[ARM] Make Simd 128 bit load/store more like existing load/store.

- Renames kArmSimd128Load, kArmSimd128Store to kArmVld1S128,
  kArmVst1S128
- Handles the unaligned load/store cases.
LOG=N
BUG=v8:6020

Review-Url: https://codereview.chromium.org/2769083003
Cr-Commit-Position: refs/heads/master@{#44117}

src/compiler/arm/code-generator-arm.cc

@@ -1444,7 +1444,7 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmVld1F64: {
-      __ vld1(NeonSize::Neon8, NeonListOperand(i.OutputDoubleRegister()),
+      __ vld1(Neon8, NeonListOperand(i.OutputDoubleRegister()),
               NeonMemOperand(i.InputRegister(0)));
       break;
     }
@@ -1453,6 +1453,16 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
               NeonMemOperand(i.InputRegister(1)));
       break;
     }
+    case kArmVld1S128: {
+      __ vld1(Neon8, NeonListOperand(i.OutputSimd128Register()),
+              NeonMemOperand(i.InputRegister(0)));
+      break;
+    }
+    case kArmVst1S128: {
+      __ vst1(Neon8, NeonListOperand(i.InputSimd128Register(0)),
+              NeonMemOperand(i.InputRegister(1)));
+      break;
+    }
     case kArmVldrF64:
       __ vldr(i.OutputDoubleRegister(), i.InputOffset());
       DCHECK_EQ(LeaveCC, i.OutputSBit());
@@ -1992,18 +2002,6 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
               i.OutputSimd128Register());
       break;
     }
-    case kArmSimd128Load: {
-      MemOperand src = i.InputOffset();
-      __ vld1(Neon8, NeonListOperand(i.OutputSimd128Register()),
-              NeonMemOperand(src.rn(), src.rm()));
-      break;
-    }
-    case kArmSimd128Store: {
-      MemOperand src = i.InputOffset(1);
-      __ vst1(Neon8, NeonListOperand(i.InputSimd128Register(0)),
-              NeonMemOperand(src.rn(), src.rm()));
-      break;
-    }
     case kArmSimd128And: {
       __ vand(i.OutputSimd128Register(), i.InputSimd128Register(0),
               i.InputSimd128Register(1));

src/compiler/arm/instruction-codes-arm.h

@@ -107,6 +107,8 @@ namespace compiler {
   V(ArmVld1F64)                    \
   V(ArmVstrF64)                    \
   V(ArmVst1F64)                    \
+  V(ArmVld1S128)                   \
+  V(ArmVst1S128)                   \
   V(ArmFloat32Max)                 \
   V(ArmFloat64Max)                 \
   V(ArmFloat32Min)                 \
@@ -213,8 +215,6 @@ namespace compiler {
   V(ArmUint8x16LessThan)           \
   V(ArmUint8x16LessThanOrEqual)    \
   V(ArmSimd128Zero)                \
-  V(ArmSimd128Load)                \
-  V(ArmSimd128Store)               \
   V(ArmSimd128And)                 \
   V(ArmSimd128Or)                  \
   V(ArmSimd128Xor)                 \

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

@@ -199,8 +199,6 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kArmUint8x16LessThan:
     case kArmUint8x16LessThanOrEqual:
     case kArmSimd128Zero:
-    case kArmSimd128Load:
-    case kArmSimd128Store:
     case kArmSimd128And:
     case kArmSimd128Or:
     case kArmSimd128Xor:
@@ -217,6 +215,7 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kArmVldrF32:
     case kArmVldrF64:
     case kArmVld1F64:
+    case kArmVld1S128:
     case kArmLdrb:
     case kArmLdrsb:
     case kArmLdrh:
@@ -227,6 +226,7 @@ int InstructionScheduler::GetTargetInstructionFlags(
     case kArmVstrF32:
     case kArmVstrF64:
     case kArmVst1F64:
+    case kArmVst1S128:
     case kArmStrb:
     case kArmStrh:
     case kArmStr:

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

@@ -427,7 +427,7 @@ void InstructionSelector::VisitLoad(Node* node) {
       opcode = kArmLdr;
       break;
     case MachineRepresentation::kSimd128:
-      opcode = kArmSimd128Load;
+      opcode = kArmVld1S128;
       break;
     case MachineRepresentation::kWord64:   // Fall through.
     case MachineRepresentation::kSimd1x4:  // Fall through.
@@ -517,7 +517,7 @@ void InstructionSelector::VisitStore(Node* node) {
         opcode = kArmStr;
         break;
       case MachineRepresentation::kSimd128:
-        opcode = kArmSimd128Store;
+        opcode = kArmVst1S128;
         break;
       case MachineRepresentation::kWord64:   // Fall through.
       case MachineRepresentation::kSimd1x4:  // Fall through.
@@ -542,8 +542,8 @@ void InstructionSelector::VisitProtectedStore(Node* node) {
 }
 
 void InstructionSelector::VisitUnalignedLoad(Node* node) {
-  UnalignedLoadRepresentation load_rep =
+  MachineRepresentation load_rep =
-      UnalignedLoadRepresentationOf(node->op());
+      UnalignedLoadRepresentationOf(node->op()).representation();
   ArmOperandGenerator g(this);
   Node* base = node->InputAt(0);
   Node* index = node->InputAt(1);
@@ -551,16 +551,18 @@ void InstructionSelector::VisitUnalignedLoad(Node* node) {
   InstructionCode opcode = kArmLdr;
   // Only floating point loads need to be specially handled; integer loads
   // support unaligned access. We support unaligned FP loads by loading to
-  // integer registers first, then moving to the destination FP register.
+  // integer registers first, then moving to the destination FP register. If
-  switch (load_rep.representation()) {
+  // NEON is supported, we use the vld1.8 instruction.
+  switch (load_rep) {
     case MachineRepresentation::kFloat32: {
       InstructionOperand temp = g.TempRegister();
       EmitLoad(this, opcode, &temp, base, index);
       Emit(kArmVmovF32U32, g.DefineAsRegister(node), temp);
       return;
     }
-    case MachineRepresentation::kFloat64: {
+    case MachineRepresentation::kFloat64:
-      // Compute the address of the least-significant half of the FP value.
+    case MachineRepresentation::kSimd128: {
+      // Compute the address of the least-significant byte 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.
@@ -585,8 +587,12 @@ void InstructionSelector::VisitUnalignedLoad(Node* node) {
 
       if (CpuFeatures::IsSupported(NEON)) {
         // With NEON we can load directly from the calculated address.
-        Emit(kArmVld1F64, g.DefineAsRegister(node), addr);
+        ArchOpcode op = load_rep == MachineRepresentation::kFloat64
+                            ? kArmVld1F64
+                            : kArmVld1S128;
+        Emit(op, g.DefineAsRegister(node), addr);
       } else {
+        DCHECK_NE(MachineRepresentation::kSimd128, load_rep);
         // Load both halves and move to an FP register.
         InstructionOperand fp_lo = g.TempRegister();
         InstructionOperand fp_hi = g.TempRegister();
@@ -619,6 +625,7 @@ void InstructionSelector::VisitUnalignedStore(Node* node) {
   // Only floating point stores need to be specially handled; integer stores
   // support unaligned access. We support unaligned FP stores by moving the
   // value to integer registers first, then storing to the destination address.
+  // If NEON is supported, we use the vst1.8 instruction.
   switch (store_rep) {
     case MachineRepresentation::kFloat32: {
       inputs[input_count++] = g.TempRegister();
@@ -627,7 +634,8 @@ void InstructionSelector::VisitUnalignedStore(Node* node) {
       EmitStore(this, kArmStr, input_count, inputs, index);
       return;
     }
-    case MachineRepresentation::kFloat64: {
+    case MachineRepresentation::kFloat64:
+    case MachineRepresentation::kSimd128: {
       if (CpuFeatures::IsSupported(NEON)) {
         InstructionOperand address = g.TempRegister();
         {
@@ -653,8 +661,12 @@ void InstructionSelector::VisitUnalignedStore(Node* node) {
 
         inputs[input_count++] = g.UseRegister(value);
         inputs[input_count++] = address;
-        Emit(kArmVst1F64, 0, nullptr, input_count, inputs);
+        ArchOpcode op = store_rep == MachineRepresentation::kFloat64
+                            ? kArmVst1F64
+                            : kArmVst1S128;
+        Emit(op, 0, nullptr, input_count, inputs);
       } else {
+        DCHECK_NE(MachineRepresentation::kSimd128, store_rep);
         // 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