Add SIMD integer min and max instructions (#2416)

As proposed in https://github.com/WebAssembly/simd/pull/27.

1 files changed, 155 insertions, 83 deletions

diff --git a/test/simd.wast.from-wast b/test/simd.wast.from-wast
index 4fab1e544..9cbf42d71 100644
--- a/test/simd.wast.from-wast
+++ b/test/simd.wast.from-wast
@@ -515,462 +515,534 @@
    (local.get $1)
   )
  )
- (func $i16x8.neg (; 89 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i8x16.min_s (; 89 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i8x16.min_s
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i8x16.min_u (; 90 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i8x16.min_u
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i8x16.max_s (; 91 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i8x16.max_s
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i8x16.max_u (; 92 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i8x16.max_u
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i16x8.neg (; 93 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i16x8.neg
    (local.get $0)
   )
  )
- (func $i16x8.any_true (; 90 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
+ (func $i16x8.any_true (; 94 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
   (i16x8.any_true
    (local.get $0)
   )
  )
- (func $i16x8.all_true (; 91 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
+ (func $i16x8.all_true (; 95 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
   (i16x8.all_true
    (local.get $0)
   )
  )
- (func $i16x8.shl (; 92 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i16x8.shl (; 96 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i16x8.shl
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.shr_s (; 93 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i16x8.shr_s (; 97 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i16x8.shr_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.shr_u (; 94 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i16x8.shr_u (; 98 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i16x8.shr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.add (; 95 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.add (; 99 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.add_saturate_s (; 96 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.add_saturate_s (; 100 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.add_saturate_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.add_saturate_u (; 97 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.add_saturate_u (; 101 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.add_saturate_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.sub (; 98 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.sub (; 102 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.sub_saturate_s (; 99 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.sub_saturate_s (; 103 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.sub_saturate_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.sub_saturate_u (; 100 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.sub_saturate_u (; 104 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.sub_saturate_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.mul (; 101 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.mul (; 105 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i32x4.neg (; 102 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i16x8.min_s (; 106 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i16x8.min_s
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i16x8.min_u (; 107 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i16x8.min_u
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i16x8.max_s (; 108 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i16x8.max_s
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i16x8.max_u (; 109 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i16x8.max_u
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i32x4.neg (; 110 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i32x4.neg
    (local.get $0)
   )
  )
- (func $i32x4.any_true (; 103 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
+ (func $i32x4.any_true (; 111 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
   (i32x4.any_true
    (local.get $0)
   )
  )
- (func $i32x4.all_true (; 104 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
+ (func $i32x4.all_true (; 112 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
   (i32x4.all_true
    (local.get $0)
   )
  )
- (func $i32x4.shl (; 105 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i32x4.shl (; 113 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i32x4.shl
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i32x4.shr_s (; 106 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i32x4.shr_s (; 114 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i32x4.shr_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i32x4.shr_u (; 107 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i32x4.shr_u (; 115 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i32x4.shr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i32x4.add (; 108 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i32x4.add (; 116 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i32x4.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i32x4.sub (; 109 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i32x4.sub (; 117 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i32x4.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i32x4.mul (; 110 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i32x4.mul (; 118 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i32x4.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i64x2.neg (; 111 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i32x4.min_s (; 119 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i32x4.min_s
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i32x4.min_u (; 120 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i32x4.min_u
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i32x4.max_s (; 121 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i32x4.max_s
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i32x4.max_u (; 122 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+  (i32x4.max_u
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $i64x2.neg (; 123 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i64x2.neg
    (local.get $0)
   )
  )
- (func $i64x2.any_true (; 112 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
+ (func $i64x2.any_true (; 124 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
   (i64x2.any_true
    (local.get $0)
   )
  )
- (func $i64x2.all_true (; 113 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
+ (func $i64x2.all_true (; 125 ;) (type $FUNCSIG$iV) (param $0 v128) (result i32)
   (i64x2.all_true
    (local.get $0)
   )
  )
- (func $i64x2.shl (; 114 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i64x2.shl (; 126 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shl
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i64x2.shr_s (; 115 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i64x2.shr_s (; 127 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shr_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i64x2.shr_u (; 116 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
+ (func $i64x2.shr_u (; 128 ;) (type $FUNCSIG$VVi) (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i64x2.add (; 117 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i64x2.add (; 129 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i64x2.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i64x2.sub (; 118 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i64x2.sub (; 130 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i64x2.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f32x4.add (; 119 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f32x4.add (; 131 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f32x4.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f32x4.sub (; 120 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f32x4.sub (; 132 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f32x4.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f32x4.mul (; 121 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f32x4.mul (; 133 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f32x4.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f32x4.div (; 122 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f32x4.div (; 134 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f32x4.div
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f32x4.min (; 123 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f32x4.min (; 135 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f32x4.min
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f32x4.max (; 124 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f32x4.max (; 136 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f32x4.max
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f32x4.abs (; 125 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f32x4.abs (; 137 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f32x4.abs
    (local.get $0)
   )
  )
- (func $f32x4.neg (; 126 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f32x4.neg (; 138 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f32x4.neg
    (local.get $0)
   )
  )
- (func $f32x4.sqrt (; 127 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f32x4.sqrt (; 139 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f32x4.sqrt
    (local.get $0)
   )
  )
- (func $f32x4.qfma (; 128 ;) (type $FUNCSIG$VVVV) (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $f32x4.qfma (; 140 ;) (type $FUNCSIG$VVVV) (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f32x4.qfma
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $f32x4.qfms (; 129 ;) (type $FUNCSIG$VVVV) (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $f32x4.qfms (; 141 ;) (type $FUNCSIG$VVVV) (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f32x4.qfms
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $f64x2.add (; 130 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f64x2.add (; 142 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f64x2.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f64x2.sub (; 131 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f64x2.sub (; 143 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f64x2.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f64x2.mul (; 132 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f64x2.mul (; 144 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f64x2.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f64x2.div (; 133 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f64x2.div (; 145 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f64x2.div
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f64x2.min (; 134 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f64x2.min (; 146 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f64x2.min
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f64x2.max (; 135 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $f64x2.max (; 147 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (f64x2.max
    (local.get $0)
    (local.get $1)
   )
  )
- (func $f64x2.abs (; 136 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f64x2.abs (; 148 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f64x2.abs
    (local.get $0)
   )
  )
- (func $f64x2.neg (; 137 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f64x2.neg (; 149 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f64x2.neg
    (local.get $0)
   )
  )
- (func $f64x2.sqrt (; 138 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f64x2.sqrt (; 150 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f64x2.sqrt
    (local.get $0)
   )
  )
- (func $f64x2.qfma (; 139 ;) (type $FUNCSIG$VVVV) (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $f64x2.qfma (; 151 ;) (type $FUNCSIG$VVVV) (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f64x2.qfma
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $f64x2.qfms (; 140 ;) (type $FUNCSIG$VVVV) (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $f64x2.qfms (; 152 ;) (type $FUNCSIG$VVVV) (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f64x2.qfms
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $i32x4.trunc_sat_f32x4_s (; 141 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i32x4.trunc_sat_f32x4_s (; 153 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i32x4.trunc_sat_f32x4_s
    (local.get $0)
   )
  )
- (func $i32x4.trunc_sat_f32x4_u (; 142 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i32x4.trunc_sat_f32x4_u (; 154 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i32x4.trunc_sat_f32x4_u
    (local.get $0)
   )
  )
- (func $i64x2.trunc_sat_f64x2_s (; 143 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i64x2.trunc_sat_f64x2_s (; 155 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i64x2.trunc_sat_f64x2_s
    (local.get $0)
   )
  )
- (func $i64x2.trunc_sat_f64x2_u (; 144 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i64x2.trunc_sat_f64x2_u (; 156 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i64x2.trunc_sat_f64x2_u
    (local.get $0)
   )
  )
- (func $f32x4.convert_i32x4_s (; 145 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f32x4.convert_i32x4_s (; 157 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f32x4.convert_i32x4_s
    (local.get $0)
   )
  )
- (func $f32x4.convert_i32x4_u (; 146 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f32x4.convert_i32x4_u (; 158 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f32x4.convert_i32x4_u
    (local.get $0)
   )
  )
- (func $f64x2.convert_i64x2_s (; 147 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f64x2.convert_i64x2_s (; 159 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f64x2.convert_i64x2_s
    (local.get $0)
   )
  )
- (func $f64x2.convert_i64x2_u (; 148 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $f64x2.convert_i64x2_u (; 160 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (f64x2.convert_i64x2_u
    (local.get $0)
   )
  )
- (func $v8x16.load_splat (; 149 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $v8x16.load_splat (; 161 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (v8x16.load_splat
    (local.get $0)
   )
  )
- (func $v16x8.load_splat (; 150 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $v16x8.load_splat (; 162 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (v16x8.load_splat
    (local.get $0)
   )
  )
- (func $v32x4.load_splat (; 151 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $v32x4.load_splat (; 163 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (v32x4.load_splat
    (local.get $0)
   )
  )
- (func $v64x2.load_splat (; 152 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $v64x2.load_splat (; 164 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (v64x2.load_splat
    (local.get $0)
   )
  )
- (func $i8x16.narrow_i16x8_s (; 153 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i8x16.narrow_i16x8_s (; 165 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i8x16.narrow_i16x8_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i8x16.narrow_i16x8_u (; 154 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i8x16.narrow_i16x8_u (; 166 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i8x16.narrow_i16x8_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.narrow_i32x4_s (; 155 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.narrow_i32x4_s (; 167 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.narrow_i32x4_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.narrow_i32x4_u (; 156 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $i16x8.narrow_i32x4_u (; 168 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (i16x8.narrow_i32x4_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $i16x8.widen_low_i8x16_s (; 157 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i16x8.widen_low_i8x16_s (; 169 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i16x8.widen_low_i8x16_s
    (local.get $0)
   )
  )
- (func $i16x8.widen_high_i8x16_s (; 158 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i16x8.widen_high_i8x16_s (; 170 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i16x8.widen_high_i8x16_s
    (local.get $0)
   )
  )
- (func $i16x8.widen_low_i8x16_u (; 159 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i16x8.widen_low_i8x16_u (; 171 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i16x8.widen_low_i8x16_u
    (local.get $0)
   )
  )
- (func $i16x8.widen_high_i8x16_u (; 160 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i16x8.widen_high_i8x16_u (; 172 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i16x8.widen_high_i8x16_u
    (local.get $0)
   )
  )
- (func $i32x4.widen_low_i16x8_s (; 161 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i32x4.widen_low_i16x8_s (; 173 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i32x4.widen_low_i16x8_s
    (local.get $0)
   )
  )
- (func $i32x4.widen_high_i16x8_s (; 162 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i32x4.widen_high_i16x8_s (; 174 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i32x4.widen_high_i16x8_s
    (local.get $0)
   )
  )
- (func $i32x4.widen_low_i16x8_u (; 163 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i32x4.widen_low_i16x8_u (; 175 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i32x4.widen_low_i16x8_u
    (local.get $0)
   )
  )
- (func $i32x4.widen_high_i16x8_u (; 164 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
+ (func $i32x4.widen_high_i16x8_u (; 176 ;) (type $FUNCSIG$VV) (param $0 v128) (result v128)
   (i32x4.widen_high_i16x8_u
    (local.get $0)
   )
  )
- (func $i16x8.load8x8_u (; 165 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $i16x8.load8x8_u (; 177 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (i16x8.load8x8_u
    (local.get $0)
   )
  )
- (func $i16x8.load8x8_s (; 166 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $i16x8.load8x8_s (; 178 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (i16x8.load8x8_s
    (local.get $0)
   )
  )
- (func $i32x4.load16x4_s (; 167 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $i32x4.load16x4_s (; 179 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (i32x4.load16x4_s
    (local.get $0)
   )
  )
- (func $i32x4.load16x4_u (; 168 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $i32x4.load16x4_u (; 180 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (i32x4.load16x4_u
    (local.get $0)
   )
  )
- (func $i64x2.load32x2_s (; 169 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $i64x2.load32x2_s (; 181 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (i64x2.load32x2_s
    (local.get $0)
   )
  )
- (func $i64x2.load32x2_u (; 170 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
+ (func $i64x2.load32x2_u (; 182 ;) (type $FUNCSIG$Vi) (param $0 i32) (result v128)
   (i64x2.load32x2_u
    (local.get $0)
   )
  )
- (func $v8x16.swizzle (; 171 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
+ (func $v8x16.swizzle (; 183 ;) (type $FUNCSIG$VVV) (param $0 v128) (param $1 v128) (result v128)
   (v8x16.swizzle
    (local.get $0)
    (local.get $1)