Add i32x4.dot_i16x8_s (#2420)

This experimental instruction is specified in
https://github.com/WebAssembly/simd/pull/127 and is being implemented
to enable further investigation of its performance impact.

15 files changed, 43 insertions, 0 deletions

diff --git a/src/binaryen-c.cpp b/src/binaryen-c.cpp
index 2b9b9d2ae..4e5ad22f1 100644
--- a/src/binaryen-c.cpp
+++ b/src/binaryen-c.cpp
@@ -842,6 +842,9 @@ BinaryenOp BinaryenMinSVecI32x4(void) { return MinSVecI32x4; }
 BinaryenOp BinaryenMinUVecI32x4(void) { return MinUVecI32x4; }
 BinaryenOp BinaryenMaxSVecI32x4(void) { return MaxSVecI32x4; }
 BinaryenOp BinaryenMaxUVecI32x4(void) { return MaxUVecI32x4; }
+BinaryenOp BinaryenDotSVecI16x8ToVecI32x4(void) {
+  return DotSVecI16x8ToVecI32x4;
+}
 BinaryenOp BinaryenNegVecI64x2(void) { return NegVecI64x2; }
 BinaryenOp BinaryenAnyTrueVecI64x2(void) { return AnyTrueVecI64x2; }
 BinaryenOp BinaryenAllTrueVecI64x2(void) { return AllTrueVecI64x2; }
diff --git a/src/binaryen-c.h b/src/binaryen-c.h
index 936cfb0f6..dd5acf325 100644
--- a/src/binaryen-c.h
+++ b/src/binaryen-c.h
@@ -504,6 +504,7 @@ BINARYEN_API BinaryenOp BinaryenMinSVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenMinUVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenMaxSVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenMaxUVecI32x4(void);
+BINARYEN_API BinaryenOp BinaryenDotSVecI16x8ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenNegVecI64x2(void);
 BINARYEN_API BinaryenOp BinaryenAnyTrueVecI64x2(void);
 BINARYEN_API BinaryenOp BinaryenAllTrueVecI64x2(void);
diff --git a/src/gen-s-parser.inc b/src/gen-s-parser.inc
index 9a8d9b88c..5e6e37f5a 100644
--- a/src/gen-s-parser.inc
+++ b/src/gen-s-parser.inc
@@ -1409,6 +1409,9 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 'd':
+                if (strcmp(op, "i32x4.dot_i16x8_s") == 0) { return makeBinary(s, BinaryOp::DotSVecI16x8ToVecI32x4); }
+                goto parse_error;
               case 'e': {
                 switch (op[7]) {
                   case 'q':
diff --git a/src/ir/cost.h b/src/ir/cost.h
index 26c120675..60eb84b08 100644
--- a/src/ir/cost.h
+++ b/src/ir/cost.h
@@ -648,6 +648,9 @@ struct CostAnalyzer : public Visitor<CostAnalyzer, Index> {
       case MaxUVecI32x4:
         ret = 1;
         break;
+      case DotSVecI16x8ToVecI32x4:
+        ret = 1;
+        break;
       case AddVecI64x2:
         ret = 1;
         break;
diff --git a/src/js/binaryen.js-post.js b/src/js/binaryen.js-post.js
index f2d679900..37d2432cb 100644
--- a/src/js/binaryen.js-post.js
+++ b/src/js/binaryen.js-post.js
@@ -357,6 +357,7 @@ Module['MinSVecI16x8'] = Module['_BinaryenMinSVecI16x8']();
 Module['MinUVecI16x8'] = Module['_BinaryenMinUVecI16x8']();
 Module['MaxSVecI16x8'] = Module['_BinaryenMaxSVecI16x8']();
 Module['MaxUVecI16x8'] = Module['_BinaryenMaxUVecI16x8']();
+Module['DotSVecI16x8ToVecI32x4'] = Module['_BinaryenDotSVecI16x8ToVecI32x4']();
 Module['NegVecI32x4'] = Module['_BinaryenNegVecI32x4']();
 Module['AnyTrueVecI32x4'] = Module['_BinaryenAnyTrueVecI32x4']();
 Module['AllTrueVecI32x4'] = Module['_BinaryenAllTrueVecI32x4']();
@@ -1676,6 +1677,9 @@ function wrapModule(module, self) {
     'max_u': function(left, right) {
       return Module['_BinaryenBinary'](module, Module['MaxUVecI32x4'], left, right);
     },
+    'dot_i16x8_s': function(left, right) {
+      return Module['_BinaryenBinary'](module, Module['DotSVecI16x8ToVecI32x4'], left, right);
+    },
     'trunc_sat_f32x4_s': function(value) {
       return Module['_BinaryenUnary'](module, Module['TruncSatSVecF32x4ToVecI32x4'], value);
     },
diff --git a/src/literal.h b/src/literal.h
index 007a2cde8..1106d1c27 100644
--- a/src/literal.h
+++ b/src/literal.h
@@ -364,6 +364,7 @@ public:
   Literal minUI32x4(const Literal& other) const;
   Literal maxSI32x4(const Literal& other) const;
   Literal maxUI32x4(const Literal& other) const;
+  Literal dotSI16x8toI32x4(const Literal& other) const;
   Literal negI64x2() const;
   Literal anyTrueI64x2() const;
   Literal allTrueI64x2() const;
diff --git a/src/passes/Print.cpp b/src/passes/Print.cpp
index e383974d6..bfd12dc94 100644
--- a/src/passes/Print.cpp
+++ b/src/passes/Print.cpp
@@ -1222,6 +1222,9 @@ struct PrintExpressionContents
       case MaxUVecI32x4:
         o << "i32x4.max_u";
         break;
+      case DotSVecI16x8ToVecI32x4:
+        o << "i32x4.dot_i16x8_s";
+        break;
       case AddVecI64x2:
         o << "i64x2.add";
         break;
diff --git a/src/tools/fuzzing.h b/src/tools/fuzzing.h
index c6cce4427..c66795c11 100644
--- a/src/tools/fuzzing.h
+++ b/src/tools/fuzzing.h
@@ -2169,6 +2169,7 @@ private:
                                  MinUVecI32x4,
                                  MaxSVecI32x4,
                                  MaxUVecI32x4,
+                                 DotSVecI16x8ToVecI32x4,
                                  AddVecI64x2,
                                  SubVecI64x2,
                                  AddVecF32x4,
diff --git a/src/wasm-binary.h b/src/wasm-binary.h
index 2a457f5a7..7014fd875 100644
--- a/src/wasm-binary.h
+++ b/src/wasm-binary.h
@@ -831,6 +831,7 @@ enum ASTNodes {
   I32x4MinU = 0x81,
   I32x4MaxS = 0x82,
   I32x4MaxU = 0x83,
+  I32x4DotSVecI16x8 = 0xd9,
   I64x2Neg = 0x84,
   I64x2AnyTrue = 0x85,
   I64x2AllTrue = 0x86,
diff --git a/src/wasm-interpreter.h b/src/wasm-interpreter.h
index f8cf3ae42..dc2696716 100644
--- a/src/wasm-interpreter.h
+++ b/src/wasm-interpreter.h
@@ -801,6 +801,8 @@ public:
         return left.maxSI32x4(right);
       case MaxUVecI32x4:
         return left.maxUI32x4(right);
+      case DotSVecI16x8ToVecI32x4:
+        return left.dotSI16x8toI32x4(right);
       case AddVecI64x2:
         return left.addI64x2(right);
       case SubVecI64x2:
diff --git a/src/wasm.h b/src/wasm.h
index ca1114182..77eeea5d3 100644
--- a/src/wasm.h
+++ b/src/wasm.h
@@ -383,6 +383,7 @@ enum BinaryOp {
   MinUVecI32x4,
   MaxSVecI32x4,
   MaxUVecI32x4,
+  DotSVecI16x8ToVecI32x4,
   AddVecI64x2,
   SubVecI64x2,
   AddVecF32x4,
diff --git a/src/wasm/literal.cpp b/src/wasm/literal.cpp
index f4ddf69cc..41d98dea3 100644
--- a/src/wasm/literal.cpp
+++ b/src/wasm/literal.cpp
@@ -1829,6 +1829,17 @@ Literal Literal::maxF64x2(const Literal& other) const {
   return binary<2, &Literal::getLanesF64x2, &Literal::max>(*this, other);
 }
 
+Literal Literal::dotSI16x8toI32x4(const Literal& other) const {
+  LaneArray<8> lhs = getLanesSI16x8();
+  LaneArray<8> rhs = other.getLanesSI16x8();
+  LaneArray<4> result;
+  for (size_t i = 0; i < 4; ++i) {
+    result[i] = Literal(lhs[i * 2].geti32() * rhs[i * 2].geti32() +
+                        lhs[i * 2 + 1].geti32() * rhs[i * 2 + 1].geti32());
+  }
+  return Literal(result);
+}
+
 Literal Literal::bitselectV128(const Literal& left,
                                const Literal& right) const {
   return andV128(left).orV128(notV128().andV128(right));
diff --git a/src/wasm/wasm-binary.cpp b/src/wasm/wasm-binary.cpp
index cb67d1556..8d6721c9c 100644
--- a/src/wasm/wasm-binary.cpp
+++ b/src/wasm/wasm-binary.cpp
@@ -3939,6 +3939,10 @@ bool WasmBinaryBuilder::maybeVisitSIMDBinary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Binary>();
       curr->op = MaxUVecI32x4;
       break;
+    case BinaryConsts::I32x4DotSVecI16x8:
+      curr = allocator.alloc<Binary>();
+      curr->op = DotSVecI16x8ToVecI32x4;
+      break;
     case BinaryConsts::I64x2Add:
       curr = allocator.alloc<Binary>();
       curr->op = AddVecI64x2;
diff --git a/src/wasm/wasm-stack.cpp b/src/wasm/wasm-stack.cpp
index d4af126b7..c1d4f1222 100644
--- a/src/wasm/wasm-stack.cpp
+++ b/src/wasm/wasm-stack.cpp
@@ -1459,6 +1459,10 @@ void BinaryInstWriter::visitBinary(Binary* curr) {
     case MaxUVecI32x4:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::I32x4MaxU);
       break;
+    case DotSVecI16x8ToVecI32x4:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I32x4DotSVecI16x8);
+      break;
     case AddVecI64x2:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::I64x2Add);
       break;
diff --git a/src/wasm/wasm-validator.cpp b/src/wasm/wasm-validator.cpp
index ed54173ed..82cd6d6f0 100644
--- a/src/wasm/wasm-validator.cpp
+++ b/src/wasm/wasm-validator.cpp
@@ -1361,6 +1361,7 @@ void FunctionValidator::visitBinary(Binary* curr) {
     case MinUVecI32x4:
     case MaxSVecI32x4:
     case MaxUVecI32x4:
+    case DotSVecI16x8ToVecI32x4:
     case AddVecI64x2:
     case SubVecI64x2:
     case AddVecF32x4: