallSpecializer( |
| 34 | FlowGraph* flow_graph, |
| 35 | SpeculativeInliningPolicy* speculative_policy) |
| 36 | : CallSpecializer(flow_graph, |
| 37 | speculative_policy, |
| 38 | Field::ShouldCloneFields()) {} |
| 39 | |
| 40 | bool JitCallSpecializer::IsAllowedForInlining(intptr_t deopt_id) const { |
| 41 | return true; |
| 42 | } |
| 43 | |
| 44 | bool JitCallSpecializer::TryOptimizeStaticCallUsingStaticTypes( |
| 45 | StaticCallInstr* call) { |
| 46 | return false; |
| 47 | } |
| 48 | |
| 49 | void JitCallSpecializer::ReplaceWithStaticCall(InstanceCallInstr* instr, |
| 50 | const Function& target, |
| 51 | intptr_t call_count) { |
| 52 | StaticCallInstr* call = |
| 53 | StaticCallInstr::FromCall(Z, instr, target, call_count); |
| 54 | const CallTargets& targets = instr->Targets(); |
| 55 | if (targets.IsMonomorphic() && targets.MonomorphicExactness().IsExact()) { |
| 56 | if (targets.MonomorphicExactness().IsTriviallyExact()) { |
| 57 | flow_graph()->AddExactnessGuard(instr, targets.MonomorphicReceiverCid()); |
| 58 | } |
| 59 | call->set_entry_kind(Code::EntryKind::kUnchecked); |
| 60 | } |
| 61 | instr->ReplaceWith(call, current_iterator()); |
| 62 | } |
| 63 | |
| 64 | // Tries to optimize instance call by replacing it with a faster instruction |
| 65 | // (e.g, binary op, field load, ..). |
| 66 | // TODO(dartbug.com/30635) Evaluate how much this can be shared with |
| 67 | // AotCallSpecializer. |
| 68 | void JitCallSpecializer::VisitInstanceCall(InstanceCallInstr* instr) { |
| 69 | const CallTargets& targets = instr->Targets(); |
| 70 | if (targets.is_empty()) { |
| 71 | return; // No feedback. |
| 72 | } |
| 73 | |
| 74 | const Token::Kind op_kind = instr->token_kind(); |
| 75 | |
| 76 | // Type test is special as it always gets converted into inlined code. |
| 77 | if (Token::IsTypeTestOperator(op_kind)) { |
| 78 | ReplaceWithInstanceOf(instr); |
| 79 | return; |
| 80 | } |
| 81 | |
| 82 | if ((op_kind == Token::kASSIGN_INDEX) && TryReplaceWithIndexedOp(instr)) { |
| 83 | return; |
| 84 | } |
| 85 | if ((op_kind == Token::kINDEX) && TryReplaceWithIndexedOp(instr)) { |
| 86 | return; |
| 87 | } |
| 88 | |
| 89 | if (op_kind == Token::kEQ && TryReplaceWithEqualityOp(instr, op_kind)) { |
| 90 | return; |
| 91 | } |
| 92 | |
| 93 | if (Token::IsRelationalOperator(op_kind) && |
| 94 | TryReplaceWithRelationalOp(instr, op_kind)) { |
| 95 | return; |
| 96 | } |
| 97 | |
| 98 | if (Token::IsBinaryOperator(op_kind) && |
| 99 | TryReplaceWithBinaryOp(instr, op_kind)) { |
| 100 | return; |
| 101 | } |
| 102 | if (Token::IsUnaryOperator(op_kind) && |
| 103 | TryReplaceWithUnaryOp(instr, op_kind)) { |
| 104 | return; |
| 105 | } |
| 106 | if ((op_kind == Token::kGET) && TryInlineInstanceGetter(instr)) { |
| 107 | return; |
| 108 | } |
| 109 | if ((op_kind == Token::kSET) && TryInlineInstanceSetter(instr)) { |
| 110 | return; |
| 111 | } |
| 112 | if (TryInlineInstanceMethod(instr)) { |
| 113 | return; |
| 114 | } |
| 115 | |
| 116 | bool has_one_target = targets.HasSingleTarget(); |
| 117 | if (has_one_target) { |
| 118 | // Check if the single target is a polymorphic target, if it is, |
| 119 | // we don't have one target. |
| 120 | const Function& target = targets.FirstTarget(); |
| 121 | if (target.recognized_kind() == MethodRecognizer::kObjectRuntimeType) { |
| 122 | has_one_target = PolymorphicInstanceCallInstr::ComputeRuntimeType( |
| 123 | targets) != Type::null(); |
| 124 | } else { |
| 125 | has_one_target = !target.is_polymorphic_target(); |
| 126 | } |
| 127 | } |
| 128 | |
| 129 | if (has_one_target) { |
| 130 | const Function& target = targets.FirstTarget(); |
| 131 | if (flow_graph()->CheckForInstanceCall(instr, target.kind()) == |
| 132 | FlowGraph::ToCheck::kNoCheck) { |
| 133 | ReplaceWithStaticCall(instr, target, targets.AggregateCallCount()); |
| 134 | return; |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | // If there is only one target we can make this into a deopting class check, |
| 139 | // followed by a call instruction that does not check the class of the |
| 140 | // receiver. This enables a lot of optimizations because after the class |
| 141 | // check we can probably inline the call and not worry about side effects. |
| 142 | // However, this can fall down if new receiver classes arrive at this call |
| 143 | // site after we generated optimized code. This causes a deopt, and after a |
| 144 | // few deopts we won't optimize this function any more at all. Therefore for |
| 145 | // very polymorphic sites we don't make this optimization, keeping it as a |
| 146 | // regular checked PolymorphicInstanceCall, which falls back to the slow but |
| 147 | // non-deopting megamorphic call stub when it sees new receiver classes. |
| 148 | if (has_one_target && FLAG_polymorphic_with_deopt && |
| 149 | (!instr->ic_data()->HasDeoptReason(ICData::kDeoptCheckClass) || |
| 150 | targets.length() <= FLAG_max_polymorphic_checks)) { |
| 151 | // Type propagation has not run yet, we cannot eliminate the check. |
| 152 | AddReceiverCheck(instr); |
| 153 | |
| 154 | // Call can still deoptimize, do not detach environment from instr. |
| 155 | const Function& target = targets.FirstTarget(); |
| 156 | ReplaceWithStaticCall(instr, target, targets.AggregateCallCount()); |
| 157 | } else { |
| 158 | PolymorphicInstanceCallInstr* call = |
| 159 | PolymorphicInstanceCallInstr::FromCall(Z, instr, targets, |
| 160 | /* complete = */ false); |
| 161 | instr->ReplaceWith(call, current_iterator()); |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | void JitCallSpecializer::VisitStoreInstanceField( |
| 166 | StoreInstanceFieldInstr* instr) { |
| 167 | if (instr->IsUnboxedStore()) { |
| 168 | // Determine if this field should be unboxed based on the usage of getter |
| 169 | // and setter functions: The heuristic requires that the setter has a |
| 170 | // usage count of at least 1/kGetterSetterRatio of the getter usage count. |
| 171 | // This is to avoid unboxing fields where the setter is never or rarely |
| 172 | // executed. |
| 173 | const Field& field = instr->slot().field(); |
| 174 | const String& field_name = String::Handle(Z, field.name()); |
| 175 | const Class& owner = Class::Handle(Z, field.Owner()); |
| 176 | const Function& getter = |
| 177 | Function::Handle(Z, owner.LookupGetterFunction(field_name)); |
| 178 | const Function& setter = |
| 179 | Function::Handle(Z, owner.LookupSetterFunction(field_name)); |
| 180 | bool unboxed_field = false; |
| 181 | if (!getter.IsNull() && !setter.IsNull()) { |
| 182 | if (field.is_double_initialized()) { |
| 183 | unboxed_field = true; |
| 184 | } else if ((setter.usage_counter() > 0) && |
| 185 | ((FLAG_getter_setter_ratio * setter.usage_counter()) >= |
| 186 | getter.usage_counter())) { |
| 187 | unboxed_field = true; |
| 188 | } |
| 189 | } |
| 190 | if (!unboxed_field) { |
| 191 | if (Compiler::IsBackgroundCompilation()) { |
| 192 | isolate()->AddDeoptimizingBoxedField(field); |
| 193 | Compiler::AbortBackgroundCompilation( |
| 194 | DeoptId::kNone, "Unboxing instance field while compiling" ); |
| 195 | UNREACHABLE(); |
| 196 | } |
| 197 | if (FLAG_trace_optimization || FLAG_trace_field_guards) { |
| 198 | THR_Print("Disabling unboxing of %s\n" , field.ToCString()); |
| 199 | if (!setter.IsNull()) { |
| 200 | THR_Print(" setter usage count: %" Pd "\n" , setter.usage_counter()); |
| 201 | } |
| 202 | if (!getter.IsNull()) { |
| 203 | THR_Print(" getter usage count: %" Pd "\n" , getter.usage_counter()); |
| 204 | } |
| 205 | } |
| 206 | ASSERT(field.IsOriginal()); |
| 207 | field.set_is_unboxing_candidate(false); |
| 208 | field.DeoptimizeDependentCode(); |
| 209 | } else { |
| 210 | flow_graph()->parsed_function().AddToGuardedFields(&field); |
| 211 | } |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | // Replace generic context allocation or cloning with a sequence of inlined |
| 216 | // allocation and explicit initializing stores. |
| 217 | // If context_value is not NULL then newly allocated context is a populated |
| 218 | // with values copied from it, otherwise it is initialized with null. |
| 219 | void JitCallSpecializer::LowerContextAllocation( |
| 220 | Definition* alloc, |
| 221 | const ZoneGrowableArray<const Slot*>& context_variables, |
| 222 | Value* context_value) { |
| 223 | ASSERT(alloc->IsAllocateContext() || alloc->IsCloneContext()); |
| 224 | |
| 225 | AllocateUninitializedContextInstr* replacement = |
| 226 | new AllocateUninitializedContextInstr(alloc->token_pos(), |
| 227 | context_variables.length()); |
| 228 | alloc->ReplaceWith(replacement, current_iterator()); |
| 229 | |
| 230 | Instruction* cursor = replacement; |
| 231 | |
| 232 | Value* initial_value; |
| 233 | if (context_value != NULL) { |
| 234 | LoadFieldInstr* load = |
| 235 | new (Z) LoadFieldInstr(context_value->CopyWithType(Z), |
| 236 | Slot::Context_parent(), alloc->token_pos()); |
| 237 | flow_graph()->InsertAfter(cursor, load, NULL, FlowGraph::kValue); |
| 238 | cursor = load; |
| 239 | initial_value = new (Z) Value(load); |
| 240 | } else { |
| 241 | initial_value = new (Z) Value(flow_graph()->constant_null()); |
| 242 | } |
| 243 | StoreInstanceFieldInstr* store = new (Z) StoreInstanceFieldInstr( |
| 244 | Slot::Context_parent(), new (Z) Value(replacement), initial_value, |
| 245 | kNoStoreBarrier, alloc->token_pos(), |
| 246 | StoreInstanceFieldInstr::Kind::kInitializing); |
| 247 | flow_graph()->InsertAfter(cursor, store, nullptr, FlowGraph::kEffect); |
| 248 | cursor = replacement; |
| 249 | |
| 250 | for (auto& slot : context_variables) { |
| 251 | if (context_value != nullptr) { |
| 252 | LoadFieldInstr* load = new (Z) LoadFieldInstr( |
| 253 | context_value->CopyWithType(Z), *slot, alloc->token_pos()); |
| 254 | flow_graph()->InsertAfter(cursor, load, nullptr, FlowGraph::kValue); |
| 255 | cursor = load; |
| 256 | initial_value = new (Z) Value(load); |
| 257 | } else { |
| 258 | initial_value = new (Z) Value(flow_graph()->constant_null()); |
| 259 | } |
| 260 | |
| 261 | store = new (Z) StoreInstanceFieldInstr( |
| 262 | *slot, new (Z) Value(replacement), initial_value, kNoStoreBarrier, |
| 263 | alloc->token_pos(), StoreInstanceFieldInstr::Kind::kInitializing); |
| 264 | flow_graph()->InsertAfter(cursor, store, nullptr, FlowGraph::kEffect); |
| 265 | cursor = store; |
| 266 | } |
| 267 | } |
| 268 | |
| 269 | void JitCallSpecializer::VisitAllocateContext(AllocateContextInstr* instr) { |
| 270 | LowerContextAllocation(instr, instr->context_slots(), nullptr); |
| 271 | } |
| 272 | |
| 273 | void JitCallSpecializer::VisitCloneContext(CloneContextInstr* instr) { |
| 274 | LowerContextAllocation(instr, instr->context_slots(), instr->context_value()); |
| 275 | } |
| 276 | |
| 277 | } // namespace dart |
| 278 | |