ARM64: Instruction simplification for array accesses.

HArrayGet and HArraySet with variable indexes generate two
instructions on arm64, like

    add temp, obj, #data_offset
    ldr out, [temp, index LSL #shift_amount]

When we have multiple accesses to the same array, the initial `add`
instruction is redundant.

This patch introduces the first instruction simplification in the
arm64-specific instruction simplification pass. It splits HArrayGet
and HArraySet using the new arm64-specific IR HIntermediateAddress.
After that we run GVN again to squash the multiple occurrences of
HIntermediateAddress.

Change-Id: I2e3d12fbb07fed07b2cb2f3f47f99f5a032f8312
diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc
index ffb9b79..5dda394 100644
--- a/compiler/optimizing/code_generator_arm64.cc
+++ b/compiler/optimizing/code_generator_arm64.cc
@@ -1580,6 +1580,21 @@
   HandleBinaryOp(instruction);
 }
 
+void LocationsBuilderARM64::VisitArm64IntermediateAddress(HArm64IntermediateAddress* instruction) {
+  LocationSummary* locations =
+      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
+  locations->SetInAt(0, Location::RequiresRegister());
+  locations->SetInAt(1, ARM64EncodableConstantOrRegister(instruction->GetOffset(), instruction));
+  locations->SetOut(Location::RequiresRegister());
+}
+
+void InstructionCodeGeneratorARM64::VisitArm64IntermediateAddress(
+    HArm64IntermediateAddress* instruction) {
+  __ Add(OutputRegister(instruction),
+         InputRegisterAt(instruction, 0),
+         Operand(InputOperandAt(instruction, 1)));
+}
+
 void LocationsBuilderARM64::VisitArrayGet(HArrayGet* instruction) {
   LocationSummary* locations =
       new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
@@ -1593,14 +1608,16 @@
 }
 
 void InstructionCodeGeneratorARM64::VisitArrayGet(HArrayGet* instruction) {
-  LocationSummary* locations = instruction->GetLocations();
   Primitive::Type type = instruction->GetType();
   Register obj = InputRegisterAt(instruction, 0);
-  Location index = locations->InAt(1);
+  Location index = instruction->GetLocations()->InAt(1);
   size_t offset = mirror::Array::DataOffset(Primitive::ComponentSize(type)).Uint32Value();
   MemOperand source = HeapOperand(obj);
+  CPURegister dest = OutputCPURegister(instruction);
+
   MacroAssembler* masm = GetVIXLAssembler();
   UseScratchRegisterScope temps(masm);
+  // Block pools between `Load` and `MaybeRecordImplicitNullCheck`.
   BlockPoolsScope block_pools(masm);
 
   if (index.IsConstant()) {
@@ -1608,15 +1625,26 @@
     source = HeapOperand(obj, offset);
   } else {
     Register temp = temps.AcquireSameSizeAs(obj);
-    __ Add(temp, obj, offset);
+    if (instruction->GetArray()->IsArm64IntermediateAddress()) {
+      // We do not need to compute the intermediate address from the array: the
+      // input instruction has done it already. See the comment in
+      // `InstructionSimplifierArm64::TryExtractArrayAccessAddress()`.
+      if (kIsDebugBuild) {
+        HArm64IntermediateAddress* tmp = instruction->GetArray()->AsArm64IntermediateAddress();
+        DCHECK(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64() == offset);
+      }
+      temp = obj;
+    } else {
+      __ Add(temp, obj, offset);
+    }
     source = HeapOperand(temp, XRegisterFrom(index), LSL, Primitive::ComponentSizeShift(type));
   }
 
-  codegen_->Load(type, OutputCPURegister(instruction), source);
+  codegen_->Load(type, dest, source);
   codegen_->MaybeRecordImplicitNullCheck(instruction);
 
-  if (type == Primitive::kPrimNot) {
-    GetAssembler()->MaybeUnpoisonHeapReference(OutputCPURegister(instruction).W());
+  if (instruction->GetType() == Primitive::kPrimNot) {
+    GetAssembler()->MaybeUnpoisonHeapReference(dest.W());
   }
 }
 
@@ -1670,7 +1698,18 @@
     } else {
       UseScratchRegisterScope temps(masm);
       Register temp = temps.AcquireSameSizeAs(array);
-      __ Add(temp, array, offset);
+      if (instruction->GetArray()->IsArm64IntermediateAddress()) {
+        // We do not need to compute the intermediate address from the array: the
+        // input instruction has done it already. See the comment in
+        // `InstructionSimplifierArm64::TryExtractArrayAccessAddress()`.
+        if (kIsDebugBuild) {
+          HArm64IntermediateAddress* tmp = instruction->GetArray()->AsArm64IntermediateAddress();
+          DCHECK(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64() == offset);
+        }
+        temp = array;
+      } else {
+        __ Add(temp, array, offset);
+      }
       destination = HeapOperand(temp,
                                 XRegisterFrom(index),
                                 LSL,
@@ -1680,6 +1719,7 @@
     codegen_->MaybeRecordImplicitNullCheck(instruction);
   } else {
     DCHECK(needs_write_barrier);
+    DCHECK(!instruction->GetArray()->IsArm64IntermediateAddress());
     vixl::Label done;
     SlowPathCodeARM64* slow_path = nullptr;
     {
diff --git a/compiler/optimizing/code_generator_arm64.h b/compiler/optimizing/code_generator_arm64.h
index a068b48..799f1bd 100644
--- a/compiler/optimizing/code_generator_arm64.h
+++ b/compiler/optimizing/code_generator_arm64.h
@@ -382,7 +382,7 @@
                      uint32_t dex_pc,
                      SlowPathCode* slow_path);
 
-  ParallelMoveResolverARM64* GetMoveResolver() { return &move_resolver_; }
+  ParallelMoveResolverARM64* GetMoveResolver() OVERRIDE { return &move_resolver_; }
 
   bool NeedsTwoRegisters(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
     return false;
diff --git a/compiler/optimizing/common_arm64.h b/compiler/optimizing/common_arm64.h
index 4abe5e9..e1a8c9c 100644
--- a/compiler/optimizing/common_arm64.h
+++ b/compiler/optimizing/common_arm64.h
@@ -203,19 +203,23 @@
 
   int64_t value = CodeGenerator::GetInt64ValueOf(constant);
 
-  if (instr->IsAdd() || instr->IsSub() || instr->IsCondition() ||
-      instr->IsCompare() || instr->IsBoundsCheck()) {
+  if (instr->IsAnd() || instr->IsOr() || instr->IsXor()) {
+    // Uses logical operations.
+    return vixl::Assembler::IsImmLogical(value, vixl::kXRegSize);
+  } else if (instr->IsNeg()) {
+    // Uses mov -immediate.
+    return vixl::Assembler::IsImmMovn(value, vixl::kXRegSize);
+  } else {
+    DCHECK(instr->IsAdd() ||
+           instr->IsArm64IntermediateAddress() ||
+           instr->IsBoundsCheck() ||
+           instr->IsCompare() ||
+           instr->IsCondition() ||
+           instr->IsSub());
     // Uses aliases of ADD/SUB instructions.
     // If `value` does not fit but `-value` does, VIXL will automatically use
     // the 'opposite' instruction.
     return vixl::Assembler::IsImmAddSub(value) || vixl::Assembler::IsImmAddSub(-value);
-  } else if (instr->IsAnd() || instr->IsOr() || instr->IsXor()) {
-    // Uses logical operations.
-    return vixl::Assembler::IsImmLogical(value, vixl::kXRegSize);
-  } else {
-    DCHECK(instr->IsNeg());
-    // Uses mov -immediate.
-    return vixl::Assembler::IsImmMovn(value, vixl::kXRegSize);
   }
 }
 
diff --git a/compiler/optimizing/instruction_simplifier_arm64.cc b/compiler/optimizing/instruction_simplifier_arm64.cc
index 4b2d36f..eb79f46 100644
--- a/compiler/optimizing/instruction_simplifier_arm64.cc
+++ b/compiler/optimizing/instruction_simplifier_arm64.cc
@@ -16,8 +16,65 @@
 
 #include "instruction_simplifier_arm64.h"
 
+#include "mirror/array-inl.h"
+
 namespace art {
 namespace arm64 {
 
+void InstructionSimplifierArm64Visitor::TryExtractArrayAccessAddress(HInstruction* access,
+                                                                     HInstruction* array,
+                                                                     HInstruction* index,
+                                                                     int access_size) {
+  if (index->IsConstant() ||
+      (index->IsBoundsCheck() && index->AsBoundsCheck()->GetIndex()->IsConstant())) {
+    // When the index is a constant all the addressing can be fitted in the
+    // memory access instruction, so do not split the access.
+    return;
+  }
+  if (access->IsArraySet() &&
+      access->AsArraySet()->GetValue()->GetType() == Primitive::kPrimNot) {
+    // The access may require a runtime call or the original array pointer.
+    return;
+  }
+
+  // Proceed to extract the base address computation.
+  ArenaAllocator* arena = GetGraph()->GetArena();
+
+  HIntConstant* offset =
+      GetGraph()->GetIntConstant(mirror::Array::DataOffset(access_size).Uint32Value());
+  HArm64IntermediateAddress* address =
+      new (arena) HArm64IntermediateAddress(array, offset, kNoDexPc);
+  access->GetBlock()->InsertInstructionBefore(address, access);
+  access->ReplaceInput(address, 0);
+  // Both instructions must depend on GC to prevent any instruction that can
+  // trigger GC to be inserted between the two.
+  access->AddSideEffects(SideEffects::DependsOnGC());
+  DCHECK(address->GetSideEffects().Includes(SideEffects::DependsOnGC()));
+  DCHECK(access->GetSideEffects().Includes(SideEffects::DependsOnGC()));
+  // TODO: Code generation for HArrayGet and HArraySet will check whether the input address
+  // is an HArm64IntermediateAddress and generate appropriate code.
+  // We would like to replace the `HArrayGet` and `HArraySet` with custom instructions (maybe
+  // `HArm64Load` and `HArm64Store`). We defer these changes because these new instructions would
+  // not bring any advantages yet.
+  // Also see the comments in
+  // `InstructionCodeGeneratorARM64::VisitArrayGet()` and
+  // `InstructionCodeGeneratorARM64::VisitArraySet()`.
+  RecordSimplification();
+}
+
+void InstructionSimplifierArm64Visitor::VisitArrayGet(HArrayGet* instruction) {
+  TryExtractArrayAccessAddress(instruction,
+                               instruction->GetArray(),
+                               instruction->GetIndex(),
+                               Primitive::ComponentSize(instruction->GetType()));
+}
+
+void InstructionSimplifierArm64Visitor::VisitArraySet(HArraySet* instruction) {
+  TryExtractArrayAccessAddress(instruction,
+                               instruction->GetArray(),
+                               instruction->GetIndex(),
+                               Primitive::ComponentSize(instruction->GetComponentType()));
+}
+
 }  // namespace arm64
 }  // namespace art
diff --git a/compiler/optimizing/instruction_simplifier_arm64.h b/compiler/optimizing/instruction_simplifier_arm64.h
index d7f4eae..4b697db 100644
--- a/compiler/optimizing/instruction_simplifier_arm64.h
+++ b/compiler/optimizing/instruction_simplifier_arm64.h
@@ -35,6 +35,14 @@
     }
   }
 
+  void TryExtractArrayAccessAddress(HInstruction* access,
+                                    HInstruction* array,
+                                    HInstruction* index,
+                                    int access_size);
+
+  void VisitArrayGet(HArrayGet* instruction) OVERRIDE;
+  void VisitArraySet(HArraySet* instruction) OVERRIDE;
+
   OptimizingCompilerStats* stats_;
 };
 
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index 939e62c..490bfc8 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -1079,14 +1079,23 @@
 
 #define FOR_EACH_CONCRETE_INSTRUCTION_ARM(M)
 
+#ifndef ART_ENABLE_CODEGEN_arm64
 #define FOR_EACH_CONCRETE_INSTRUCTION_ARM64(M)
+#else
+#define FOR_EACH_CONCRETE_INSTRUCTION_ARM64(M)                          \
+  M(Arm64IntermediateAddress, Instruction)
+#endif
 
 #define FOR_EACH_CONCRETE_INSTRUCTION_MIPS64(M)
 
+#ifndef ART_ENABLE_CODEGEN_x86
+#define FOR_EACH_CONCRETE_INSTRUCTION_X86(M)
+#else
 #define FOR_EACH_CONCRETE_INSTRUCTION_X86(M)                            \
   M(X86ComputeBaseMethodAddress, Instruction)                           \
   M(X86LoadFromConstantTable, Instruction)                              \
   M(X86PackedSwitch, Instruction)
+#endif
 
 #define FOR_EACH_CONCRETE_INSTRUCTION_X86_64(M)
 
@@ -1370,6 +1379,10 @@
     return SideEffects(flags_ & ~other.flags_);
   }
 
+  void Add(SideEffects other) {
+    flags_ |= other.flags_;
+  }
+
   bool Includes(SideEffects other) const {
     return (other.flags_ & flags_) == other.flags_;
   }
@@ -1943,6 +1956,7 @@
   }
 
   SideEffects GetSideEffects() const { return side_effects_; }
+  void AddSideEffects(SideEffects other) { side_effects_.Add(other); }
 
   size_t GetLifetimePosition() const { return lifetime_position_; }
   void SetLifetimePosition(size_t position) { lifetime_position_ = position; }
@@ -2012,7 +2026,7 @@
   // order of blocks where this instruction's live interval start.
   size_t lifetime_position_;
 
-  const SideEffects side_effects_;
+  SideEffects side_effects_;
 
   // TODO: for primitive types this should be marked as invalid.
   ReferenceTypeInfo reference_type_info_;
@@ -4424,8 +4438,11 @@
   HArrayGet(HInstruction* array,
             HInstruction* index,
             Primitive::Type type,
-            uint32_t dex_pc)
-      : HExpression(type, SideEffects::ArrayReadOfType(type), dex_pc) {
+            uint32_t dex_pc,
+            SideEffects additional_side_effects = SideEffects::None())
+      : HExpression(type,
+                    SideEffects::ArrayReadOfType(type).Union(additional_side_effects),
+                    dex_pc) {
     SetRawInputAt(0, array);
     SetRawInputAt(1, index);
   }
@@ -4460,10 +4477,13 @@
             HInstruction* index,
             HInstruction* value,
             Primitive::Type expected_component_type,
-            uint32_t dex_pc)
+            uint32_t dex_pc,
+            SideEffects additional_side_effects = SideEffects::None())
       : HTemplateInstruction(
             SideEffects::ArrayWriteOfType(expected_component_type).Union(
-                SideEffectsForArchRuntimeCalls(value->GetType())), dex_pc),
+                SideEffectsForArchRuntimeCalls(value->GetType())).Union(
+                    additional_side_effects),
+            dex_pc),
         expected_component_type_(expected_component_type),
         needs_type_check_(value->GetType() == Primitive::kPrimNot),
         value_can_be_null_(true),
@@ -4518,6 +4538,10 @@
         : expected_component_type_;
   }
 
+  Primitive::Type GetRawExpectedComponentType() const {
+    return expected_component_type_;
+  }
+
   static SideEffects SideEffectsForArchRuntimeCalls(Primitive::Type value_type) {
     return (value_type == Primitive::kPrimNot) ? SideEffects::CanTriggerGC() : SideEffects::None();
   }
@@ -4576,6 +4600,7 @@
 
   bool CanThrow() const OVERRIDE { return true; }
 
+  HInstruction* GetIndex() const { return InputAt(0); }
 
   DECLARE_INSTRUCTION(BoundsCheck);
 
@@ -5389,6 +5414,9 @@
 
 }  // namespace art
 
+#ifdef ART_ENABLE_CODEGEN_arm64
+#include "nodes_arm64.h"
+#endif
 #ifdef ART_ENABLE_CODEGEN_x86
 #include "nodes_x86.h"
 #endif
diff --git a/compiler/optimizing/nodes_arm64.h b/compiler/optimizing/nodes_arm64.h
new file mode 100644
index 0000000..885d3a2
--- /dev/null
+++ b/compiler/optimizing/nodes_arm64.h
@@ -0,0 +1,47 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_OPTIMIZING_NODES_ARM64_H_
+#define ART_COMPILER_OPTIMIZING_NODES_ARM64_H_
+
+namespace art {
+
+// This instruction computes an intermediate address pointing in the 'middle' of an object. The
+// result pointer cannot be handled by GC, so extra care is taken to make sure that this value is
+// never used across anything that can trigger GC.
+class HArm64IntermediateAddress : public HExpression<2> {
+ public:
+  HArm64IntermediateAddress(HInstruction* base_address, HInstruction* offset, uint32_t dex_pc)
+      : HExpression(Primitive::kPrimNot, SideEffects::DependsOnGC(), dex_pc) {
+    SetRawInputAt(0, base_address);
+    SetRawInputAt(1, offset);
+  }
+
+  bool CanBeMoved() const OVERRIDE { return true; }
+  bool InstructionDataEquals(HInstruction* other ATTRIBUTE_UNUSED) const OVERRIDE { return true; }
+
+  HInstruction* GetBaseAddress() const { return InputAt(0); }
+  HInstruction* GetOffset() const { return InputAt(1); }
+
+  DECLARE_INSTRUCTION(Arm64IntermediateAddress);
+
+ private:
+  DISALLOW_COPY_AND_ASSIGN(HArm64IntermediateAddress);
+};
+
+}  // namespace art
+
+#endif  // ART_COMPILER_OPTIMIZING_NODES_ARM64_H_