[simd.js] Add SIMD store 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 7a43c847365ab62737ec2de1b5ba03000b00b93a..18faeb8741da6f3a757de254aef325cd9167bcb0 100644
--- a/src/runtime/runtime.h
+++ b/src/runtime/runtime.h
@@ -615,6 +615,10 @@ namespace internal {
   F(Float32x4Load1, 2, 1)                      \
   F(Float32x4Load2, 2, 1)                      \
   F(Float32x4Load3, 2, 1)                      \
+  F(Float32x4Store, 3, 1)                      \
+  F(Float32x4Store1, 3, 1)                     \
+  F(Float32x4Store2, 3, 1)                     \
+  F(Float32x4Store3, 3, 1)                     \
   F(Int32x4Check, 1, 1)                        \
   F(Int32x4ExtractLane, 2, 1)                  \
   F(Int32x4ReplaceLane, 3, 1)                  \
@@ -651,6 +655,10 @@ namespace internal {
   F(Int32x4Load1, 2, 1)                        \
   F(Int32x4Load2, 2, 1)                        \
   F(Int32x4Load3, 2, 1)                        \
+  F(Int32x4Store, 3, 1)                        \
+  F(Int32x4Store1, 3, 1)                       \
+  F(Int32x4Store2, 3, 1)                       \
+  F(Int32x4Store3, 3, 1)                       \
   F(Uint32x4Check, 1, 1)                       \
   F(Uint32x4ExtractLane, 2, 1)                 \
   F(Uint32x4ReplaceLane, 3, 1)                 \
@@ -687,6 +695,10 @@ namespace internal {
   F(Uint32x4Load1, 2, 1)                       \
   F(Uint32x4Load2, 2, 1)                       \
   F(Uint32x4Load3, 2, 1)                       \
+  F(Uint32x4Store, 3, 1)                       \
+  F(Uint32x4Store1, 3, 1)                      \
+  F(Uint32x4Store2, 3, 1)                      \
+  F(Uint32x4Store3, 3, 1)                      \
   F(Bool32x4Check, 1, 1)                       \
   F(Bool32x4ExtractLane, 2, 1)                 \
   F(Bool32x4ReplaceLane, 3, 1)                 \
@@ -732,6 +744,7 @@ namespace internal {
   F(Int16x8FromInt8x16Bits, 1, 1)              \
   F(Int16x8FromUint8x16Bits, 1, 1)             \
   F(Int16x8Load, 2, 1)                         \
+  F(Int16x8Store, 3, 1)                        \
   F(Uint16x8Check, 1, 1)                       \
   F(Uint16x8ExtractLane, 2, 1)                 \
   F(Uint16x8ReplaceLane, 3, 1)                 \
@@ -768,6 +781,7 @@ namespace internal {
   F(Uint16x8FromInt8x16Bits, 1, 1)             \
   F(Uint16x8FromUint8x16Bits, 1, 1)            \
   F(Uint16x8Load, 2, 1)                        \
+  F(Uint16x8Store, 3, 1)                       \
   F(Bool16x8Check, 1, 1)                       \
   F(Bool16x8ExtractLane, 2, 1)                 \
   F(Bool16x8ReplaceLane, 3, 1)                 \
@@ -813,6 +827,7 @@ namespace internal {
   F(Int8x16FromUint16x8Bits, 1, 1)             \
   F(Int8x16FromUint8x16Bits, 1, 1)             \
   F(Int8x16Load, 2, 1)                         \
+  F(Int8x16Store, 3, 1)                        \
   F(Uint8x16Check, 1, 1)                       \
   F(Uint8x16ExtractLane, 2, 1)                 \
   F(Uint8x16ReplaceLane, 3, 1)                 \
@@ -849,6 +864,7 @@ namespace internal {
   F(Uint8x16FromUint16x8Bits, 1, 1)            \
   F(Uint8x16FromInt8x16Bits, 1, 1)             \
   F(Uint8x16Load, 2, 1)                        \
+  F(Uint8x16Store, 3, 1)                       \
   F(Bool8x16Check, 1, 1)                       \
   F(Bool8x16ExtractLane, 2, 1)                 \
   F(Bool8x16ReplaceLane, 3, 1)                 \