25 files changed, 669 insertions, 60 deletions

diff --git a/scripts/gen-s-parser.py b/scripts/gen-s-parser.py
index 836ca4dba..80f6911b4 100755
--- a/scripts/gen-s-parser.py
+++ b/scripts/gen-s-parser.py
@@ -437,6 +437,10 @@ instructions = [
     ("f32x4.max",            "makeBinary(s, BinaryOp::MaxVecF32x4)"),
     ("f32x4.pmin",           "makeBinary(s, BinaryOp::PMinVecF32x4)"),
     ("f32x4.pmax",           "makeBinary(s, BinaryOp::PMaxVecF32x4)"),
+    ("f32x4.ceil",           "makeUnary(s, UnaryOp::CeilVecF32x4)"),
+    ("f32x4.floor",          "makeUnary(s, UnaryOp::FloorVecF32x4)"),
+    ("f32x4.trunc",          "makeUnary(s, UnaryOp::TruncVecF32x4)"),
+    ("f32x4.nearest",        "makeUnary(s, UnaryOp::NearestVecF32x4)"),
     ("f64x2.abs",            "makeUnary(s, UnaryOp::AbsVecF64x2)"),
     ("f64x2.neg",            "makeUnary(s, UnaryOp::NegVecF64x2)"),
     ("f64x2.sqrt",           "makeUnary(s, UnaryOp::SqrtVecF64x2)"),
@@ -450,6 +454,10 @@ instructions = [
     ("f64x2.max",            "makeBinary(s, BinaryOp::MaxVecF64x2)"),
     ("f64x2.pmin",           "makeBinary(s, BinaryOp::PMinVecF64x2)"),
     ("f64x2.pmax",           "makeBinary(s, BinaryOp::PMaxVecF64x2)"),
+    ("f64x2.ceil",           "makeUnary(s, UnaryOp::CeilVecF64x2)"),
+    ("f64x2.floor",          "makeUnary(s, UnaryOp::FloorVecF64x2)"),
+    ("f64x2.trunc",          "makeUnary(s, UnaryOp::TruncVecF64x2)"),
+    ("f64x2.nearest",        "makeUnary(s, UnaryOp::NearestVecF64x2)"),
     ("i32x4.trunc_sat_f32x4_s",  "makeUnary(s, UnaryOp::TruncSatSVecF32x4ToVecI32x4)"),
     ("i32x4.trunc_sat_f32x4_u",  "makeUnary(s, UnaryOp::TruncSatUVecF32x4ToVecI32x4)"),
     ("i64x2.trunc_sat_f64x2_s",  "makeUnary(s, UnaryOp::TruncSatSVecF64x2ToVecI64x2)"),
diff --git a/src/binaryen-c.cpp b/src/binaryen-c.cpp
index 96edb4c21..e6155bad3 100644
--- a/src/binaryen-c.cpp
+++ b/src/binaryen-c.cpp
@@ -682,6 +682,10 @@ BinaryenOp BinaryenDivVecF32x4(void) { return DivVecF32x4; }
 BinaryenOp BinaryenMinVecF32x4(void) { return MinVecF32x4; }
 BinaryenOp BinaryenMaxVecF32x4(void) { return MaxVecF32x4; }
 BinaryenOp BinaryenPMinVecF32x4(void) { return PMinVecF32x4; }
+BinaryenOp BinaryenCeilVecF32x4(void) { return CeilVecF32x4; }
+BinaryenOp BinaryenFloorVecF32x4(void) { return FloorVecF32x4; }
+BinaryenOp BinaryenTruncVecF32x4(void) { return TruncVecF32x4; }
+BinaryenOp BinaryenNearestVecF32x4(void) { return NearestVecF32x4; }
 BinaryenOp BinaryenPMaxVecF32x4(void) { return PMaxVecF32x4; }
 BinaryenOp BinaryenAbsVecF64x2(void) { return AbsVecF64x2; }
 BinaryenOp BinaryenNegVecF64x2(void) { return NegVecF64x2; }
@@ -696,6 +700,10 @@ BinaryenOp BinaryenMinVecF64x2(void) { return MinVecF64x2; }
 BinaryenOp BinaryenMaxVecF64x2(void) { return MaxVecF64x2; }
 BinaryenOp BinaryenPMinVecF64x2(void) { return PMinVecF64x2; }
 BinaryenOp BinaryenPMaxVecF64x2(void) { return PMaxVecF64x2; }
+BinaryenOp BinaryenCeilVecF64x2(void) { return CeilVecF64x2; }
+BinaryenOp BinaryenFloorVecF64x2(void) { return FloorVecF64x2; }
+BinaryenOp BinaryenTruncVecF64x2(void) { return TruncVecF64x2; }
+BinaryenOp BinaryenNearestVecF64x2(void) { return NearestVecF64x2; }
 BinaryenOp BinaryenTruncSatSVecF32x4ToVecI32x4(void) {
   return TruncSatSVecF32x4ToVecI32x4;
 }
diff --git a/src/binaryen-c.h b/src/binaryen-c.h
index 9fd08f968..e52d5f46a 100644
--- a/src/binaryen-c.h
+++ b/src/binaryen-c.h
@@ -543,6 +543,10 @@ BINARYEN_API BinaryenOp BinaryenMinVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenMaxVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenPMinVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenPMaxVecF32x4(void);
+BINARYEN_API BinaryenOp BinaryenCeilVecF32x4(void);
+BINARYEN_API BinaryenOp BinaryenFloorVecF32x4(void);
+BINARYEN_API BinaryenOp BinaryenTruncVecF32x4(void);
+BINARYEN_API BinaryenOp BinaryenNearestVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenAbsVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenNegVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenSqrtVecF64x2(void);
@@ -556,6 +560,10 @@ BINARYEN_API BinaryenOp BinaryenMinVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenMaxVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenPMinVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenPMaxVecF64x2(void);
+BINARYEN_API BinaryenOp BinaryenCeilVecF64x2(void);
+BINARYEN_API BinaryenOp BinaryenFloorVecF64x2(void);
+BINARYEN_API BinaryenOp BinaryenTruncVecF64x2(void);
+BINARYEN_API BinaryenOp BinaryenNearestVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenTruncSatSVecF32x4ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenTruncSatUVecF32x4ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenTruncSatSVecF64x2ToVecI64x2(void);
diff --git a/src/gen-s-parser.inc b/src/gen-s-parser.inc
index 3a7569a61..c7a703aea 100644
--- a/src/gen-s-parser.inc
+++ b/src/gen-s-parser.inc
@@ -267,13 +267,21 @@ switch (op[0]) {
                 }
               }
               case 'c': {
-                switch (op[20]) {
-                  case 's':
-                    if (strcmp(op, "f32x4.convert_i32x4_s") == 0) { return makeUnary(s, UnaryOp::ConvertSVecI32x4ToVecF32x4); }
-                    goto parse_error;
-                  case 'u':
-                    if (strcmp(op, "f32x4.convert_i32x4_u") == 0) { return makeUnary(s, UnaryOp::ConvertUVecI32x4ToVecF32x4); }
+                switch (op[7]) {
+                  case 'e':
+                    if (strcmp(op, "f32x4.ceil") == 0) { return makeUnary(s, UnaryOp::CeilVecF32x4); }
                     goto parse_error;
+                  case 'o': {
+                    switch (op[20]) {
+                      case 's':
+                        if (strcmp(op, "f32x4.convert_i32x4_s") == 0) { return makeUnary(s, UnaryOp::ConvertSVecI32x4ToVecF32x4); }
+                        goto parse_error;
+                      case 'u':
+                        if (strcmp(op, "f32x4.convert_i32x4_u") == 0) { return makeUnary(s, UnaryOp::ConvertUVecI32x4ToVecF32x4); }
+                        goto parse_error;
+                      default: goto parse_error;
+                    }
+                  }
                   default: goto parse_error;
                 }
               }
@@ -291,6 +299,9 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 'f':
+                if (strcmp(op, "f32x4.floor") == 0) { return makeUnary(s, UnaryOp::FloorVecF32x4); }
+                goto parse_error;
               case 'g': {
                 switch (op[7]) {
                   case 'e':
@@ -332,6 +343,9 @@ switch (op[0]) {
                   case '\0':
                     if (strcmp(op, "f32x4.ne") == 0) { return makeBinary(s, BinaryOp::NeVecF32x4); }
                     goto parse_error;
+                  case 'a':
+                    if (strcmp(op, "f32x4.nearest") == 0) { return makeUnary(s, UnaryOp::NearestVecF32x4); }
+                    goto parse_error;
                   case 'g':
                     if (strcmp(op, "f32x4.neg") == 0) { return makeUnary(s, UnaryOp::NegVecF32x4); }
                     goto parse_error;
@@ -377,6 +391,9 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 't':
+                if (strcmp(op, "f32x4.trunc") == 0) { return makeUnary(s, UnaryOp::TruncVecF32x4); }
+                goto parse_error;
               default: goto parse_error;
             }
           }
@@ -559,13 +576,21 @@ switch (op[0]) {
                 }
               }
               case 'c': {
-                switch (op[20]) {
-                  case 's':
-                    if (strcmp(op, "f64x2.convert_i64x2_s") == 0) { return makeUnary(s, UnaryOp::ConvertSVecI64x2ToVecF64x2); }
-                    goto parse_error;
-                  case 'u':
-                    if (strcmp(op, "f64x2.convert_i64x2_u") == 0) { return makeUnary(s, UnaryOp::ConvertUVecI64x2ToVecF64x2); }
+                switch (op[7]) {
+                  case 'e':
+                    if (strcmp(op, "f64x2.ceil") == 0) { return makeUnary(s, UnaryOp::CeilVecF64x2); }
                     goto parse_error;
+                  case 'o': {
+                    switch (op[20]) {
+                      case 's':
+                        if (strcmp(op, "f64x2.convert_i64x2_s") == 0) { return makeUnary(s, UnaryOp::ConvertSVecI64x2ToVecF64x2); }
+                        goto parse_error;
+                      case 'u':
+                        if (strcmp(op, "f64x2.convert_i64x2_u") == 0) { return makeUnary(s, UnaryOp::ConvertUVecI64x2ToVecF64x2); }
+                        goto parse_error;
+                      default: goto parse_error;
+                    }
+                  }
                   default: goto parse_error;
                 }
               }
@@ -583,6 +608,9 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 'f':
+                if (strcmp(op, "f64x2.floor") == 0) { return makeUnary(s, UnaryOp::FloorVecF64x2); }
+                goto parse_error;
               case 'g': {
                 switch (op[7]) {
                   case 'e':
@@ -624,6 +652,9 @@ switch (op[0]) {
                   case '\0':
                     if (strcmp(op, "f64x2.ne") == 0) { return makeBinary(s, BinaryOp::NeVecF64x2); }
                     goto parse_error;
+                  case 'a':
+                    if (strcmp(op, "f64x2.nearest") == 0) { return makeUnary(s, UnaryOp::NearestVecF64x2); }
+                    goto parse_error;
                   case 'g':
                     if (strcmp(op, "f64x2.neg") == 0) { return makeUnary(s, UnaryOp::NegVecF64x2); }
                     goto parse_error;
@@ -669,6 +700,9 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 't':
+                if (strcmp(op, "f64x2.trunc") == 0) { return makeUnary(s, UnaryOp::TruncVecF64x2); }
+                goto parse_error;
               default: goto parse_error;
             }
           }
diff --git a/src/ir/cost.h b/src/ir/cost.h
index f890e8910..ea3e32909 100644
--- a/src/ir/cost.h
+++ b/src/ir/cost.h
@@ -173,9 +173,17 @@ struct CostAnalyzer : public Visitor<CostAnalyzer, Index> {
       case AbsVecF32x4:
       case NegVecF32x4:
       case SqrtVecF32x4:
+      case CeilVecF32x4:
+      case FloorVecF32x4:
+      case TruncVecF32x4:
+      case NearestVecF32x4:
       case AbsVecF64x2:
       case NegVecF64x2:
       case SqrtVecF64x2:
+      case CeilVecF64x2:
+      case FloorVecF64x2:
+      case TruncVecF64x2:
+      case NearestVecF64x2:
       case TruncSatSVecF32x4ToVecI32x4:
       case TruncSatUVecF32x4ToVecI32x4:
       case TruncSatSVecF64x2ToVecI64x2:
diff --git a/src/js/binaryen.js-post.js b/src/js/binaryen.js-post.js
index 2d7bd3f36..dfdb05457 100644
--- a/src/js/binaryen.js-post.js
+++ b/src/js/binaryen.js-post.js
@@ -423,6 +423,10 @@ function initializeConstants() {
     'MaxVecF32x4',
     'PMinVecF32x4',
     'PMaxVecF32x4',
+    'CeilVecF32x4',
+    'FloorVecF32x4',
+    'TruncVecF32x4',
+    'NearestVecF32x4',
     'AbsVecF64x2',
     'NegVecF64x2',
     'SqrtVecF64x2',
@@ -436,6 +440,10 @@ function initializeConstants() {
     'MaxVecF64x2',
     'PMinVecF64x2',
     'PMaxVecF64x2',
+    'CeilVecF64x2',
+    'FloorVecF64x2',
+    'TruncVecF64x2',
+    'NearestVecF64x2',
     'TruncSatSVecF32x4ToVecI32x4',
     'TruncSatUVecF32x4ToVecI32x4',
     'TruncSatSVecF64x2ToVecI64x2',
@@ -1918,6 +1926,18 @@ function wrapModule(module, self) {
     'pmax': function(left, right) {
       return Module['_BinaryenBinary'](module, Module['PMaxVecF32x4'], left, right);
     },
+    'ceil': function(value) {
+      return Module['_BinaryenUnary'](module, Module['CeilVecF32x4'], value);
+    },
+    'floor': function(value) {
+      return Module['_BinaryenUnary'](module, Module['FloorVecF32x4'], value);
+    },
+    'trunc': function(value) {
+      return Module['_BinaryenUnary'](module, Module['TruncVecF32x4'], value);
+    },
+    'nearest': function(value) {
+      return Module['_BinaryenUnary'](module, Module['NearestVecF32x4'], value);
+    },
     'convert_i32x4_s': function(value) {
       return Module['_BinaryenUnary'](module, Module['ConvertSVecI32x4ToVecF32x4'], value);
     },
@@ -1993,6 +2013,18 @@ function wrapModule(module, self) {
     'pmax': function(left, right) {
       return Module['_BinaryenBinary'](module, Module['PMaxVecF64x2'], left, right);
     },
+    'ceil': function(value) {
+      return Module['_BinaryenUnary'](module, Module['CeilVecF64x2'], value);
+    },
+    'floor': function(value) {
+      return Module['_BinaryenUnary'](module, Module['FloorVecF64x2'], value);
+    },
+    'trunc': function(value) {
+      return Module['_BinaryenUnary'](module, Module['TruncVecF64x2'], value);
+    },
+    'nearest': function(value) {
+      return Module['_BinaryenUnary'](module, Module['NearestVecF64x2'], value);
+    },
     'convert_i64x2_s': function(value) {
       return Module['_BinaryenUnary'](module, Module['ConvertSVecI64x2ToVecF64x2'], value);
     },
diff --git a/src/literal.h b/src/literal.h
index 2295a126e..8bfee38c1 100644
--- a/src/literal.h
+++ b/src/literal.h
@@ -429,6 +429,10 @@ public:
   Literal maxF32x4(const Literal& other) const;
   Literal pminF32x4(const Literal& other) const;
   Literal pmaxF32x4(const Literal& other) const;
+  Literal ceilF32x4() const;
+  Literal floorF32x4() const;
+  Literal truncF32x4() const;
+  Literal nearestF32x4() const;
   Literal absF64x2() const;
   Literal negF64x2() const;
   Literal sqrtF64x2() const;
@@ -440,6 +444,10 @@ public:
   Literal maxF64x2(const Literal& other) const;
   Literal pminF64x2(const Literal& other) const;
   Literal pmaxF64x2(const Literal& other) const;
+  Literal ceilF64x2() const;
+  Literal floorF64x2() const;
+  Literal truncF64x2() const;
+  Literal nearestF64x2() const;
   Literal truncSatToSI32x4() const;
   Literal truncSatToUI32x4() const;
   Literal truncSatToSI64x2() const;
diff --git a/src/passes/Print.cpp b/src/passes/Print.cpp
index efa31cac0..6246a5f5a 100644
--- a/src/passes/Print.cpp
+++ b/src/passes/Print.cpp
@@ -780,6 +780,18 @@ struct PrintExpressionContents
       case SqrtVecF32x4:
         o << "f32x4.sqrt";
         break;
+      case CeilVecF32x4:
+        o << "f32x4.ceil";
+        break;
+      case FloorVecF32x4:
+        o << "f32x4.floor";
+        break;
+      case TruncVecF32x4:
+        o << "f32x4.trunc";
+        break;
+      case NearestVecF32x4:
+        o << "f32x4.nearest";
+        break;
       case AbsVecF64x2:
         o << "f64x2.abs";
         break;
@@ -789,6 +801,18 @@ struct PrintExpressionContents
       case SqrtVecF64x2:
         o << "f64x2.sqrt";
         break;
+      case CeilVecF64x2:
+        o << "f64x2.ceil";
+        break;
+      case FloorVecF64x2:
+        o << "f64x2.floor";
+        break;
+      case TruncVecF64x2:
+        o << "f64x2.trunc";
+        break;
+      case NearestVecF64x2:
+        o << "f64x2.nearest";
+        break;
       case TruncSatSVecF32x4ToVecI32x4:
         o << "i32x4.trunc_sat_f32x4_s";
         break;
diff --git a/src/wasm-binary.h b/src/wasm-binary.h
index 995adde2a..b8d14377d 100644
--- a/src/wasm-binary.h
+++ b/src/wasm-binary.h
@@ -862,6 +862,15 @@ enum ASTNodes {
   F32x4PMin = 0xea,
   F32x4PMax = 0xeb,
 
+  F32x4Ceil = 0xd8,
+  F32x4Floor = 0xd9,
+  F32x4Trunc = 0xda,
+  F32x4Nearest = 0xdb,
+  F64x2Ceil = 0xdc,
+  F64x2Floor = 0xdd,
+  F64x2Trunc = 0xde,
+  F64x2Nearest = 0xdf,
+
   F64x2Abs = 0xec,
   F64x2Neg = 0xed,
   F64x2Sqrt = 0xef,
diff --git a/src/wasm-interpreter.h b/src/wasm-interpreter.h
index c47786a08..86ea7e146 100644
--- a/src/wasm-interpreter.h
+++ b/src/wasm-interpreter.h
@@ -495,12 +495,28 @@ public:
         return value.negF32x4();
       case SqrtVecF32x4:
         return value.sqrtF32x4();
+      case CeilVecF32x4:
+        return value.ceilF32x4();
+      case FloorVecF32x4:
+        return value.floorF32x4();
+      case TruncVecF32x4:
+        return value.truncF32x4();
+      case NearestVecF32x4:
+        return value.nearestF32x4();
       case AbsVecF64x2:
         return value.absF64x2();
       case NegVecF64x2:
         return value.negF64x2();
       case SqrtVecF64x2:
         return value.sqrtF64x2();
+      case CeilVecF64x2:
+        return value.ceilF64x2();
+      case FloorVecF64x2:
+        return value.floorF64x2();
+      case TruncVecF64x2:
+        return value.truncF64x2();
+      case NearestVecF64x2:
+        return value.nearestF64x2();
       case TruncSatSVecF32x4ToVecI32x4:
         return value.truncSatToSI32x4();
       case TruncSatUVecF32x4ToVecI32x4:
diff --git a/src/wasm.h b/src/wasm.h
index 0afe5ab28..a1c386289 100644
--- a/src/wasm.h
+++ b/src/wasm.h
@@ -177,9 +177,17 @@ enum UnaryOp {
   AbsVecF32x4,
   NegVecF32x4,
   SqrtVecF32x4,
+  CeilVecF32x4,
+  FloorVecF32x4,
+  TruncVecF32x4,
+  NearestVecF32x4,
   AbsVecF64x2,
   NegVecF64x2,
   SqrtVecF64x2,
+  CeilVecF64x2,
+  FloorVecF64x2,
+  TruncVecF64x2,
+  NearestVecF64x2,
 
   // SIMD conversions
   TruncSatSVecF32x4ToVecI32x4,
diff --git a/src/wasm/literal.cpp b/src/wasm/literal.cpp
index 3326adc74..5aebf30a3 100644
--- a/src/wasm/literal.cpp
+++ b/src/wasm/literal.cpp
@@ -1531,6 +1531,18 @@ Literal Literal::negF32x4() const {
 Literal Literal::sqrtF32x4() const {
   return unary<4, &Literal::getLanesF32x4, &Literal::sqrt>(*this);
 }
+Literal Literal::ceilF32x4() const {
+  return unary<4, &Literal::getLanesF32x4, &Literal::ceil>(*this);
+}
+Literal Literal::floorF32x4() const {
+  return unary<4, &Literal::getLanesF32x4, &Literal::floor>(*this);
+}
+Literal Literal::truncF32x4() const {
+  return unary<4, &Literal::getLanesF32x4, &Literal::trunc>(*this);
+}
+Literal Literal::nearestF32x4() const {
+  return unary<4, &Literal::getLanesF32x4, &Literal::nearbyint>(*this);
+}
 Literal Literal::absF64x2() const {
   return unary<2, &Literal::getLanesF64x2, &Literal::abs>(*this);
 }
@@ -1540,6 +1552,18 @@ Literal Literal::negF64x2() const {
 Literal Literal::sqrtF64x2() const {
   return unary<2, &Literal::getLanesF64x2, &Literal::sqrt>(*this);
 }
+Literal Literal::ceilF64x2() const {
+  return unary<2, &Literal::getLanesF64x2, &Literal::ceil>(*this);
+}
+Literal Literal::floorF64x2() const {
+  return unary<2, &Literal::getLanesF64x2, &Literal::floor>(*this);
+}
+Literal Literal::truncF64x2() const {
+  return unary<2, &Literal::getLanesF64x2, &Literal::trunc>(*this);
+}
+Literal Literal::nearestF64x2() const {
+  return unary<2, &Literal::getLanesF64x2, &Literal::nearbyint>(*this);
+}
 Literal Literal::truncSatToSI32x4() const {
   return unary<4, &Literal::getLanesF32x4, &Literal::truncSatToSI32>(*this);
 }
diff --git a/src/wasm/wasm-binary.cpp b/src/wasm/wasm-binary.cpp
index 3441953a7..a766e9b63 100644
--- a/src/wasm/wasm-binary.cpp
+++ b/src/wasm/wasm-binary.cpp
@@ -4194,6 +4194,22 @@ bool WasmBinaryBuilder::maybeVisitSIMDUnary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Unary>();
       curr->op = SqrtVecF32x4;
       break;
+    case BinaryConsts::F32x4Ceil:
+      curr = allocator.alloc<Unary>();
+      curr->op = CeilVecF32x4;
+      break;
+    case BinaryConsts::F32x4Floor:
+      curr = allocator.alloc<Unary>();
+      curr->op = FloorVecF32x4;
+      break;
+    case BinaryConsts::F32x4Trunc:
+      curr = allocator.alloc<Unary>();
+      curr->op = TruncVecF32x4;
+      break;
+    case BinaryConsts::F32x4Nearest:
+      curr = allocator.alloc<Unary>();
+      curr->op = NearestVecF32x4;
+      break;
     case BinaryConsts::F64x2Abs:
       curr = allocator.alloc<Unary>();
       curr->op = AbsVecF64x2;
@@ -4206,6 +4222,22 @@ bool WasmBinaryBuilder::maybeVisitSIMDUnary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Unary>();
       curr->op = SqrtVecF64x2;
       break;
+    case BinaryConsts::F64x2Ceil:
+      curr = allocator.alloc<Unary>();
+      curr->op = CeilVecF64x2;
+      break;
+    case BinaryConsts::F64x2Floor:
+      curr = allocator.alloc<Unary>();
+      curr->op = FloorVecF64x2;
+      break;
+    case BinaryConsts::F64x2Trunc:
+      curr = allocator.alloc<Unary>();
+      curr->op = TruncVecF64x2;
+      break;
+    case BinaryConsts::F64x2Nearest:
+      curr = allocator.alloc<Unary>();
+      curr->op = NearestVecF64x2;
+      break;
     case BinaryConsts::I32x4TruncSatSF32x4:
       curr = allocator.alloc<Unary>();
       curr->op = TruncSatSVecF32x4ToVecI32x4;
diff --git a/src/wasm/wasm-stack.cpp b/src/wasm/wasm-stack.cpp
index 58fa3f2ef..9ebf39fb8 100644
--- a/src/wasm/wasm-stack.cpp
+++ b/src/wasm/wasm-stack.cpp
@@ -981,6 +981,19 @@ void BinaryInstWriter::visitUnary(Unary* curr) {
     case SqrtVecF32x4:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4Sqrt);
       break;
+    case CeilVecF32x4:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4Ceil);
+      break;
+    case FloorVecF32x4:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4Floor);
+      break;
+    case TruncVecF32x4:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4Trunc);
+      break;
+    case NearestVecF32x4:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::F32x4Nearest);
+      break;
     case AbsVecF64x2:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2Abs);
       break;
@@ -990,6 +1003,19 @@ void BinaryInstWriter::visitUnary(Unary* curr) {
     case SqrtVecF64x2:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2Sqrt);
       break;
+    case CeilVecF64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2Ceil);
+      break;
+    case FloorVecF64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2Floor);
+      break;
+    case TruncVecF64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2Trunc);
+      break;
+    case NearestVecF64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::F64x2Nearest);
+      break;
     case TruncSatSVecF32x4ToVecI32x4:
       o << int8_t(BinaryConsts::SIMDPrefix)
         << U32LEB(BinaryConsts::I32x4TruncSatSF32x4);
diff --git a/src/wasm/wasm-validator.cpp b/src/wasm/wasm-validator.cpp
index 606cb47a8..97e45d471 100644
--- a/src/wasm/wasm-validator.cpp
+++ b/src/wasm/wasm-validator.cpp
@@ -1704,9 +1704,17 @@ void FunctionValidator::visitUnary(Unary* curr) {
     case AbsVecF32x4:
     case NegVecF32x4:
     case SqrtVecF32x4:
+    case CeilVecF32x4:
+    case FloorVecF32x4:
+    case TruncVecF32x4:
+    case NearestVecF32x4:
     case AbsVecF64x2:
     case NegVecF64x2:
     case SqrtVecF64x2:
+    case CeilVecF64x2:
+    case FloorVecF64x2:
+    case TruncVecF64x2:
+    case NearestVecF64x2:
     case TruncSatSVecF32x4ToVecI32x4:
     case TruncSatUVecF32x4ToVecI32x4:
     case TruncSatSVecF64x2ToVecI64x2:
diff --git a/src/wasm/wasm.cpp b/src/wasm/wasm.cpp
index 704e48e88..6f86a13ac 100644
--- a/src/wasm/wasm.cpp
+++ b/src/wasm/wasm.cpp
@@ -746,9 +746,17 @@ void Unary::finalize() {
     case AbsVecF32x4:
     case NegVecF32x4:
     case SqrtVecF32x4:
+    case CeilVecF32x4:
+    case FloorVecF32x4:
+    case TruncVecF32x4:
+    case NearestVecF32x4:
     case AbsVecF64x2:
     case NegVecF64x2:
     case SqrtVecF64x2:
+    case CeilVecF64x2:
+    case FloorVecF64x2:
+    case TruncVecF64x2:
+    case NearestVecF64x2:
     case TruncSatSVecF32x4ToVecI32x4:
     case TruncSatUVecF32x4ToVecI32x4:
     case TruncSatSVecF64x2ToVecI64x2:
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))