[wasm] Implement 128-bit endian swap for simd type

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 158c70fc5c03d9e1489ace784e16d2aac5e48dc8..dbe22b8e92cb9ef367b4b8df8fd5690a38b3d733 100644
--- a/src/compiler/wasm-compiler.cc
+++ b/src/compiler/wasm-compiler.cc
@@ -1066,6 +1066,7 @@ static bool ReverseBytesSupported(MachineOperatorBuilder* m,
                                   size_t size_in_bytes) {
   switch (size_in_bytes) {
     case 4:
+    case 16:
       return m->Word32ReverseBytes().IsSupported();
     case 8:
       return m->Word64ReverseBytes().IsSupported();
@@ -1102,6 +1103,9 @@ Node* WasmGraphBuilder::BuildChangeEndianness(Node* node, MachineType memtype,
       // No need to change endianness for byte size, return original node
       return node;
       break;
+    case MachineRepresentation::kSimd128:
+      DCHECK(ReverseBytesSupported(m, 16));
+      break;
     default:
       UNREACHABLE();
       break;
@@ -1124,6 +1128,27 @@ Node* WasmGraphBuilder::BuildChangeEndianness(Node* node, MachineType memtype,
       case 8:
         result = graph()->NewNode(m->Word64ReverseBytes().op(), value);
         break;
+      case 16: {
+        Node* byte_reversed_lanes[4];
+        for (int lane = 0; lane < 4; lane++) {
+          byte_reversed_lanes[lane] = graph()->NewNode(
+              m->Word32ReverseBytes().op(),
+              graph()->NewNode(jsgraph()->machine()->I32x4ExtractLane(lane),
+                               value));
+        }
+
+        // This is making a copy of the value.
+        result =
+            graph()->NewNode(jsgraph()->machine()->S128And(), value, value);

[Clemens Hammacher, 2017/04/28 09:45:47]

Wouldn't "result = value;" suffice? The I32x4ReplaceLane should generate a new
value, no?

[john.yan, 2017/04/28 15:01:28]

That's exactly what I did for the first time. I think it's fine if SIMD is
enabled. SIMD lowering for I32x4ReplaceLane would actually replaced the old lane
with the new value, but doesn't generate a new vector. So I was running in a
problem ending up messing with order with I32x4ExtractLane above.

+
+        for (int lane = 0; lane < 4; lane++) {
+          result =
+              graph()->NewNode(jsgraph()->machine()->I32x4ReplaceLane(3 - lane),
+                               result, byte_reversed_lanes[lane]);
+        }
+
+        break;
+      }
       default:
         UNREACHABLE();
     }