Add new min/max/dot SIMD instructions

package.json

@@ -21,7 +21,7 @@
     "url": "https://github.com/AssemblyScript/assemblyscript/issues"
   },
   "dependencies": {
-    "binaryen": "89.0.0-nightly.20191012",
+    "binaryen": "89.0.0-nightly.20191113",
     "long": "^4.0.0",
     "source-map-support": "^0.5.16",
     "ts-node": "^6.2.0",

src/builtins.ts

@@ -323,6 +323,7 @@ export namespace BuiltinSymbols {
   export const v128_all_true = "~lib/builtins/v128.all_true";
   export const v128_min = "~lib/builtins/v128.min";
   export const v128_max = "~lib/builtins/v128.max";
+  export const v128_dot = "~lib/builtins/v128.dot";
   export const v128_abs = "~lib/builtins/v128.abs";
   export const v128_sqrt = "~lib/builtins/v128.sqrt";
   export const v128_eq = "~lib/builtins/v128.eq";
@@ -353,6 +354,10 @@ export namespace BuiltinSymbols {
   export const i8x16_add = "~lib/builtins/i8x16.add";
   export const i8x16_sub = "~lib/builtins/i8x16.sub";
   export const i8x16_mul = "~lib/builtins/i8x16.mul";
+  export const i8x16_min_s = "~lib/builtins/i8x16.min_s";
+  export const i8x16_min_u = "~lib/builtins/i8x16.min_u";
+  export const i8x16_max_s = "~lib/builtins/i8x16.max_s";
+  export const i8x16_max_u = "~lib/builtins/i8x16.max_u";
   export const i8x16_neg = "~lib/builtins/i8x16.neg";
   export const i8x16_add_saturate_s = "~lib/builtins/i8x16.add_saturate_s";
   export const i8x16_add_saturate_u = "~lib/builtins/i8x16.add_saturate_u";
@@ -383,6 +388,10 @@ export namespace BuiltinSymbols {
   export const i16x8_add = "~lib/builtins/i16x8.add";
   export const i16x8_sub = "~lib/builtins/i16x8.sub";
   export const i16x8_mul = "~lib/builtins/i16x8.mul";
+  export const i16x8_min_s = "~lib/builtins/i16x8.min_s";
+  export const i16x8_min_u = "~lib/builtins/i16x8.min_u";
+  export const i16x8_max_s = "~lib/builtins/i16x8.max_s";
+  export const i16x8_max_u = "~lib/builtins/i16x8.max_u";
   export const i16x8_neg = "~lib/builtins/i16x8.neg";
   export const i16x8_add_saturate_s = "~lib/builtins/i16x8.add_saturate_s";
   export const i16x8_add_saturate_u = "~lib/builtins/i16x8.add_saturate_u";
@@ -418,6 +427,11 @@ export namespace BuiltinSymbols {
   export const i32x4_add = "~lib/builtins/i32x4.add";
   export const i32x4_sub = "~lib/builtins/i32x4.sub";
   export const i32x4_mul = "~lib/builtins/i32x4.mul";
+  export const i32x4_min_s = "~lib/builtins/i32x4.min_s";
+  export const i32x4_min_u = "~lib/builtins/i32x4.min_u";
+  export const i32x4_max_s = "~lib/builtins/i32x4.max_s";
+  export const i32x4_max_u = "~lib/builtins/i32x4.max_u";
+  export const i32x4_dot_i16x8_s = "~lib/builtins/i32x4.dot_i16x8_s";
   export const i32x4_neg = "~lib/builtins/i32x4.neg";
   export const i32x4_shl = "~lib/builtins/i32x4.shl";
   export const i32x4_shr_s = "~lib/builtins/i32x4.shr_s";
@@ -3332,6 +3346,20 @@ export function compileCall(
       let arg1 = compiler.compileExpression(operands[1], Type.v128, Constraints.CONV_IMPLICIT);
       if (!type.is(TypeFlags.REFERENCE)) {
         switch (type.kind) {
+          case TypeKind.I8: return module.binary(BinaryOp.MinI8x16, arg0, arg1);
+          case TypeKind.U8: return module.binary(BinaryOp.MinU8x16, arg0, arg1);
+          case TypeKind.I16: return module.binary(BinaryOp.MinI16x8, arg0, arg1);
+          case TypeKind.U16: return module.binary(BinaryOp.MinU16x8, arg0, arg1);
+          case TypeKind.ISIZE: {
+            if (compiler.options.isWasm64) break;
+            // fall-through
+          }
+          case TypeKind.I32: return module.binary(BinaryOp.MinI32x4, arg0, arg1);
+          case TypeKind.USIZE: {
+            if (compiler.options.isWasm64) break;
+            // fall-through
+          }
+          case TypeKind.U32: return module.binary(BinaryOp.MinU32x4, arg0, arg1);
           case TypeKind.F32: return module.binary(BinaryOp.MinF32x4, arg0, arg1);
           case TypeKind.F64: return module.binary(BinaryOp.MinF64x2, arg0, arg1);
         }
@@ -3356,6 +3384,20 @@ export function compileCall(
       let arg1 = compiler.compileExpression(operands[1], Type.v128, Constraints.CONV_IMPLICIT);
       if (!type.is(TypeFlags.REFERENCE)) {
         switch (type.kind) {
+          case TypeKind.I8: return module.binary(BinaryOp.MaxI8x16, arg0, arg1);
+          case TypeKind.U8: return module.binary(BinaryOp.MaxU8x16, arg0, arg1);
+          case TypeKind.I16: return module.binary(BinaryOp.MaxI16x8, arg0, arg1);
+          case TypeKind.U16: return module.binary(BinaryOp.MaxU16x8, arg0, arg1);
+          case TypeKind.ISIZE: {
+            if (compiler.options.isWasm64) break;
+            // fall-through
+          }
+          case TypeKind.I32: return module.binary(BinaryOp.MaxI32x4, arg0, arg1);
+          case TypeKind.USIZE: {
+            if (compiler.options.isWasm64) break;
+            // fall-through
+          }
+          case TypeKind.U32: return module.binary(BinaryOp.MaxU32x4, arg0, arg1);
           case TypeKind.F32: return module.binary(BinaryOp.MaxF32x4, arg0, arg1);
           case TypeKind.F64: return module.binary(BinaryOp.MaxF64x2, arg0, arg1);
         }
@@ -3366,6 +3408,29 @@ export function compileCall(
       );
       return module.unreachable();
     }
+    case BuiltinSymbols.v128_dot: { // dot<T!>(a: v128, b: v128) -> v128
+      if (
+        checkFeatureEnabled(Feature.SIMD, reportNode, compiler) |
+        checkTypeRequired(typeArguments, reportNode, compiler) |
+        checkArgsRequired(operands, 2, reportNode, compiler)
+      ) {
+        compiler.currentType = Type.v128;
+        return module.unreachable();
+      }
+      let type = typeArguments![0];
+      let arg0 = compiler.compileExpression(operands[0], Type.v128, Constraints.CONV_IMPLICIT);
+      let arg1 = compiler.compileExpression(operands[1], Type.v128, Constraints.CONV_IMPLICIT);
+      if (!type.is(TypeFlags.REFERENCE)) {
+        switch (type.kind) {
+          case TypeKind.I16: return module.binary(BinaryOp.DotI16x8, arg0, arg1);
+        }
+      }
+      compiler.error(
+        DiagnosticCode.Operation_0_cannot_be_applied_to_type_1,
+        reportNode.typeArgumentsRange, "v128.dot", type.toString()
+      );
+      return module.unreachable();
+    }
     case BuiltinSymbols.v128_eq: { // eq<T!>(a: v128, b: v128) -> v128
       if (
         checkFeatureEnabled(Feature.SIMD, reportNode, compiler) |
@@ -4501,6 +4566,10 @@ function tryDeferASM(
     case BuiltinSymbols.i8x16_add: return deferASM(BuiltinSymbols.v128_add, compiler, Type.i8, operands, Type.v128, reportNode);
     case BuiltinSymbols.i8x16_sub: return deferASM(BuiltinSymbols.v128_sub, compiler, Type.i8, operands, Type.v128, reportNode);
     case BuiltinSymbols.i8x16_mul: return deferASM(BuiltinSymbols.v128_mul, compiler, Type.i8, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i8x16_min_s: return deferASM(BuiltinSymbols.v128_min, compiler, Type.i8, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i8x16_min_u: return deferASM(BuiltinSymbols.v128_min, compiler, Type.u8, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i8x16_max_s: return deferASM(BuiltinSymbols.v128_max, compiler, Type.i8, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i8x16_max_u: return deferASM(BuiltinSymbols.v128_max, compiler, Type.u8, operands, Type.v128, reportNode);
     case BuiltinSymbols.i8x16_neg: return deferASM(BuiltinSymbols.v128_neg, compiler, Type.i8, operands, Type.v128, reportNode);
     case BuiltinSymbols.i8x16_add_saturate_s: return deferASM(BuiltinSymbols.v128_add_saturate, compiler, Type.i8, operands, Type.v128, reportNode);
     case BuiltinSymbols.i8x16_add_saturate_u: return deferASM(BuiltinSymbols.v128_add_saturate, compiler, Type.u8, operands, Type.v128, reportNode);
@@ -4531,6 +4600,10 @@ function tryDeferASM(
     case BuiltinSymbols.i16x8_add: return deferASM(BuiltinSymbols.v128_add, compiler, Type.i16, operands, Type.v128, reportNode);
     case BuiltinSymbols.i16x8_sub: return deferASM(BuiltinSymbols.v128_sub, compiler, Type.i16, operands, Type.v128, reportNode);
     case BuiltinSymbols.i16x8_mul: return deferASM(BuiltinSymbols.v128_mul, compiler, Type.i16, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i16x8_min_s: return deferASM(BuiltinSymbols.v128_min, compiler, Type.i16, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i16x8_min_u: return deferASM(BuiltinSymbols.v128_min, compiler, Type.u16, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i16x8_max_s: return deferASM(BuiltinSymbols.v128_max, compiler, Type.i16, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i16x8_max_u: return deferASM(BuiltinSymbols.v128_max, compiler, Type.u16, operands, Type.v128, reportNode);
     case BuiltinSymbols.i16x8_neg: return deferASM(BuiltinSymbols.v128_neg, compiler, Type.i16, operands, Type.v128, reportNode);
     case BuiltinSymbols.i16x8_add_saturate_s: return deferASM(BuiltinSymbols.v128_add_saturate, compiler, Type.i16, operands, Type.v128, reportNode);
     case BuiltinSymbols.i16x8_add_saturate_u: return deferASM(BuiltinSymbols.v128_add_saturate, compiler, Type.u16, operands, Type.v128, reportNode);
@@ -4566,6 +4639,11 @@ function tryDeferASM(
     case BuiltinSymbols.i32x4_add: return deferASM(BuiltinSymbols.v128_add, compiler, Type.i32, operands, Type.v128, reportNode);
     case BuiltinSymbols.i32x4_sub: return deferASM(BuiltinSymbols.v128_sub, compiler, Type.i32, operands, Type.v128, reportNode);
     case BuiltinSymbols.i32x4_mul: return deferASM(BuiltinSymbols.v128_mul, compiler, Type.i32, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i32x4_min_s: return deferASM(BuiltinSymbols.v128_min, compiler, Type.i32, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i32x4_min_u: return deferASM(BuiltinSymbols.v128_min, compiler, Type.u32, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i32x4_max_s: return deferASM(BuiltinSymbols.v128_max, compiler, Type.i32, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i32x4_max_u: return deferASM(BuiltinSymbols.v128_max, compiler, Type.u32, operands, Type.v128, reportNode);
+    case BuiltinSymbols.i32x4_dot_i16x8_s: return deferASM(BuiltinSymbols.v128_dot, compiler, Type.i16, operands, Type.v128, reportNode);
     case BuiltinSymbols.i32x4_neg: return deferASM(BuiltinSymbols.v128_neg, compiler, Type.i32, operands, Type.v128, reportNode);
     case BuiltinSymbols.i32x4_shl: return deferASM(BuiltinSymbols.v128_shl, compiler, Type.i32, operands, Type.v128, reportNode);
     case BuiltinSymbols.i32x4_shr_s: return deferASM(BuiltinSymbols.v128_shr, compiler, Type.i32, operands, Type.v128, reportNode);

src/glue/binaryen.d.ts

@@ -331,6 +331,10 @@ declare function _BinaryenSubVecI8x16(): BinaryenOp;
 declare function _BinaryenSubSatSVecI8x16(): BinaryenOp;
 declare function _BinaryenSubSatUVecI8x16(): BinaryenOp;
 declare function _BinaryenMulVecI8x16(): BinaryenOp;
+declare function _BinaryenMinSVecI8x16(): BinaryenOp;
+declare function _BinaryenMinUVecI8x16(): BinaryenOp;
+declare function _BinaryenMaxSVecI8x16(): BinaryenOp;
+declare function _BinaryenMaxUVecI8x16(): BinaryenOp;
 declare function _BinaryenNegVecI16x8(): BinaryenOp;
 declare function _BinaryenAnyTrueVecI16x8(): BinaryenOp;
 declare function _BinaryenAllTrueVecI16x8(): BinaryenOp;
@@ -344,6 +348,10 @@ declare function _BinaryenSubVecI16x8(): BinaryenOp;
 declare function _BinaryenSubSatSVecI16x8(): BinaryenOp;
 declare function _BinaryenSubSatUVecI16x8(): BinaryenOp;
 declare function _BinaryenMulVecI16x8(): BinaryenOp;
+declare function _BinaryenMinSVecI16x8(): BinaryenOp;
+declare function _BinaryenMinUVecI16x8(): BinaryenOp;
+declare function _BinaryenMaxSVecI16x8(): BinaryenOp;
+declare function _BinaryenMaxUVecI16x8(): BinaryenOp;
 declare function _BinaryenNegVecI32x4(): BinaryenOp;
 declare function _BinaryenAnyTrueVecI32x4(): BinaryenOp;
 declare function _BinaryenAllTrueVecI32x4(): BinaryenOp;
@@ -353,6 +361,11 @@ declare function _BinaryenShrUVecI32x4(): BinaryenOp;
 declare function _BinaryenAddVecI32x4(): BinaryenOp;
 declare function _BinaryenSubVecI32x4(): BinaryenOp;
 declare function _BinaryenMulVecI32x4(): BinaryenOp;
+declare function _BinaryenMinSVecI32x4(): BinaryenOp;
+declare function _BinaryenMinUVecI32x4(): BinaryenOp;
+declare function _BinaryenMaxSVecI32x4(): BinaryenOp;
+declare function _BinaryenMaxUVecI32x4(): BinaryenOp;
+declare function _BinaryenDotSVecI16x8ToVecI32x4(): BinaryenOp;
 declare function _BinaryenNegVecI64x2(): BinaryenOp;
 declare function _BinaryenAnyTrueVecI64x2(): BinaryenOp;
 declare function _BinaryenAllTrueVecI64x2(): BinaryenOp;
@@ -650,6 +663,8 @@ declare type BinaryenFunctionRef = usize;
 declare function _BinaryenAddFunction(module: BinaryenModuleRef, name: usize, type: BinaryenFunctionTypeRef, varTypes: usize, numVarTypes: BinaryenIndex, body: BinaryenExpressionRef): BinaryenFunctionRef;
 declare function _BinaryenGetFunction(module: BinaryenModuleRef, name: usize): BinaryenFunctionRef;
 declare function _BinaryenRemoveFunction(module: BinaryenModuleRef, name: usize): void;
+declare function _BinaryenGetNumFunctions(module: BinaryenModuleRef): BinaryenIndex;
+declare function _BinaryenGetFunctionByIndex(module: BinaryenModuleRef, index: BinaryenIndex): BinaryenFunctionRef;
 
 declare function _BinaryenFunctionGetName(func: BinaryenFunctionRef): usize;
 declare function _BinaryenFunctionGetType(func: BinaryenFunctionRef): BinaryenFunctionTypeRef;
@@ -679,6 +694,8 @@ declare function _BinaryenAddMemoryExport(module: BinaryenModuleRef, internalNam
 declare function _BinaryenAddGlobalExport(module: BinaryenModuleRef, internalName: usize, externalName: usize): BinaryenExportRef;
 declare function _BinaryenAddEventExport(module: BinaryenModuleRef, internalName: usize, externalName: usize): BinaryenExportRef;
 declare function _BinaryenRemoveExport(module: BinaryenModuleRef, externalName: usize): void;
+declare function _BinaryenGetNumExports(module: BinaryenModuleRef): BinaryenIndex;
+declare function _BinaryenGetExportByIndex(module: BinaryenModuleRef, index: BinaryenIndex): BinaryenExportRef;
 
 declare type BinaryenGlobalRef = usize;
 
@@ -700,6 +717,10 @@ declare function _BinaryenEventGetParam(event: BinaryenEventRef, index: Binaryen
 declare function _BinaryenSetFunctionTable(module: BinaryenModuleRef, initial: BinaryenIndex, maximum: BinaryenIndex, funcs: usize, numFuncs: BinaryenIndex, offset: BinaryenExpressionRef): void;
 
 declare function _BinaryenSetMemory(module: BinaryenModuleRef, initial: BinaryenIndex, maximum: BinaryenIndex, exportName: usize, segments: usize, segmentPassive: usize, segmentOffsets: usize, segmentSizes: usize, numSegments: BinaryenIndex, shared: bool): void;
+declare function _BinaryenGetNumMemorySegments(module: BinaryenModuleRef): BinaryenIndex;
+declare function _BinaryenGetMemorySegmentByteOffset(module: BinaryenModuleRef, index: BinaryenIndex): u32;
+declare function _BinaryenGetMemorySegmentByteLength(module: BinaryenModuleRef, id: BinaryenIndex): usize;
+declare function _BinaryenCopyMemorySegmentData(module: BinaryenModuleRef, id: BinaryenIndex, buffer: usize): void;
 
 declare function _BinaryenSetStart(module: BinaryenModuleRef, start: BinaryenFunctionRef): void;
 

src/module.ts

@@ -333,16 +333,29 @@ export enum BinaryOp {
   SubSatI8x16 = _BinaryenSubSatSVecI8x16(),
   SubSatU8x16 = _BinaryenSubSatUVecI8x16(),
   MulI8x16 = _BinaryenMulVecI8x16(),
+  MinI8x16 = _BinaryenMinSVecI8x16(),
+  MinU8x16 = _BinaryenMinUVecI8x16(),
+  MaxI8x16 = _BinaryenMaxSVecI8x16(),
+  MaxU8x16 = _BinaryenMaxUVecI8x16(),
   AddI16x8 = _BinaryenAddVecI16x8(),
   AddSatI16x8 = _BinaryenAddSatSVecI16x8(),
   AddSatU16x8 = _BinaryenAddSatUVecI16x8(),
   SubI16x8 = _BinaryenSubVecI16x8(),
   SubSatI16x8 = _BinaryenSubSatSVecI16x8(),
   SubSatU16x8 = _BinaryenSubSatUVecI16x8(),
   MulI16x8 = _BinaryenMulVecI16x8(),
+  MinI16x8 = _BinaryenMinSVecI16x8(),
+  MinU16x8 = _BinaryenMinUVecI16x8(),
+  MaxI16x8 = _BinaryenMaxSVecI16x8(),
+  MaxU16x8 = _BinaryenMaxUVecI16x8(),
   AddI32x4 = _BinaryenAddVecI32x4(),
   SubI32x4 = _BinaryenSubVecI32x4(),
   MulI32x4 = _BinaryenMulVecI32x4(),
+  MinI32x4 = _BinaryenMinSVecI32x4(),
+  MinU32x4 = _BinaryenMinUVecI32x4(),
+  MaxI32x4 = _BinaryenMaxSVecI32x4(),
+  MaxU32x4 = _BinaryenMaxUVecI32x4(),
+  DotI16x8 = _BinaryenDotSVecI16x8ToVecI32x4(),
   AddI64x2 = _BinaryenAddVecI64x2(),
   SubI64x2 = _BinaryenSubVecI64x2(),
   AddF32x4 = _BinaryenAddVecF32x4(),

std/assembly/builtins.ts

@@ -1091,11 +1091,15 @@ export namespace v128 {
 
   // @ts-ignore: decorator
   @builtin
-  export declare function min<T>(a: v128, b: v128): v128; // f32, f64 only
+  export declare function min<T>(a: v128, b: v128): v128;
 
   // @ts-ignore: decorator
   @builtin
-  export declare function max<T>(a: v128, b: v128): v128; // f32, f64 only
+  export declare function max<T>(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function dot<T>(a: v128, b: v128): v128; // i16 only
 
   // @ts-ignore: decorator
   @builtin
@@ -1195,6 +1199,22 @@ export namespace i8x16 {
   @builtin
   export declare function mul(a: v128, b: v128): v128;
 
+  // @ts-ignore: decorator
+  @builtin
+  export declare function min_s(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function min_u(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function max_s(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function max_u(a: v128, b: v128): v128;
+
   // @ts-ignore: decorator
   @builtin
   export declare function neg(a: v128): v128;
@@ -1318,6 +1338,22 @@ export namespace i16x8 {
   @builtin
   export declare function mul(a: v128, b: v128): v128;
 
+  // @ts-ignore: decorator
+  @builtin
+  export declare function min_s(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function min_u(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function max_s(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function max_u(a: v128, b: v128): v128;
+
   // @ts-ignore: decorator
   @builtin
   export declare function neg(a: v128): v128;
@@ -1461,6 +1497,26 @@ export namespace i32x4 {
   @builtin
   export declare function mul(a: v128, b: v128): v128;
 
+  // @ts-ignore: decorator
+  @builtin
+  export declare function min_s(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function min_u(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function max_s(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function max_u(a: v128, b: v128): v128;
+
+  // @ts-ignore: decorator
+  @builtin
+  export declare function dot_i16x8_s(a: v128, b: v128): v128;
+
   // @ts-ignore: decorator
   @builtin
   export declare function neg(a: v128): v128;