Implement pseudo-min/max SIMD instructions (#2847)

As specified in https://github.com/WebAssembly/simd/pull/122.

14 files changed, 156 insertions, 0 deletions

diff --git a/src/binaryen-c.cpp b/src/binaryen-c.cpp
index a64c46fbf..5f4ad2c40 100644
--- a/src/binaryen-c.cpp
+++ b/src/binaryen-c.cpp
@@ -681,6 +681,8 @@ BinaryenOp BinaryenMulVecF32x4(void) { return MulVecF32x4; }
 BinaryenOp BinaryenDivVecF32x4(void) { return DivVecF32x4; }
 BinaryenOp BinaryenMinVecF32x4(void) { return MinVecF32x4; }
 BinaryenOp BinaryenMaxVecF32x4(void) { return MaxVecF32x4; }
+BinaryenOp BinaryenPMinVecF32x4(void) { return PMinVecF32x4; }
+BinaryenOp BinaryenPMaxVecF32x4(void) { return PMaxVecF32x4; }
 BinaryenOp BinaryenAbsVecF64x2(void) { return AbsVecF64x2; }
 BinaryenOp BinaryenNegVecF64x2(void) { return NegVecF64x2; }
 BinaryenOp BinaryenSqrtVecF64x2(void) { return SqrtVecF64x2; }
@@ -692,6 +694,8 @@ BinaryenOp BinaryenMulVecF64x2(void) { return MulVecF64x2; }
 BinaryenOp BinaryenDivVecF64x2(void) { return DivVecF64x2; }
 BinaryenOp BinaryenMinVecF64x2(void) { return MinVecF64x2; }
 BinaryenOp BinaryenMaxVecF64x2(void) { return MaxVecF64x2; }
+BinaryenOp BinaryenPMinVecF64x2(void) { return PMinVecF64x2; }
+BinaryenOp BinaryenPMaxVecF64x2(void) { return PMaxVecF64x2; }
 BinaryenOp BinaryenTruncSatSVecF32x4ToVecI32x4(void) {
   return TruncSatSVecF32x4ToVecI32x4;
 }
diff --git a/src/binaryen-c.h b/src/binaryen-c.h
index e460671b9..7a33a4c3c 100644
--- a/src/binaryen-c.h
+++ b/src/binaryen-c.h
@@ -541,6 +541,8 @@ BINARYEN_API BinaryenOp BinaryenMulVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenDivVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenMinVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenMaxVecF32x4(void);
+BINARYEN_API BinaryenOp BinaryenPMinVecF32x4(void);
+BINARYEN_API BinaryenOp BinaryenPMaxVecF32x4(void);
 BINARYEN_API BinaryenOp BinaryenAbsVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenNegVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenSqrtVecF64x2(void);
@@ -552,6 +554,8 @@ BINARYEN_API BinaryenOp BinaryenMulVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenDivVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenMinVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenMaxVecF64x2(void);
+BINARYEN_API BinaryenOp BinaryenPMinVecF64x2(void);
+BINARYEN_API BinaryenOp BinaryenPMaxVecF64x2(void);
 BINARYEN_API BinaryenOp BinaryenTruncSatSVecF32x4ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenTruncSatUVecF32x4ToVecI32x4(void);
 BINARYEN_API BinaryenOp BinaryenTruncSatSVecF64x2ToVecI64x2(void);
diff --git a/src/gen-s-parser.inc b/src/gen-s-parser.inc
index 09a5070b9..351a3119f 100644
--- a/src/gen-s-parser.inc
+++ b/src/gen-s-parser.inc
@@ -338,6 +338,17 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 'p': {
+                switch (op[8]) {
+                  case 'a':
+                    if (strcmp(op, "f32x4.pmax") == 0) { return makeBinary(s, BinaryOp::PMaxVecF32x4); }
+                    goto parse_error;
+                  case 'i':
+                    if (strcmp(op, "f32x4.pmin") == 0) { return makeBinary(s, BinaryOp::PMinVecF32x4); }
+                    goto parse_error;
+                  default: goto parse_error;
+                }
+              }
               case 'q': {
                 switch (op[9]) {
                   case 'a':
@@ -619,6 +630,17 @@ switch (op[0]) {
                   default: goto parse_error;
                 }
               }
+              case 'p': {
+                switch (op[8]) {
+                  case 'a':
+                    if (strcmp(op, "f64x2.pmax") == 0) { return makeBinary(s, BinaryOp::PMaxVecF64x2); }
+                    goto parse_error;
+                  case 'i':
+                    if (strcmp(op, "f64x2.pmin") == 0) { return makeBinary(s, BinaryOp::PMinVecF64x2); }
+                    goto parse_error;
+                  default: goto parse_error;
+                }
+              }
               case 'q': {
                 switch (op[9]) {
                   case 'a':
diff --git a/src/ir/cost.h b/src/ir/cost.h
index 2c503da02..8092d4ba2 100644
--- a/src/ir/cost.h
+++ b/src/ir/cost.h
@@ -687,6 +687,12 @@ struct CostAnalyzer : public Visitor<CostAnalyzer, Index> {
       case MaxVecF32x4:
         ret = 1;
         break;
+      case PMinVecF32x4:
+        ret = 1;
+        break;
+      case PMaxVecF32x4:
+        ret = 1;
+        break;
       case AddVecF64x2:
         ret = 1;
         break;
@@ -705,6 +711,12 @@ struct CostAnalyzer : public Visitor<CostAnalyzer, Index> {
       case MaxVecF64x2:
         ret = 1;
         break;
+      case PMinVecF64x2:
+        ret = 1;
+        break;
+      case PMaxVecF64x2:
+        ret = 1;
+        break;
       case NarrowSVecI16x8ToVecI8x16:
         ret = 1;
         break;
diff --git a/src/js/binaryen.js-post.js b/src/js/binaryen.js-post.js
index 1706d2b97..993f8c129 100644
--- a/src/js/binaryen.js-post.js
+++ b/src/js/binaryen.js-post.js
@@ -421,6 +421,8 @@ function initializeConstants() {
     'DivVecF32x4',
     'MinVecF32x4',
     'MaxVecF32x4',
+    'PMinVecF32x4',
+    'PMaxVecF32x4',
     'AbsVecF64x2',
     'NegVecF64x2',
     'SqrtVecF64x2',
@@ -432,6 +434,8 @@ function initializeConstants() {
     'DivVecF64x2',
     'MinVecF64x2',
     'MaxVecF64x2',
+    'PMinVecF64x2',
+    'PMaxVecF64x2',
     'TruncSatSVecF32x4ToVecI32x4',
     'TruncSatUVecF32x4ToVecI32x4',
     'TruncSatSVecF64x2ToVecI64x2',
@@ -1904,6 +1908,12 @@ function wrapModule(module, self) {
     'max': function(left, right) {
       return Module['_BinaryenBinary'](module, Module['MaxVecF32x4'], left, right);
     },
+    'pmin': function(left, right) {
+      return Module['_BinaryenBinary'](module, Module['PMinVecF32x4'], left, right);
+    },
+    'pmax': function(left, right) {
+      return Module['_BinaryenBinary'](module, Module['PMaxVecF32x4'], left, right);
+    },
     'convert_i32x4_s': function(value) {
       return Module['_BinaryenUnary'](module, Module['ConvertSVecI32x4ToVecF32x4'], value);
     },
@@ -1973,6 +1983,12 @@ function wrapModule(module, self) {
     'max': function(left, right) {
       return Module['_BinaryenBinary'](module, Module['MaxVecF64x2'], left, right);
     },
+    'pmin': function(left, right) {
+      return Module['_BinaryenBinary'](module, Module['PMinVecF64x2'], left, right);
+    },
+    'pmax': function(left, right) {
+      return Module['_BinaryenBinary'](module, Module['PMaxVecF64x2'], left, right);
+    },
     'convert_i64x2_s': function(value) {
       return Module['_BinaryenUnary'](module, Module['ConvertSVecI64x2ToVecF64x2'], value);
     },
diff --git a/src/literal.h b/src/literal.h
index b21f08995..bba2b24d8 100644
--- a/src/literal.h
+++ b/src/literal.h
@@ -271,6 +271,8 @@ public:
 
   Literal min(const Literal& other) const;
   Literal max(const Literal& other) const;
+  Literal pmin(const Literal& other) const;
+  Literal pmax(const Literal& other) const;
   Literal copysign(const Literal& other) const;
 
   std::array<Literal, 16> getLanesSI8x16() const;
@@ -424,6 +426,8 @@ public:
   Literal divF32x4(const Literal& other) const;
   Literal minF32x4(const Literal& other) const;
   Literal maxF32x4(const Literal& other) const;
+  Literal pminF32x4(const Literal& other) const;
+  Literal pmaxF32x4(const Literal& other) const;
   Literal absF64x2() const;
   Literal negF64x2() const;
   Literal sqrtF64x2() const;
@@ -433,6 +437,8 @@ public:
   Literal divF64x2(const Literal& other) const;
   Literal minF64x2(const Literal& other) const;
   Literal maxF64x2(const Literal& other) const;
+  Literal pminF64x2(const Literal& other) const;
+  Literal pmaxF64x2(const Literal& other) const;
   Literal truncSatToSI32x4() const;
   Literal truncSatToUI32x4() const;
   Literal truncSatToSI64x2() const;
diff --git a/src/passes/Print.cpp b/src/passes/Print.cpp
index 0839b3feb..3ed6edcb5 100644
--- a/src/passes/Print.cpp
+++ b/src/passes/Print.cpp
@@ -1337,6 +1337,12 @@ struct PrintExpressionContents
       case MaxVecF32x4:
         o << "f32x4.max";
         break;
+      case PMinVecF32x4:
+        o << "f32x4.pmin";
+        break;
+      case PMaxVecF32x4:
+        o << "f32x4.pmax";
+        break;
       case AddVecF64x2:
         o << "f64x2.add";
         break;
@@ -1355,6 +1361,12 @@ struct PrintExpressionContents
       case MaxVecF64x2:
         o << "f64x2.max";
         break;
+      case PMinVecF64x2:
+        o << "f64x2.pmin";
+        break;
+      case PMaxVecF64x2:
+        o << "f64x2.pmax";
+        break;
 
       case NarrowSVecI16x8ToVecI8x16:
         o << "i8x16.narrow_i16x8_s";
diff --git a/src/wasm-binary.h b/src/wasm-binary.h
index 1ee9253d0..7a2e33864 100644
--- a/src/wasm-binary.h
+++ b/src/wasm-binary.h
@@ -859,6 +859,8 @@ enum ASTNodes {
   F32x4Div = 0xe7,
   F32x4Min = 0xe8,
   F32x4Max = 0xe9,
+  F32x4PMin = 0xea,
+  F32x4PMax = 0xeb,
 
   F64x2Abs = 0xec,
   F64x2Neg = 0xed,
@@ -869,6 +871,8 @@ enum ASTNodes {
   F64x2Div = 0xf3,
   F64x2Min = 0xf4,
   F64x2Max = 0xf5,
+  F64x2PMin = 0xf6,
+  F64x2PMax = 0xf7,
 
   I32x4TruncSatSF32x4 = 0xf8,
   I32x4TruncSatUF32x4 = 0xf9,
diff --git a/src/wasm-interpreter.h b/src/wasm-interpreter.h
index 2841dc432..257a0670f 100644
--- a/src/wasm-interpreter.h
+++ b/src/wasm-interpreter.h
@@ -893,6 +893,10 @@ public:
         return left.minF32x4(right);
       case MaxVecF32x4:
         return left.maxF32x4(right);
+      case PMinVecF32x4:
+        return left.pminF32x4(right);
+      case PMaxVecF32x4:
+        return left.pmaxF32x4(right);
       case AddVecF64x2:
         return left.addF64x2(right);
       case SubVecF64x2:
@@ -905,6 +909,10 @@ public:
         return left.minF64x2(right);
       case MaxVecF64x2:
         return left.maxF64x2(right);
+      case PMinVecF64x2:
+        return left.pminF64x2(right);
+      case PMaxVecF64x2:
+        return left.pmaxF64x2(right);
 
       case NarrowSVecI16x8ToVecI8x16:
         return left.narrowSToVecI8x16(right);
diff --git a/src/wasm.h b/src/wasm.h
index b9b06a92c..81f9f4ba0 100644
--- a/src/wasm.h
+++ b/src/wasm.h
@@ -400,12 +400,16 @@ enum BinaryOp {
   DivVecF32x4,
   MinVecF32x4,
   MaxVecF32x4,
+  PMinVecF32x4,
+  PMaxVecF32x4,
   AddVecF64x2,
   SubVecF64x2,
   MulVecF64x2,
   DivVecF64x2,
   MinVecF64x2,
   MaxVecF64x2,
+  PMinVecF64x2,
+  PMaxVecF64x2,
 
   // SIMD Conversion
   NarrowSVecI16x8ToVecI8x16,
diff --git a/src/wasm/literal.cpp b/src/wasm/literal.cpp
index 33e53ca48..5ad82acaf 100644
--- a/src/wasm/literal.cpp
+++ b/src/wasm/literal.cpp
@@ -1321,6 +1321,26 @@ Literal Literal::max(const Literal& other) const {
   }
 }
 
+Literal Literal::pmin(const Literal& other) const {
+  switch (type.getSingle()) {
+    case Type::f32:
+    case Type::f64:
+      return other.lt(*this).geti32() ? other : *this;
+    default:
+      WASM_UNREACHABLE("unexpected type");
+  }
+}
+
+Literal Literal::pmax(const Literal& other) const {
+  switch (type.getSingle()) {
+    case Type::f32:
+    case Type::f64:
+      return this->lt(other).geti32() ? other : *this;
+    default:
+      WASM_UNREACHABLE("unexpected type");
+  }
+}
+
 Literal Literal::copysign(const Literal& other) const {
   // operate on bits directly, to avoid signalling bit being set on a float
   switch (type.getSingle()) {
@@ -1958,6 +1978,12 @@ Literal Literal::minF32x4(const Literal& other) const {
 Literal Literal::maxF32x4(const Literal& other) const {
   return binary<4, &Literal::getLanesF32x4, &Literal::max>(*this, other);
 }
+Literal Literal::pminF32x4(const Literal& other) const {
+  return binary<4, &Literal::getLanesF32x4, &Literal::pmin>(*this, other);
+}
+Literal Literal::pmaxF32x4(const Literal& other) const {
+  return binary<4, &Literal::getLanesF32x4, &Literal::pmax>(*this, other);
+}
 Literal Literal::addF64x2(const Literal& other) const {
   return binary<2, &Literal::getLanesF64x2, &Literal::add>(*this, other);
 }
@@ -1976,6 +2002,12 @@ Literal Literal::minF64x2(const Literal& other) const {
 Literal Literal::maxF64x2(const Literal& other) const {
   return binary<2, &Literal::getLanesF64x2, &Literal::max>(*this, other);
 }
+Literal Literal::pminF64x2(const Literal& other) const {
+  return binary<2, &Literal::getLanesF64x2, &Literal::pmin>(*this, other);
+}
+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();
diff --git a/src/wasm/wasm-binary.cpp b/src/wasm/wasm-binary.cpp
index b3314134d..3181288f8 100644
--- a/src/wasm/wasm-binary.cpp
+++ b/src/wasm/wasm-binary.cpp
@@ -4009,6 +4009,14 @@ bool WasmBinaryBuilder::maybeVisitSIMDBinary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Binary>();
       curr->op = MaxVecF32x4;
       break;
+    case BinaryConsts::F32x4PMin:
+      curr = allocator.alloc<Binary>();
+      curr->op = PMinVecF32x4;
+      break;
+    case BinaryConsts::F32x4PMax:
+      curr = allocator.alloc<Binary>();
+      curr->op = PMaxVecF32x4;
+      break;
     case BinaryConsts::F64x2Add:
       curr = allocator.alloc<Binary>();
       curr->op = AddVecF64x2;
@@ -4033,6 +4041,14 @@ bool WasmBinaryBuilder::maybeVisitSIMDBinary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Binary>();
       curr->op = MaxVecF64x2;
       break;
+    case BinaryConsts::F64x2PMin:
+      curr = allocator.alloc<Binary>();
+      curr->op = PMinVecF64x2;
+      break;
+    case BinaryConsts::F64x2PMax:
+      curr = allocator.alloc<Binary>();
+      curr->op = PMaxVecF64x2;
+      break;
     case BinaryConsts::I8x16NarrowSI16x8:
       curr = allocator.alloc<Binary>();
       curr->op = NarrowSVecI16x8ToVecI8x16;
diff --git a/src/wasm/wasm-stack.cpp b/src/wasm/wasm-stack.cpp
index 48fe79fa1..a9415612d 100644
--- a/src/wasm/wasm-stack.cpp
+++ b/src/wasm/wasm-stack.cpp
@@ -1561,6 +1561,12 @@ void BinaryInstWriter::visitBinary(Binary* curr) {
     case MaxVecF32x4:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4Max);
       break;
+    case PMinVecF32x4:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4PMin);
+      break;
+    case PMaxVecF32x4:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F32x4PMax);
+      break;
     case AddVecF64x2:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2Add);
       break;
@@ -1579,6 +1585,12 @@ void BinaryInstWriter::visitBinary(Binary* curr) {
     case MaxVecF64x2:
       o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2Max);
       break;
+    case PMinVecF64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2PMin);
+      break;
+    case PMaxVecF64x2:
+      o << int8_t(BinaryConsts::SIMDPrefix) << U32LEB(BinaryConsts::F64x2PMax);
+      break;
 
     case NarrowSVecI16x8ToVecI8x16:
       o << int8_t(BinaryConsts::SIMDPrefix)
diff --git a/src/wasm/wasm-validator.cpp b/src/wasm/wasm-validator.cpp
index 1aee9a15d..f5e134dba 100644
--- a/src/wasm/wasm-validator.cpp
+++ b/src/wasm/wasm-validator.cpp
@@ -1455,12 +1455,16 @@ void FunctionValidator::visitBinary(Binary* curr) {
     case DivVecF32x4:
     case MinVecF32x4:
     case MaxVecF32x4:
+    case PMinVecF32x4:
+    case PMaxVecF32x4:
     case AddVecF64x2:
     case SubVecF64x2:
     case MulVecF64x2:
     case DivVecF64x2:
     case MinVecF64x2:
     case MaxVecF64x2:
+    case PMinVecF64x2:
+    case PMaxVecF64x2:
     case NarrowSVecI16x8ToVecI8x16:
     case NarrowUVecI16x8ToVecI8x16:
     case NarrowSVecI32x4ToVecI16x8: