[simd.js] Add SIMD load functions for Phase 1.

src/runtime/runtime.h

Index: src/runtime/runtime.h
diff --git a/src/runtime/runtime.h b/src/runtime/runtime.h
index 0ff2b62c5735bf925d2ceb53ececca34a7e33646..574b0637d6af88b06ea112956a5a0f6cf0f4d755 100644
--- a/src/runtime/runtime.h
+++ b/src/runtime/runtime.h
@@ -614,6 +614,10 @@ namespace internal {
   F(Float32x4FromUint16x8Bits, 1, 1)           \
   F(Float32x4FromInt8x16Bits, 1, 1)            \
   F(Float32x4FromUint8x16Bits, 1, 1)           \
+  F(Float32x4Load, 2, 1)                       \
+  F(Float32x4Load1, 2, 1)                      \
+  F(Float32x4Load2, 2, 1)                      \
+  F(Float32x4Load3, 2, 1)                      \
   F(Int32x4Check, 1, 1)                        \
   F(Int32x4ExtractLane, 2, 1)                  \
   F(Int32x4ReplaceLane, 3, 1)                  \
@@ -646,6 +650,10 @@ namespace internal {
   F(Int32x4FromUint16x8Bits, 1, 1)             \
   F(Int32x4FromInt8x16Bits, 1, 1)              \
   F(Int32x4FromUint8x16Bits, 1, 1)             \
+  F(Int32x4Load, 2, 1)                         \
+  F(Int32x4Load1, 2, 1)                        \
+  F(Int32x4Load2, 2, 1)                        \
+  F(Int32x4Load3, 2, 1)                        \
   F(Uint32x4Check, 1, 1)                       \
   F(Uint32x4ExtractLane, 2, 1)                 \
   F(Uint32x4ReplaceLane, 3, 1)                 \
@@ -678,6 +686,10 @@ namespace internal {
   F(Uint32x4FromUint16x8Bits, 1, 1)            \
   F(Uint32x4FromInt8x16Bits, 1, 1)             \
   F(Uint32x4FromUint8x16Bits, 1, 1)            \
+  F(Uint32x4Load, 2, 1)                        \
+  F(Uint32x4Load1, 2, 1)                       \
+  F(Uint32x4Load2, 2, 1)                       \
+  F(Uint32x4Load3, 2, 1)                       \
   F(Bool32x4Check, 1, 1)                       \
   F(Bool32x4ExtractLane, 2, 1)                 \
   F(Bool32x4ReplaceLane, 3, 1)                 \
@@ -722,6 +734,7 @@ namespace internal {
   F(Int16x8FromUint16x8Bits, 1, 1)             \
   F(Int16x8FromInt8x16Bits, 1, 1)              \
   F(Int16x8FromUint8x16Bits, 1, 1)             \
+  F(Int16x8Load, 2, 1)                         \
   F(Uint16x8Check, 1, 1)                       \
   F(Uint16x8ExtractLane, 2, 1)                 \
   F(Uint16x8ReplaceLane, 3, 1)                 \
@@ -757,6 +770,7 @@ namespace internal {
   F(Uint16x8FromInt16x8Bits, 1, 1)             \
   F(Uint16x8FromInt8x16Bits, 1, 1)             \
   F(Uint16x8FromUint8x16Bits, 1, 1)            \
+  F(Uint16x8Load, 2, 1)                        \
   F(Bool16x8Check, 1, 1)                       \
   F(Bool16x8ExtractLane, 2, 1)                 \
   F(Bool16x8ReplaceLane, 3, 1)                 \
@@ -801,6 +815,7 @@ namespace internal {
   F(Int8x16FromInt16x8Bits, 1, 1)              \
   F(Int8x16FromUint16x8Bits, 1, 1)             \
   F(Int8x16FromUint8x16Bits, 1, 1)             \
+  F(Int8x16Load, 2, 1)                         \
   F(Uint8x16Check, 1, 1)                       \
   F(Uint8x16ExtractLane, 2, 1)                 \
   F(Uint8x16ReplaceLane, 3, 1)                 \
@@ -836,6 +851,7 @@ namespace internal {
   F(Uint8x16FromInt16x8Bits, 1, 1)             \
   F(Uint8x16FromUint16x8Bits, 1, 1)            \
   F(Uint8x16FromInt8x16Bits, 1, 1)             \
+  F(Uint8x16Load, 2, 1)                        \
   F(Bool8x16Check, 1, 1)                       \
   F(Bool8x16ExtractLane, 2, 1)                 \
   F(Bool8x16ReplaceLane, 3, 1)                 \