Implement WASM big-endian support

src/compiler/wasm-compiler.cc

 // Copyright 2015 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/compiler/wasm-compiler.h"
 
 #include "src/isolate-inl.h"
 
 #include "src/base/platform/elapsed-timer.h"
 #include "src/base/platform/platform.h"
@@ skipping 974 matching lines @@
       int64_t masked = (match.Value() & kMask64);
       if (match.Value() != masked) node = jsgraph()->Int64Constant(masked);
     } else {
       node = graph()->NewNode(jsgraph()->machine()->Word64And(), node,
                               jsgraph()->Int64Constant(kMask64));
     }
   }
   return node;
 }
 
+Node* WasmGraphBuilder::BuildChangeEndianness(Node* node, MachineType memtype,
+                                              wasm::LocalType wasmtype) {
+  Node* result;
+  Node* value = node;
+  wasm::WasmOpcode shiftLeftOpcode;
+  wasm::WasmOpcode shiftRightOpcode;
+  wasm::WasmOpcode andOpcode, orOpcode;
+  int valueSizeInBytes = 1 << ElementSizeLog2Of(memtype.representation());
+  int valueSizeInBits = 8 * valueSizeInBytes;
+  bool isFloat = false;
+
+  switch (memtype.representation()) {
+    case MachineRepresentation::kFloat64:
+      value = Unop(wasm::kExprI64ReinterpretF64, node);
+      isFloat = true;
+    case MachineRepresentation::kWord64:
+      shiftLeftOpcode = wasm::kExprI64Shl;
+      shiftRightOpcode = wasm::kExprI64ShrU;
+      andOpcode = wasm::kExprI64And;
+      orOpcode = wasm::kExprI64Ior;
+      result = jsgraph()->Int64Constant(0);
+      break;
+    case MachineRepresentation::kFloat32:
+      value = Unop(wasm::kExprI32ReinterpretF32, node);
+      isFloat = true;
+    case MachineRepresentation::kWord32:
+    case MachineRepresentation::kWord16:
+      shiftLeftOpcode = wasm::kExprI32Shl;
+      shiftRightOpcode = wasm::kExprI32ShrU;
+      andOpcode = wasm::kExprI32And;
+      orOpcode = wasm::kExprI32Ior;
+      result = jsgraph()->Int32Constant(0);
+      break;
+    case MachineRepresentation::kWord8:
+      // No need to change endianness for byte size, return original node
+      return node;
+      break;
+    default:
+      UNREACHABLE();
+      break;
+  }
+
+  int i;
+  uint32_t shiftCount;
+
+  for (i = 0, shiftCount = valueSizeInBits - 8; i < valueSizeInBits / 2;
+       i += 8, shiftCount -= 16) {
+    DCHECK(shiftCount > 0);
+    DCHECK((shiftCount + 8) % 16 == 0);
+
+    Node* shiftLower =
+        Binop(shiftLeftOpcode, value, jsgraph()->Int32Constant(shiftCount));

[titzer, 2016/06/06 07:44:00]

Reusing the Binop() routine is a little inefficient, since it goes through
another switch and also the logic to mask shift counts if necessary. Can you
build the shift node directly here?

+    Node* shiftHigher =
+        Binop(shiftRightOpcode, value, jsgraph()->Int32Constant(shiftCount));
+
+    Node* lowerByte;
+    Node* higherByte;
+
+    if (valueSizeInBits > 32) {
+      lowerByte =
+          Binop(andOpcode, shiftLower,
+                jsgraph()->Int64Constant(0xFFl << (valueSizeInBits - 8 - i)));
+      higherByte =
+          Binop(andOpcode, shiftHigher, jsgraph()->Int64Constant(0xFFl << i));
+    } else {
+      lowerByte =
+          Binop(andOpcode, shiftLower,
+                jsgraph()->Int32Constant(0xFF << (valueSizeInBits - 8 - i)));
+      higherByte =
+          Binop(andOpcode, shiftHigher, jsgraph()->Int32Constant(0xFF << i));
+    }
+
+    result = Binop(orOpcode, result, lowerByte);
+    result = Binop(orOpcode, result, higherByte);
+  }
+
+  if (isFloat) {
+    switch (memtype.representation()) {
+      case MachineRepresentation::kFloat64:
+        result = Unop(wasm::kExprF64ReinterpretI64, result);
+        break;
+      case MachineRepresentation::kFloat32:
+        result = Unop(wasm::kExprF32ReinterpretI32, result);
+        break;
+      default:
+        UNREACHABLE();
+        break;
+    }
+  }
+
+  // We need to sign extend the value
+  if (memtype.IsSigned()) {
+    DCHECK(!isFloat);
+    if ((wasmtype == wasm::kAstI64 || wasmtype == wasm::kAstI32) &&
+        valueSizeInBits < 32) {
+      uint32_t mask = 1 << (valueSizeInBits - 1);
+
+      Node* signBit =
+          Binop(wasm::kExprI32And, result, jsgraph()->Int32Constant(mask));
+      Diamond d(graph(), jsgraph()->common(), Unop(wasm::kExprI32Eqz, signBit));

[titzer, 2016/06/06 07:44:00]

What about using (x << machine_width - width) >> (machine_width - width)?

+
+      result =
+          d.Phi(MachineRepresentation::kWord32, result,
+                Binop(wasm::kExprI32Ior, result,
+                      jsgraph()->Int32Constant(0xFFFFFFFF << valueSizeInBits)));
+    }
+  }
+
+  return result;
+}
+
 Node* WasmGraphBuilder::BuildF32Neg(Node* input) {
   Node* result =
       Unop(wasm::kExprF32ReinterpretI32,
            Binop(wasm::kExprI32Xor, Unop(wasm::kExprI32ReinterpretF32, input),
                  jsgraph()->Int32Constant(0x80000000)));
 
   return result;
 }
 
 Node* WasmGraphBuilder::BuildF64Neg(Node* input) {
@@ skipping 1756 matching lines @@
                  ElementSizeLog2Of(memtype.representation());
 
   if (aligned ||
       jsgraph()->machine()->UnalignedLoadSupported(memtype, alignment)) {
     load = graph()->NewNode(jsgraph()->machine()->Load(memtype),
                             MemBuffer(offset), index, *effect_, *control_);
     *effect_ = load;
   } else {
     load = BuildUnalignedLoad(type, memtype, index, offset, alignment);
   }
+#if defined(V8_TARGET_BIG_ENDIAN)
+  // TODO(mips-team) Implement byte swap turbofan operator

[john.yan, 2016/06/07 17:47:00]

FYI: https://codereview.chromium.org/2045943002/

+  // and use it if available for better performance
+  load = BuildChangeEndianness(load, memtype, type);
+#endif
 
   if (type == wasm::kAstI64 &&
       ElementSizeLog2Of(memtype.representation()) < 3) {
     // TODO(titzer): TF zeroes the upper bits of 64-bit loads for subword sizes.
     if (memtype.IsSigned()) {
       // sign extend
       load = graph()->NewNode(jsgraph()->machine()->ChangeInt32ToInt64(), load);
     } else {
       // zero extend
       load =
@@ skipping 100 matching lines @@
                                  uint32_t offset, uint32_t alignment, Node* val,
                                  wasm::WasmCodePosition position) {
   Node* store;
 
   // WASM semantics throw on OOB. Introduce explicit bounds check.
   BoundsCheckMem(memtype, index, offset, position);
   StoreRepresentation rep(memtype.representation(), kNoWriteBarrier);
   bool aligned = static_cast<int>(alignment) >=
                  ElementSizeLog2Of(memtype.representation());
 
+#if defined(V8_TARGET_BIG_ENDIAN)
+  // TODO(mips-team) Implement byte swap turbofan operator
+  // and use it if available for better performance
+  val = BuildChangeEndianness(val, memtype);
+#endif
+
   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_);
     *effect_ = store;
   } else {
     store = BuildUnalignedStore(memtype, index, offset, alignment, val);
   }
@@ skipping 404 matching lines @@
                             function_->code_start_offset),
            compile_ms);
   }
 
   return code;
 }
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8