[FP16] Implement f32x4.promote_low_f16x8.

scripts/gen-s-parser.py

@@ -549,6 +549,7 @@
     ("i16x8.trunc_sat_f16x8_u",  "makeUnary(UnaryOp::TruncSatUVecF16x8ToVecI16x8)"),
     ("f16x8.convert_i16x8_s",    "makeUnary(UnaryOp::ConvertSVecI16x8ToVecF16x8)"),
     ("f16x8.convert_i16x8_u",    "makeUnary(UnaryOp::ConvertUVecI16x8ToVecF16x8)"),
+    ("f32x4.promote_low_f16x8",  "makeUnary(UnaryOp::PromoteLowVecF16x8ToVecF32x4)"),
     ("f16x8.madd",               "makeSIMDTernary(SIMDTernaryOp::MaddVecF16x8)"),
     ("f16x8.nmadd",              "makeSIMDTernary(SIMDTernaryOp::NmaddVecF16x8)"),
 

src/binaryen-c.cpp

@@ -1021,6 +1021,9 @@ BinaryenOp BinaryenDemoteZeroVecF64x2ToVecF32x4(void) {
 BinaryenOp BinaryenPromoteLowVecF32x4ToVecF64x2(void) {
   return PromoteLowVecF32x4ToVecF64x2;
 }
+BinaryenOp BinaryenPromoteLowVecF16x8ToVecF32x4(void) {
+  return PromoteLowVecF16x8ToVecF32x4;
+}
 BinaryenOp BinaryenRelaxedTruncSVecF32x4ToVecI32x4(void) {
   return RelaxedTruncSVecF32x4ToVecI32x4;
 }

src/binaryen-c.h

@@ -684,6 +684,7 @@ BINARYEN_API BinaryenOp BinaryenTruncSatZeroSVecF64x2ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenTruncSatZeroUVecF64x2ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenDemoteZeroVecF64x2ToVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenPromoteLowVecF32x4ToVecF64x2(void);
+BINARYEN_API BinaryenOp BinaryenPromoteLowVecF16x8ToVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenRelaxedTruncSVecF32x4ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenRelaxedTruncUVecF32x4ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenRelaxedTruncZeroSVecF64x2ToVecI32x4(void);

src/gen-s-parser.inc

@@ -1127,16 +1127,27 @@ switch (buf[0]) {
                 }
               }
               case 'p': {
-                switch (buf[8]) {
-                  case 'a':
-                    if (op == "f32x4.pmax"sv) {
-                      CHECK_ERR(makeBinary(ctx, pos, annotations, BinaryOp::PMaxVecF32x4));
-                      return Ok{};
+                switch (buf[7]) {
+                  case 'm': {
+                    switch (buf[8]) {
+                      case 'a':
+                        if (op == "f32x4.pmax"sv) {
+                          CHECK_ERR(makeBinary(ctx, pos, annotations, BinaryOp::PMaxVecF32x4));
+                          return Ok{};
+                        }
+                        goto parse_error;
+                      case 'i':
+                        if (op == "f32x4.pmin"sv) {
+                          CHECK_ERR(makeBinary(ctx, pos, annotations, BinaryOp::PMinVecF32x4));
+                          return Ok{};
+                        }
+                        goto parse_error;
+                      default: goto parse_error;
                     }
-                    goto parse_error;
-                  case 'i':
-                    if (op == "f32x4.pmin"sv) {
-                      CHECK_ERR(makeBinary(ctx, pos, annotations, BinaryOp::PMinVecF32x4));
+                  }
+                  case 'r':
+                    if (op == "f32x4.promote_low_f16x8"sv) {
+                      CHECK_ERR(makeUnary(ctx, pos, annotations, UnaryOp::PromoteLowVecF16x8ToVecF32x4));
                       return Ok{};
                     }
                     goto parse_error;

src/ir/child-typer.h

@@ -447,6 +447,7 @@ template<typename Subtype> struct ChildTyper : OverriddenVisitor<Subtype> {
       case TruncSatUVecF16x8ToVecI16x8:
       case ConvertSVecI16x8ToVecF16x8:
       case ConvertUVecI16x8ToVecF16x8:
+      case PromoteLowVecF16x8ToVecF32x4:
       case AnyTrueVec128:
       case AllTrueVecI8x16:
       case AllTrueVecI16x8:

src/ir/cost.h

@@ -284,6 +284,7 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, CostType> {
       case TruncSatUVecF16x8ToVecI16x8:
       case ConvertSVecI16x8ToVecF16x8:
       case ConvertUVecI16x8ToVecF16x8:
+      case PromoteLowVecF16x8ToVecF32x4:
         ret = 1;
         break;
       case InvalidUnary:

src/literal.h

@@ -723,6 +723,7 @@ class Literal {
   Literal truncSatZeroUToI32x4() const;
   Literal demoteZeroToF32x4() const;
   Literal promoteLowToF64x2() const;
+  Literal promoteLowF16x8ToF32x4() const;
   Literal truncSatToSI16x8() const;
   Literal truncSatToUI16x8() const;
   Literal convertSToF16x8() const;

src/passes/Print.cpp

@@ -1401,6 +1401,9 @@ struct PrintExpressionContents
       case ConvertUVecI16x8ToVecF16x8:
         o << "f16x8.convert_i16x8_u";
         break;
+      case PromoteLowVecF16x8ToVecF32x4:
+        o << "f32x4.promote_low_f16x8";
+        break;
       case InvalidUnary:
         WASM_UNREACHABLE("unvalid unary operator");
     }

src/tools/fuzzing/fuzzing.cpp

@@ -4526,7 +4526,8 @@ Expression* TranslateToFuzzReader::makeUnary(Type type) {
                                          TruncSatSVecF16x8ToVecI16x8,
                                          TruncSatUVecF16x8ToVecI16x8,
                                          ConvertSVecI16x8ToVecF16x8,
-                                         ConvertUVecI16x8ToVecF16x8)),
+                                         ConvertUVecI16x8ToVecF16x8,
+                                         PromoteLowVecF16x8ToVecF32x4)),
                              make(Type::v128)});
       }
       WASM_UNREACHABLE("invalid value");

src/wasm-binary.h

@@ -1126,6 +1126,7 @@ enum ASTNodes {
   I16x8TruncSatF16x8U = 0x146,
   F16x8ConvertI16x8S = 0x147,
   F16x8ConvertI16x8U = 0x148,
+  F32x4PromoteLowF16x8 = 0x14b,
 
   // bulk memory opcodes
 

src/wasm-interpreter.h

@@ -1164,6 +1164,8 @@ class ExpressionRunner : public OverriddenVisitor<SubType, Flow> {
         return value.convertSToF16x8();
       case ConvertUVecI16x8ToVecF16x8:
         return value.convertUToF16x8();
+      case PromoteLowVecF16x8ToVecF32x4:
+        return value.promoteLowF16x8ToF32x4();
       case InvalidUnary:
         WASM_UNREACHABLE("invalid unary op");
     }

src/wasm.h

@@ -251,6 +251,7 @@ enum UnaryOp {
   TruncSatUVecF16x8ToVecI16x8,
   ConvertSVecI16x8ToVecF16x8,
   ConvertUVecI16x8ToVecF16x8,
+  PromoteLowVecF16x8ToVecF32x4,
 
   InvalidUnary
 };

src/wasm/literal.cpp

@@ -2915,6 +2915,14 @@ Literal Literal::demoteZeroToF32x4() const {
 Literal Literal::promoteLowToF64x2() const {
   return extendF32<LaneOrder::Low>(*this);
 }
+Literal Literal::promoteLowF16x8ToF32x4() const {
+  auto lanes = getLanesF16x8();
+  LaneArray<4> result;
+  for (size_t i = 0; i < 4; ++i) {
+    result[i] = lanes[i];
+  }
+  return Literal(result);
+}
 
 Literal Literal::swizzleI8x16(const Literal& other) const {
   auto lanes = getLanesUI8x16();

src/wasm/wasm-binary.cpp

@@ -4474,6 +4474,8 @@ Result<> WasmBinaryReader::readInst() {
           return builder.makeUnary(ConvertSVecI16x8ToVecF16x8);
         case BinaryConsts::F16x8ConvertI16x8U:
           return builder.makeUnary(ConvertUVecI16x8ToVecF16x8);
+        case BinaryConsts::F32x4PromoteLowF16x8:
+          return builder.makeUnary(PromoteLowVecF16x8ToVecF32x4);
         case BinaryConsts::I8x16ExtractLaneS:
           return builder.makeSIMDExtract(ExtractLaneSVecI8x16,
                                          getLaneIndex(16));

src/wasm/wasm-stack.cpp

@@ -1459,6 +1459,10 @@ void BinaryInstWriter::visitUnary(Unary* curr) {
       o << static_cast<int8_t>(BinaryConsts::SIMDPrefix)
         << U32LEB(BinaryConsts::F16x8ConvertI16x8U);
       break;
+    case PromoteLowVecF16x8ToVecF32x4:
+      o << static_cast<int8_t>(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::F32x4PromoteLowF16x8);
+      break;
     case InvalidUnary:
       WASM_UNREACHABLE("invalid unary op");
   }

src/wasm/wasm-validator.cpp

@@ -2380,6 +2380,7 @@ void FunctionValidator::visitUnary(Unary* curr) {
     case TruncSatZeroUVecF64x2ToVecI32x4:
     case DemoteZeroVecF64x2ToVecF32x4:
     case PromoteLowVecF32x4ToVecF64x2:
+    case PromoteLowVecF16x8ToVecF32x4:
     case RelaxedTruncSVecF32x4ToVecI32x4:
     case RelaxedTruncUVecF32x4ToVecI32x4:
     case RelaxedTruncZeroSVecF64x2ToVecI32x4:

src/wasm/wasm.cpp

@@ -714,6 +714,7 @@ void Unary::finalize() {
     case TruncSatUVecF16x8ToVecI16x8:
     case ConvertSVecI16x8ToVecF16x8:
     case ConvertUVecI16x8ToVecF16x8:
+    case PromoteLowVecF16x8ToVecF32x4:
       type = Type::v128;
       break;
     case AnyTrueVec128:

test/lit/basic/f16.wast

@@ -597,6 +597,22 @@
    (local.get $0)
   )
  )
+
+ ;; CHECK-TEXT:      (func $f32x4.promote_low_f16x8 (type $1) (param $0 v128) (result v128)
+ ;; CHECK-TEXT-NEXT:  (f32x4.promote_low_f16x8
+ ;; CHECK-TEXT-NEXT:   (local.get $0)
+ ;; CHECK-TEXT-NEXT:  )
+ ;; CHECK-TEXT-NEXT: )
+ ;; CHECK-BIN:      (func $f32x4.promote_low_f16x8 (type $1) (param $0 v128) (result v128)
+ ;; CHECK-BIN-NEXT:  (f32x4.promote_low_f16x8
+ ;; CHECK-BIN-NEXT:   (local.get $0)
+ ;; CHECK-BIN-NEXT:  )
+ ;; CHECK-BIN-NEXT: )
+ (func $f32x4.promote_low_f16x8 (param $0 v128) (result v128)
+  (f32x4.promote_low_f16x8
+   (local.get $0)
+  )
+ )
 )
 ;; CHECK-BIN-NODEBUG:      (type $0 (func (param v128 v128) (result v128)))
 
@@ -827,3 +843,9 @@
 ;; CHECK-BIN-NODEBUG-NEXT:   (local.get $0)
 ;; CHECK-BIN-NODEBUG-NEXT:  )
 ;; CHECK-BIN-NODEBUG-NEXT: )
+
+;; CHECK-BIN-NODEBUG:      (func $32 (type $1) (param $0 v128) (result v128)
+;; CHECK-BIN-NODEBUG-NEXT:  (f32x4.promote_low_f16x8
+;; CHECK-BIN-NODEBUG-NEXT:   (local.get $0)
+;; CHECK-BIN-NODEBUG-NEXT:  )
+;; CHECK-BIN-NODEBUG-NEXT: )

test/spec/f16.wast

@@ -38,6 +38,7 @@
   (func (export "i16x8.trunc_sat_f16x8_u") (param $0 v128) (result v128) (i16x8.trunc_sat_f16x8_u (local.get $0)))
   (func (export "f16x8.convert_i16x8_s") (param $0 v128) (result v128) (f16x8.convert_i16x8_s (local.get $0)))
   (func (export "f16x8.convert_i16x8_u") (param $0 v128) (result v128) (f16x8.convert_i16x8_u (local.get $0)))
+  (func (export "f32x4.promote_low_f16x8") (param $0 v128) (result v128) (f32x4.promote_low_f16x8 (local.get $0)))
   ;; Multiple operation tests:
   (func (export "splat_replace") (result v128) (f16x8.replace_lane 0 (f16x8.splat (f32.const 1)) (f32.const 99))
  )
@@ -247,3 +248,23 @@
     (v128.const i16x8 0 1      -1     -32    0 0 0 0))
     ;;                  1      inf    65504
     (v128.const i16x8 0 0x3c00 0x7c00 0x7bff 0 0 0 0))
+
+(assert_return (invoke "f32x4.promote_low_f16x8"
+    ;;                1.0    -1.0   2.0    -2.0   0 0 0 0
+    (v128.const i16x8 0x3c00 0xbc00 0x4000 0xc000 0 0 0 0))
+    ;;                1.0        -1.0       2.0        -2.0
+    (v128.const i32x4 0x3f800000 0xbf800000 0x40000000 0xc0000000))
+
+;; Edge cases: Infinities, NaNs, Zeros
+(assert_return (invoke "f32x4.promote_low_f16x8"
+    ;;                inf    -inf   nan    -0.0   0 0 0 0
+    (v128.const i16x8 0x7c00 0xfc00 0x7e00 0x8000 0 0 0 0))
+    ;;                inf        -inf       nan        -0.0
+    (v128.const i32x4 0x7f800000 0xff800000 0x7fc00000 0x80000000))
+
+;; Edge cases: Denormal
+(assert_return (invoke "f32x4.promote_low_f16x8"
+    ;;                denormal
+    (v128.const i16x8 0x0001 0      0 0 0 0 0 0))
+    ;;                2^-24
+    (v128.const i32x4 0x33800000 0      0 0))