[arm][turbofan] Use NEON for unaligned float64 memory accesses

src/compiler/arm/instruction-selector-arm.cc

 // Copyright 2014 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/base/adapters.h"
 #include "src/base/bits.h"
 #include "src/compiler/instruction-selector-impl.h"
 #include "src/compiler/node-matchers.h"
 #include "src/compiler/node-properties.h"
 
@@ skipping 542 matching lines @@
   // support unaligned access. We support unaligned FP loads by loading to
   // integer registers first, then moving to the destination FP register.
   switch (load_rep.representation()) {
     case MachineRepresentation::kFloat32: {
       InstructionOperand temp = g.TempRegister();
       EmitLoad(this, opcode, &temp, base, index);
       Emit(kArmVmovF32U32, g.DefineAsRegister(node), temp);
       return;
     }
     case MachineRepresentation::kFloat64: {
-      // TODO(arm): use vld1.8 for this when NEON is available.
-      // Compute the address of the least-significant half of the FP value.
-      // We assume that the base node is unlikely to be an encodable immediate
-      // or the result of a shift operation, so only consider the addressing
-      // mode that should be used for the index node.
-      InstructionCode add_opcode = kArmAdd;
-      InstructionOperand inputs[3];
-      inputs[0] = g.UseRegister(base);
+      if (CpuFeatures::IsSupported(NEON)) {
+        InstructionOperand output = g.DefineAsRegister(node);
+        EmitLoad(this, kArmVld1F64, &output, base, index);
+      } else {
+        // Compute the address of the least-significant half of the FP value.
+        // We assume that the base node is unlikely to be an encodable immediate
+        // or the result of a shift operation, so only consider the addressing
+        // mode that should be used for the index node.
+        InstructionCode add_opcode = kArmAdd;
+        InstructionOperand inputs[3];
+        inputs[0] = g.UseRegister(base);
 
-      size_t input_count;
-      if (TryMatchImmediateOrShift(this, &add_opcode, index, &input_count,
-                                   &inputs[1])) {
-        // input_count has been set by TryMatchImmediateOrShift(), so increment
-        // it to account for the base register in inputs[0].
-        input_count++;
-      } else {
-        add_opcode |= AddressingModeField::encode(kMode_Operand2_R);
-        inputs[1] = g.UseRegister(index);
-        input_count = 2;  // Base register and index.
+        size_t input_count;
+        if (TryMatchImmediateOrShift(this, &add_opcode, index, &input_count,
+                                     &inputs[1])) {
+          // input_count has been set by TryMatchImmediateOrShift(), so
+          // increment it to account for the base register in inputs[0].
+          input_count++;
+        } else {
+          add_opcode |= AddressingModeField::encode(kMode_Operand2_R);
+          inputs[1] = g.UseRegister(index);
+          input_count = 2;  // Base register and index.
+        }
+
+        InstructionOperand addr = g.TempRegister();
+        Emit(add_opcode, 1, &addr, input_count, inputs);
+
+        // Load both halves and move to an FP register.
+        InstructionOperand fp_lo = g.TempRegister();
+        InstructionOperand fp_hi = g.TempRegister();
+        opcode |= AddressingModeField::encode(kMode_Offset_RI);
+        Emit(opcode, fp_lo, addr, g.TempImmediate(0));
+        Emit(opcode, fp_hi, addr, g.TempImmediate(4));
+        Emit(kArmVmovF64U32U32, g.DefineAsRegister(node), fp_lo, fp_hi);
       }
-
-      InstructionOperand addr = g.TempRegister();
-      Emit(add_opcode, 1, &addr, input_count, inputs);
-
-      // Load both halves and move to an FP register.
-      InstructionOperand fp_lo = g.TempRegister();
-      InstructionOperand fp_hi = g.TempRegister();
-      opcode |= AddressingModeField::encode(kMode_Offset_RI);
-      Emit(opcode, fp_lo, addr, g.TempImmediate(0));
-      Emit(opcode, fp_hi, addr, g.TempImmediate(4));
-      Emit(kArmVmovF64U32U32, g.DefineAsRegister(node), fp_lo, fp_hi);
       return;
     }
     default:
       // All other cases should support unaligned accesses.
       UNREACHABLE();
       return;
   }
 }
 
 void InstructionSelector::VisitUnalignedStore(Node* node) {
@@ skipping 13 matching lines @@
   // value to integer registers first, then storing to the destination address.
   switch (store_rep) {
     case MachineRepresentation::kFloat32: {
       inputs[input_count++] = g.TempRegister();
       Emit(kArmVmovU32F32, inputs[0], g.UseRegister(value));
       inputs[input_count++] = g.UseRegister(base);
       EmitStore(this, kArmStr, input_count, inputs, index);
       return;
     }
     case MachineRepresentation::kFloat64: {
-      // TODO(arm): use vst1.8 for this when NEON is available.
-      // Store a 64-bit floating point value using two 32-bit integer stores.
-      // Computing the store address here would require three live temporary
-      // registers (fp<63:32>, fp<31:0>, address), so compute base + 4 after
-      // storing the least-significant half of the value.
+      if (CpuFeatures::IsSupported(NEON)) {
+        inputs[input_count++] = g.UseRegister(value);
+        inputs[input_count++] = g.UseRegister(base);
+        EmitStore(this, kArmVst1F64, input_count, inputs, index);
+      } else {
+        // Store a 64-bit floating point value using two 32-bit integer stores.
+        // Computing the store address here would require three live temporary
+        // registers (fp<63:32>, fp<31:0>, address), so compute base + 4 after
+        // storing the least-significant half of the value.
 
-      // First, move the 64-bit FP value into two temporary integer registers.
-      InstructionOperand fp[] = {g.TempRegister(), g.TempRegister()};
-      inputs[input_count++] = g.UseRegister(value);
-      Emit(kArmVmovU32U32F64, arraysize(fp), fp, input_count,
-           inputs);
+        // First, move the 64-bit FP value into two temporary integer registers.
+        InstructionOperand fp[] = {g.TempRegister(), g.TempRegister()};
+        inputs[input_count++] = g.UseRegister(value);
+        Emit(kArmVmovU32U32F64, arraysize(fp), fp, input_count, inputs);
 
-      // Store the least-significant half.
-      inputs[0] = fp[0];  // Low 32-bits of FP value.
-      inputs[input_count++] = g.UseRegister(base);  // First store base address.
-      EmitStore(this, kArmStr, input_count, inputs, index);
+        // Store the least-significant half.
+        inputs[0] = fp[0];  // Low 32-bits of FP value.
+        inputs[input_count++] =
+            g.UseRegister(base);  // First store base address.
+        EmitStore(this, kArmStr, input_count, inputs, index);
 
-      // Store the most-significant half.
-      InstructionOperand base4 = g.TempRegister();
-      Emit(kArmAdd | AddressingModeField::encode(kMode_Operand2_I), base4,
-           g.UseRegister(base), g.TempImmediate(4));  // Compute base + 4.
-      inputs[0] = fp[1];  // High 32-bits of FP value.
-      inputs[1] = base4;  // Second store base + 4 address.
-      EmitStore(this, kArmStr, input_count, inputs, index);
+        // Store the most-significant half.
+        InstructionOperand base4 = g.TempRegister();
+        Emit(kArmAdd | AddressingModeField::encode(kMode_Operand2_I), base4,
+             g.UseRegister(base), g.TempImmediate(4));  // Compute base + 4.
+        inputs[0] = fp[1];  // High 32-bits of FP value.
+        inputs[1] = base4;  // Second store base + 4 address.
+        EmitStore(this, kArmStr, input_count, inputs, index);
+      }
       return;
     }
     default:
       // All other cases should support unaligned accesses.
       UNREACHABLE();
       return;
   }
 }
 
 void InstructionSelector::VisitCheckedLoad(Node* node) {
@@ skipping 1812 matching lines @@
   Vector<MachineType> req_aligned = Vector<MachineType>::New(2);
   req_aligned[0] = MachineType::Float32();
   req_aligned[1] = MachineType::Float64();
   return MachineOperatorBuilder::AlignmentRequirements::
       SomeUnalignedAccessUnsupported(req_aligned, req_aligned);
 }
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8