Implement relaxed SIMD dot product instructions (#4586)

As proposed in https://github.com/WebAssembly/relaxed-simd/issues/52.

13 files changed, 235 insertions, 13 deletions

diff --git a/scripts/gen-s-parser.py b/scripts/gen-s-parser.py
index 26a96fdd6..29ce83820 100755
--- a/scripts/gen-s-parser.py
+++ b/scripts/gen-s-parser.py
@@ -533,7 +533,11 @@ instructions = [
     ("f32x4.relaxed_max", "makeBinary(s, BinaryOp::RelaxedMaxVecF32x4)"),
     ("f64x2.relaxed_min", "makeBinary(s, BinaryOp::RelaxedMinVecF64x2)"),
     ("f64x2.relaxed_max", "makeBinary(s, BinaryOp::RelaxedMaxVecF64x2)"),
-    ("i16x8.relaxed_q15mulr_s",   "makeBinary(s, BinaryOp::RelaxedQ15MulrSVecI16x8)"),
+    ("i16x8.relaxed_q15mulr_s", "makeBinary(s, BinaryOp::RelaxedQ15MulrSVecI16x8)"),
+    ("i16x8.dot_i8x16_i7x16_s", "makeBinary(s, BinaryOp::DotI8x16I7x16SToVecI16x8)"),
+    ("i16x8.dot_i8x16_i7x16_u", "makeBinary(s, BinaryOp::DotI8x16I7x16UToVecI16x8)"),
+    ("i32x4.dot_i8x16_i7x16_add_s", "makeSIMDTernary(s, SIMDTernaryOp::DotI8x16I7x16AddSToVecI32x4)"),
+    ("i32x4.dot_i8x16_i7x16_add_u", "makeSIMDTernary(s, SIMDTernaryOp::DotI8x16I7x16AddUToVecI32x4)"),
 
     # reference types instructions
     ("ref.null",             "makeRefNull(s)"),
diff --git a/src/gen-s-parser.inc b/src/gen-s-parser.inc
index 895676fa8..d9505bcdd 100644
--- a/src/gen-s-parser.inc
+++ b/src/gen-s-parser.inc
@@ -933,6 +933,17 @@ switch (op[0]) {
           case 'b':
             if (strcmp(op, "i16x8.bitmask") == 0) { return makeUnary(s, UnaryOp::BitmaskVecI16x8); }
             goto parse_error;
+          case 'd': {
+            switch (op[22]) {
+              case 's':
+                if (strcmp(op, "i16x8.dot_i8x16_i7x16_s") == 0) { return makeBinary(s, BinaryOp::DotI8x16I7x16SToVecI16x8); }
+                goto parse_error;
+              case 'u':
+                if (strcmp(op, "i16x8.dot_i8x16_i7x16_u") == 0) { return makeBinary(s, BinaryOp::DotI8x16I7x16UToVecI16x8); }
+                goto parse_error;
+              default: goto parse_error;
+            }
+          }
           case 'e': {
             switch (op[7]) {
               case 'q':
@@ -1692,9 +1703,25 @@ switch (op[0]) {
                   case 'b':
                     if (strcmp(op, "i32x4.bitmask") == 0) { return makeUnary(s, UnaryOp::BitmaskVecI32x4); }
                     goto parse_error;
-                  case 'd':
-                    if (strcmp(op, "i32x4.dot_i16x8_s") == 0) { return makeBinary(s, BinaryOp::DotSVecI16x8ToVecI32x4); }
-                    goto parse_error;
+                  case 'd': {
+                    switch (op[11]) {
+                      case '1':
+                        if (strcmp(op, "i32x4.dot_i16x8_s") == 0) { return makeBinary(s, BinaryOp::DotSVecI16x8ToVecI32x4); }
+                        goto parse_error;
+                      case '8': {
+                        switch (op[26]) {
+                          case 's':
+                            if (strcmp(op, "i32x4.dot_i8x16_i7x16_add_s") == 0) { return makeSIMDTernary(s, SIMDTernaryOp::DotI8x16I7x16AddSToVecI32x4); }
+                            goto parse_error;
+                          case 'u':
+                            if (strcmp(op, "i32x4.dot_i8x16_i7x16_add_u") == 0) { return makeSIMDTernary(s, SIMDTernaryOp::DotI8x16I7x16AddUToVecI32x4); }
+                            goto parse_error;
+                          default: goto parse_error;
+                        }
+                      }
+                      default: goto parse_error;
+                    }
+                  }
                   case 'e': {
                     switch (op[7]) {
                       case 'q':
diff --git a/src/ir/cost.h b/src/ir/cost.h
index 9423101d3..c0a8a76e9 100644
--- a/src/ir/cost.h
+++ b/src/ir/cost.h
@@ -494,6 +494,8 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, CostType> {
       case SwizzleVecI8x16:
       case RelaxedSwizzleVecI8x16:
       case RelaxedQ15MulrSVecI16x8:
+      case DotI8x16I7x16SToVecI16x8:
+      case DotI8x16I7x16UToVecI16x8:
         ret = 1;
         break;
       case InvalidBinary:
@@ -541,6 +543,8 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, CostType> {
       case RelaxedFmsVecF32x4:
       case RelaxedFmaVecF64x2:
       case RelaxedFmsVecF64x2:
+      case DotI8x16I7x16AddSToVecI32x4:
+      case DotI8x16I7x16AddUToVecI32x4:
         ret = 1;
         break;
     }
diff --git a/src/literal.h b/src/literal.h
index 4ad81268e..0438e0686 100644
--- a/src/literal.h
+++ b/src/literal.h
@@ -572,6 +572,8 @@ public:
   Literal minUI32x4(const Literal& other) const;
   Literal maxSI32x4(const Literal& other) const;
   Literal maxUI32x4(const Literal& other) const;
+  Literal dotSI8x16toI16x8(const Literal& other) const;
+  Literal dotUI8x16toI16x8(const Literal& other) const;
   Literal dotSI16x8toI32x4(const Literal& other) const;
   Literal extMulLowSI32x4(const Literal& other) const;
   Literal extMulHighSI32x4(const Literal& other) const;
diff --git a/src/passes/Print.cpp b/src/passes/Print.cpp
index 240db975b..9a4235cbf 100644
--- a/src/passes/Print.cpp
+++ b/src/passes/Print.cpp
@@ -683,6 +683,12 @@ struct PrintExpressionContents
       case RelaxedFmsVecF64x2:
         o << "f64x2.relaxed_fms";
         break;
+      case DotI8x16I7x16AddSToVecI32x4:
+        o << "i32x4.dot_i8x16_i7x16_add_s";
+        break;
+      case DotI8x16I7x16AddUToVecI32x4:
+        o << "i32x4.dot_i8x16_i7x16_add_u";
+        break;
     }
     restoreNormalColor(o);
   }
@@ -1854,6 +1860,12 @@ struct PrintExpressionContents
       case RelaxedQ15MulrSVecI16x8:
         o << "i16x8.relaxed_q15mulr_s";
         break;
+      case DotI8x16I7x16SToVecI16x8:
+        o << "i16x8.dot_i8x16_i7x16_s";
+        break;
+      case DotI8x16I7x16UToVecI16x8:
+        o << "i16x8.dot_i8x16_i7x16_u";
+        break;
 
       case InvalidBinary:
         WASM_UNREACHABLE("unvalid binary operator");
diff --git a/src/wasm-binary.h b/src/wasm-binary.h
index 859efee54..784321c33 100644
--- a/src/wasm-binary.h
+++ b/src/wasm-binary.h
@@ -1038,6 +1038,10 @@ enum ASTNodes {
   F64x2RelaxedMin = 0xd4,
   F64x2RelaxedMax = 0xee,
   I16x8RelaxedQ15MulrS = 0x111,
+  I16x8DotI8x16I7x16S = 0x112,
+  I16x8DotI8x16I7x16U = 0x113,
+  I32x4DotI8x16I7x16AddS = 0x114,
+  I32x4DotI8x16I7x16AddU = 0x115,
 
   // bulk memory opcodes
 
diff --git a/src/wasm-interpreter.h b/src/wasm-interpreter.h
index 416647137..3b0ceea99 100644
--- a/src/wasm-interpreter.h
+++ b/src/wasm-interpreter.h
@@ -1016,6 +1016,11 @@ public:
       case RelaxedSwizzleVecI8x16:
         return left.swizzleI8x16(right);
 
+      case DotI8x16I7x16SToVecI16x8:
+        return left.dotSI8x16toI16x8(right);
+      case DotI8x16I7x16UToVecI16x8:
+        return left.dotUI8x16toI16x8(right);
+
       case InvalidBinary:
         WASM_UNREACHABLE("invalid binary op");
     }
@@ -1124,7 +1129,7 @@ public:
       case RelaxedFmsVecF64x2:
         return a.relaxedFmsF64x2(b, c);
       default:
-        // TODO: implement signselect
+        // TODO: implement signselect and dot_add
         WASM_UNREACHABLE("not implemented");
     }
   }
diff --git a/src/wasm.h b/src/wasm.h
index 2bf057c7b..3b736fcbc 100644
--- a/src/wasm.h
+++ b/src/wasm.h
@@ -473,6 +473,8 @@ enum BinaryOp {
   RelaxedMinVecF64x2,
   RelaxedMaxVecF64x2,
   RelaxedQ15MulrSVecI16x8,
+  DotI8x16I7x16SToVecI16x8,
+  DotI8x16I7x16UToVecI16x8,
 
   InvalidBinary
 };
@@ -552,6 +554,8 @@ enum SIMDTernaryOp {
   LaneselectI16x8,
   LaneselectI32x4,
   LaneselectI64x2,
+  DotI8x16I7x16AddSToVecI32x4,
+  DotI8x16I7x16AddUToVecI32x4,
 };
 
 enum RefIsOp {
diff --git a/src/wasm/literal.cpp b/src/wasm/literal.cpp
index ea398acd3..fe763680e 100644
--- a/src/wasm/literal.cpp
+++ b/src/wasm/literal.cpp
@@ -2372,17 +2372,33 @@ Literal Literal::pmaxF64x2(const Literal& other) const {
   return binary<2, &Literal::getLanesF64x2, &Literal::pmax>(*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());
+template<size_t Lanes,
+         size_t Factor,
+         LaneArray<Lanes * Factor> (Literal::*IntoLanes)() const>
+static Literal dot(const Literal& left, const Literal& right) {
+  LaneArray<Lanes* Factor> lhs = (left.*IntoLanes)();
+  LaneArray<Lanes* Factor> rhs = (right.*IntoLanes)();
+  LaneArray<Lanes> result;
+  for (size_t i = 0; i < Lanes; ++i) {
+    result[i] = Literal(int32_t(0));
+    for (size_t j = 0; j < Factor; ++j) {
+      result[i] = Literal(result[i].geti32() + lhs[i * Factor + j].geti32() *
+                                                 rhs[i * Factor + j].geti32());
+    }
   }
   return Literal(result);
 }
 
+Literal Literal::dotSI8x16toI16x8(const Literal& other) const {
+  return dot<8, 2, &Literal::getLanesSI8x16>(*this, other);
+}
+Literal Literal::dotUI8x16toI16x8(const Literal& other) const {
+  return dot<8, 2, &Literal::getLanesUI8x16>(*this, other);
+}
+Literal Literal::dotSI16x8toI32x4(const Literal& other) const {
+  return dot<4, 2, &Literal::getLanesSI16x8>(*this, other);
+}
+
 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 026e277fb..52af3151f 100644
--- a/src/wasm/wasm-binary.cpp
+++ b/src/wasm/wasm-binary.cpp
@@ -5601,6 +5601,14 @@ bool WasmBinaryBuilder::maybeVisitSIMDBinary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Binary>();
       curr->op = RelaxedQ15MulrSVecI16x8;
       break;
+    case BinaryConsts::I16x8DotI8x16I7x16S:
+      curr = allocator.alloc<Binary>();
+      curr->op = DotI8x16I7x16SToVecI16x8;
+      break;
+    case BinaryConsts::I16x8DotI8x16I7x16U:
+      curr = allocator.alloc<Binary>();
+      curr->op = DotI8x16I7x16UToVecI16x8;
+      break;
     default:
       return false;
   }
@@ -6075,6 +6083,14 @@ bool WasmBinaryBuilder::maybeVisitSIMDTernary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<SIMDTernary>();
       curr->op = RelaxedFmsVecF64x2;
       break;
+    case BinaryConsts::I32x4DotI8x16I7x16AddS:
+      curr = allocator.alloc<SIMDTernary>();
+      curr->op = DotI8x16I7x16AddSToVecI32x4;
+      break;
+    case BinaryConsts::I32x4DotI8x16I7x16AddU:
+      curr = allocator.alloc<SIMDTernary>();
+      curr->op = DotI8x16I7x16AddUToVecI32x4;
+      break;
     default:
       return false;
   }
diff --git a/src/wasm/wasm-stack.cpp b/src/wasm/wasm-stack.cpp
index 67104a8f1..176f2f53f 100644
--- a/src/wasm/wasm-stack.cpp
+++ b/src/wasm/wasm-stack.cpp
@@ -569,6 +569,12 @@ void BinaryInstWriter::visitSIMDTernary(SIMDTernary* curr) {
     case RelaxedFmsVecF64x2:
       o << U32LEB(BinaryConsts::F64x2RelaxedFms);
       break;
+    case DotI8x16I7x16AddSToVecI32x4:
+      o << U32LEB(BinaryConsts::I32x4DotI8x16I7x16AddS);
+      break;
+    case DotI8x16I7x16AddUToVecI32x4:
+      o << U32LEB(BinaryConsts::I32x4DotI8x16I7x16AddU);
+      break;
   }
 }
 
@@ -1846,6 +1852,14 @@ void BinaryInstWriter::visitBinary(Binary* curr) {
       o << int8_t(BinaryConsts::SIMDPrefix)
         << U32LEB(BinaryConsts::I16x8RelaxedQ15MulrS);
       break;
+    case DotI8x16I7x16SToVecI16x8:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I16x8DotI8x16I7x16S);
+      break;
+    case DotI8x16I7x16UToVecI16x8:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I16x8DotI8x16I7x16U);
+      break;
 
     case InvalidBinary:
       WASM_UNREACHABLE("invalid binary op");
diff --git a/src/wasm/wasm-validator.cpp b/src/wasm/wasm-validator.cpp
index dd896e139..bebb3eabe 100644
--- a/src/wasm/wasm-validator.cpp
+++ b/src/wasm/wasm-validator.cpp
@@ -1635,7 +1635,9 @@ void FunctionValidator::visitBinary(Binary* curr) {
     case NarrowUVecI32x4ToVecI16x8:
     case SwizzleVecI8x16:
     case RelaxedSwizzleVecI8x16:
-    case RelaxedQ15MulrSVecI16x8: {
+    case RelaxedQ15MulrSVecI16x8:
+    case DotI8x16I7x16SToVecI16x8:
+    case DotI8x16I7x16UToVecI16x8: {
       shouldBeEqualOrFirstIsUnreachable(
         curr->left->type, Type(Type::v128), curr, "v128 op");
       shouldBeEqualOrFirstIsUnreachable(
diff --git a/test/lit/relaxed-simd.wast b/test/lit/relaxed-simd.wast
index e1d936c17..9d957ff60 100644
--- a/test/lit/relaxed-simd.wast
+++ b/test/lit/relaxed-simd.wast
@@ -369,6 +369,88 @@
   )
  )
 
+ ;; CHECK-BINARY:      (func $i16x8.dot_i8x16_i7x16_s (param $0 v128) (param $1 v128) (result v128)
+ ;; CHECK-BINARY-NEXT:  (i16x8.dot_i8x16_i7x16_s
+ ;; CHECK-BINARY-NEXT:   (local.get $0)
+ ;; CHECK-BINARY-NEXT:   (local.get $1)
+ ;; CHECK-BINARY-NEXT:  )
+ ;; CHECK-BINARY-NEXT: )
+ ;; CHECK-TEXT:      (func $i16x8.dot_i8x16_i7x16_s (param $0 v128) (param $1 v128) (result v128)
+ ;; CHECK-TEXT-NEXT:  (i16x8.dot_i8x16_i7x16_s
+ ;; CHECK-TEXT-NEXT:   (local.get $0)
+ ;; CHECK-TEXT-NEXT:   (local.get $1)
+ ;; CHECK-TEXT-NEXT:  )
+ ;; CHECK-TEXT-NEXT: )
+ (func $i16x8.dot_i8x16_i7x16_s (param $0 v128) (param $1 v128) (result v128)
+  (i16x8.dot_i8x16_i7x16_s
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+
+ ;; CHECK-BINARY:      (func $i16x8.dot_i8x16_i7x16_u (param $0 v128) (param $1 v128) (result v128)
+ ;; CHECK-BINARY-NEXT:  (i16x8.dot_i8x16_i7x16_u
+ ;; CHECK-BINARY-NEXT:   (local.get $0)
+ ;; CHECK-BINARY-NEXT:   (local.get $1)
+ ;; CHECK-BINARY-NEXT:  )
+ ;; CHECK-BINARY-NEXT: )
+ ;; CHECK-TEXT:      (func $i16x8.dot_i8x16_i7x16_u (param $0 v128) (param $1 v128) (result v128)
+ ;; CHECK-TEXT-NEXT:  (i16x8.dot_i8x16_i7x16_u
+ ;; CHECK-TEXT-NEXT:   (local.get $0)
+ ;; CHECK-TEXT-NEXT:   (local.get $1)
+ ;; CHECK-TEXT-NEXT:  )
+ ;; CHECK-TEXT-NEXT: )
+ (func $i16x8.dot_i8x16_i7x16_u (param $0 v128) (param $1 v128) (result v128)
+  (i16x8.dot_i8x16_i7x16_u
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+
+;; CHECK-BINARY:      (func $i32x4.dot_i8x16_i7x16_add_s (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+;; CHECK-BINARY-NEXT:  (i32x4.dot_i8x16_i7x16_add_s
+;; CHECK-BINARY-NEXT:   (local.get $0)
+;; CHECK-BINARY-NEXT:   (local.get $1)
+;; CHECK-BINARY-NEXT:   (local.get $2)
+;; CHECK-BINARY-NEXT:  )
+;; CHECK-BINARY-NEXT: )
+;; CHECK-TEXT:      (func $i32x4.dot_i8x16_i7x16_add_s (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+;; CHECK-TEXT-NEXT:  (i32x4.dot_i8x16_i7x16_add_s
+;; CHECK-TEXT-NEXT:   (local.get $0)
+;; CHECK-TEXT-NEXT:   (local.get $1)
+;; CHECK-TEXT-NEXT:   (local.get $2)
+;; CHECK-TEXT-NEXT:  )
+;; CHECK-TEXT-NEXT: )
+(func $i32x4.dot_i8x16_i7x16_add_s (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (i32x4.dot_i8x16_i7x16_add_s
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+
+;; CHECK-BINARY:      (func $i32x4.dot_i8x16_i7x16_add_u (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+;; CHECK-BINARY-NEXT:  (i32x4.dot_i8x16_i7x16_add_u
+;; CHECK-BINARY-NEXT:   (local.get $0)
+;; CHECK-BINARY-NEXT:   (local.get $1)
+;; CHECK-BINARY-NEXT:   (local.get $2)
+;; CHECK-BINARY-NEXT:  )
+;; CHECK-BINARY-NEXT: )
+;; CHECK-TEXT:      (func $i32x4.dot_i8x16_i7x16_add_u (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+;; CHECK-TEXT-NEXT:  (i32x4.dot_i8x16_i7x16_add_u
+;; CHECK-TEXT-NEXT:   (local.get $0)
+;; CHECK-TEXT-NEXT:   (local.get $1)
+;; CHECK-TEXT-NEXT:   (local.get $2)
+;; CHECK-TEXT-NEXT:  )
+;; CHECK-TEXT-NEXT: )
+(func $i32x4.dot_i8x16_i7x16_add_u (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+  (i32x4.dot_i8x16_i7x16_add_u
+   (local.get $0)
+   (local.get $1)
+   (local.get $2)
+  )
+ )
+
 )
 ;; CHECK-NODEBUG:      (type $v128_v128_v128_=>_v128 (func (param v128 v128 v128) (result v128)))
 
@@ -507,3 +589,33 @@
 ;; CHECK-NODEBUG-NEXT:   (local.get $1)
 ;; CHECK-NODEBUG-NEXT:  )
 ;; CHECK-NODEBUG-NEXT: )
+
+;; CHECK-NODEBUG:      (func $18 (param $0 v128) (param $1 v128) (result v128)
+;; CHECK-NODEBUG-NEXT:  (i16x8.dot_i8x16_i7x16_s
+;; CHECK-NODEBUG-NEXT:   (local.get $0)
+;; CHECK-NODEBUG-NEXT:   (local.get $1)
+;; CHECK-NODEBUG-NEXT:  )
+;; CHECK-NODEBUG-NEXT: )
+
+;; CHECK-NODEBUG:      (func $19 (param $0 v128) (param $1 v128) (result v128)
+;; CHECK-NODEBUG-NEXT:  (i16x8.dot_i8x16_i7x16_u
+;; CHECK-NODEBUG-NEXT:   (local.get $0)
+;; CHECK-NODEBUG-NEXT:   (local.get $1)
+;; CHECK-NODEBUG-NEXT:  )
+;; CHECK-NODEBUG-NEXT: )
+
+;; CHECK-NODEBUG:      (func $20 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+;; CHECK-NODEBUG-NEXT:  (i32x4.dot_i8x16_i7x16_add_s
+;; CHECK-NODEBUG-NEXT:   (local.get $0)
+;; CHECK-NODEBUG-NEXT:   (local.get $1)
+;; CHECK-NODEBUG-NEXT:   (local.get $2)
+;; CHECK-NODEBUG-NEXT:  )
+;; CHECK-NODEBUG-NEXT: )
+
+;; CHECK-NODEBUG:      (func $21 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+;; CHECK-NODEBUG-NEXT:  (i32x4.dot_i8x16_i7x16_add_u
+;; CHECK-NODEBUG-NEXT:   (local.get $0)
+;; CHECK-NODEBUG-NEXT:   (local.get $1)
+;; CHECK-NODEBUG-NEXT:   (local.get $2)
+;; CHECK-NODEBUG-NEXT:  )
+;; CHECK-NODEBUG-NEXT: )