Implement UnalignedLoad and UnalignedStore in WASM using LoadByte/Shift/Or and StoreByte/Shift/And.

src/compiler/wasm-compiler.cc

Index: src/compiler/wasm-compiler.cc
diff --git a/src/compiler/wasm-compiler.cc b/src/compiler/wasm-compiler.cc
index 96557b8386b17eb98318e7682503f2d6c16315a9..07f861bb7f2b08afd0e5300bfe844f577ae0c5fa 100644
--- a/src/compiler/wasm-compiler.cc
+++ b/src/compiler/wasm-compiler.cc
@@ -2591,10 +2591,127 @@ void WasmGraphBuilder::BoundsCheckMem(MachineType memtype, Node* index,
   trap_->AddTrapIfFalse(wasm::kTrapMemOutOfBounds, cond, position);
 }
 
+Node* WasmGraphBuilder::BuildUnalignedLoad(wasm::LocalType type,
+                                           MachineType memtype, Node* index,
+                                           uint32_t offset,
+                                           uint32_t alignment) {
+  Node* result;
+  Node* load;
+  bool extendTo64Bit = false;
+
+  wasm::WasmOpcode shiftOpcode;
+  wasm::WasmOpcode orOpcode;
+  Node* shiftConst;
+
+  auto determineType = [](bool signExtend, uint32_t alignment) -> MachineType {

[titzer, 2016/05/11 15:22:35]

Can we lift this out to a helper method instead of a lambda?

+    switch (alignment) {
+      case 0:
+        return signExtend ? MachineType::Int8() : MachineType::Uint8();
+      case 1:
+        return signExtend ? MachineType::Int16() : MachineType::Uint16();
+      case 2:
+        return signExtend ? MachineType::Int32() : MachineType::Uint32();
+      default:
+        UNREACHABLE();
+    }
+  };
+
+  auto calculateOffset = [this](Node* baseOffset, int numberOfBytes, int stride,
+                                int current) -> Node* {
+    int offset;
+
+#if defined(V8_TARGET_LITTLE_ENDIAN)
+    offset = numberOfBytes - stride - current;
+#elif defined(V8_TARGET_BIG_ENDIAN)
+    offset = current;
+#else
+#error Unsupported endianness
+#endif
+
+    if (offset == 0) {
+      return baseOffset;
+    } else {
+      return Binop(wasm::kExprI32Add, baseOffset,
+                   jsgraph()->Int32Constant(offset));
+    }
+  };
+
+  bool signExtend = memtype.IsSigned();
+
+  bool isFloat = IsFloatingPoint(memtype.representation());
+  int stride =
+      1 << ElementSizeLog2Of(determineType(false, alignment).representation());
+  int numberOfBytes = 1 << ElementSizeLog2Of(memtype.representation());
+  int numberOfLoads = numberOfBytes / stride;
+
+  DCHECK(numberOfBytes % stride == 0);
+  DCHECK(numberOfLoads >= 2);
+
+  switch (type) {
+    case wasm::kAstI64:
+    case wasm::kAstF64:
+      shiftOpcode = wasm::kExprI64Shl;
+      orOpcode = wasm::kExprI64Ior;
+      result = jsgraph()->Int64Constant(0);
+      shiftConst = jsgraph()->Int64Constant(8 * stride);
+      extendTo64Bit = true;
+      break;
+    case wasm::kAstI32:
+    case wasm::kAstF32:
+      shiftOpcode = wasm::kExprI32Shl;
+      orOpcode = wasm::kExprI32Ior;
+      result = jsgraph()->Int32Constant(0);
+      shiftConst = jsgraph()->Int32Constant(8 * stride);
+      break;
+    default:
+      UNREACHABLE();
+  }
+
+  Node* baseOffset = MemBuffer(offset);
+
+  for (int i = 0; i < numberOfBytes; i += stride) {
+    result = Binop(shiftOpcode, result, shiftConst);
+    load = graph()->NewNode(
+        jsgraph()->machine()->Load(determineType(signExtend, alignment)),
+        calculateOffset(baseOffset, numberOfBytes, stride, i), index, *effect_,
+        *control_);
+    *effect_ = load;
+    if (extendTo64Bit) {
+      if (signExtend) {
+        load =
+            graph()->NewNode(jsgraph()->machine()->ChangeInt32ToInt64(), load);
+      } else {
+        load = graph()->NewNode(jsgraph()->machine()->ChangeUint32ToUint64(),
+                                load);
+      }
+    }
+    signExtend = false;
+    result = Binop(orOpcode, result, load);
+  }
+
+  // Convert to float
+  if (isFloat) {
+    switch (type) {
+      case wasm::kAstF32:
+        result = Unop(wasm::kExprF32ReinterpretI32, result);
+        break;
+      case wasm::kAstF64:
+        result = Unop(wasm::kExprF64ReinterpretI64, result);
+        break;
+      default:
+        UNREACHABLE();
+    }
+  }
+
+  return result;
+}
+
 Node* WasmGraphBuilder::LoadMem(wasm::LocalType type, MachineType memtype,
                                 Node* index, uint32_t offset,
+                                uint32_t alignment,
                                 wasm::WasmCodePosition position) {
   Node* load;
+  bool skip = false;
 
   if (module_ && module_->asm_js()) {
     // asm.js semantics use CheckedLoad (i.e. OOB reads return 0ish).
@@ -2605,11 +2722,21 @@ Node* WasmGraphBuilder::LoadMem(wasm::LocalType type, MachineType memtype,
   } else {
     // WASM semantics throw on OOB. Introduce explicit bounds check.
     BoundsCheckMem(memtype, index, offset, position);
-    load = graph()->NewNode(jsgraph()->machine()->Load(memtype),
-                            MemBuffer(offset), index, *effect_, *control_);
+    bool aligned = alignment >= ElementSizeLog2Of(memtype.representation());
+
+    if (aligned ||
+        jsgraph()->machine()->UnalignedLoadSupported(memtype, alignment)) {
+      load = graph()->NewNode(jsgraph()->machine()->Load(memtype),
+                              MemBuffer(offset), index, *effect_, *control_);
+    } else {
+      load = BuildUnalignedLoad(type, memtype, index, offset, alignment);
+      skip = true;
+    }
   }
 
-  *effect_ = load;
+  if (!skip) {

[titzer, 2016/05/11 15:22:35]

Can you lift this up into the branch instead of using a boolean?

+    *effect_ = load;
+  }
 
   if (type == wasm::kAstI64 &&
       ElementSizeLog2Of(memtype.representation()) < 3) {
@@ -2627,10 +2754,110 @@ Node* WasmGraphBuilder::LoadMem(wasm::LocalType type, MachineType memtype,
   return load;
 }
 
+Node* WasmGraphBuilder::BuildUnalignedStore(MachineType memtype, Node* index,
+                                            uint32_t offset, uint32_t alignment,
+                                            Node* val) {
+  Node* store;
+  Node* newValue;
+
+  wasm::WasmOpcode shiftOpcode;
+
+  Node* shiftConst;
+  bool extendTo64Bit = false;
+
+  auto determineType = [](uint32_t alignment) -> MachineType {
+    switch (alignment) {
+      case 0:
+        return MachineType::Uint8();
+      case 1:
+        return MachineType::Uint16();
+      case 2:
+        return MachineType::Uint32();
+      default:
+        UNREACHABLE();
+    }
+  };
+
+  auto calculateOffset = [this](Node* baseOffset, int numberOfBytes, int stride,
+                                int current) -> Node* {
+    int offset;
+#if defined(V8_TARGET_LITTLE_ENDIAN)
+    offset = current;
+#elif defined(V8_TARGET_BIG_ENDIAN)
+    offset = numberOfBytes - stride - current;
+#else
+#error Unsupported endianness
+#endif
+    if (offset == 0) {
+      return baseOffset;
+    } else {
+      return Binop(wasm::kExprI32Add, baseOffset,
+                   jsgraph()->Int32Constant(offset));
+    }
+  };
+
+  bool isFloat = IsFloatingPoint(memtype.representation());
+  int stride =
+      1 << ElementSizeLog2Of(determineType(alignment).representation());
+  int numberOfBytes = 1 << ElementSizeLog2Of(memtype.representation());
+  int numberOfLoads = numberOfBytes / stride;
+
+  DCHECK(numberOfBytes % stride == 0);
+  DCHECK(numberOfLoads >= 2);
+
+  StoreRepresentation rep(determineType(alignment).representation(),
+                          kNoWriteBarrier);
+
+  if (ElementSizeLog2Of(memtype.representation()) <= 2) {
+    shiftOpcode = wasm::kExprI32ShrU;
+    shiftConst = jsgraph()->Int32Constant(8 * stride);
+  } else {
+    shiftOpcode = wasm::kExprI64ShrU;
+    shiftConst = jsgraph()->Int64Constant(8 * stride);
+    extendTo64Bit = true;
+  }
+
+  newValue = val;
+  if (isFloat) {
+    switch (memtype.representation()) {
+      case MachineRepresentation::kFloat64:
+        newValue = Unop(wasm::kExprI64ReinterpretF64, val);
+        break;
+      case MachineRepresentation::kFloat32:
+        newValue = Unop(wasm::kExprI32ReinterpretF32, val);
+        break;
+      default:
+        UNREACHABLE();
+    }
+  }
+
+  Node* baseOffset = MemBuffer(offset);
+
+  for (int i = 0; i < numberOfBytes - stride; i += stride) {
+    store = graph()->NewNode(
+        jsgraph()->machine()->Store(rep),
+        calculateOffset(baseOffset, numberOfBytes, stride, i), index,
+        extendTo64Bit ? Unop(wasm::kExprI32ConvertI64, newValue) : newValue,
+        *effect_, *control_);
+    newValue = Binop(shiftOpcode, newValue, shiftConst);
+    *effect_ = store;
+  }
+  store = graph()->NewNode(
+      jsgraph()->machine()->Store(rep),
+      calculateOffset(baseOffset, numberOfBytes, stride,
+                      numberOfBytes - stride),
+      index,
+      extendTo64Bit ? Unop(wasm::kExprI32ConvertI64, newValue) : newValue,
+      *effect_, *control_);
+  *effect_ = store;
+  return val;
+}
+
 Node* WasmGraphBuilder::StoreMem(MachineType memtype, Node* index,
-                                 uint32_t offset, Node* val,
+                                 uint32_t offset, uint32_t alignment, Node* val,
                                  wasm::WasmCodePosition position) {
   Node* store;
+  bool skip = false;
   if (module_ && module_->asm_js()) {
     // asm.js semantics use CheckedStore (i.e. ignore OOB writes).
     DCHECK_EQ(0, offset);
@@ -2642,11 +2869,24 @@ Node* WasmGraphBuilder::StoreMem(MachineType memtype, Node* index,
     // WASM semantics throw on OOB. Introduce explicit bounds check.
     BoundsCheckMem(memtype, index, offset, position);
     StoreRepresentation rep(memtype.representation(), kNoWriteBarrier);
-    store =
-        graph()->NewNode(jsgraph()->machine()->Store(rep), MemBuffer(offset),
-                         index, val, *effect_, *control_);
+    bool aligned = alignment >= ElementSizeLog2Of(memtype.representation());
+
+    if (aligned ||
+        jsgraph()->machine()->UnalignedStoreSupported(memtype, alignment)) {
+      StoreRepresentation rep(memtype.representation(), kNoWriteBarrier);
+      store =
+          graph()->NewNode(jsgraph()->machine()->Store(rep), MemBuffer(offset),
+                           index, val, *effect_, *control_);
+    } else {
+      store = BuildUnalignedStore(memtype, index, offset, alignment, val);
+      skip = true;
+    }
   }
-  *effect_ = store;
+
+  if (!skip) {
+    *effect_ = store;
+  }
+
   return store;
 }
 
@@ -2877,7 +3117,8 @@ std::pair<JSGraph*, SourcePositionTable*> BuildGraphForWasmFunction(
   CommonOperatorBuilder* common = new (zone) CommonOperatorBuilder(zone);
   MachineOperatorBuilder* machine = new (zone) MachineOperatorBuilder(
       zone, MachineType::PointerRepresentation(),
-      InstructionSelector::SupportedMachineOperatorFlags());
+      InstructionSelector::SupportedMachineOperatorFlags(),
+      InstructionSelector::AlignmentConfiguration());
   JSGraph* jsgraph =
       new (zone) JSGraph(isolate, graph, common, nullptr, nullptr, machine);
   SourcePositionTable* source_position_table =