[compiler] [wasm] Introduce Word32/64ReverseBytes as TF Optional Opcode

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 6b7c99327b17dd7552cba36b5cab84ac8be5ce60..f3ec3468fe9a3e1468c389af555ed0f93c51e768 100644
--- a/src/compiler/wasm-compiler.cc
+++ b/src/compiler/wasm-compiler.cc
@@ -1053,44 +1053,65 @@ Node* WasmGraphBuilder::BuildChangeEndianness(Node* node, MachineType memtype,
   int i;
   uint32_t shiftCount;
 
-  for (i = 0, shiftCount = valueSizeInBits - 8; i < valueSizeInBits / 2;
-       i += 8, shiftCount -= 16) {
-    Node* shiftLower;
-    Node* shiftHigher;
-    Node* lowerByte;
-    Node* higherByte;
-
-    DCHECK(shiftCount > 0);
-    DCHECK((shiftCount + 8) % 16 == 0);
-
-    if (valueSizeInBits > 32) {
-      shiftLower = graph()->NewNode(shiftLeftOpcode, value,
-                                    jsgraph()->Int64Constant(shiftCount));
-      shiftHigher = graph()->NewNode(shiftRightOpcode, value,
-                                     jsgraph()->Int64Constant(shiftCount));
-      lowerByte = graph()->NewNode(
-          andOpcode, shiftLower,
-          jsgraph()->Int64Constant(static_cast<uint64_t>(0xFF)
-                                   << (valueSizeInBits - 8 - i)));
-      higherByte = graph()->NewNode(
-          andOpcode, shiftHigher,
-          jsgraph()->Int64Constant(static_cast<uint64_t>(0xFF) << i));
-    } else {
-      shiftLower = graph()->NewNode(shiftLeftOpcode, value,
-                                    jsgraph()->Int32Constant(shiftCount));
-      shiftHigher = graph()->NewNode(shiftRightOpcode, value,
-                                     jsgraph()->Int32Constant(shiftCount));
-      lowerByte = graph()->NewNode(
-          andOpcode, shiftLower,
-          jsgraph()->Int32Constant(static_cast<uint32_t>(0xFF)
-                                   << (valueSizeInBits - 8 - i)));
-      higherByte = graph()->NewNode(
-          andOpcode, shiftHigher,
-          jsgraph()->Int32Constant(static_cast<uint32_t>(0xFF) << i));
+  if (jsgraph()->machine()->ReverseBytesSupported(memtype)) {
+    switch (valueSizeInBytes) {
+      case 1:
+        break;
+      case 2:
+        result = graph()->NewNode(
+            jsgraph()->machine()->Word16ReverseBytes().op(), value);
+        break;
+      case 4:
+        result = graph()->NewNode(
+            jsgraph()->machine()->Word32ReverseBytes().op(), value);
+        break;
+      case 8:
+        result = graph()->NewNode(
+            jsgraph()->machine()->Word64ReverseBytes().op(), value);
+        break;
+      default:
+        UNREACHABLE();
     }
+  } else {
+    for (i = 0, shiftCount = valueSizeInBits - 8; i < valueSizeInBits / 2;
+         i += 8, shiftCount -= 16) {
+      Node* shiftLower;
+      Node* shiftHigher;
+      Node* lowerByte;
+      Node* higherByte;
+
+      DCHECK(shiftCount > 0);
+      DCHECK((shiftCount + 8) % 16 == 0);
+
+      if (valueSizeInBits > 32) {
+        shiftLower = graph()->NewNode(shiftLeftOpcode, value,
+                                      jsgraph()->Int64Constant(shiftCount));
+        shiftHigher = graph()->NewNode(shiftRightOpcode, value,
+                                       jsgraph()->Int64Constant(shiftCount));
+        lowerByte = graph()->NewNode(
+            andOpcode, shiftLower,
+            jsgraph()->Int64Constant(static_cast<uint64_t>(0xFF)
+                                     << (valueSizeInBits - 8 - i)));
+        higherByte = graph()->NewNode(
+            andOpcode, shiftHigher,
+            jsgraph()->Int64Constant(static_cast<uint64_t>(0xFF) << i));
+      } else {
+        shiftLower = graph()->NewNode(shiftLeftOpcode, value,
+                                      jsgraph()->Int32Constant(shiftCount));
+        shiftHigher = graph()->NewNode(shiftRightOpcode, value,
+                                       jsgraph()->Int32Constant(shiftCount));
+        lowerByte = graph()->NewNode(
+            andOpcode, shiftLower,
+            jsgraph()->Int32Constant(static_cast<uint32_t>(0xFF)
+                                     << (valueSizeInBits - 8 - i)));
+        higherByte = graph()->NewNode(
+            andOpcode, shiftHigher,
+            jsgraph()->Int32Constant(static_cast<uint32_t>(0xFF) << i));
+      }
 
-    result = graph()->NewNode(orOpcode, result, lowerByte);
-    result = graph()->NewNode(orOpcode, result, higherByte);
+      result = graph()->NewNode(orOpcode, result, lowerByte);
+      result = graph()->NewNode(orOpcode, result, higherByte);
+    }
   }
 
   if (isFloat) {
@@ -1119,7 +1140,11 @@ Node* WasmGraphBuilder::BuildChangeEndianness(Node* node, MachineType memtype,
         shiftBitCount = jsgraph()->Int32Constant(64 - valueSizeInBits);
         result = graph()->NewNode(
             m->Word64Sar(),
-            graph()->NewNode(m->Word64Shl(), result, shiftBitCount),
+            graph()->NewNode(
+                m->Word64Shl(),
+                graph()->NewNode(jsgraph()->machine()->ChangeInt32ToInt64(),
+                                 result),
+                shiftBitCount),
             shiftBitCount);
       } else if (wasmtype == wasm::kAstI32) {
         shiftBitCount = jsgraph()->Int32Constant(32 - valueSizeInBits);
@@ -3082,8 +3107,6 @@ Node* WasmGraphBuilder::LoadMem(wasm::LocalType type, MachineType memtype,
     load = BuildUnalignedLoad(type, memtype, index, offset, alignment);
   }
 #if defined(V8_TARGET_BIG_ENDIAN)
-  // TODO(john.yan) Implement byte swap turbofan operator
-  // and use it if available for better performance
   load = BuildChangeEndianness(load, memtype, type);
 #endif
 
@@ -3203,12 +3226,11 @@ 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);
+
   bool aligned = static_cast<int>(alignment) >=
                  ElementSizeLog2Of(memtype.representation());
 
 #if defined(V8_TARGET_BIG_ENDIAN)
-  // TODO(john.yan) Implement byte swap turbofan operator
-  // and use it if available for better performance
   val = BuildChangeEndianness(val, memtype);
 #endif