Add prototype SIMD rounding instructions (#2895)

As specified in https://github.com/WebAssembly/simd/pull/232.

9 files changed, 396 insertions, 48 deletions

diff --git a/test/binaryen.js/kitchen-sink.js b/test/binaryen.js/kitchen-sink.js
index 6c5afa169..cd3af6750 100644
--- a/test/binaryen.js/kitchen-sink.js
+++ b/test/binaryen.js/kitchen-sink.js
@@ -391,6 +391,10 @@ function test_core() {
     module.f32x4.max(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f32x4.pmin(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f32x4.pmax(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.f32x4.ceil(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.f32x4.floor(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.f32x4.trunc(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.f32x4.nearest(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f64x2.add(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f64x2.sub(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f64x2.mul(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
@@ -399,6 +403,10 @@ function test_core() {
     module.f64x2.max(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f64x2.pmin(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f64x2.pmax(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.f64x2.ceil(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.f64x2.floor(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.f64x2.trunc(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.f64x2.nearest(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.i8x16.narrow_i16x8_s(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.i8x16.narrow_i16x8_u(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.i16x8.narrow_i32x4_s(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
diff --git a/test/binaryen.js/kitchen-sink.js.txt b/test/binaryen.js/kitchen-sink.js.txt
index 14f8c0d03..d81ea5312 100644
--- a/test/binaryen.js/kitchen-sink.js.txt
+++ b/test/binaryen.js/kitchen-sink.js.txt
@@ -1297,6 +1297,26 @@ getExpressionInfo(tuple[3])={"id":14,"type":5,"value":3.7}
        )
       )
       (drop
+       (f32x4.ceil
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f32x4.floor
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f32x4.trunc
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f32x4.nearest
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (f64x2.add
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
@@ -1345,6 +1365,26 @@ getExpressionInfo(tuple[3])={"id":14,"type":5,"value":3.7}
        )
       )
       (drop
+       (f64x2.ceil
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f64x2.floor
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f64x2.trunc
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f64x2.nearest
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (i8x16.narrow_i16x8_s
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
@@ -3095,6 +3135,26 @@ getExpressionInfo(tuple[3])={"id":14,"type":5,"value":3.7}
        )
       )
       (drop
+       (f32x4.ceil
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f32x4.floor
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f32x4.trunc
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f32x4.nearest
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (f64x2.add
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
@@ -3143,6 +3203,26 @@ getExpressionInfo(tuple[3])={"id":14,"type":5,"value":3.7}
        )
       )
       (drop
+       (f64x2.ceil
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f64x2.floor
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f64x2.trunc
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (f64x2.nearest
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (i8x16.narrow_i16x8_s
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
diff --git a/test/example/c-api-kitchen-sink.c b/test/example/c-api-kitchen-sink.c
index 764cc0ba6..1f6ffc9fc 100644
--- a/test/example/c-api-kitchen-sink.c
+++ b/test/example/c-api-kitchen-sink.c
@@ -564,6 +564,10 @@ void test_core() {
     makeBinary(module, BinaryenMaxVecF32x4(), v128),
     makeBinary(module, BinaryenPMinVecF32x4(), v128),
     makeBinary(module, BinaryenPMaxVecF32x4(), v128),
+    makeBinary(module, BinaryenCeilVecF32x4(), v128),
+    makeBinary(module, BinaryenFloorVecF32x4(), v128),
+    makeBinary(module, BinaryenTruncVecF32x4(), v128),
+    makeBinary(module, BinaryenNearestVecF32x4(), v128),
     makeBinary(module, BinaryenAddVecF64x2(), v128),
     makeBinary(module, BinaryenSubVecF64x2(), v128),
     makeBinary(module, BinaryenMulVecF64x2(), v128),
@@ -572,6 +576,10 @@ void test_core() {
     makeBinary(module, BinaryenMaxVecF64x2(), v128),
     makeBinary(module, BinaryenPMinVecF64x2(), v128),
     makeBinary(module, BinaryenPMaxVecF64x2(), v128),
+    makeBinary(module, BinaryenCeilVecF64x2(), v128),
+    makeBinary(module, BinaryenFloorVecF64x2(), v128),
+    makeBinary(module, BinaryenTruncVecF64x2(), v128),
+    makeBinary(module, BinaryenNearestVecF64x2(), v128),
     makeBinary(module, BinaryenNarrowSVecI16x8ToVecI8x16(), v128),
     makeBinary(module, BinaryenNarrowUVecI16x8ToVecI8x16(), v128),
     makeBinary(module, BinaryenNarrowSVecI32x4ToVecI16x8(), v128),
diff --git a/test/example/c-api-kitchen-sink.txt b/test/example/c-api-kitchen-sink.txt
index fda846fcb..ff9256fda 100644
--- a/test/example/c-api-kitchen-sink.txt
+++ b/test/example/c-api-kitchen-sink.txt
@@ -1231,6 +1231,30 @@ BinaryenFeatureAll: 1023
        )
       )
       (drop
+       (i16x8.lt_s
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (i16x8.lt_u
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (i16x8.gt_s
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (i16x8.gt_u
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (f64x2.add
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
@@ -1279,6 +1303,30 @@ BinaryenFeatureAll: 1023
        )
       )
       (drop
+       (i16x8.ge_u
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (i32x4.eq
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (i32x4.ne
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
+       (i32x4.lt_s
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (i8x16.narrow_i16x8_s
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
diff --git a/test/simd.wast b/test/simd.wast
index edd9a2c43..7ee8760e1 100644
--- a/test/simd.wast
+++ b/test/simd.wast
@@ -844,6 +844,26 @@
    (local.get $1)
   )
  )
+ (func $f32x4.ceil (param $0 v128) (result v128)
+  (f32x4.ceil
+   (local.get $0)
+  )
+ )
+ (func $f32x4.floor (param $0 v128) (result v128)
+  (f32x4.floor
+   (local.get $0)
+  )
+ )
+ (func $f32x4.trunc (param $0 v128) (result v128)
+  (f32x4.trunc
+   (local.get $0)
+  )
+ )
+ (func $f32x4.nearest (param $0 v128) (result v128)
+  (f32x4.nearest
+   (local.get $0)
+  )
+ )
  (func $f32x4.abs (param $0 v128) (result v128)
   (f32x4.abs
    (local.get $0)
@@ -921,6 +941,26 @@
    (local.get $1)
   )
  )
+ (func $f64x2.ceil (param $0 v128) (result v128)
+  (f64x2.ceil
+   (local.get $0)
+  )
+ )
+ (func $f64x2.floor (param $0 v128) (result v128)
+  (f64x2.floor
+   (local.get $0)
+  )
+ )
+ (func $f64x2.trunc (param $0 v128) (result v128)
+  (f64x2.trunc
+   (local.get $0)
+  )
+ )
+ (func $f64x2.nearest (param $0 v128) (result v128)
+  (f64x2.nearest
+   (local.get $0)
+  )
+ )
  (func $f64x2.abs (param $0 v128) (result v128)
   (f64x2.abs
    (local.get $0)
diff --git a/test/simd.wast.from-wast b/test/simd.wast.from-wast
index 35643ec04..508d44c7c 100644
--- a/test/simd.wast.from-wast
+++ b/test/simd.wast.from-wast
@@ -860,6 +860,26 @@
    (local.get $1)
   )
  )
+ (func $f32x4.ceil (param $0 v128) (result v128)
+  (f32x4.ceil
+   (local.get $0)
+  )
+ )
+ (func $f32x4.floor (param $0 v128) (result v128)
+  (f32x4.floor
+   (local.get $0)
+  )
+ )
+ (func $f32x4.trunc (param $0 v128) (result v128)
+  (f32x4.trunc
+   (local.get $0)
+  )
+ )
+ (func $f32x4.nearest (param $0 v128) (result v128)
+  (f32x4.nearest
+   (local.get $0)
+  )
+ )
  (func $f32x4.abs (param $0 v128) (result v128)
   (f32x4.abs
    (local.get $0)
@@ -937,6 +957,26 @@
    (local.get $1)
   )
  )
+ (func $f64x2.ceil (param $0 v128) (result v128)
+  (f64x2.ceil
+   (local.get $0)
+  )
+ )
+ (func $f64x2.floor (param $0 v128) (result v128)
+  (f64x2.floor
+   (local.get $0)
+  )
+ )
+ (func $f64x2.trunc (param $0 v128) (result v128)
+  (f64x2.trunc
+   (local.get $0)
+  )
+ )
+ (func $f64x2.nearest (param $0 v128) (result v128)
+  (f64x2.nearest
+   (local.get $0)
+  )
+ )
  (func $f64x2.abs (param $0 v128) (result v128)
   (f64x2.abs
    (local.get $0)
diff --git a/test/simd.wast.fromBinary b/test/simd.wast.fromBinary
index 40ce99ec2..d53376883 100644
--- a/test/simd.wast.fromBinary
+++ b/test/simd.wast.fromBinary
@@ -860,6 +860,26 @@
    (local.get $1)
   )
  )
+ (func $f32x4.ceil (param $0 v128) (result v128)
+  (f32x4.ceil
+   (local.get $0)
+  )
+ )
+ (func $f32x4.floor (param $0 v128) (result v128)
+  (f32x4.floor
+   (local.get $0)
+  )
+ )
+ (func $f32x4.trunc (param $0 v128) (result v128)
+  (f32x4.trunc
+   (local.get $0)
+  )
+ )
+ (func $f32x4.nearest (param $0 v128) (result v128)
+  (f32x4.nearest
+   (local.get $0)
+  )
+ )
  (func $f32x4.abs (param $0 v128) (result v128)
   (f32x4.abs
    (local.get $0)
@@ -937,6 +957,26 @@
    (local.get $1)
   )
  )
+ (func $f64x2.ceil (param $0 v128) (result v128)
+  (f64x2.ceil
+   (local.get $0)
+  )
+ )
+ (func $f64x2.floor (param $0 v128) (result v128)
+  (f64x2.floor
+   (local.get $0)
+  )
+ )
+ (func $f64x2.trunc (param $0 v128) (result v128)
+  (f64x2.trunc
+   (local.get $0)
+  )
+ )
+ (func $f64x2.nearest (param $0 v128) (result v128)
+  (f64x2.nearest
+   (local.get $0)
+  )
+ )
  (func $f64x2.abs (param $0 v128) (result v128)
   (f64x2.abs
    (local.get $0)
diff --git a/test/simd.wast.fromBinary.noDebugInfo b/test/simd.wast.fromBinary.noDebugInfo
index 85ca0a9f4..474dfc7c5 100644
--- a/test/simd.wast.fromBinary.noDebugInfo
+++ b/test/simd.wast.fromBinary.noDebugInfo
@@ -861,266 +861,306 @@
   )
  )
  (func $149 (param $0 v128) (result v128)
-  (f32x4.abs
+  (f32x4.ceil
    (local.get $0)
   )
  )
  (func $150 (param $0 v128) (result v128)
-  (f32x4.neg
+  (f32x4.floor
    (local.get $0)
   )
  )
  (func $151 (param $0 v128) (result v128)
+  (f32x4.trunc
+   (local.get $0)
+  )
+ )
+ (func $152 (param $0 v128) (result v128)
+  (f32x4.nearest
+   (local.get $0)
+  )
+ )
+ (func $153 (param $0 v128) (result v128)
+  (f32x4.abs
+   (local.get $0)
+  )
+ )
+ (func $154 (param $0 v128) (result v128)
+  (f32x4.neg
+   (local.get $0)
+  )
+ )
+ (func $155 (param $0 v128) (result v128)
   (f32x4.sqrt
    (local.get $0)
   )
  )
- (func $152 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $156 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f32x4.qfma
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $153 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $157 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f32x4.qfms
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $154 (param $0 v128) (param $1 v128) (result v128)
+ (func $158 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $155 (param $0 v128) (param $1 v128) (result v128)
+ (func $159 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $156 (param $0 v128) (param $1 v128) (result v128)
+ (func $160 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $157 (param $0 v128) (param $1 v128) (result v128)
+ (func $161 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.div
    (local.get $0)
    (local.get $1)
   )
  )
- (func $158 (param $0 v128) (param $1 v128) (result v128)
+ (func $162 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.min
    (local.get $0)
    (local.get $1)
   )
  )
- (func $159 (param $0 v128) (param $1 v128) (result v128)
+ (func $163 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.max
    (local.get $0)
    (local.get $1)
   )
  )
- (func $160 (param $0 v128) (param $1 v128) (result v128)
+ (func $164 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.pmin
    (local.get $0)
    (local.get $1)
   )
  )
- (func $161 (param $0 v128) (param $1 v128) (result v128)
+ (func $165 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.pmax
    (local.get $0)
    (local.get $1)
   )
  )
- (func $162 (param $0 v128) (result v128)
+ (func $166 (param $0 v128) (result v128)
+  (f64x2.ceil
+   (local.get $0)
+  )
+ )
+ (func $167 (param $0 v128) (result v128)
+  (f64x2.floor
+   (local.get $0)
+  )
+ )
+ (func $168 (param $0 v128) (result v128)
+  (f64x2.trunc
+   (local.get $0)
+  )
+ )
+ (func $169 (param $0 v128) (result v128)
+  (f64x2.nearest
+   (local.get $0)
+  )
+ )
+ (func $170 (param $0 v128) (result v128)
   (f64x2.abs
    (local.get $0)
   )
  )
- (func $163 (param $0 v128) (result v128)
+ (func $171 (param $0 v128) (result v128)
   (f64x2.neg
    (local.get $0)
   )
  )
- (func $164 (param $0 v128) (result v128)
+ (func $172 (param $0 v128) (result v128)
   (f64x2.sqrt
    (local.get $0)
   )
  )
- (func $165 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $173 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f64x2.qfma
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $166 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $174 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f64x2.qfms
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $167 (param $0 v128) (result v128)
+ (func $175 (param $0 v128) (result v128)
   (i32x4.trunc_sat_f32x4_s
    (local.get $0)
   )
  )
- (func $168 (param $0 v128) (result v128)
+ (func $176 (param $0 v128) (result v128)
   (i32x4.trunc_sat_f32x4_u
    (local.get $0)
   )
  )
- (func $169 (param $0 v128) (result v128)
+ (func $177 (param $0 v128) (result v128)
   (i64x2.trunc_sat_f64x2_s
    (local.get $0)
   )
  )
- (func $170 (param $0 v128) (result v128)
+ (func $178 (param $0 v128) (result v128)
   (i64x2.trunc_sat_f64x2_u
    (local.get $0)
   )
  )
- (func $171 (param $0 v128) (result v128)
+ (func $179 (param $0 v128) (result v128)
   (f32x4.convert_i32x4_s
    (local.get $0)
   )
  )
- (func $172 (param $0 v128) (result v128)
+ (func $180 (param $0 v128) (result v128)
   (f32x4.convert_i32x4_u
    (local.get $0)
   )
  )
- (func $173 (param $0 v128) (result v128)
+ (func $181 (param $0 v128) (result v128)
   (f64x2.convert_i64x2_s
    (local.get $0)
   )
  )
- (func $174 (param $0 v128) (result v128)
+ (func $182 (param $0 v128) (result v128)
   (f64x2.convert_i64x2_u
    (local.get $0)
   )
  )
- (func $175 (param $0 i32) (result v128)
+ (func $183 (param $0 i32) (result v128)
   (v8x16.load_splat
    (local.get $0)
   )
  )
- (func $176 (param $0 i32) (result v128)
+ (func $184 (param $0 i32) (result v128)
   (v16x8.load_splat
    (local.get $0)
   )
  )
- (func $177 (param $0 i32) (result v128)
+ (func $185 (param $0 i32) (result v128)
   (v32x4.load_splat
    (local.get $0)
   )
  )
- (func $178 (param $0 i32) (result v128)
+ (func $186 (param $0 i32) (result v128)
   (v64x2.load_splat
    (local.get $0)
   )
  )
- (func $179 (param $0 v128) (param $1 v128) (result v128)
+ (func $187 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.narrow_i16x8_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $180 (param $0 v128) (param $1 v128) (result v128)
+ (func $188 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.narrow_i16x8_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $181 (param $0 v128) (param $1 v128) (result v128)
+ (func $189 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.narrow_i32x4_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $182 (param $0 v128) (param $1 v128) (result v128)
+ (func $190 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.narrow_i32x4_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $183 (param $0 v128) (result v128)
+ (func $191 (param $0 v128) (result v128)
   (i16x8.widen_low_i8x16_s
    (local.get $0)
   )
  )
- (func $184 (param $0 v128) (result v128)
+ (func $192 (param $0 v128) (result v128)
   (i16x8.widen_high_i8x16_s
    (local.get $0)
   )
  )
- (func $185 (param $0 v128) (result v128)
+ (func $193 (param $0 v128) (result v128)
   (i16x8.widen_low_i8x16_u
    (local.get $0)
   )
  )
- (func $186 (param $0 v128) (result v128)
+ (func $194 (param $0 v128) (result v128)
   (i16x8.widen_high_i8x16_u
    (local.get $0)
   )
  )
- (func $187 (param $0 v128) (result v128)
+ (func $195 (param $0 v128) (result v128)
   (i32x4.widen_low_i16x8_s
    (local.get $0)
   )
  )
- (func $188 (param $0 v128) (result v128)
+ (func $196 (param $0 v128) (result v128)
   (i32x4.widen_high_i16x8_s
    (local.get $0)
   )
  )
- (func $189 (param $0 v128) (result v128)
+ (func $197 (param $0 v128) (result v128)
   (i32x4.widen_low_i16x8_u
    (local.get $0)
   )
  )
- (func $190 (param $0 v128) (result v128)
+ (func $198 (param $0 v128) (result v128)
   (i32x4.widen_high_i16x8_u
    (local.get $0)
   )
  )
- (func $191 (param $0 i32) (result v128)
+ (func $199 (param $0 i32) (result v128)
   (i16x8.load8x8_u
    (local.get $0)
   )
  )
- (func $192 (param $0 i32) (result v128)
+ (func $200 (param $0 i32) (result v128)
   (i16x8.load8x8_s
    (local.get $0)
   )
  )
- (func $193 (param $0 i32) (result v128)
+ (func $201 (param $0 i32) (result v128)
   (i32x4.load16x4_s
    (local.get $0)
   )
  )
- (func $194 (param $0 i32) (result v128)
+ (func $202 (param $0 i32) (result v128)
   (i32x4.load16x4_u
    (local.get $0)
   )
  )
- (func $195 (param $0 i32) (result v128)
+ (func $203 (param $0 i32) (result v128)
   (i64x2.load32x2_s
    (local.get $0)
   )
  )
- (func $196 (param $0 i32) (result v128)
+ (func $204 (param $0 i32) (result v128)
   (i64x2.load32x2_u
    (local.get $0)
   )
  )
- (func $197 (param $0 v128) (param $1 v128) (result v128)
+ (func $205 (param $0 v128) (param $1 v128) (result v128)
   (v8x16.swizzle
    (local.get $0)
    (local.get $1)
diff --git a/test/spec/simd.wast b/test/spec/simd.wast
index 26da927a9..4fe9c84e2 100644
--- a/test/spec/simd.wast
+++ b/test/spec/simd.wast
@@ -181,6 +181,10 @@
  (func (export "f32x4.max") (param $0 v128) (param $1 v128) (result v128) (f32x4.max (local.get $0) (local.get $1)))
  (func (export "f32x4.pmin") (param $0 v128) (param $1 v128) (result v128) (f32x4.pmin (local.get $0) (local.get $1)))
  (func (export "f32x4.pmax") (param $0 v128) (param $1 v128) (result v128) (f32x4.pmax (local.get $0) (local.get $1)))
+ (func (export "f32x4.ceil") (param $0 v128) (result v128) (f32x4.ceil (local.get $0)))
+ (func (export "f32x4.floor") (param $0 v128) (result v128) (f32x4.floor (local.get $0)))
+ (func (export "f32x4.trunc") (param $0 v128) (result v128) (f32x4.trunc (local.get $0)))
+ (func (export "f32x4.nearest") (param $0 v128) (result v128) (f32x4.nearest (local.get $0)))
  (func (export "f64x2.abs") (param $0 v128) (result v128) (f64x2.abs (local.get $0)))
  (func (export "f64x2.neg") (param $0 v128) (result v128) (f64x2.neg (local.get $0)))
  (func (export "f64x2.sqrt") (param $0 v128) (result v128) (f64x2.sqrt (local.get $0)))
@@ -194,6 +198,10 @@
  (func (export "f64x2.max") (param $0 v128) (param $1 v128) (result v128) (f64x2.max (local.get $0) (local.get $1)))
  (func (export "f64x2.pmin") (param $0 v128) (param $1 v128) (result v128) (f64x2.pmin (local.get $0) (local.get $1)))
  (func (export "f64x2.pmax") (param $0 v128) (param $1 v128) (result v128) (f64x2.pmax (local.get $0) (local.get $1)))
+ (func (export "f64x2.ceil") (param $0 v128) (result v128) (f64x2.ceil (local.get $0)))
+ (func (export "f64x2.floor") (param $0 v128) (result v128) (f64x2.floor (local.get $0)))
+ (func (export "f64x2.trunc") (param $0 v128) (result v128) (f64x2.trunc (local.get $0)))
+ (func (export "f64x2.nearest") (param $0 v128) (result v128) (f64x2.nearest (local.get $0)))
  (func (export "i32x4.trunc_sat_f32x4_s") (param $0 v128) (result v128) (i32x4.trunc_sat_f32x4_s (local.get $0)))
  (func (export "i32x4.trunc_sat_f32x4_u") (param $0 v128) (result v128) (i32x4.trunc_sat_f32x4_u (local.get $0)))
  (func (export "i64x2.trunc_sat_f64x2_s") (param $0 v128) (result v128) (i64x2.trunc_sat_f64x2_s (local.get $0)))
@@ -799,6 +807,18 @@
 (assert_return (invoke "f32x4.max" (v128.const f32x4 -0 0 nan 5) (v128.const f32x4 0 -0 5 nan)) (v128.const f32x4 0 0 nan nan))
 (assert_return (invoke "f32x4.pmin" (v128.const f32x4 -0 0 nan 5) (v128.const f32x4 0 -0 5 nan)) (v128.const f32x4 -0 0 nan 5))
 (assert_return (invoke "f32x4.pmax" (v128.const f32x4 -0 0 nan 5) (v128.const f32x4 0 -0 5 nan)) (v128.const f32x4 -0 0 nan 5))
+(assert_return (invoke "f32x4.ceil" (v128.const f32x4 -0 0 infinity -infinity)) (v128.const f32x4 -0 0 infinity -infinity))
+(assert_return (invoke "f32x4.ceil" (v128.const f32x4 nan 42 0.5 -0.5)) (v128.const f32x4 nan 42 1 -0))
+(assert_return (invoke "f32x4.ceil" (v128.const f32x4 1.5 -1.5 4.2 -4.2)) (v128.const f32x4 2 -1 5 -4))
+(assert_return (invoke "f32x4.floor" (v128.const f32x4 -0 0 infinity -infinity)) (v128.const f32x4 -0 0 infinity -infinity))
+(assert_return (invoke "f32x4.floor" (v128.const f32x4 nan 42 0.5 -0.5)) (v128.const f32x4 nan 42 0 -1))
+(assert_return (invoke "f32x4.floor" (v128.const f32x4 1.5 -1.5 4.2 -4.2)) (v128.const f32x4 1 -2 4 -5))
+(assert_return (invoke "f32x4.trunc" (v128.const f32x4 -0 0 infinity -infinity)) (v128.const f32x4 -0 0 infinity -infinity))
+(assert_return (invoke "f32x4.trunc" (v128.const f32x4 nan 42 0.5 -0.5)) (v128.const f32x4 nan 42 0 -0))
+(assert_return (invoke "f32x4.trunc" (v128.const f32x4 1.5 -1.5 4.2 -4.2)) (v128.const f32x4 1 -1 4 -4))
+(assert_return (invoke "f32x4.nearest" (v128.const f32x4 -0 0 infinity -infinity)) (v128.const f32x4 -0 0 infinity -infinity))
+(assert_return (invoke "f32x4.nearest" (v128.const f32x4 nan 42 0.5 -0.5)) (v128.const f32x4 nan 42 0 -0))
+(assert_return (invoke "f32x4.nearest" (v128.const f32x4 1.5 -1.5 4.2 -4.2)) (v128.const f32x4 2 -2 4 -4))
 
 ;; f64x2 arithmetic
 (assert_return (invoke "f64x2.abs" (v128.const f64x2 -0 nan)) (v128.const f64x2 0 nan))
@@ -824,6 +844,30 @@
 (assert_return (invoke "f64x2.pmin" (v128.const f64x2 nan 5) (v128.const f64x2 5 nan)) (v128.const f64x2 nan 5))
 (assert_return (invoke "f64x2.pmax" (v128.const f64x2 -0 0) (v128.const f64x2 0 -0)) (v128.const f64x2 -0 0))
 (assert_return (invoke "f64x2.pmax" (v128.const f64x2 nan 5) (v128.const f64x2 5 nan)) (v128.const f64x2 nan 5))
+(assert_return (invoke "f64x2.ceil" (v128.const f64x2 -0 0)) (v128.const f64x2 -0 0))
+(assert_return (invoke "f64x2.ceil" (v128.const f64x2 infinity -infinity)) (v128.const f64x2 infinity -infinity))
+(assert_return (invoke "f64x2.ceil" (v128.const f64x2 nan 42)) (v128.const f64x2 nan 42))
+(assert_return (invoke "f64x2.ceil" (v128.const f64x2 0.5 -0.5)) (v128.const f64x2 1 -0))
+(assert_return (invoke "f64x2.ceil" (v128.const f64x2 1.5 -1.5)) (v128.const f64x2 2 -1))
+(assert_return (invoke "f64x2.ceil" (v128.const f64x2 4.2 -4.2)) (v128.const f64x2 5 -4))
+(assert_return (invoke "f64x2.floor" (v128.const f64x2 -0 0)) (v128.const f64x2 -0 0))
+(assert_return (invoke "f64x2.floor" (v128.const f64x2 infinity -infinity)) (v128.const f64x2 infinity -infinity))
+(assert_return (invoke "f64x2.floor" (v128.const f64x2 nan 42)) (v128.const f64x2 nan 42))
+(assert_return (invoke "f64x2.floor" (v128.const f64x2 0.5 -0.5)) (v128.const f64x2 0 -1))
+(assert_return (invoke "f64x2.floor" (v128.const f64x2 1.5 -1.5)) (v128.const f64x2 1 -2))
+(assert_return (invoke "f64x2.floor" (v128.const f64x2 4.2 -4.2)) (v128.const f64x2 4 -5))
+(assert_return (invoke "f64x2.trunc" (v128.const f64x2 -0 0)) (v128.const f64x2 -0 0))
+(assert_return (invoke "f64x2.trunc" (v128.const f64x2 infinity -infinity)) (v128.const f64x2 infinity -infinity))
+(assert_return (invoke "f64x2.trunc" (v128.const f64x2 nan 42)) (v128.const f64x2 nan 42))
+(assert_return (invoke "f64x2.trunc" (v128.const f64x2 0.5 -0.5)) (v128.const f64x2 0 -0))
+(assert_return (invoke "f64x2.trunc" (v128.const f64x2 1.5 -1.5)) (v128.const f64x2 1 -1))
+(assert_return (invoke "f64x2.trunc" (v128.const f64x2 4.2 -4.2)) (v128.const f64x2 4 -4))
+(assert_return (invoke "f64x2.nearest" (v128.const f64x2 -0 0)) (v128.const f64x2 -0 0))
+(assert_return (invoke "f64x2.nearest" (v128.const f64x2 infinity -infinity)) (v128.const f64x2 infinity -infinity))
+(assert_return (invoke "f64x2.nearest" (v128.const f64x2 nan 42)) (v128.const f64x2 nan 42))
+(assert_return (invoke "f64x2.nearest" (v128.const f64x2 0.5 -0.5)) (v128.const f64x2 0 -0))
+(assert_return (invoke "f64x2.nearest" (v128.const f64x2 1.5 -1.5)) (v128.const f64x2 2 -2))
+(assert_return (invoke "f64x2.nearest" (v128.const f64x2 4.2 -4.2)) (v128.const f64x2 4 -4))
 
 ;; conversions
 (assert_return (invoke "i32x4.trunc_sat_f32x4_s" (v128.const f32x4 42 nan infinity -infinity)) (v128.const i32x4 42 0 2147483647 -2147483648))