Prototype SIMD instructions implemented in LLVM (#3440)

- i64x2.eq (https://github.com/WebAssembly/simd/pull/381) - i64x2 widens (https://github.com/WebAssembly/simd/pull/290) - i64x2.bitmask (https://github.com/WebAssembly/simd/pull/368) - signselect ops (https://github.com/WebAssembly/simd/pull/124)
author: Thomas Lively <7121787+tlively@users.noreply.github.com> 2020-12-11 19:06:34 -0500
committer: GitHub <noreply@github.com> 2020-12-11 19:06:34 -0500
commit: dd33ae0512aa97f537d40e61631d268c8a8fd17e (patch)
tree: e5c219ba349e56940669bb873bb57a3f2626bf77
parent: 290147d8d43a7448d68939ec711b524ba4fb3fbd (diff)
download: binaryen-dd33ae0512aa97f537d40e61631d268c8a8fd17e.tar.gz
binaryen-dd33ae0512aa97f537d40e61631d268c8a8fd17e.tar.bz2
binaryen-dd33ae0512aa97f537d40e61631d268c8a8fd17e.zip
18 files changed, 717 insertions, 215 deletions
diff --git a/scripts/gen-s-parser.py b/scripts/gen-s-parser.py
index e94c58086..545d48f9c 100755
--- a/scripts/gen-s-parser.py
+++ b/scripts/gen-s-parser.py
@@ -332,6 +332,7 @@ instructions = [
     ("i32x4.le_u",           "makeBinary(s, BinaryOp::LeUVecI32x4)"),
     ("i32x4.ge_s",           "makeBinary(s, BinaryOp::GeSVecI32x4)"),
     ("i32x4.ge_u",           "makeBinary(s, BinaryOp::GeUVecI32x4)"),
+    ("i64x2.eq",             "makeBinary(s, BinaryOp::EqVecI64x2)"),
     ("f32x4.eq",             "makeBinary(s, BinaryOp::EqVecF32x4)"),
     ("f32x4.ne",             "makeBinary(s, BinaryOp::NeVecF32x4)"),
     ("f32x4.lt",             "makeBinary(s, BinaryOp::LtVecF32x4)"),
@@ -350,6 +351,10 @@ instructions = [
     ("v128.xor",             "makeBinary(s, BinaryOp::XorVec128)"),
     ("v128.andnot",          "makeBinary(s, BinaryOp::AndNotVec128)"),
     ("v128.bitselect",       "makeSIMDTernary(s, SIMDTernaryOp::Bitselect)"),
+    ("v8x16.signselect",     "makeSIMDTernary(s, SIMDTernaryOp::SignSelectVec8x16)"),
+    ("v16x8.signselect",     "makeSIMDTernary(s, SIMDTernaryOp::SignSelectVec16x8)"),
+    ("v32x4.signselect",     "makeSIMDTernary(s, SIMDTernaryOp::SignSelectVec32x4)"),
+    ("v64x2.signselect",     "makeSIMDTernary(s, SIMDTernaryOp::SignSelectVec64x2)"),
     ("v128.load8_lane",      "makeSIMDLoadStoreLane(s, LoadLaneVec8x16)"),
     ("v128.load16_lane",     "makeSIMDLoadStoreLane(s, LoadLaneVec16x8)"),
     ("v128.load32_lane",     "makeSIMDLoadStoreLane(s, LoadLaneVec32x4)"),
@@ -427,6 +432,7 @@ instructions = [
     ("i64x2.neg",            "makeUnary(s, UnaryOp::NegVecI64x2)"),
     ("i64x2.any_true",       "makeUnary(s, UnaryOp::AnyTrueVecI64x2)"),
     ("i64x2.all_true",       "makeUnary(s, UnaryOp::AllTrueVecI64x2)"),
+    ("i64x2.bitmask",        "makeUnary(s, UnaryOp::BitmaskVecI64x2)"),
     ("i64x2.shl",            "makeSIMDShift(s, SIMDShiftOp::ShlVecI64x2)"),
     ("i64x2.shr_s",          "makeSIMDShift(s, SIMDShiftOp::ShrSVecI64x2)"),
     ("i64x2.shr_u",          "makeSIMDShift(s, SIMDShiftOp::ShrUVecI64x2)"),
@@ -503,6 +509,10 @@ instructions = [
     ("i32x4.widen_high_i16x8_s", "makeUnary(s, UnaryOp::WidenHighSVecI16x8ToVecI32x4)"),
     ("i32x4.widen_low_i16x8_u",  "makeUnary(s, UnaryOp::WidenLowUVecI16x8ToVecI32x4)"),
     ("i32x4.widen_high_i16x8_u", "makeUnary(s, UnaryOp::WidenHighUVecI16x8ToVecI32x4)"),
+    ("i64x2.widen_low_i32x4_s",  "makeUnary(s, UnaryOp::WidenLowSVecI32x4ToVecI64x2)"),
+    ("i64x2.widen_high_i32x4_s", "makeUnary(s, UnaryOp::WidenHighSVecI32x4ToVecI64x2)"),
+    ("i64x2.widen_low_i32x4_u",  "makeUnary(s, UnaryOp::WidenLowUVecI32x4ToVecI64x2)"),
+    ("i64x2.widen_high_i32x4_u", "makeUnary(s, UnaryOp::WidenHighUVecI32x4ToVecI64x2)"),
     ("v8x16.swizzle",            "makeBinary(s, BinaryOp::SwizzleVec8x16)"),
     # reference types instructions
     # TODO Add table instructions
diff --git a/src/gen-s-parser.inc b/src/gen-s-parser.inc
index 8afcea917..d4e1afe38 100644
--- a/src/gen-s-parser.inc
+++ b/src/gen-s-parser.inc
@@ -2312,38 +2312,49 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 'b':
+                if (strcmp(op, "i64x2.bitmask") == 0) { return makeUnary(s, UnaryOp::BitmaskVecI64x2); }
+                goto parse_error;
               case 'e': {
-                switch (op[9]) {
-                  case 'm': {
-                    switch (op[13]) {
-                      case 'h': {
-                        switch (op[24]) {
-                          case 's':
-                            if (strcmp(op, "i64x2.extmul_high_i32x4_s") == 0) { return makeBinary(s, BinaryOp::ExtMulHighSVecI64x2); }
-                            goto parse_error;
-                          case 'u':
-                            if (strcmp(op, "i64x2.extmul_high_i32x4_u") == 0) { return makeBinary(s, BinaryOp::ExtMulHighUVecI64x2); }
-                            goto parse_error;
-                          default: goto parse_error;
-                        }
-                      }
-                      case 'l': {
-                        switch (op[23]) {
-                          case 's':
-                            if (strcmp(op, "i64x2.extmul_low_i32x4_s") == 0) { return makeBinary(s, BinaryOp::ExtMulLowSVecI64x2); }
-                            goto parse_error;
-                          case 'u':
-                            if (strcmp(op, "i64x2.extmul_low_i32x4_u") == 0) { return makeBinary(s, BinaryOp::ExtMulLowUVecI64x2); }
-                            goto parse_error;
+                switch (op[7]) {
+                  case 'q':
+                    if (strcmp(op, "i64x2.eq") == 0) { return makeBinary(s, BinaryOp::EqVecI64x2); }
+                    goto parse_error;
+                  case 'x': {
+                    switch (op[9]) {
+                      case 'm': {
+                        switch (op[13]) {
+                          case 'h': {
+                            switch (op[24]) {
+                              case 's':
+                                if (strcmp(op, "i64x2.extmul_high_i32x4_s") == 0) { return makeBinary(s, BinaryOp::ExtMulHighSVecI64x2); }
+                                goto parse_error;
+                              case 'u':
+                                if (strcmp(op, "i64x2.extmul_high_i32x4_u") == 0) { return makeBinary(s, BinaryOp::ExtMulHighUVecI64x2); }
+                                goto parse_error;
+                              default: goto parse_error;
+                            }
+                          }
+                          case 'l': {
+                            switch (op[23]) {
+                              case 's':
+                                if (strcmp(op, "i64x2.extmul_low_i32x4_s") == 0) { return makeBinary(s, BinaryOp::ExtMulLowSVecI64x2); }
+                                goto parse_error;
+                              case 'u':
+                                if (strcmp(op, "i64x2.extmul_low_i32x4_u") == 0) { return makeBinary(s, BinaryOp::ExtMulLowUVecI64x2); }
+                                goto parse_error;
+                              default: goto parse_error;
+                            }
+                          }
                           default: goto parse_error;
                         }
                       }
+                      case 'r':
+                        if (strcmp(op, "i64x2.extract_lane") == 0) { return makeSIMDExtract(s, SIMDExtractOp::ExtractLaneVecI64x2, 2); }
+                        goto parse_error;
                       default: goto parse_error;
                     }
                   }
-                  case 'r':
-                    if (strcmp(op, "i64x2.extract_lane") == 0) { return makeSIMDExtract(s, SIMDExtractOp::ExtractLaneVecI64x2, 2); }
-                    goto parse_error;
                   default: goto parse_error;
                 }
               }
@@ -2408,6 +2419,33 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 'w': {
+                switch (op[12]) {
+                  case 'h': {
+                    switch (op[23]) {
+                      case 's':
+                        if (strcmp(op, "i64x2.widen_high_i32x4_s") == 0) { return makeUnary(s, UnaryOp::WidenHighSVecI32x4ToVecI64x2); }
+                        goto parse_error;
+                      case 'u':
+                        if (strcmp(op, "i64x2.widen_high_i32x4_u") == 0) { return makeUnary(s, UnaryOp::WidenHighUVecI32x4ToVecI64x2); }
+                        goto parse_error;
+                      default: goto parse_error;
+                    }
+                  }
+                  case 'l': {
+                    switch (op[22]) {
+                      case 's':
+                        if (strcmp(op, "i64x2.widen_low_i32x4_s") == 0) { return makeUnary(s, UnaryOp::WidenLowSVecI32x4ToVecI64x2); }
+                        goto parse_error;
+                      case 'u':
+                        if (strcmp(op, "i64x2.widen_low_i32x4_u") == 0) { return makeUnary(s, UnaryOp::WidenLowUVecI32x4ToVecI64x2); }
+                        goto parse_error;
+                      default: goto parse_error;
+                    }
+                  }
+                  default: goto parse_error;
+                }
+              }
               default: goto parse_error;
             }
           }
@@ -2962,18 +3000,42 @@ switch (op[0]) {
               default: goto parse_error;
             }
           }
-          case '6':
-            if (strcmp(op, "v16x8.load_splat") == 0) { return makeSIMDLoad(s, SIMDLoadOp::LoadSplatVec16x8); }
+          case '6': {
+            switch (op[6]) {
+              case 'l':
+                if (strcmp(op, "v16x8.load_splat") == 0) { return makeSIMDLoad(s, SIMDLoadOp::LoadSplatVec16x8); }
+                goto parse_error;
+              case 's':
+                if (strcmp(op, "v16x8.signselect") == 0) { return makeSIMDTernary(s, SIMDTernaryOp::SignSelectVec16x8); }
+                goto parse_error;
+              default: goto parse_error;
+            }
+          }
+          default: goto parse_error;
+        }
+      }
+      case '3': {
+        switch (op[6]) {
+          case 'l':
+            if (strcmp(op, "v32x4.load_splat") == 0) { return makeSIMDLoad(s, SIMDLoadOp::LoadSplatVec32x4); }
+            goto parse_error;
+          case 's':
+            if (strcmp(op, "v32x4.signselect") == 0) { return makeSIMDTernary(s, SIMDTernaryOp::SignSelectVec32x4); }
+            goto parse_error;
+          default: goto parse_error;
+        }
+      }
+      case '6': {
+        switch (op[6]) {
+          case 'l':
+            if (strcmp(op, "v64x2.load_splat") == 0) { return makeSIMDLoad(s, SIMDLoadOp::LoadSplatVec64x2); }
+            goto parse_error;
+          case 's':
+            if (strcmp(op, "v64x2.signselect") == 0) { return makeSIMDTernary(s, SIMDTernaryOp::SignSelectVec64x2); }
             goto parse_error;
           default: goto parse_error;
         }
       }
-      case '3':
-        if (strcmp(op, "v32x4.load_splat") == 0) { return makeSIMDLoad(s, SIMDLoadOp::LoadSplatVec32x4); }
-        goto parse_error;
-      case '6':
-        if (strcmp(op, "v64x2.load_splat") == 0) { return makeSIMDLoad(s, SIMDLoadOp::LoadSplatVec64x2); }
-        goto parse_error;
       case '8': {
         switch (op[6]) {
           case 'l':
@@ -2984,6 +3046,9 @@ switch (op[0]) {
               case 'h':
                 if (strcmp(op, "v8x16.shuffle") == 0) { return makeSIMDShuffle(s); }
                 goto parse_error;
+              case 'i':
+                if (strcmp(op, "v8x16.signselect") == 0) { return makeSIMDTernary(s, SIMDTernaryOp::SignSelectVec8x16); }
+                goto parse_error;
               case 'w':
                 if (strcmp(op, "v8x16.swizzle") == 0) { return makeBinary(s, BinaryOp::SwizzleVec8x16); }
                 goto parse_error;
diff --git a/src/ir/cost.h b/src/ir/cost.h
index 66a316bf6..c5b77ea01 100644
--- a/src/ir/cost.h
+++ b/src/ir/cost.h
@@ -191,6 +191,7 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, Index> {
       case NegVecI64x2:
       case AnyTrueVecI64x2:
       case AllTrueVecI64x2:
+      case BitmaskVecI64x2:
       case AbsVecF32x4:
       case NegVecF32x4:
       case SqrtVecF32x4:
@@ -221,6 +222,10 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, Index> {
       case WidenHighSVecI16x8ToVecI32x4:
       case WidenLowUVecI16x8ToVecI32x4:
       case WidenHighUVecI16x8ToVecI32x4:
+      case WidenLowSVecI32x4ToVecI64x2:
+      case WidenHighSVecI32x4ToVecI64x2:
+      case WidenLowUVecI32x4ToVecI64x2:
+      case WidenHighUVecI32x4ToVecI64x2:
         ret = 1;
         break;
       case InvalidUnary:
@@ -363,6 +368,7 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, Index> {
       case GtUVecI32x4:
       case GeSVecI32x4:
       case GeUVecI32x4:
+      case EqVecI64x2:
       case EqVecF32x4:
       case NeVecF32x4:
       case LtVecF32x4:
@@ -508,6 +514,10 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, Index> {
     Index ret = 0;
     switch (curr->op) {
       case Bitselect:
+      case SignSelectVec8x16:
+      case SignSelectVec16x8:
+      case SignSelectVec32x4:
+      case SignSelectVec64x2:
         ret = 1;
         break;
       case QFMAF32x4:
diff --git a/src/literal.h b/src/literal.h
index 80a7cea7a..8a1829d10 100644
--- a/src/literal.h
+++ b/src/literal.h
@@ -480,6 +480,7 @@ public:
   Literal leUI32x4(const Literal& other) const;
   Literal geSI32x4(const Literal& other) const;
   Literal geUI32x4(const Literal& other) const;
+  Literal eqI64x2(const Literal& other) const;
   Literal eqF32x4(const Literal& other) const;
   Literal neF32x4(const Literal& other) const;
   Literal ltF32x4(const Literal& other) const;
diff --git a/src/passes/Print.cpp b/src/passes/Print.cpp
index 4007b6daa..b61a72720 100644
--- a/src/passes/Print.cpp
+++ b/src/passes/Print.cpp
@@ -533,6 +533,18 @@ struct PrintExpressionContents
       case QFMSF64x2:
         o << "f64x2.qfms";
         break;
+      case SignSelectVec8x16:
+        o << "v8x16.signselect";
+        break;
+      case SignSelectVec16x8:
+        o << "v16x8.signselect";
+        break;
+      case SignSelectVec32x4:
+        o << "v32x4.signselect";
+        break;
+      case SignSelectVec64x2:
+        o << "v64x2.signselect";
+        break;
     }
   }
   void visitSIMDShift(SIMDShift* curr) {
@@ -941,6 +953,9 @@ struct PrintExpressionContents
       case AllTrueVecI64x2:
         o << "i64x2.all_true";
         break;
+      case BitmaskVecI64x2:
+        o << "i64x2.bitmask";
+        break;
       case AbsVecF32x4:
         o << "f32x4.abs";
         break;
@@ -1031,6 +1046,18 @@ struct PrintExpressionContents
       case WidenHighUVecI16x8ToVecI32x4:
         o << "i32x4.widen_high_i16x8_u";
         break;
+      case WidenLowSVecI32x4ToVecI64x2:
+        o << "i64x2.widen_low_i32x4_s";
+        break;
+      case WidenHighSVecI32x4ToVecI64x2:
+        o << "i64x2.widen_high_i32x4_s";
+        break;
+      case WidenLowUVecI32x4ToVecI64x2:
+        o << "i64x2.widen_low_i32x4_u";
+        break;
+      case WidenHighUVecI32x4ToVecI64x2:
+        o << "i64x2.widen_high_i32x4_u";
+        break;
       case InvalidUnary:
         WASM_UNREACHABLE("unvalid unary operator");
     }
@@ -1360,6 +1387,9 @@ struct PrintExpressionContents
       case GeUVecI32x4:
         o << "i32x4.ge_u";
         break;
+      case EqVecI64x2:
+        o << "i64x2.eq";
+        break;
       case EqVecF32x4:
         o << "f32x4.eq";
         break;
diff --git a/src/wasm-binary.h b/src/wasm-binary.h
index dffc1507c..1d250aa2b 100644
--- a/src/wasm-binary.h
+++ b/src/wasm-binary.h
@@ -782,6 +782,7 @@ enum ASTNodes {
   I32x4LeU = 0x3e,
   I32x4GeS = 0x3f,
   I32x4GeU = 0x40,
+  I64x2Eq = 0xc0,
   F32x4Eq = 0x41,
   F32x4Ne = 0x42,
   F32x4Lt = 0x43,
@@ -802,6 +803,11 @@ enum ASTNodes {
   V128Xor = 0x51,
   V128Bitselect = 0x52,
 
+  V8x16SignSelect = 0x7d,
+  V16x8SignSelect = 0x7e,
+  V32x4SignSelect = 0x7f,
+  V64x2SignSelect = 0x94,
+
   V128Load8Lane = 0x58,
   V128Load16Lane = 0x59,
   V128Load32Lane = 0x5a,
@@ -885,6 +891,11 @@ enum ASTNodes {
   I32x4MaxU = 0xb9,
   I32x4DotSVecI16x8 = 0xba,
 
+  I64x2Bitmask = 0xc4,
+  I64x2WidenLowSI32x4 = 0xc7,
+  I64x2WidenHighSI32x4 = 0xc8,
+  I64x2WidenLowUI32x4 = 0xc9,
+  I64x2WidenHighUI32x4 = 0xca,
   I64x2Neg = 0xc1,
   I64x2AnyTrue = 0xc2,
   I64x2AllTrue = 0xc3,
diff --git a/src/wasm-interpreter.h b/src/wasm-interpreter.h
index e4c20082e..dbc14de5f 100644
--- a/src/wasm-interpreter.h
+++ b/src/wasm-interpreter.h
@@ -491,6 +491,8 @@ public:
         return value.anyTrueI64x2();
       case AllTrueVecI64x2:
         return value.allTrueI64x2();
+      case BitmaskVecI64x2:
+        WASM_UNREACHABLE("unimp");
       case AbsVecF32x4:
         return value.absF32x4();
       case NegVecF32x4:
@@ -551,6 +553,11 @@ public:
         return value.widenLowUToVecI32x4();
       case WidenHighUVecI16x8ToVecI32x4:
         return value.widenHighUToVecI32x4();
+      case WidenLowSVecI32x4ToVecI64x2:
+      case WidenHighSVecI32x4ToVecI64x2:
+      case WidenLowUVecI32x4ToVecI64x2:
+      case WidenHighUVecI32x4ToVecI64x2:
+        WASM_UNREACHABLE("unimp");
       case InvalidUnary:
         WASM_UNREACHABLE("invalid unary op");
     }
@@ -796,6 +803,8 @@ public:
         return left.geSI32x4(right);
       case GeUVecI32x4:
         return left.geUI32x4(right);
+      case EqVecI64x2:
+        return left.eqI64x2(right);
       case EqVecF32x4:
         return left.eqF32x4(right);
       case NeVecF32x4:
@@ -1067,7 +1076,7 @@ public:
       case Bitselect:
         return c.bitselectV128(a, b);
       default:
-        // TODO: implement qfma/qfms
+        // TODO: implement qfma/qfms and signselect
         WASM_UNREACHABLE("not implemented");
     }
   }
diff --git a/src/wasm.h b/src/wasm.h
index 0c6566407..bc39eae7b 100644
--- a/src/wasm.h
+++ b/src/wasm.h
@@ -175,6 +175,7 @@ enum UnaryOp {
   NegVecI64x2,
   AnyTrueVecI64x2,
   AllTrueVecI64x2,
+  BitmaskVecI64x2,
   AbsVecF32x4,
   NegVecF32x4,
   SqrtVecF32x4,
@@ -207,6 +208,10 @@ enum UnaryOp {
   WidenHighSVecI16x8ToVecI32x4,
   WidenLowUVecI16x8ToVecI32x4,
   WidenHighUVecI16x8ToVecI32x4,
+  WidenLowSVecI32x4ToVecI64x2,
+  WidenHighSVecI32x4ToVecI64x2,
+  WidenLowUVecI32x4ToVecI64x2,
+  WidenHighUVecI32x4ToVecI64x2,
 
   InvalidUnary
 };
@@ -351,6 +356,7 @@ enum BinaryOp {
   LeUVecI32x4,
   GeSVecI32x4,
   GeUVecI32x4,
+  EqVecI64x2,
   EqVecF32x4,
   NeVecF32x4,
   LtVecF32x4,
@@ -509,7 +515,17 @@ enum SIMDLoadStoreLaneOp {
   StoreLaneVec64x2,
 };
 
-enum SIMDTernaryOp { Bitselect, QFMAF32x4, QFMSF32x4, QFMAF64x2, QFMSF64x2 };
+enum SIMDTernaryOp {
+  Bitselect,
+  QFMAF32x4,
+  QFMSF32x4,
+  QFMAF64x2,
+  QFMSF64x2,
+  SignSelectVec8x16,
+  SignSelectVec16x8,
+  SignSelectVec32x4,
+  SignSelectVec64x2
+};
 
 //
 // Expressions
diff --git a/src/wasm/literal.cpp b/src/wasm/literal.cpp
index c19efa225..9bdf886b3 100644
--- a/src/wasm/literal.cpp
+++ b/src/wasm/literal.cpp
@@ -1986,6 +1986,10 @@ Literal Literal::geSI32x4(const Literal& other) const {
 Literal Literal::geUI32x4(const Literal& other) const {
   return compare<4, &Literal::getLanesI32x4, &Literal::geU>(*this, other);
 }
+Literal Literal::eqI64x2(const Literal& other) const {
+  return compare<2, &Literal::getLanesI64x2, &Literal::eq, int64_t>(*this,
+                                                                    other);
+}
 Literal Literal::eqF32x4(const Literal& other) const {
   return compare<4, &Literal::getLanesF32x4, &Literal::eq>(*this, other);
 }
diff --git a/src/wasm/wasm-binary.cpp b/src/wasm/wasm-binary.cpp
index 1b3d50a41..796ba4b4e 100644
--- a/src/wasm/wasm-binary.cpp
+++ b/src/wasm/wasm-binary.cpp
@@ -4371,6 +4371,10 @@ bool WasmBinaryBuilder::maybeVisitSIMDBinary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Binary>();
       curr->op = GeUVecI32x4;
       break;
+    case BinaryConsts::I64x2Eq:
+      curr = allocator.alloc<Binary>();
+      curr->op = EqVecI64x2;
+      break;
     case BinaryConsts::F32x4Eq:
       curr = allocator.alloc<Binary>();
       curr->op = EqVecF32x4;
@@ -4828,6 +4832,10 @@ bool WasmBinaryBuilder::maybeVisitSIMDUnary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Unary>();
       curr->op = AllTrueVecI64x2;
       break;
+    case BinaryConsts::I64x2Bitmask:
+      curr = allocator.alloc<Unary>();
+      curr->op = BitmaskVecI64x2;
+      break;
     case BinaryConsts::F32x4Abs:
       curr = allocator.alloc<Unary>();
       curr->op = AbsVecF32x4;
@@ -4948,6 +4956,22 @@ bool WasmBinaryBuilder::maybeVisitSIMDUnary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Unary>();
       curr->op = WidenHighUVecI16x8ToVecI32x4;
       break;
+    case BinaryConsts::I64x2WidenLowSI32x4:
+      curr = allocator.alloc<Unary>();
+      curr->op = WidenLowSVecI32x4ToVecI64x2;
+      break;
+    case BinaryConsts::I64x2WidenHighSI32x4:
+      curr = allocator.alloc<Unary>();
+      curr->op = WidenHighSVecI32x4ToVecI64x2;
+      break;
+    case BinaryConsts::I64x2WidenLowUI32x4:
+      curr = allocator.alloc<Unary>();
+      curr->op = WidenLowUVecI32x4ToVecI64x2;
+      break;
+    case BinaryConsts::I64x2WidenHighUI32x4:
+      curr = allocator.alloc<Unary>();
+      curr->op = WidenHighUVecI32x4ToVecI64x2;
+      break;
     default:
       return false;
   }
@@ -5101,6 +5125,22 @@ bool WasmBinaryBuilder::maybeVisitSIMDTernary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<SIMDTernary>();
       curr->op = Bitselect;
       break;
+    case BinaryConsts::V8x16SignSelect:
+      curr = allocator.alloc<SIMDTernary>();
+      curr->op = SignSelectVec8x16;
+      break;
+    case BinaryConsts::V16x8SignSelect:
+      curr = allocator.alloc<SIMDTernary>();
+      curr->op = SignSelectVec16x8;
+      break;
+    case BinaryConsts::V32x4SignSelect:
+      curr = allocator.alloc<SIMDTernary>();
+      curr->op = SignSelectVec32x4;
+      break;
+    case BinaryConsts::V64x2SignSelect:
+      curr = allocator.alloc<SIMDTernary>();
+      curr->op = SignSelectVec64x2;
+      break;
     case BinaryConsts::F32x4QFMA:
       curr = allocator.alloc<SIMDTernary>();
       curr->op = QFMAF32x4;
diff --git a/src/wasm/wasm-stack.cpp b/src/wasm/wasm-stack.cpp
index 741aef5ce..3a24d0cc0 100644
--- a/src/wasm/wasm-stack.cpp
+++ b/src/wasm/wasm-stack.cpp
@@ -560,6 +560,18 @@ void BinaryInstWriter::visitSIMDTernary(SIMDTernary* curr) {
     case QFMSF64x2:
       o << U32LEB(BinaryConsts::F64x2QFMS);
       break;
+    case SignSelectVec8x16:
+      o << U32LEB(BinaryConsts::V8x16SignSelect);
+      break;
+    case SignSelectVec16x8:
+      o << U32LEB(BinaryConsts::V16x8SignSelect);
+      break;
+    case SignSelectVec32x4:
+      o << U32LEB(BinaryConsts::V32x4SignSelect);
+      break;
+    case SignSelectVec64x2:
+      o << U32LEB(BinaryConsts::V64x2SignSelect);
+      break;
   }
 }
 
@@ -1024,6 +1036,10 @@ void BinaryInstWriter::visitUnary(Unary* curr) {
       o << int8_t(BinaryConsts::SIMDPrefix)
         << U32LEB(BinaryConsts::I64x2AllTrue);
       break;
+    case BitmaskVecI64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I64x2Bitmask);
+      break;
     case AbsVecF32x4:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4Abs);
       break;
@@ -1132,6 +1148,22 @@ void BinaryInstWriter::visitUnary(Unary* curr) {
       o << int8_t(BinaryConsts::SIMDPrefix)
         << U32LEB(BinaryConsts::I32x4WidenHighUI16x8);
       break;
+    case WidenLowSVecI32x4ToVecI64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I64x2WidenLowSI32x4);
+      break;
+    case WidenHighSVecI32x4ToVecI64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I64x2WidenHighSI32x4);
+      break;
+    case WidenLowUVecI32x4ToVecI64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I64x2WidenLowUI32x4);
+      break;
+    case WidenHighUVecI32x4ToVecI64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I64x2WidenHighUI32x4);
+      break;
     case InvalidUnary:
       WASM_UNREACHABLE("invalid unary op");
   }
@@ -1461,6 +1493,9 @@ void BinaryInstWriter::visitBinary(Binary* curr) {
     case GeUVecI32x4:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::I32x4GeU);
       break;
+    case EqVecI64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::I64x2Eq);
+      break;
     case EqVecF32x4:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4Eq);
       break;
diff --git a/src/wasm/wasm-validator.cpp b/src/wasm/wasm-validator.cpp
index 56cf9c114..b294c953c 100644
--- a/src/wasm/wasm-validator.cpp
+++ b/src/wasm/wasm-validator.cpp
@@ -1534,6 +1534,7 @@ void FunctionValidator::visitBinary(Binary* curr) {
     case GtUVecI32x4:
     case GeSVecI32x4:
     case GeUVecI32x4:
+    case EqVecI64x2:
     case EqVecF32x4:
     case NeVecF32x4:
     case LtVecF32x4:
@@ -1880,6 +1881,10 @@ void FunctionValidator::visitUnary(Unary* curr) {
     case WidenHighSVecI16x8ToVecI32x4:
     case WidenLowUVecI16x8ToVecI32x4:
     case WidenHighUVecI16x8ToVecI32x4:
+    case WidenLowSVecI32x4ToVecI64x2:
+    case WidenHighSVecI32x4ToVecI64x2:
+    case WidenLowUVecI32x4ToVecI64x2:
+    case WidenHighUVecI32x4ToVecI64x2:
       shouldBeEqual(curr->type, Type(Type::v128), curr, "expected v128 type");
       shouldBeEqual(
         curr->value->type, Type(Type::v128), curr, "expected v128 operand");
@@ -1895,6 +1900,7 @@ void FunctionValidator::visitUnary(Unary* curr) {
     case BitmaskVecI8x16:
     case BitmaskVecI16x8:
     case BitmaskVecI32x4:
+    case BitmaskVecI64x2:
       shouldBeEqual(curr->type, Type(Type::i32), curr, "expected i32 type");
       shouldBeEqual(
         curr->value->type, Type(Type::v128), curr, "expected v128 operand");
diff --git a/src/wasm/wasm.cpp b/src/wasm/wasm.cpp
index 93a14aecc..8e6550c27 100644
--- a/src/wasm/wasm.cpp
+++ b/src/wasm/wasm.cpp
@@ -869,6 +869,10 @@ void Unary::finalize() {
     case WidenHighSVecI16x8ToVecI32x4:
     case WidenLowUVecI16x8ToVecI32x4:
     case WidenHighUVecI16x8ToVecI32x4:
+    case WidenLowSVecI32x4ToVecI64x2:
+    case WidenHighSVecI32x4ToVecI64x2:
+    case WidenLowUVecI32x4ToVecI64x2:
+    case WidenHighUVecI32x4ToVecI64x2:
       type = Type::v128;
       break;
     case AnyTrueVecI8x16:
@@ -882,6 +886,7 @@ void Unary::finalize() {
     case BitmaskVecI8x16:
     case BitmaskVecI16x8:
     case BitmaskVecI32x4:
+    case BitmaskVecI64x2:
       type = Type::i32;
       break;
 
diff --git a/test/simd.wast b/test/simd.wast
index 8ff2298a8..7296a075e 100644
--- a/test/simd.wast
+++ b/test/simd.wast
@@ -316,6 +316,12 @@
    (local.get $1)
   )
  )
+ (func $i64x2.eq (param $0 v128) (param $1 v128) (result v128)
+  (i64x2.eq
+   (local.get $0)
+   (local.get $1)
+  )
+ )
  (func $f32x4.eq (param $0 v128) (param $1 v128) (result v128)
   (f32x4.eq
    (local.get $0)
@@ -424,6 +430,34 @@
    (local.get $2)
   )
  )
+ (func $v8x16.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v8x16.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v16x8.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v16x8.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v32x4.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v32x4.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v64x2.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v64x2.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
  (func $v128.load8_lane (param $0 i32) (param $1 v128) (result v128)
   (v128.load8_lane 0
    (local.get $0)
@@ -903,6 +937,11 @@
    (local.get $0)
   )
  )
+ (func $i64x2.bitmask (param $0 v128) (result i32)
+  (i64x2.bitmask
+   (local.get $0)
+  )
+ )
  (func $i64x2.shl (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shl
    (local.get $0)
@@ -1281,6 +1320,26 @@
    (local.get $0)
   )
  )
+ (func $i64x2.widen_low_i32x4_s (param $0 v128) (result v128)
+  (i64x2.widen_low_i32x4_s
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_high_i32x4_s (param $0 v128) (result v128)
+  (i64x2.widen_high_i32x4_s
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_low_i32x4_u (param $0 v128) (result v128)
+  (i64x2.widen_low_i32x4_u
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_high_i32x4_u (param $0 v128) (result v128)
+  (i64x2.widen_high_i32x4_u
+   (local.get $0)
+  )
+ )
  (func $i16x8.load8x8_u (param $0 i32) (result v128)
   (i16x8.load8x8_u
    (local.get $0)
diff --git a/test/simd.wast.from-wast b/test/simd.wast.from-wast
index bf71a0ac3..ab30772d3 100644
--- a/test/simd.wast.from-wast
+++ b/test/simd.wast.from-wast
@@ -4,10 +4,10 @@
  (type $v128_=>_i32 (func (param v128) (result i32)))
  (type $i32_=>_v128 (func (param i32) (result v128)))
  (type $v128_i32_=>_v128 (func (param v128 i32) (result v128)))
+ (type $v128_v128_v128_=>_v128 (func (param v128 v128 v128) (result v128)))
  (type $i32_v128_=>_none (func (param i32 v128)))
  (type $i32_v128_=>_v128 (func (param i32 v128) (result v128)))
  (type $none_=>_v128 (func (result v128)))
- (type $v128_v128_v128_=>_v128 (func (param v128 v128 v128) (result v128)))
  (type $v128_=>_i64 (func (param v128) (result i64)))
  (type $v128_=>_f32 (func (param v128) (result f32)))
  (type $v128_=>_f64 (func (param v128) (result f64)))
@@ -333,6 +333,12 @@
    (local.get $1)
   )
  )
+ (func $i64x2.eq (param $0 v128) (param $1 v128) (result v128)
+  (i64x2.eq
+   (local.get $0)
+   (local.get $1)
+  )
+ )
  (func $f32x4.eq (param $0 v128) (param $1 v128) (result v128)
   (f32x4.eq
    (local.get $0)
@@ -441,6 +447,34 @@
    (local.get $2)
   )
  )
+ (func $v8x16.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v8x16.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v16x8.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v16x8.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v32x4.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v32x4.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v64x2.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v64x2.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
  (func $v128.load8_lane (param $0 i32) (param $1 v128) (result v128)
   (v128.load8_lane 0
    (local.get $0)
@@ -920,6 +954,11 @@
    (local.get $0)
   )
  )
+ (func $i64x2.bitmask (param $0 v128) (result i32)
+  (i64x2.bitmask
+   (local.get $0)
+  )
+ )
  (func $i64x2.shl (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shl
    (local.get $0)
@@ -1298,6 +1337,26 @@
    (local.get $0)
   )
  )
+ (func $i64x2.widen_low_i32x4_s (param $0 v128) (result v128)
+  (i64x2.widen_low_i32x4_s
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_high_i32x4_s (param $0 v128) (result v128)
+  (i64x2.widen_high_i32x4_s
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_low_i32x4_u (param $0 v128) (result v128)
+  (i64x2.widen_low_i32x4_u
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_high_i32x4_u (param $0 v128) (result v128)
+  (i64x2.widen_high_i32x4_u
+   (local.get $0)
+  )
+ )
  (func $i16x8.load8x8_u (param $0 i32) (result v128)
   (i16x8.load8x8_u
    (local.get $0)
diff --git a/test/simd.wast.fromBinary b/test/simd.wast.fromBinary
index 94435e153..935de0c09 100644
--- a/test/simd.wast.fromBinary
+++ b/test/simd.wast.fromBinary
@@ -4,10 +4,10 @@
  (type $v128_=>_i32 (func (param v128) (result i32)))
  (type $i32_=>_v128 (func (param i32) (result v128)))
  (type $v128_i32_=>_v128 (func (param v128 i32) (result v128)))
+ (type $v128_v128_v128_=>_v128 (func (param v128 v128 v128) (result v128)))
  (type $i32_v128_=>_none (func (param i32 v128)))
  (type $i32_v128_=>_v128 (func (param i32 v128) (result v128)))
  (type $none_=>_v128 (func (result v128)))
- (type $v128_v128_v128_=>_v128 (func (param v128 v128 v128) (result v128)))
  (type $v128_=>_i64 (func (param v128) (result i64)))
  (type $v128_=>_f32 (func (param v128) (result f32)))
  (type $v128_=>_f64 (func (param v128) (result f64)))
@@ -333,6 +333,12 @@
    (local.get $1)
   )
  )
+ (func $i64x2.eq (param $0 v128) (param $1 v128) (result v128)
+  (i64x2.eq
+   (local.get $0)
+   (local.get $1)
+  )
+ )
  (func $f32x4.eq (param $0 v128) (param $1 v128) (result v128)
   (f32x4.eq
    (local.get $0)
@@ -441,6 +447,34 @@
    (local.get $2)
   )
  )
+ (func $v8x16.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v8x16.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v16x8.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v16x8.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v32x4.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v32x4.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $v64x2.signselect (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v64x2.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
  (func $v128.load8_lane (param $0 i32) (param $1 v128) (result v128)
   (v128.load8_lane 0
    (local.get $0)
@@ -920,6 +954,11 @@
    (local.get $0)
   )
  )
+ (func $i64x2.bitmask (param $0 v128) (result i32)
+  (i64x2.bitmask
+   (local.get $0)
+  )
+ )
  (func $i64x2.shl (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shl
    (local.get $0)
@@ -1298,6 +1337,26 @@
    (local.get $0)
   )
  )
+ (func $i64x2.widen_low_i32x4_s (param $0 v128) (result v128)
+  (i64x2.widen_low_i32x4_s
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_high_i32x4_s (param $0 v128) (result v128)
+  (i64x2.widen_high_i32x4_s
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_low_i32x4_u (param $0 v128) (result v128)
+  (i64x2.widen_low_i32x4_u
+   (local.get $0)
+  )
+ )
+ (func $i64x2.widen_high_i32x4_u (param $0 v128) (result v128)
+  (i64x2.widen_high_i32x4_u
+   (local.get $0)
+  )
+ )
  (func $i16x8.load8x8_u (param $0 i32) (result v128)
   (i16x8.load8x8_u
    (local.get $0)
diff --git a/test/simd.wast.fromBinary.noDebugInfo b/test/simd.wast.fromBinary.noDebugInfo
index 25d3ef4e2..42543f448 100644
--- a/test/simd.wast.fromBinary.noDebugInfo
+++ b/test/simd.wast.fromBinary.noDebugInfo
@@ -4,10 +4,10 @@
  (type $v128_=>_i32 (func (param v128) (result i32)))
  (type $i32_=>_v128 (func (param i32) (result v128)))
  (type $v128_i32_=>_v128 (func (param v128 i32) (result v128)))
+ (type $v128_v128_v128_=>_v128 (func (param v128 v128 v128) (result v128)))
  (type $i32_v128_=>_none (func (param i32 v128)))
  (type $i32_v128_=>_v128 (func (param i32 v128) (result v128)))
  (type $none_=>_v128 (func (result v128)))
- (type $v128_v128_v128_=>_v128 (func (param v128 v128 v128) (result v128)))
  (type $v128_=>_i64 (func (param v128) (result i64)))
  (type $v128_=>_f32 (func (param v128) (result f32)))
  (type $v128_=>_f64 (func (param v128) (result f64)))
@@ -334,1011 +334,1070 @@
   )
  )
  (func $58 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.eq
+  (i64x2.eq
    (local.get $0)
    (local.get $1)
   )
  )
  (func $59 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.ne
+  (f32x4.eq
    (local.get $0)
    (local.get $1)
   )
  )
  (func $60 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.lt
+  (f32x4.ne
    (local.get $0)
    (local.get $1)
   )
  )
  (func $61 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.gt
+  (f32x4.lt
    (local.get $0)
    (local.get $1)
   )
  )
  (func $62 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.le
+  (f32x4.gt
    (local.get $0)
    (local.get $1)
   )
  )
  (func $63 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.ge
+  (f32x4.le
    (local.get $0)
    (local.get $1)
   )
  )
  (func $64 (param $0 v128) (param $1 v128) (result v128)
-  (f64x2.eq
+  (f32x4.ge
    (local.get $0)
    (local.get $1)
   )
  )
  (func $65 (param $0 v128) (param $1 v128) (result v128)
-  (f64x2.ne
+  (f64x2.eq
    (local.get $0)
    (local.get $1)
   )
  )
  (func $66 (param $0 v128) (param $1 v128) (result v128)
-  (f64x2.lt
+  (f64x2.ne
    (local.get $0)
    (local.get $1)
   )
  )
  (func $67 (param $0 v128) (param $1 v128) (result v128)
-  (f64x2.gt
+  (f64x2.lt
    (local.get $0)
    (local.get $1)
   )
  )
  (func $68 (param $0 v128) (param $1 v128) (result v128)
-  (f64x2.le
+  (f64x2.gt
    (local.get $0)
    (local.get $1)
   )
  )
  (func $69 (param $0 v128) (param $1 v128) (result v128)
+  (f64x2.le
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $70 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.ge
    (local.get $0)
    (local.get $1)
   )
  )
- (func $70 (param $0 v128) (result v128)
+ (func $71 (param $0 v128) (result v128)
   (v128.not
    (local.get $0)
   )
  )
- (func $71 (param $0 v128) (param $1 v128) (result v128)
+ (func $72 (param $0 v128) (param $1 v128) (result v128)
   (v128.and
    (local.get $0)
    (local.get $1)
   )
  )
- (func $72 (param $0 v128) (param $1 v128) (result v128)
+ (func $73 (param $0 v128) (param $1 v128) (result v128)
   (v128.or
    (local.get $0)
    (local.get $1)
   )
  )
- (func $73 (param $0 v128) (param $1 v128) (result v128)
+ (func $74 (param $0 v128) (param $1 v128) (result v128)
   (v128.xor
    (local.get $0)
    (local.get $1)
   )
  )
- (func $74 (param $0 v128) (param $1 v128) (result v128)
+ (func $75 (param $0 v128) (param $1 v128) (result v128)
   (v128.andnot
    (local.get $0)
    (local.get $1)
   )
  )
- (func $75 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $76 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (v128.bitselect
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $76 (param $0 i32) (param $1 v128) (result v128)
+ (func $77 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v8x16.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $78 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v16x8.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $79 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v32x4.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $80 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (v64x2.signselect
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+ (func $81 (param $0 i32) (param $1 v128) (result v128)
   (v128.load8_lane 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $77 (param $0 i32) (param $1 v128) (result v128)
+ (func $82 (param $0 i32) (param $1 v128) (result v128)
   (v128.load16_lane 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $78 (param $0 i32) (param $1 v128) (result v128)
+ (func $83 (param $0 i32) (param $1 v128) (result v128)
   (v128.load32_lane 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $79 (param $0 i32) (param $1 v128) (result v128)
+ (func $84 (param $0 i32) (param $1 v128) (result v128)
   (v128.load64_lane 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $80 (param $0 i32) (param $1 v128) (result v128)
+ (func $85 (param $0 i32) (param $1 v128) (result v128)
   (v128.load64_lane align=1 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $81 (param $0 i32) (param $1 v128) (result v128)
+ (func $86 (param $0 i32) (param $1 v128) (result v128)
   (v128.load64_lane offset=32 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $82 (param $0 i32) (param $1 v128) (result v128)
+ (func $87 (param $0 i32) (param $1 v128) (result v128)
   (v128.load64_lane offset=32 align=1 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $83 (param $0 i32) (param $1 v128)
+ (func $88 (param $0 i32) (param $1 v128)
   (v128.store8_lane 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $84 (param $0 i32) (param $1 v128)
+ (func $89 (param $0 i32) (param $1 v128)
   (v128.store16_lane 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $85 (param $0 i32) (param $1 v128)
+ (func $90 (param $0 i32) (param $1 v128)
   (v128.store32_lane 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $86 (param $0 i32) (param $1 v128)
+ (func $91 (param $0 i32) (param $1 v128)
   (v128.store64_lane 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $87 (param $0 i32) (param $1 v128)
+ (func $92 (param $0 i32) (param $1 v128)
   (v128.store64_lane align=1 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $88 (param $0 i32) (param $1 v128)
+ (func $93 (param $0 i32) (param $1 v128)
   (v128.store64_lane offset=32 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $89 (param $0 i32) (param $1 v128)
+ (func $94 (param $0 i32) (param $1 v128)
   (v128.store64_lane offset=32 align=1 0
    (local.get $0)
    (local.get $1)
   )
  )
- (func $90 (param $0 v128) (result v128)
+ (func $95 (param $0 v128) (result v128)
   (i8x16.popcnt
    (local.get $0)
   )
  )
- (func $91 (param $0 v128) (result v128)
+ (func $96 (param $0 v128) (result v128)
   (i8x16.abs
    (local.get $0)
   )
  )
- (func $92 (param $0 v128) (result v128)
+ (func $97 (param $0 v128) (result v128)
   (i8x16.neg
    (local.get $0)
   )
  )
- (func $93 (param $0 v128) (result i32)
+ (func $98 (param $0 v128) (result i32)
   (i8x16.any_true
    (local.get $0)
   )
  )
- (func $94 (param $0 v128) (result i32)
+ (func $99 (param $0 v128) (result i32)
   (i8x16.all_true
    (local.get $0)
   )
  )
- (func $95 (param $0 v128) (result i32)
+ (func $100 (param $0 v128) (result i32)
   (i8x16.bitmask
    (local.get $0)
   )
  )
- (func $96 (param $0 v128) (param $1 i32) (result v128)
+ (func $101 (param $0 v128) (param $1 i32) (result v128)
   (i8x16.shl
    (local.get $0)
    (local.get $1)
   )
  )
- (func $97 (param $0 v128) (param $1 i32) (result v128)
+ (func $102 (param $0 v128) (param $1 i32) (result v128)
   (i8x16.shr_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $98 (param $0 v128) (param $1 i32) (result v128)
+ (func $103 (param $0 v128) (param $1 i32) (result v128)
   (i8x16.shr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $99 (param $0 v128) (param $1 v128) (result v128)
+ (func $104 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $100 (param $0 v128) (param $1 v128) (result v128)
+ (func $105 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.add_saturate_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $101 (param $0 v128) (param $1 v128) (result v128)
+ (func $106 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.add_saturate_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $102 (param $0 v128) (param $1 v128) (result v128)
+ (func $107 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $103 (param $0 v128) (param $1 v128) (result v128)
+ (func $108 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.sub_saturate_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $104 (param $0 v128) (param $1 v128) (result v128)
+ (func $109 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.sub_saturate_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $105 (param $0 v128) (param $1 v128) (result v128)
+ (func $110 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $106 (param $0 v128) (param $1 v128) (result v128)
+ (func $111 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.min_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $107 (param $0 v128) (param $1 v128) (result v128)
+ (func $112 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.min_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $108 (param $0 v128) (param $1 v128) (result v128)
+ (func $113 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.max_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $109 (param $0 v128) (param $1 v128) (result v128)
+ (func $114 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.max_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $110 (param $0 v128) (param $1 v128) (result v128)
+ (func $115 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.avgr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $111 (param $0 v128) (result v128)
+ (func $116 (param $0 v128) (result v128)
   (i16x8.abs
    (local.get $0)
   )
  )
- (func $112 (param $0 v128) (result v128)
+ (func $117 (param $0 v128) (result v128)
   (i16x8.neg
    (local.get $0)
   )
  )
- (func $113 (param $0 v128) (result i32)
+ (func $118 (param $0 v128) (result i32)
   (i16x8.any_true
    (local.get $0)
   )
  )
- (func $114 (param $0 v128) (result i32)
+ (func $119 (param $0 v128) (result i32)
   (i16x8.all_true
    (local.get $0)
   )
  )
- (func $115 (param $0 v128) (result i32)
+ (func $120 (param $0 v128) (result i32)
   (i16x8.bitmask
    (local.get $0)
   )
  )
- (func $116 (param $0 v128) (param $1 i32) (result v128)
+ (func $121 (param $0 v128) (param $1 i32) (result v128)
   (i16x8.shl
    (local.get $0)
    (local.get $1)
   )
  )
- (func $117 (param $0 v128) (param $1 i32) (result v128)
+ (func $122 (param $0 v128) (param $1 i32) (result v128)
   (i16x8.shr_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $118 (param $0 v128) (param $1 i32) (result v128)
+ (func $123 (param $0 v128) (param $1 i32) (result v128)
   (i16x8.shr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $119 (param $0 v128) (param $1 v128) (result v128)
+ (func $124 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $120 (param $0 v128) (param $1 v128) (result v128)
+ (func $125 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.add_saturate_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $121 (param $0 v128) (param $1 v128) (result v128)
+ (func $126 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.add_saturate_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $122 (param $0 v128) (param $1 v128) (result v128)
+ (func $127 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $123 (param $0 v128) (param $1 v128) (result v128)
+ (func $128 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.sub_saturate_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $124 (param $0 v128) (param $1 v128) (result v128)
+ (func $129 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.sub_saturate_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $125 (param $0 v128) (param $1 v128) (result v128)
+ (func $130 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $126 (param $0 v128) (param $1 v128) (result v128)
+ (func $131 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.min_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $127 (param $0 v128) (param $1 v128) (result v128)
+ (func $132 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.min_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $128 (param $0 v128) (param $1 v128) (result v128)
+ (func $133 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.max_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $129 (param $0 v128) (param $1 v128) (result v128)
+ (func $134 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.max_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $130 (param $0 v128) (param $1 v128) (result v128)
+ (func $135 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.avgr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $131 (param $0 v128) (param $1 v128) (result v128)
+ (func $136 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.q15mulr_sat_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $132 (param $0 v128) (param $1 v128) (result v128)
+ (func $137 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.extmul_low_i8x16_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $133 (param $0 v128) (param $1 v128) (result v128)
+ (func $138 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.extmul_high_i8x16_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $134 (param $0 v128) (param $1 v128) (result v128)
+ (func $139 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.extmul_low_i8x16_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $135 (param $0 v128) (param $1 v128) (result v128)
+ (func $140 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.extmul_high_i8x16_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $136 (param $0 v128) (result v128)
+ (func $141 (param $0 v128) (result v128)
   (i32x4.abs
    (local.get $0)
   )
  )
- (func $137 (param $0 v128) (result v128)
+ (func $142 (param $0 v128) (result v128)
   (i32x4.neg
    (local.get $0)
   )
  )
- (func $138 (param $0 v128) (result i32)
+ (func $143 (param $0 v128) (result i32)
   (i32x4.any_true
    (local.get $0)
   )
  )
- (func $139 (param $0 v128) (result i32)
+ (func $144 (param $0 v128) (result i32)
   (i32x4.all_true
    (local.get $0)
   )
  )
- (func $140 (param $0 v128) (result i32)
+ (func $145 (param $0 v128) (result i32)
   (i32x4.bitmask
    (local.get $0)
   )
  )
- (func $141 (param $0 v128) (param $1 i32) (result v128)
+ (func $146 (param $0 v128) (param $1 i32) (result v128)
   (i32x4.shl
    (local.get $0)
    (local.get $1)
   )
  )
- (func $142 (param $0 v128) (param $1 i32) (result v128)
+ (func $147 (param $0 v128) (param $1 i32) (result v128)
   (i32x4.shr_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $143 (param $0 v128) (param $1 i32) (result v128)
+ (func $148 (param $0 v128) (param $1 i32) (result v128)
   (i32x4.shr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $144 (param $0 v128) (param $1 v128) (result v128)
+ (func $149 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $145 (param $0 v128) (param $1 v128) (result v128)
+ (func $150 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $146 (param $0 v128) (param $1 v128) (result v128)
+ (func $151 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $147 (param $0 v128) (param $1 v128) (result v128)
+ (func $152 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.min_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $148 (param $0 v128) (param $1 v128) (result v128)
+ (func $153 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.min_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $149 (param $0 v128) (param $1 v128) (result v128)
+ (func $154 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.max_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $150 (param $0 v128) (param $1 v128) (result v128)
+ (func $155 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.max_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $151 (param $0 v128) (param $1 v128) (result v128)
+ (func $156 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.dot_i16x8_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $152 (param $0 v128) (param $1 v128) (result v128)
+ (func $157 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.extmul_low_i16x8_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $153 (param $0 v128) (param $1 v128) (result v128)
+ (func $158 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.extmul_high_i16x8_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $154 (param $0 v128) (param $1 v128) (result v128)
+ (func $159 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.extmul_low_i16x8_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $155 (param $0 v128) (param $1 v128) (result v128)
+ (func $160 (param $0 v128) (param $1 v128) (result v128)
   (i32x4.extmul_high_i16x8_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $156 (param $0 v128) (result v128)
+ (func $161 (param $0 v128) (result v128)
   (i64x2.neg
    (local.get $0)
   )
  )
- (func $157 (param $0 v128) (result i32)
+ (func $162 (param $0 v128) (result i32)
   (i64x2.any_true
    (local.get $0)
   )
  )
- (func $158 (param $0 v128) (result i32)
+ (func $163 (param $0 v128) (result i32)
   (i64x2.all_true
    (local.get $0)
   )
  )
- (func $159 (param $0 v128) (param $1 i32) (result v128)
+ (func $164 (param $0 v128) (result i32)
+  (i64x2.bitmask
+   (local.get $0)
+  )
+ )
+ (func $165 (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shl
    (local.get $0)
    (local.get $1)
   )
  )
- (func $160 (param $0 v128) (param $1 i32) (result v128)
+ (func $166 (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shr_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $161 (param $0 v128) (param $1 i32) (result v128)
+ (func $167 (param $0 v128) (param $1 i32) (result v128)
   (i64x2.shr_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $162 (param $0 v128) (param $1 v128) (result v128)
+ (func $168 (param $0 v128) (param $1 v128) (result v128)
   (i64x2.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $163 (param $0 v128) (param $1 v128) (result v128)
+ (func $169 (param $0 v128) (param $1 v128) (result v128)
   (i64x2.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $164 (param $0 v128) (param $1 v128) (result v128)
+ (func $170 (param $0 v128) (param $1 v128) (result v128)
   (i64x2.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $165 (param $0 v128) (param $1 v128) (result v128)
+ (func $171 (param $0 v128) (param $1 v128) (result v128)
   (i64x2.extmul_low_i32x4_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $166 (param $0 v128) (param $1 v128) (result v128)
+ (func $172 (param $0 v128) (param $1 v128) (result v128)
   (i64x2.extmul_high_i32x4_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $167 (param $0 v128) (param $1 v128) (result v128)
+ (func $173 (param $0 v128) (param $1 v128) (result v128)
   (i64x2.extmul_low_i32x4_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $168 (param $0 v128) (param $1 v128) (result v128)
+ (func $174 (param $0 v128) (param $1 v128) (result v128)
   (i64x2.extmul_high_i32x4_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $169 (param $0 v128) (param $1 v128) (result v128)
+ (func $175 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $170 (param $0 v128) (param $1 v128) (result v128)
+ (func $176 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $171 (param $0 v128) (param $1 v128) (result v128)
+ (func $177 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $172 (param $0 v128) (param $1 v128) (result v128)
+ (func $178 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.div
    (local.get $0)
    (local.get $1)
   )
  )
- (func $173 (param $0 v128) (param $1 v128) (result v128)
+ (func $179 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.min
    (local.get $0)
    (local.get $1)
   )
  )
- (func $174 (param $0 v128) (param $1 v128) (result v128)
+ (func $180 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.max
    (local.get $0)
    (local.get $1)
   )
  )
- (func $175 (param $0 v128) (param $1 v128) (result v128)
+ (func $181 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.pmin
    (local.get $0)
    (local.get $1)
   )
  )
- (func $176 (param $0 v128) (param $1 v128) (result v128)
+ (func $182 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.pmax
    (local.get $0)
    (local.get $1)
   )
  )
- (func $177 (param $0 v128) (result v128)
+ (func $183 (param $0 v128) (result v128)
   (f32x4.ceil
    (local.get $0)
   )
  )
- (func $178 (param $0 v128) (result v128)
+ (func $184 (param $0 v128) (result v128)
   (f32x4.floor
    (local.get $0)
   )
  )
- (func $179 (param $0 v128) (result v128)
+ (func $185 (param $0 v128) (result v128)
   (f32x4.trunc
    (local.get $0)
   )
  )
- (func $180 (param $0 v128) (result v128)
+ (func $186 (param $0 v128) (result v128)
   (f32x4.nearest
    (local.get $0)
   )
  )
- (func $181 (param $0 v128) (result v128)
+ (func $187 (param $0 v128) (result v128)
   (f32x4.abs
    (local.get $0)
   )
  )
- (func $182 (param $0 v128) (result v128)
+ (func $188 (param $0 v128) (result v128)
   (f32x4.neg
    (local.get $0)
   )
  )
- (func $183 (param $0 v128) (result v128)
+ (func $189 (param $0 v128) (result v128)
   (f32x4.sqrt
    (local.get $0)
   )
  )
- (func $184 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $190 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f32x4.qfma
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $185 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $191 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f32x4.qfms
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $186 (param $0 v128) (param $1 v128) (result v128)
+ (func $192 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $187 (param $0 v128) (param $1 v128) (result v128)
+ (func $193 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $188 (param $0 v128) (param $1 v128) (result v128)
+ (func $194 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $189 (param $0 v128) (param $1 v128) (result v128)
+ (func $195 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.div
    (local.get $0)
    (local.get $1)
   )
  )
- (func $190 (param $0 v128) (param $1 v128) (result v128)
+ (func $196 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.min
    (local.get $0)
    (local.get $1)
   )
  )
- (func $191 (param $0 v128) (param $1 v128) (result v128)
+ (func $197 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.max
    (local.get $0)
    (local.get $1)
   )
  )
- (func $192 (param $0 v128) (param $1 v128) (result v128)
+ (func $198 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.pmin
    (local.get $0)
    (local.get $1)
   )
  )
- (func $193 (param $0 v128) (param $1 v128) (result v128)
+ (func $199 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.pmax
    (local.get $0)
    (local.get $1)
   )
  )
- (func $194 (param $0 v128) (result v128)
+ (func $200 (param $0 v128) (result v128)
   (f64x2.ceil
    (local.get $0)
   )
  )
- (func $195 (param $0 v128) (result v128)
+ (func $201 (param $0 v128) (result v128)
   (f64x2.floor
    (local.get $0)
   )
  )
- (func $196 (param $0 v128) (result v128)
+ (func $202 (param $0 v128) (result v128)
   (f64x2.trunc
    (local.get $0)
   )
  )
- (func $197 (param $0 v128) (result v128)
+ (func $203 (param $0 v128) (result v128)
   (f64x2.nearest
    (local.get $0)
   )
  )
- (func $198 (param $0 v128) (result v128)
+ (func $204 (param $0 v128) (result v128)
   (f64x2.abs
    (local.get $0)
   )
  )
- (func $199 (param $0 v128) (result v128)
+ (func $205 (param $0 v128) (result v128)
   (f64x2.neg
    (local.get $0)
   )
  )
- (func $200 (param $0 v128) (result v128)
+ (func $206 (param $0 v128) (result v128)
   (f64x2.sqrt
    (local.get $0)
   )
  )
- (func $201 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $207 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f64x2.qfma
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $202 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $208 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f64x2.qfms
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $203 (param $0 v128) (result v128)
+ (func $209 (param $0 v128) (result v128)
   (i32x4.trunc_sat_f32x4_s
    (local.get $0)
   )
  )
- (func $204 (param $0 v128) (result v128)
+ (func $210 (param $0 v128) (result v128)
   (i32x4.trunc_sat_f32x4_u
    (local.get $0)
   )
  )
- (func $205 (param $0 v128) (result v128)
+ (func $211 (param $0 v128) (result v128)
   (i64x2.trunc_sat_f64x2_s
    (local.get $0)
   )
  )
- (func $206 (param $0 v128) (result v128)
+ (func $212 (param $0 v128) (result v128)
   (i64x2.trunc_sat_f64x2_u
    (local.get $0)
   )
  )
- (func $207 (param $0 v128) (result v128)
+ (func $213 (param $0 v128) (result v128)
   (f32x4.convert_i32x4_s
    (local.get $0)
   )
  )
- (func $208 (param $0 v128) (result v128)
+ (func $214 (param $0 v128) (result v128)
   (f32x4.convert_i32x4_u
    (local.get $0)
   )
  )
- (func $209 (param $0 v128) (result v128)
+ (func $215 (param $0 v128) (result v128)
   (f64x2.convert_i64x2_s
    (local.get $0)
   )
  )
- (func $210 (param $0 v128) (result v128)
+ (func $216 (param $0 v128) (result v128)
   (f64x2.convert_i64x2_u
    (local.get $0)
   )
  )
- (func $211 (param $0 i32) (result v128)
+ (func $217 (param $0 i32) (result v128)
   (v8x16.load_splat
    (local.get $0)
   )
  )
- (func $212 (param $0 i32) (result v128)
+ (func $218 (param $0 i32) (result v128)
   (v16x8.load_splat
    (local.get $0)
   )
  )
- (func $213 (param $0 i32) (result v128)
+ (func $219 (param $0 i32) (result v128)
   (v32x4.load_splat
    (local.get $0)
   )
  )
- (func $214 (param $0 i32) (result v128)
+ (func $220 (param $0 i32) (result v128)
   (v64x2.load_splat
    (local.get $0)
   )
  )
- (func $215 (param $0 v128) (param $1 v128) (result v128)
+ (func $221 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.narrow_i16x8_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $216 (param $0 v128) (param $1 v128) (result v128)
+ (func $222 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.narrow_i16x8_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $217 (param $0 v128) (param $1 v128) (result v128)
+ (func $223 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.narrow_i32x4_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $218 (param $0 v128) (param $1 v128) (result v128)
+ (func $224 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.narrow_i32x4_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $219 (param $0 v128) (result v128)
+ (func $225 (param $0 v128) (result v128)
   (i16x8.widen_low_i8x16_s
    (local.get $0)
   )
  )
- (func $220 (param $0 v128) (result v128)
+ (func $226 (param $0 v128) (result v128)
   (i16x8.widen_high_i8x16_s
    (local.get $0)
   )
  )
- (func $221 (param $0 v128) (result v128)
+ (func $227 (param $0 v128) (result v128)
   (i16x8.widen_low_i8x16_u
    (local.get $0)
   )
  )
- (func $222 (param $0 v128) (result v128)
+ (func $228 (param $0 v128) (result v128)
   (i16x8.widen_high_i8x16_u
    (local.get $0)
   )
  )
- (func $223 (param $0 v128) (result v128)
+ (func $229 (param $0 v128) (result v128)
   (i32x4.widen_low_i16x8_s
    (local.get $0)
   )
  )
- (func $224 (param $0 v128) (result v128)
+ (func $230 (param $0 v128) (result v128)
   (i32x4.widen_high_i16x8_s
    (local.get $0)
   )
  )
- (func $225 (param $0 v128) (result v128)
+ (func $231 (param $0 v128) (result v128)
   (i32x4.widen_low_i16x8_u
    (local.get $0)
   )
  )
- (func $226 (param $0 v128) (result v128)
+ (func $232 (param $0 v128) (result v128)
   (i32x4.widen_high_i16x8_u
    (local.get $0)
   )
  )
- (func $227 (param $0 i32) (result v128)
+ (func $233 (param $0 v128) (result v128)
+  (i64x2.widen_low_i32x4_s
+   (local.get $0)
+  )
+ )
+ (func $234 (param $0 v128) (result v128)
+  (i64x2.widen_high_i32x4_s
+   (local.get $0)
+  )
+ )
+ (func $235 (param $0 v128) (result v128)
+  (i64x2.widen_low_i32x4_u
+   (local.get $0)
+  )
+ )
+ (func $236 (param $0 v128) (result v128)
+  (i64x2.widen_high_i32x4_u
+   (local.get $0)
+  )
+ )
+ (func $237 (param $0 i32) (result v128)
   (i16x8.load8x8_u
    (local.get $0)
   )
  )
- (func $228 (param $0 i32) (result v128)
+ (func $238 (param $0 i32) (result v128)
   (i16x8.load8x8_s
    (local.get $0)
   )
  )
- (func $229 (param $0 i32) (result v128)
+ (func $239 (param $0 i32) (result v128)
   (i32x4.load16x4_s
    (local.get $0)
   )
  )
- (func $230 (param $0 i32) (result v128)
+ (func $240 (param $0 i32) (result v128)
   (i32x4.load16x4_u
    (local.get $0)
   )
  )
- (func $231 (param $0 i32) (result v128)
+ (func $241 (param $0 i32) (result v128)
   (i64x2.load32x2_s
    (local.get $0)
   )
  )
- (func $232 (param $0 i32) (result v128)
+ (func $242 (param $0 i32) (result v128)
   (i64x2.load32x2_u
    (local.get $0)
   )
  )
- (func $233 (param $0 i32) (result v128)
+ (func $243 (param $0 i32) (result v128)
   (v128.load32_zero
    (local.get $0)
   )
  )
- (func $234 (param $0 i32) (result v128)
+ (func $244 (param $0 i32) (result v128)
   (v128.load64_zero
    (local.get $0)
   )
  )
- (func $235 (param $0 v128) (param $1 v128) (result v128)
+ (func $245 (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 332d48d01..51bea4deb 100644
--- a/test/spec/simd.wast
+++ b/test/spec/simd.wast
@@ -84,6 +84,7 @@
  (func (export "i32x4.le_u") (param $0 v128) (param $1 v128) (result v128) (i32x4.le_u (local.get $0) (local.get $1)))
  (func (export "i32x4.ge_s") (param $0 v128) (param $1 v128) (result v128) (i32x4.ge_s (local.get $0) (local.get $1)))
  (func (export "i32x4.ge_u") (param $0 v128) (param $1 v128) (result v128) (i32x4.ge_u (local.get $0) (local.get $1)))
+ (func (export "i64x2.eq") (param $0 v128) (param $1 v128) (result v128) (i64x2.eq (local.get $0) (local.get $1)))
  (func (export "f32x4.eq") (param $0 v128) (param $1 v128) (result v128) (f32x4.eq (local.get $0) (local.get $1)))
  (func (export "f32x4.ne") (param $0 v128) (param $1 v128) (result v128) (f32x4.ne (local.get $0) (local.get $1)))
  (func (export "f32x4.lt") (param $0 v128) (param $1 v128) (result v128) (f32x4.lt (local.get $0) (local.get $1)))
@@ -104,6 +105,18 @@
  (func (export "v128.bitselect") (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
    (v128.bitselect (local.get $0) (local.get $1) (local.get $2))
  )
+ (func (export "v8x16.signselect") (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+   (v8x16.signselect (local.get $0) (local.get $1) (local.get $2))
+ )
+ (func (export "v16x8.signselect") (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+   (v16x8.signselect (local.get $0) (local.get $1) (local.get $2))
+ )
+ (func (export "v32x4.signselect") (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+   (v32x4.signselect (local.get $0) (local.get $1) (local.get $2))
+ )
+ (func (export "v64x2.signselect") (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+   (v64x2.signselect (local.get $0) (local.get $1) (local.get $2))
+ )
  (func (export "v128.load8_lane") (param $0 i32) (param $1 v128) (result v128) (v128.load8_lane 0 (local.get $0) (local.get $1)))
  (func (export "v128.load16_lane") (param $0 i32) (param $1 v128) (result v128) (v128.load16_lane 0 (local.get $0) (local.get $1)))
  (func (export "v128.load32_lane") (param $0 i32) (param $1 v128) (result v128) (v128.load32_lane 0 (local.get $0) (local.get $1)))
@@ -174,6 +187,7 @@
  (func (export "i64x2.neg") (param $0 v128) (result v128) (i64x2.neg (local.get $0)))
  (func (export "i64x2.any_true") (param $0 v128) (result i32) (i64x2.any_true (local.get $0)))
  (func (export "i64x2.all_true") (param $0 v128) (result i32) (i64x2.all_true (local.get $0)))
+ (func (export "i64x2.bitmask") (param $0 v128) (result i32) (i64x2.bitmask (local.get $0)))
  (func (export "i64x2.shl") (param $0 v128) (param $1 i32) (result v128) (i64x2.shl (local.get $0) (local.get $1)))
  (func (export "i64x2.shr_s") (param $0 v128) (param $1 i32) (result v128) (i64x2.shr_s (local.get $0) (local.get $1)))
  (func (export "i64x2.shr_u") (param $0 v128) (param $1 i32) (result v128) (i64x2.shr_u (local.get $0) (local.get $1)))
@@ -238,6 +252,10 @@
  (func (export "i32x4.widen_high_i16x8_s") (param $0 v128) (result v128) (i32x4.widen_high_i16x8_s (local.get $0)))
  (func (export "i32x4.widen_low_i16x8_u") (param $0 v128) (result v128) (i32x4.widen_low_i16x8_u (local.get $0)))
  (func (export "i32x4.widen_high_i16x8_u") (param $0 v128) (result v128) (i32x4.widen_high_i16x8_u (local.get $0)))
+ (func (export "i64x2.widen_low_i32x4_s") (param $0 v128) (result v128) (i64x2.widen_low_i32x4_s (local.get $0)))
+ (func (export "i64x2.widen_high_i32x4_s") (param $0 v128) (result v128) (i64x2.widen_high_i32x4_s (local.get $0)))
+ (func (export "i64x2.widen_low_i32x4_u") (param $0 v128) (result v128) (i64x2.widen_low_i32x4_u (local.get $0)))
+ (func (export "i64x2.widen_high_i32x4_u") (param $0 v128) (result v128) (i64x2.widen_high_i32x4_u (local.get $0)))
  (func (export "i16x8.load8x8_u") (param $0 i32) (result v128) (i16x8.load8x8_u (local.get $0)))
  (func (export "i16x8.load8x8_s") (param $0 i32) (result v128) (i16x8.load8x8_s (local.get $0)))
  (func (export "i32x4.load16x4_u") (param $0 i32) (result v128) (i32x4.load16x4_u (local.get $0)))
@@ -474,6 +492,9 @@
 (assert_return (invoke "i32x4.ge_s" (v128.const i32x4 0 -1 53 -7) (v128.const i32x4 0 53 -7 -1)) (v128.const i32x4 -1 0 -1 0))
 (assert_return (invoke "i32x4.ge_u" (v128.const i32x4 0 -1 53 -7) (v128.const i32x4 0 53 -7 -1)) (v128.const i32x4 -1 -1 0 0))
 
+;; i64x2 comparisons
+(assert_return (invoke "i64x2.eq" (v128.const i64x2 0 -1) (v128.const i64x2 -1 -1)) (v128.const i64x2 0 -1))
+
 ;; f32x4 comparisons
 (assert_return (invoke "f32x4.eq" (v128.const f32x4 0 -1 1 0) (v128.const f32x4 0 0 -1 1)) (v128.const i32x4 -1 0 0 0))
 (assert_return (invoke "f32x4.ne" (v128.const f32x4 0 -1 1 0) (v128.const f32x4 0 0 -1 1)) (v128.const i32x4 0 -1 -1 -1))
@@ -521,6 +542,7 @@
   )
   (v128.const i32x4 0xABABABAB 0xAAAAAAAA 0xBBBBBBBB 0xAABBAABB)
 )
+;; TODO: signselect tests
 
 ;; load/store lane
 (assert_return (invoke "v128.load8_lane"
@@ -851,6 +873,7 @@
 (assert_return (invoke "i64x2.all_true" (v128.const i64x2 0 0)) (i32.const 0))
 (assert_return (invoke "i64x2.all_true" (v128.const i64x2 1 0)) (i32.const 0))
 (assert_return (invoke "i64x2.all_true" (v128.const i64x2 1 1)) (i32.const 1))
+;; TODO: test i64x2.bitmask
 (assert_return (invoke "i64x2.shl" (v128.const i64x2 1 0x8000000000000000) (i32.const 1)) (v128.const i64x2 2 0))
 (assert_return (invoke "i64x2.shl" (v128.const i64x2 1 0x8000000000000000) (i32.const 64)) (v128.const i64x2 1 0x8000000000000000))
 (assert_return (invoke "i64x2.shr_s" (v128.const i64x2 1 0x8000000000000000) (i32.const 1)) (v128.const i64x2 0 0xc000000000000000))
@@ -1005,6 +1028,7 @@
 (assert_return (invoke "i32x4.widen_high_i16x8_s" (v128.const i16x8 0 1 -1 32768 32767 32769 16384 -16384)) (v128.const i32x4 32767 -32767 16384 -16384))
 (assert_return (invoke "i32x4.widen_low_i16x8_u" (v128.const i16x8 0 1 -1 32768 32767 32769 16384 -16384)) (v128.const i32x4 0 1 65535 32768))
 (assert_return (invoke "i32x4.widen_high_i16x8_u" (v128.const i16x8 0 1 -1 32768 32767 32769 16384 -16384)) (v128.const i32x4 32767 32769 16384 49152))
+;; TODO: test i64x2 widens
 (assert_return (invoke "i16x8.load8x8_s" (i32.const 256)) (v128.const i16x8 0xff80 0xff90 0xffa0 0xffb0 0xffc0 0xffd0 0xffe0 0xfff0))
 (assert_return (invoke "i16x8.load8x8_u" (i32.const 256)) (v128.const i16x8 0x0080 0x0090 0x00a0 0x00b0 0x00c0 0x00d0 0x00e0 0x00f0))
 (assert_return (invoke "i32x4.load16x4_s" (i32.const 256)) (v128.const i32x4 0xffff9080 0xffffb0a0 0xffffd0c0 0xfffff0e0))
author	Thomas Lively <7121787+tlively@users.noreply.github.com>	2020-12-11 19:06:34 -0500
committer	GitHub <noreply@github.com>	2020-12-11 19:06:34 -0500
commit	dd33ae0512aa97f537d40e61631d268c8a8fd17e (patch)
tree	e5c219ba349e56940669bb873bb57a3f2626bf77
parent	290147d8d43a7448d68939ec711b524ba4fb3fbd (diff)
download	binaryen-dd33ae0512aa97f537d40e61631d268c8a8fd17e.tar.gz binaryen-dd33ae0512aa97f537d40e61631d268c8a8fd17e.tar.bz2 binaryen-dd33ae0512aa97f537d40e61631d268c8a8fd17e.zip