diff --git a/src/builtins.ts b/src/builtins.ts
index 810aea92e1..a31fed61b9 100644
--- a/src/builtins.ts
+++ b/src/builtins.ts
@@ -541,6 +541,8 @@ export namespace BuiltinNames {
   export const i16x8_extmul_low_i8x16_u = "~lib/builtins/i16x8.extmul_low_i8x16_u";
   export const i16x8_extmul_high_i8x16_s = "~lib/builtins/i16x8.extmul_high_i8x16_s";
   export const i16x8_extmul_high_i8x16_u = "~lib/builtins/i16x8.extmul_high_i8x16_u";
+  export const i16x8_shuffle = "~lib/builtins/i16x8.shuffle";
+  export const i16x8_swizzle = "~lib/builtins/i16x8.swizzle";
 
   export const i32x4_splat = "~lib/builtins/i32x4.splat";
   export const i32x4_extract_lane = "~lib/builtins/i32x4.extract_lane";
@@ -584,6 +586,8 @@ export namespace BuiltinNames {
   export const i32x4_extmul_low_i16x8_u = "~lib/builtins/i32x4.extmul_low_i16x8_u";
   export const i32x4_extmul_high_i16x8_s = "~lib/builtins/i32x4.extmul_high_i16x8_s";
   export const i32x4_extmul_high_i16x8_u = "~lib/builtins/i32x4.extmul_high_i16x8_u";
+  export const i32x4_shuffle = "~lib/builtins/i32x4.shuffle";
+  export const i32x4_swizzle = "~lib/builtins/i32x4.swizzle";
 
   export const i64x2_splat = "~lib/builtins/i64x2.splat";
   export const i64x2_extract_lane = "~lib/builtins/i64x2.extract_lane";
@@ -616,6 +620,8 @@ export namespace BuiltinNames {
   export const i64x2_extmul_low_i32x4_u = "~lib/builtins/i64x2.extmul_low_i32x4_u";
   export const i64x2_extmul_high_i32x4_s = "~lib/builtins/i64x2.extmul_high_i32x4_s";
   export const i64x2_extmul_high_i32x4_u = "~lib/builtins/i64x2.extmul_high_i32x4_u";
+  export const i64x2_shuffle = "~lib/builtins/i64x2.shuffle";
+  export const i64x2_swizzle = "~lib/builtins/i64x2.swizzle";
 
   export const f32x4_splat = "~lib/builtins/f32x4.splat";
   export const f32x4_extract_lane = "~lib/builtins/f32x4.extract_lane";
@@ -644,6 +650,8 @@ export namespace BuiltinNames {
   export const f32x4_convert_i32x4_s = "~lib/builtins/f32x4.convert_i32x4_s";
   export const f32x4_convert_i32x4_u = "~lib/builtins/f32x4.convert_i32x4_u";
   export const f32x4_demote_f64x2_zero = "~lib/builtins/f32x4.demote_f64x2_zero";
+  export const f32x4_shuffle = "~lib/builtins/f32x4.shuffle";
+  export const f32x4_swizzle = "~lib/builtins/f32x4.swizzle";
 
   export const f64x2_splat = "~lib/builtins/f64x2.splat";
   export const f64x2_extract_lane = "~lib/builtins/f64x2.extract_lane";
@@ -672,6 +680,8 @@ export namespace BuiltinNames {
   export const f64x2_convert_low_i32x4_s = "~lib/builtins/f64x2.convert_low_i32x4_s";
   export const f64x2_convert_low_i32x4_u = "~lib/builtins/f64x2.convert_low_i32x4_u";
   export const f64x2_promote_low_f32x4 = "~lib/builtins/f64x2.promote_low_f32x4";
+  export const f64x2_shuffle = "~lib/builtins/f64x2.shuffle";
+  export const f64x2_swizzle = "~lib/builtins/f64x2.swizzle";
 
   export const i31_new = "~lib/builtins/i31.new";
   export const i31_get = "~lib/builtins/i31.get";
@@ -8337,7 +8347,6 @@ builtins.set(BuiltinNames.i8x16_shuffle, builtin_i8x16_shuffle);
 
 // i8x16.swizzle -> v128.swizzle
 function builtin_i8x16_swizzle(ctx: BuiltinContext): ExpressionRef {
-  checkTypeAbsent(ctx);
   ctx.typeArguments = null;
   ctx.contextualType = Type.v128;
   return builtin_v128_swizzle(ctx);
@@ -8758,6 +8767,23 @@ function builtin_i16x8_extmul_high_i8x16_u(ctx: BuiltinContext): ExpressionRef {
 }
 builtins.set(BuiltinNames.i16x8_extmul_high_i8x16_u, builtin_i16x8_extmul_high_i8x16_u);
 
+// i16x8.shuffle -> v128.shuffle<i16>
+function builtin_i16x8_shuffle(ctx: BuiltinContext): ExpressionRef {
+  checkTypeAbsent(ctx);
+  ctx.typeArguments = [ Type.i16 ];
+  ctx.contextualType = Type.v128;
+  return builtin_v128_shuffle(ctx);
+}
+builtins.set(BuiltinNames.i16x8_shuffle, builtin_i16x8_shuffle);
+
+// i16x8.swizzle -> v128.swizzle
+function builtin_i16x8_swizzle(ctx: BuiltinContext): ExpressionRef {
+  ctx.typeArguments = null;
+  ctx.contextualType = Type.v128;
+  return builtin_v128_swizzle(ctx);
+}
+builtins.set(BuiltinNames.i16x8_swizzle, builtin_i16x8_swizzle);
+
 // i32x4.splat -> v128.splat<i32>
 function builtin_i32x4_splat(ctx: BuiltinContext): ExpressionRef {
   checkTypeAbsent(ctx);
@@ -9136,6 +9162,23 @@ function builtin_i32x4_extmul_high_i16x8_u(ctx: BuiltinContext): ExpressionRef {
 }
 builtins.set(BuiltinNames.i32x4_extmul_high_i16x8_u, builtin_i32x4_extmul_high_i16x8_u);
 
+// i32x4.shuffle -> v128.shuffle<i32>
+function builtin_i32x4_shuffle(ctx: BuiltinContext): ExpressionRef {
+  checkTypeAbsent(ctx);
+  ctx.typeArguments = [ Type.i32 ];
+  ctx.contextualType = Type.v128;
+  return builtin_v128_shuffle(ctx);
+}
+builtins.set(BuiltinNames.i32x4_shuffle, builtin_i32x4_shuffle);
+
+// i32x4.swizzle -> v128.swizzle
+function builtin_i32x4_swizzle(ctx: BuiltinContext): ExpressionRef {
+  ctx.typeArguments = null;
+  ctx.contextualType = Type.v128;
+  return builtin_v128_swizzle(ctx);
+}
+builtins.set(BuiltinNames.i32x4_swizzle, builtin_i32x4_swizzle);
+
 // i64x2.splat -> v128.splat<i64>
 function builtin_i64x2_splat(ctx: BuiltinContext): ExpressionRef {
   checkTypeAbsent(ctx);
@@ -9379,6 +9422,23 @@ function builtin_i64x2_extmul_high_i32x4_u(ctx: BuiltinContext): ExpressionRef {
 }
 builtins.set(BuiltinNames.i64x2_extmul_high_i32x4_u, builtin_i64x2_extmul_high_i32x4_u);
 
+// i64x2.shuffle -> v128.shuffle<i64>
+function builtin_i64x2_shuffle(ctx: BuiltinContext): ExpressionRef {
+  checkTypeAbsent(ctx);
+  ctx.typeArguments = [ Type.i64 ];
+  ctx.contextualType = Type.v128;
+  return builtin_v128_shuffle(ctx);
+}
+builtins.set(BuiltinNames.i64x2_shuffle, builtin_i64x2_shuffle);
+
+// i64x2.swizzle -> v128.swizzle
+function builtin_i64x2_swizzle(ctx: BuiltinContext): ExpressionRef {
+  ctx.typeArguments = null;
+  ctx.contextualType = Type.v128;
+  return builtin_v128_swizzle(ctx);
+}
+builtins.set(BuiltinNames.i64x2_swizzle, builtin_i64x2_swizzle);
+
 // f32x4.splat -> v128.splat<f32>
 function builtin_f32x4_splat(ctx: BuiltinContext): ExpressionRef {
   checkTypeAbsent(ctx);
@@ -9622,6 +9682,23 @@ function builtin_f32x4_demote_f64x2_zero(ctx: BuiltinContext): ExpressionRef {
 }
 builtins.set(BuiltinNames.f32x4_demote_f64x2_zero, builtin_f32x4_demote_f64x2_zero);
 
+// f32x4.shuffle -> v128.shuffle<f32>
+function builtin_f32x4_shuffle(ctx: BuiltinContext): ExpressionRef {
+  checkTypeAbsent(ctx);
+  ctx.typeArguments = [ Type.f32 ];
+  ctx.contextualType = Type.v128;
+  return builtin_v128_shuffle(ctx);
+}
+builtins.set(BuiltinNames.f32x4_shuffle, builtin_f32x4_shuffle);
+
+// f32x4.swizzle -> v128.swizzle
+function builtin_f32x4_swizzle(ctx: BuiltinContext): ExpressionRef {
+  ctx.typeArguments = null;
+  ctx.contextualType = Type.v128;
+  return builtin_v128_swizzle(ctx);
+}
+builtins.set(BuiltinNames.f32x4_swizzle, builtin_f32x4_swizzle);
+
 // f64x2.splat -> v128.splat<f64>
 function builtin_f64x2_splat(ctx: BuiltinContext): ExpressionRef {
   checkTypeAbsent(ctx);
@@ -9865,6 +9942,23 @@ function builtin_f64x4_promote_low_f32x4(ctx: BuiltinContext): ExpressionRef {
 }
 builtins.set(BuiltinNames.f64x2_promote_low_f32x4, builtin_f64x4_promote_low_f32x4);
 
+// f64x2.shuffle -> v128.shuffle<f32>
+function builtin_f64x2_shuffle(ctx: BuiltinContext): ExpressionRef {
+  checkTypeAbsent(ctx);
+  ctx.typeArguments = [ Type.f64 ];
+  ctx.contextualType = Type.v128;
+  return builtin_v128_shuffle(ctx);
+}
+builtins.set(BuiltinNames.f64x2_shuffle, builtin_f64x2_shuffle);
+
+// f64x2.swizzle -> v128.swizzle
+function builtin_f64x2_swizzle(ctx: BuiltinContext): ExpressionRef {
+  ctx.typeArguments = null;
+  ctx.contextualType = Type.v128;
+  return builtin_v128_swizzle(ctx);
+}
+builtins.set(BuiltinNames.f64x2_swizzle, builtin_f64x2_swizzle);
+
 // === Internal helpers =======================================================================
 
 /** Compiles the `visit_globals` function. */
diff --git a/std/assembly/builtins.ts b/std/assembly/builtins.ts
index 3cbe42a94c..6f27011722 100644
--- a/std/assembly/builtins.ts
+++ b/std/assembly/builtins.ts
@@ -1177,7 +1177,7 @@ export namespace v128 {
 
   // @ts-ignore: decorator
   @builtin
-  export declare function swizzle(a: v128, s: v128): v128;
+  export declare function swizzle(a: v128, b: v128): v128;
 
   // @ts-ignore: decorator
   @unsafe @builtin
@@ -1647,7 +1647,7 @@ export namespace i8x16 {
 
   // @ts-ignore: decorator
   @builtin
-  export declare function swizzle(a: v128, s: v128): v128;
+  export declare function swizzle(a: v128, b: v128): v128;
 }
 
 // @ts-ignore: decorator
@@ -1839,6 +1839,17 @@ export namespace i16x8 {
   // @ts-ignore: decorator
   @builtin
   export declare function extmul_high_i8x16_u(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function shuffle(
+    a: v128, b: v128,
+    l0: u8, l1: u8, l2: u8, l3: u8, l4: u8, l5: u8, l6: u8, l7: u8
+  ): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function swizzle(a: v128, b: v128): v128;
 }
 
 // @ts-ignore: decorator
@@ -2014,6 +2025,14 @@ export namespace i32x4 {
   // @ts-ignore: decorator
   @builtin
   export declare function extmul_high_i16x8_u(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function shuffle(a: v128, b: v128, l0: u8, l1: u8, l2: u8, l3: u8): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function swizzle(a: v128, b: v128): v128;
 }
 
 // @ts-ignore: decorator
@@ -2129,6 +2148,14 @@ export namespace i64x2 {
   // @ts-ignore: decorator
   @builtin
   export declare function extmul_high_i32x4_u(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function shuffle(a: v128, b: v128, l0: u8, l1: u8): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function swizzle(a: v128, b: v128): v128;
 }
 
 // @ts-ignore: decorator
@@ -2244,6 +2271,14 @@ export namespace f32x4 {
   // @ts-ignore: decorator
   @builtin
   export declare function demote_f64x2_zero(a: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function shuffle(a: v128, b: v128, l0: u8, l1: u8, l2: u8, l3: u8): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function swizzle(a: v128, b: v128): v128;
 }
 
 // @ts-ignore: decorator
@@ -2359,6 +2394,14 @@ export namespace f64x2 {
   // @ts-ignore: decorator
   @builtin
   export declare function promote_low_f32x4(a: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function shuffle(a: v128, b: v128, l0: u8, l1: u8): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function swizzle(a: v128, b: v128): v128;
 }
 
 @final
diff --git a/std/assembly/index.d.ts b/std/assembly/index.d.ts
index c10cf76a1e..69996d0cdf 100644
--- a/std/assembly/index.d.ts
+++ b/std/assembly/index.d.ts
@@ -1075,6 +1075,10 @@ declare namespace i16x8 {
   export function extmul_high_i8x16_s(a: v128, b: v128): v128;
   /** Performs the lane-wise 8-bit unsigned integer extended multiplication of the eight higher lanes producing twice wider 16-bit integer results. */
   export function extmul_high_i8x16_u(a: v128, b: v128): v128;
+  /** Selects 16-bit lanes from either vector according to the specified [0-7] respectively [8-15] lane indexes. */
+  export function shuffle(a: v128, b: v128, l0: u8, l1: u8, l2: u8, l3: u8, l4: u8, l5: u8, l6: u8, l7: u8): v128;
+  /** Selects 8-bit lanes from the first vector according to the indexes [0-15] specified by the 8-bit lanes of the second vector. */
+  export function swizzle(a: v128, s: v128): v128;
 }
 /** Initializes a 128-bit vector from four 32-bit integer values. Arguments must be compile-time constants. */
 declare function i32x4(a: i32, b: i32, c: i32, d: i32): v128;
@@ -1163,6 +1167,10 @@ declare namespace i32x4 {
   export function extmul_high_i16x8_s(a: v128, b: v128): v128;
   /** Performs the lane-wise 16-bit unsigned integer extended multiplication of the four higher lanes producing twice wider 32-bit integer results. */
   export function extmul_high_i16x8_u(a: v128, b: v128): v128;
+  /** Selects 32-bit lanes from either vector according to the specified [0-3] respectively [4-7] lane indexes. */
+  export function shuffle(a: v128, b: v128, l0: u8, l1: u8, l2: u8, l3: u8): v128;
+  /** Selects 8-bit lanes from the first vector according to the indexes [0-15] specified by the 8-bit lanes of the second vector. */
+  export function swizzle(a: v128, s: v128): v128;
 }
 /** Initializes a 128-bit vector from two 64-bit integer values. Arguments must be compile-time constants. */
 declare function i64x2(a: i64, b: i64): v128;
@@ -1221,6 +1229,10 @@ declare namespace i64x2 {
   export function extmul_high_i32x4_s(a: v128, b: v128): v128;
   /** Performs the lane-wise 32-bit unsigned integer extended multiplication of the two higher lanes producing twice wider 64-bit integer results. */
   export function extmul_high_i32x4_u(a: v128, b: v128): v128;
+  /** Selects 64-bit lanes from either vector according to the specified [0-1] respectively [2-3] lane indexes. */
+  export function shuffle(a: v128, b: v128, l0: u8, l1: u8): v128;
+  /** Selects 8-bit lanes from the first vector according to the indexes [0-15] specified by the 8-bit lanes of the second vector. */
+  export function swizzle(a: v128, s: v128): v128;
 }
 /** Initializes a 128-bit vector from four 32-bit float values. Arguments must be compile-time constants. */
 declare function f32x4(a: f32, b: f32, c: f32, d: f32): v128;
@@ -1279,6 +1291,10 @@ declare namespace f32x4 {
   export function convert_i32x4_u(a: v128): v128;
   /** Demotes each 64-bit float lane of a vector to single-precision. The higher lanes of the result are initialized to zero. */
   export function demote_f64x2_zero(a: v128): v128;
+  /** Selects 32-bit lanes from either vector according to the specified [0-3] respectively [4-7] lane indexes. */
+  export function shuffle(a: v128, b: v128, l0: u8, l1: u8, l2: u8, l3: u8, l4: u8): v128;
+  /** Selects 8-bit lanes from the first vector according to the indexes [0-15] specified by the 8-bit lanes of the second vector. */
+  export function swizzle(a: v128, s: v128): v128;
 }
 /** Initializes a 128-bit vector from two 64-bit float values. Arguments must be compile-time constants. */
 declare function f64x2(a: f64, b: f64): v128;
@@ -1337,6 +1353,10 @@ declare namespace f64x2 {
   export function convert_low_i32x4_u(a: v128): v128;
   /** Promotes the low 32-bit float lanes of a vector to double-precision. */
   export function promote_low_f32x4(a: v128): v128;
+  /** Selects 64-bit lanes from either vector according to the specified [0-1] respectively [2-3] lane indexes. */
+  export function shuffle(a: v128, b: v128, l0: u8, l1: u8): v128;
+  /** Selects 8-bit lanes from the first vector according to the indexes [0-15] specified by the 8-bit lanes of the second vector. */
+  export function swizzle(a: v128, s: v128): v128;
 }
 
 declare abstract class i31 {
diff --git a/tests/compiler/simd.debug.wat b/tests/compiler/simd.debug.wat
index 9c34856ad8..ea7ac48daa 100644
--- a/tests/compiler/simd.debug.wat
+++ b/tests/compiler/simd.debug.wat
@@ -4049,6 +4049,44 @@
   i32.const 0
   i32.ne
   drop
+  local.get $0
+  local.get $1
+  i8x16.shuffle 0 1 2 3 4 5 6 7 24 25 26 27 28 29 30 31
+  local.get $0
+  local.get $1
+  i8x16.shuffle 0 1 2 3 4 5 6 7 24 25 26 27 28 29 30 31
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 528
+   i32.const 579
+   i32.const 3
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  local.get $1
+  i8x16.swizzle
+  local.get $0
+  local.get $1
+  i8x16.swizzle
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 528
+   i32.const 584
+   i32.const 3
+   call $~lib/builtins/abort
+   unreachable
+  end
   i32.const 16
   call $~lib/rt/tlsf/__alloc
   local.set $7
@@ -4087,7 +4125,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 589
+   i32.const 599
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -4103,7 +4141,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 594
+   i32.const 604
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -4205,7 +4243,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 662
+   i32.const 672
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4223,7 +4261,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 664
+   i32.const 674
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4242,7 +4280,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 666
+   i32.const 676
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4259,7 +4297,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 667
+   i32.const 677
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4276,7 +4314,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 668
+   i32.const 678
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4337,7 +4375,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 709
+   i32.const 719
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4350,7 +4388,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 714
+   i32.const 724
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4363,7 +4401,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 715
+   i32.const 725
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4380,7 +4418,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 716
+   i32.const 726
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4397,7 +4435,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 721
+   i32.const 731
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4477,7 +4515,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 735
+   i32.const 745
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4494,7 +4532,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 736
+   i32.const 746
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4511,7 +4549,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 737
+   i32.const 747
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4528,7 +4566,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 738
+   i32.const 748
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4545,7 +4583,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 739
+   i32.const 749
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4562,7 +4600,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 740
+   i32.const 750
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4579,7 +4617,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 741
+   i32.const 751
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4596,7 +4634,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 742
+   i32.const 752
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4613,7 +4651,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 743
+   i32.const 753
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4630,7 +4668,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 744
+   i32.const 754
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4703,6 +4741,25 @@
   i32.const 0
   i32.ne
   drop
+  local.get $0
+  local.get $1
+  i8x16.shuffle 0 1 2 3 4 5 6 7 24 25 26 27 28 29 30 31
+  local.get $0
+  local.get $1
+  i8x16.shuffle 0 1 2 3 4 5 6 7 24 25 26 27 28 29 30 31
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 528
+   i32.const 769
+   i32.const 3
+   call $~lib/builtins/abort
+   unreachable
+  end
   i32.const 16
   call $~lib/rt/tlsf/__alloc
   local.set $7
@@ -4729,7 +4786,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 765
+   i32.const 780
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -4745,7 +4802,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 770
+   i32.const 785
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -4845,7 +4902,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 804
+   i32.const 819
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4863,7 +4920,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 806
+   i32.const 821
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4882,7 +4939,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 808
+   i32.const 823
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4899,7 +4956,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 809
+   i32.const 824
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4916,7 +4973,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 810
+   i32.const 825
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4932,7 +4989,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 811
+   i32.const 826
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4945,7 +5002,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 816
+   i32.const 831
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4958,7 +5015,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 817
+   i32.const 832
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4975,7 +5032,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 818
+   i32.const 833
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -4992,7 +5049,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 823
+   i32.const 838
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5065,7 +5122,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 837
+   i32.const 852
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -5081,7 +5138,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 842
+   i32.const 857
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -5461,7 +5518,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 915
+   i32.const 930
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5479,7 +5536,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 917
+   i32.const 932
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5498,7 +5555,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 919
+   i32.const 934
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5515,7 +5572,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 920
+   i32.const 935
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5532,7 +5589,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 921
+   i32.const 936
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5553,7 +5610,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 923
+   i32.const 938
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5570,7 +5627,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 924
+   i32.const 939
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5586,7 +5643,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 925
+   i32.const 940
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5599,7 +5656,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 926
+   i32.const 941
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5612,7 +5669,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 927
+   i32.const 942
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5629,7 +5686,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 928
+   i32.const 943
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5646,7 +5703,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 933
+   i32.const 948
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5677,7 +5734,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 942
+   i32.const 957
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5694,7 +5751,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 943
+   i32.const 958
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5711,7 +5768,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 944
+   i32.const 959
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5728,7 +5785,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 945
+   i32.const 960
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5745,7 +5802,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 946
+   i32.const 961
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5762,7 +5819,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 947
+   i32.const 962
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5779,7 +5836,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 948
+   i32.const 963
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5796,7 +5853,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 949
+   i32.const 964
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5812,7 +5869,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 950
+   i32.const 965
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5879,7 +5936,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 977
+   i32.const 992
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -5897,7 +5954,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 981
+   i32.const 996
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -5941,7 +5998,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 997
+   i32.const 1012
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5959,7 +6016,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 999
+   i32.const 1014
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5978,7 +6035,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1001
+   i32.const 1016
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -5995,7 +6052,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1002
+   i32.const 1017
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6012,7 +6069,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1003
+   i32.const 1018
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6033,7 +6090,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1005
+   i32.const 1020
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6050,7 +6107,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1006
+   i32.const 1021
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6066,7 +6123,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1007
+   i32.const 1022
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6079,7 +6136,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1008
+   i32.const 1023
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6092,7 +6149,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1009
+   i32.const 1024
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6109,7 +6166,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1010
+   i32.const 1025
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6126,7 +6183,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1015
+   i32.const 1030
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6157,7 +6214,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1024
+   i32.const 1039
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6174,7 +6231,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1025
+   i32.const 1040
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6191,7 +6248,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1026
+   i32.const 1041
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6208,7 +6265,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1027
+   i32.const 1042
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6225,7 +6282,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1028
+   i32.const 1043
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6242,7 +6299,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1029
+   i32.const 1044
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6259,7 +6316,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1030
+   i32.const 1045
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6276,7 +6333,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1031
+   i32.const 1046
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6292,7 +6349,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1032
+   i32.const 1047
    i32.const 3
    call $~lib/builtins/abort
    unreachable
@@ -6345,7 +6402,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1051
+   i32.const 1066
    i32.const 5
    call $~lib/builtins/abort
    unreachable
@@ -6363,7 +6420,7 @@
   if
    i32.const 0
    i32.const 528
-   i32.const 1055
+   i32.const 1070
    i32.const 5
    call $~lib/builtins/abort
    unreachable
diff --git a/tests/compiler/simd.release.wat b/tests/compiler/simd.release.wat
index 8ba8db8d1e..c2edfbc789 100644
--- a/tests/compiler/simd.release.wat
+++ b/tests/compiler/simd.release.wat
@@ -2051,7 +2051,7 @@
    if
     i32.const 0
     i32.const 1552
-    i32.const 589
+    i32.const 599
     i32.const 5
     call $~lib/builtins/abort
     unreachable
@@ -2065,7 +2065,7 @@
    if
     i32.const 0
     i32.const 1552
-    i32.const 594
+    i32.const 604
     i32.const 5
     call $~lib/builtins/abort
     unreachable
@@ -2103,7 +2103,7 @@
    if
     i32.const 0
     i32.const 1552
-    i32.const 765
+    i32.const 780
     i32.const 5
     call $~lib/builtins/abort
     unreachable
@@ -2117,7 +2117,7 @@
    if
     i32.const 0
     i32.const 1552
-    i32.const 770
+    i32.const 785
     i32.const 5
     call $~lib/builtins/abort
     unreachable
@@ -2149,7 +2149,7 @@
    if
     i32.const 0
     i32.const 1552
-    i32.const 837
+    i32.const 852
     i32.const 5
     call $~lib/builtins/abort
     unreachable
@@ -2163,7 +2163,7 @@
    if
     i32.const 0
     i32.const 1552
-    i32.const 842
+    i32.const 857
     i32.const 5
     call $~lib/builtins/abort
     unreachable
diff --git a/tests/compiler/simd.ts b/tests/compiler/simd.ts
index e7c2acce90..4acb95b520 100644
--- a/tests/compiler/simd.ts
+++ b/tests/compiler/simd.ts
@@ -576,6 +576,16 @@ function test_i16x8(): void {
   assert(i16x8.extend_low_i8x16_u(i8x16.replace_lane(i8x16.splat(-1), 8, 0)) == i16x8.splat(255));
   assert(i16x8.extend_high_i8x16_s(i8x16.replace_lane(i8x16.splat(-1), 0, 0)) == i16x8.splat(-1));
   assert(i16x8.extend_high_i8x16_u(i8x16.replace_lane(i8x16.splat(-1), 0, 0)) == i16x8.splat(255));
+  assert(
+    i16x8.shuffle(a, b, 0, 1, 2, 3, 12, 13, 14, 15)
+    ==
+    v128.shuffle<i16>(a, b, 0, 1, 2, 3, 12, 13, 14, 15)
+  );
+  assert(
+    i16x8.swizzle(a, b)
+    ==
+    v128.swizzle(a, b)
+  );
   {
     let ptr = __alloc(16);
     store<i8>(ptr, 1);
@@ -756,6 +766,11 @@ function test_i32x4(): void {
   assert(i32x4.extend_low_i16x8_u(i16x8.replace_lane(i16x8.splat(-1), 4, 0)) == i32x4.splat(65535));
   assert(i32x4.extend_high_i16x8_s(i16x8.replace_lane(i16x8.splat(-1), 0, 0)) == i32x4.splat(-1));
   assert(i32x4.extend_high_i16x8_u(i16x8.replace_lane(i16x8.splat(-1), 0, 0)) == i32x4.splat(65535));
+  assert(
+    i32x4.shuffle(a, b, 0, 1, 6, 7)
+    ==
+    v128.shuffle<i32>(a, b, 0, 1, 6, 7)
+  );
   {
     let ptr = __alloc(16);
     store<i16>(ptr, 1);