Implement v128 instructions (#508)

src/common.ts

@@ -118,6 +118,16 @@ export namespace CommonSymbols {
   export const f32 = "f32";
   export const f64 = "f64";
   export const v128 = "v128";
+  export const i8x16 = "i8x16";
+  export const u8x16 = "u8x16";
+  export const i16x8 = "i16x8";
+  export const u16x8 = "u16x8";
+  export const i32x4 = "i32x4";
+  export const u32x4 = "u32x4";
+  export const i64x2 = "i64x2";
+  export const u64x2 = "u64x2";
+  export const f32x4 = "f32x4";
+  export const f64x2 = "f64x2";
   export const void_ = "void";
   export const number = "number";
   export const boolean = "boolean";
@@ -146,6 +156,7 @@ export namespace LibrarySymbols {
   export const ASC_FEATURE_SIGN_EXTENSION = "ASC_FEATURE_SIGN_EXTENSION";
   export const ASC_FEATURE_BULK_MEMORY = "ASC_FEATURE_BULK_MEMORY";
   export const ASC_FEATURE_SIMD = "ASC_FEATURE_SIMD";
+  export const ASC_FEATURE_THREADS = "ASC_FEATURE_THREADS";
   // classes
   export const I8 = "I8";
   export const I16 = "I16";

src/compiler.ts

@@ -3199,6 +3199,12 @@ export class Compiler extends DiagnosticEmitter {
             expr = module.createBinary(BinaryOp.EqF64, leftExpr, rightExpr);
             break;
           }
+          case TypeKind.V128: {
+            expr = module.createUnary(UnaryOp.AllTrueVecI8x16,
+              module.createBinary(BinaryOp.EqVecI8x16, leftExpr, rightExpr)
+            );
+            break;
+          }
           default: {
             assert(false);
             expr = module.createUnreachable();
@@ -3287,6 +3293,12 @@ export class Compiler extends DiagnosticEmitter {
             expr = module.createBinary(BinaryOp.NeF64, leftExpr, rightExpr);
             break;
           }
+          case TypeKind.V128: {
+            expr = module.createUnary(UnaryOp.AnyTrueVecI8x16,
+              module.createBinary(BinaryOp.NeVecI8x16, leftExpr, rightExpr)
+            );
+            break;
+          }
           default: {
             assert(false);
             expr = module.createUnreachable();

src/diagnosticMessages.generated.ts

@@ -32,6 +32,8 @@ export enum DiagnosticCode {
   Optional_properties_are_not_supported = 219,
   Expression_must_be_a_compile_time_constant = 220,
   Module_cannot_have_multiple_start_functions = 221,
+  _0_must_be_a_value_between_1_and_2_inclusive = 222,
+  _0_must_be_a_power_of_two = 223,
   Unterminated_string_literal = 1002,
   Identifier_expected = 1003,
   _0_expected = 1005,
@@ -165,6 +167,8 @@ export function diagnosticCodeToString(code: DiagnosticCode): string {
     case 219: return "Optional properties are not supported.";
     case 220: return "Expression must be a compile-time constant.";
     case 221: return "Module cannot have multiple start functions.";
+    case 222: return "'{0}' must be a value between '{1}' and '{2}' inclusive.";
+    case 223: return "'{0}' must be a power of two.";
     case 1002: return "Unterminated string literal.";
     case 1003: return "Identifier expected.";
     case 1005: return "'{0}' expected.";

src/diagnosticMessages.json

@@ -24,6 +24,8 @@
   "Optional properties are not supported.": 219,
   "Expression must be a compile-time constant.": 220,
   "Module cannot have multiple start functions.": 221,
+  "'{0}' must be a value between '{1}' and '{2}' inclusive.": 222,
+  "'{0}' must be a power of two.": 223,
 
   "Unterminated string literal.": 1002,
   "Identifier expected.": 1003,

src/flow.ts

@@ -232,6 +232,7 @@ export class Flow {
       case NativeType.I64: { temps = parentFunction.tempI64s; break; }
       case NativeType.F32: { temps = parentFunction.tempF32s; break; }
       case NativeType.F64: { temps = parentFunction.tempF64s; break; }
+      case NativeType.V128: { temps = parentFunction.tempV128s; break; }
       default: throw new Error("concrete type expected");
     }
     var local: Local;
@@ -270,6 +271,10 @@ export class Flow {
         temps = parentFunction.tempF64s || (parentFunction.tempF64s = []);
         break;
       }
+      case NativeType.V128: {
+        temps = parentFunction.tempV128s || (parentFunction.tempV128s = []);
+        break;
+      }
       default: throw new Error("concrete type expected");
     }
     assert(local.index >= 0);
@@ -297,6 +302,10 @@ export class Flow {
         temps = parentFunction.tempF64s || (parentFunction.tempF64s = []);
         break;
       }
+      case NativeType.V128: {
+        temps = parentFunction.tempV128s || (parentFunction.tempV128s = []);
+        break;
+      }
       default: throw new Error("concrete type expected");
     }
     var local: Local;

src/module.ts

@@ -122,7 +122,7 @@ export enum UnaryOp {
   ExtendI16ToI32 = _BinaryenExtendS16Int32(),
   ExtendI8ToI64 = _BinaryenExtendS8Int64(),
   ExtendI16ToI64 = _BinaryenExtendS16Int64(),
-  ExtendI32ToI64 = _BinaryenExtendS32Int64()
+  ExtendI32ToI64 = _BinaryenExtendS32Int64(),
 
   // see: https://github.com/WebAssembly/nontrapping-float-to-int-conversions
   // TruncF32ToI32Sat
@@ -133,6 +133,41 @@ export enum UnaryOp {
   // TruncF32ToU64Sat
   // TruncF64ToI64Sat
   // TruncF64ToU64Sat
+
+  // see: https://github.com/WebAssembly/simd
+  SplatVecI8x16 = _BinaryenSplatVecI8x16(),
+  SplatVecI16x8 = _BinaryenSplatVecI16x8(),
+  SplatVecI32x4 = _BinaryenSplatVecI32x4(),
+  SplatVecI64x2 = _BinaryenSplatVecI64x2(),
+  SplatVecF32x4 = _BinaryenSplatVecF32x4(),
+  SplatVecF64x2 = _BinaryenSplatVecF64x2(),
+  NotVec128 = _BinaryenNotVec128(),
+  NegVecI8x16 = _BinaryenNegVecI8x16(),
+  AnyTrueVecI8x16 = _BinaryenAnyTrueVecI8x16(),
+  AllTrueVecI8x16 = _BinaryenAllTrueVecI8x16(),
+  NegVecI16x8 = _BinaryenNegVecI16x8(),
+  AnyTrueVecI16x8 = _BinaryenAnyTrueVecI16x8(),
+  AllTrueVecI16x8 = _BinaryenAllTrueVecI16x8(),
+  NegVecI32x4 = _BinaryenNegVecI32x4(),
+  AnyTrueVecI32x4 = _BinaryenAnyTrueVecI32x4(),
+  AllTrueVecI32x4 = _BinaryenAllTrueVecI32x4(),
+  NegVecI64x2 = _BinaryenNegVecI64x2(),
+  AnyTrueVecI64x2 = _BinaryenAnyTrueVecI64x2(),
+  AllTrueVecI64x2 = _BinaryenAllTrueVecI64x2(),
+  AbsVecF32x4 = _BinaryenAbsVecF32x4(),
+  NegVecF32x4 = _BinaryenNegVecF32x4(),
+  SqrtVecF32x4 = _BinaryenSqrtVecF32x4(),
+  AbsVecF64x2 = _BinaryenAbsVecF64x2(),
+  NegVecF64x2 = _BinaryenNegVecF64x2(),
+  SqrtVecF64x2 = _BinaryenSqrtVecF64x2(),
+  TruncSatSVecF32x4ToVecI32x4 = _BinaryenTruncSatSVecF32x4ToVecI32x4(),
+  TruncSatUVecF32x4ToVecI32x4 = _BinaryenTruncSatUVecF32x4ToVecI32x4(),
+  TruncSatSVecF64x2ToVecI64x2 = _BinaryenTruncSatSVecF64x2ToVecI64x2(),
+  TruncSatUVecF64x2ToVecI64x2 = _BinaryenTruncSatUVecF64x2ToVecI64x2(),
+  ConvertSVecI32x4ToVecF32x4 = _BinaryenConvertSVecI32x4ToVecF32x4(),
+  ConvertUVecI32x4ToVecF32x4 = _BinaryenConvertUVecI32x4ToVecF32x4(),
+  ConvertSVecI64x2ToVecF64x2 = _BinaryenConvertSVecI64x2ToVecF64x2(),
+  ConvertUVecI64x2ToVecF64x2 = _BinaryenConvertUVecI64x2ToVecF64x2()
 }
 
 export enum BinaryOp {
@@ -211,61 +246,9 @@ export enum BinaryOp {
   LtF64 = _BinaryenLtFloat64(),
   LeF64 = _BinaryenLeFloat64(),
   GtF64 = _BinaryenGtFloat64(),
-  GeF64 = _BinaryenGeFloat64()
-}
+  GeF64 = _BinaryenGeFloat64(),
 
-export enum HostOp {
-  CurrentMemory = _BinaryenCurrentMemory(),
-  GrowMemory = _BinaryenGrowMemory(),
-
-  // see: https://github.com/WebAssembly/bulk-memory-operations
-  // MoveMemory
-  // SetMemory
-}
-
-export enum AtomicRMWOp {
-  Add = _BinaryenAtomicRMWAdd(),
-  Sub = _BinaryenAtomicRMWSub(),
-  And = _BinaryenAtomicRMWAnd(),
-  Or = _BinaryenAtomicRMWOr(),
-  Xor = _BinaryenAtomicRMWXor(),
-  Xchg = _BinaryenAtomicRMWXchg()
-}
-
-export enum SIMDOp {
-  SplatVecI8x16 = _BinaryenSplatVecI8x16(),
-  SplatVecI16x8 = _BinaryenSplatVecI16x8(),
-  SplatVecI32x4 = _BinaryenSplatVecI32x4(),
-  SplatVecI64x2 = _BinaryenSplatVecI64x2(),
-  SplatVecF32x4 = _BinaryenSplatVecF32x4(),
-  SplatVecF64x2 = _BinaryenSplatVecF64x2(),
-  NotVec128 = _BinaryenNotVec128(),
-  NegVecI8x16 = _BinaryenNegVecI8x16(),
-  AnyTrueVecI8x16 = _BinaryenAnyTrueVecI8x16(),
-  AllTrueVecI8x16 = _BinaryenAllTrueVecI8x16(),
-  NegVecI16x8 = _BinaryenNegVecI16x8(),
-  AnyTrueVecI16x8 = _BinaryenAnyTrueVecI16x8(),
-  AllTrueVecI16x8 = _BinaryenAllTrueVecI16x8(),
-  NegVecI32x4 = _BinaryenNegVecI32x4(),
-  AnyTrueVecI32x4 = _BinaryenAnyTrueVecI32x4(),
-  AllTrueVecI32x4 = _BinaryenAllTrueVecI32x4(),
-  NegVecI64x2 = _BinaryenNegVecI64x2(),
-  AnyTrueVecI64x2 = _BinaryenAnyTrueVecI64x2(),
-  AllTrueVecI64x2 = _BinaryenAllTrueVecI64x2(),
-  AbsVecF32x4 = _BinaryenAbsVecF32x4(),
-  NegVecF32x4 = _BinaryenNegVecF32x4(),
-  SqrtVecF32x4 = _BinaryenSqrtVecF32x4(),
-  AbsVecF64x2 = _BinaryenAbsVecF64x2(),
-  NegVecF64x2 = _BinaryenNegVecF64x2(),
-  SqrtVecF64x2 = _BinaryenSqrtVecF64x2(),
-  TruncSatSVecF32x4ToVecI32x4 = _BinaryenTruncSatSVecF32x4ToVecI32x4(),
-  TruncSatUVecF32x4ToVecI32x4 = _BinaryenTruncSatUVecF32x4ToVecI32x4(),
-  TruncSatSVecF64x2ToVecI64x2 = _BinaryenTruncSatSVecF64x2ToVecI64x2(),
-  TruncSatUVecF64x2ToVecI64x2 = _BinaryenTruncSatUVecF64x2ToVecI64x2(),
-  ConvertSVecI32x4ToVecF32x4 = _BinaryenConvertSVecI32x4ToVecF32x4(),
-  ConvertUVecI32x4ToVecF32x4 = _BinaryenConvertUVecI32x4ToVecF32x4(),
-  ConvertSVecI64x2ToVecF64x2 = _BinaryenConvertSVecI64x2ToVecF64x2(),
-  ConvertUVecI64x2ToVecF64x2 = _BinaryenConvertUVecI64x2ToVecF64x2(),
+  // see: https://github.com/WebAssembly/simd
   EqVecI8x16 = _BinaryenEqVecI8x16(),
   NeVecI8x16 = _BinaryenNeVecI8x16(),
   LtSVecI8x16 = _BinaryenLtSVecI8x16(),
@@ -344,6 +327,59 @@ export enum SIMDOp {
   MaxVecF64x2 = _BinaryenMaxVecF64x2()
 }
 
+export enum HostOp {
+  CurrentMemory = _BinaryenCurrentMemory(),
+  GrowMemory = _BinaryenGrowMemory(),
+
+  // see: https://github.com/WebAssembly/bulk-memory-operations
+  // MoveMemory
+  // SetMemory
+}
+
+export enum AtomicRMWOp {
+  Add = _BinaryenAtomicRMWAdd(),
+  Sub = _BinaryenAtomicRMWSub(),
+  And = _BinaryenAtomicRMWAnd(),
+  Or = _BinaryenAtomicRMWOr(),
+  Xor = _BinaryenAtomicRMWXor(),
+  Xchg = _BinaryenAtomicRMWXchg()
+}
+
+export enum SIMDExtractOp {
+  ExtractLaneSVecI8x16 = _BinaryenExtractLaneSVecI8x16(),
+  ExtractLaneUVecI8x16 = _BinaryenExtractLaneUVecI8x16(),
+  ExtractLaneSVecI16x8 = _BinaryenExtractLaneSVecI16x8(),
+  ExtractLaneUVecI16x8 = _BinaryenExtractLaneUVecI16x8(),
+  ExtractLaneVecI32x4 = _BinaryenExtractLaneVecI32x4(),
+  ExtractLaneVecI64x2 = _BinaryenExtractLaneVecI64x2(),
+  ExtractLaneVecF32x4 = _BinaryenExtractLaneVecF32x4(),
+  ExtractLaneVecF64x2 = _BinaryenExtractLaneVecF64x2(),
+}
+
+export enum SIMDReplaceOp {
+  ReplaceLaneVecI8x16 = _BinaryenReplaceLaneVecI8x16(),
+  ReplaceLaneVecI16x8 = _BinaryenReplaceLaneVecI16x8(),
+  ReplaceLaneVecI32x4 = _BinaryenReplaceLaneVecI32x4(),
+  ReplaceLaneVecI64x2 = _BinaryenReplaceLaneVecI64x2(),
+  ReplaceLaneVecF32x4 = _BinaryenReplaceLaneVecF32x4(),
+  ReplaceLaneVecF64x2 = _BinaryenReplaceLaneVecF64x2()
+}
+
+export enum SIMDShiftOp { // FIXME: seems to be missing in binaryen-c.h
+  ShlVecI8x16,
+  ShrSVecI8x16,
+  ShrUVecI8x16,
+  ShlVecI16x8,
+  ShrSVecI16x8,
+  ShrUVecI16x8,
+  ShlVecI32x4,
+  ShrSVecI32x4,
+  ShrUVecI32x4,
+  ShlVecI64x2,
+  ShrSVecI64x2,
+  ShrUVecI64x2
+}
+
 export class MemorySegment {
 
   buffer: Uint8Array;
@@ -510,9 +546,10 @@ export class Module {
     signed: bool,
     ptr: ExpressionRef,
     type: NativeType,
-    offset: Index = 0
+    offset: Index = 0,
+    align: Index = bytes // naturally aligned by default
   ): ExpressionRef {
-    return _BinaryenLoad(this.ref, bytes, signed ? 1 : 0, offset, /* always aligned */ bytes, type, ptr);
+    return _BinaryenLoad(this.ref, bytes, signed ? 1 : 0, offset, align, type, ptr);
   }
 
   createStore(
@@ -520,9 +557,10 @@ export class Module {
     ptr: ExpressionRef,
     value: ExpressionRef,
     type: NativeType,
-    offset: Index = 0
+    offset: Index = 0,
+    align: Index = bytes // naturally aligned by default
   ): ExpressionRef {
-    return _BinaryenStore(this.ref, bytes, offset, /* always aligned */ bytes, ptr, value, type);
+    return _BinaryenStore(this.ref, bytes, offset, align, ptr, value, type);
   }
 
   createAtomicLoad(
@@ -732,6 +770,55 @@ export class Module {
     return _BinaryenMemoryFill(this.ref, dest, value, size);
   }
 
+  // simd
+
+  createSIMDExtract(
+    op: SIMDExtractOp,
+    vec: ExpressionRef,
+    idx: u8
+  ): ExpressionRef {
+    return _BinaryenSIMDExtract(this.ref, op, vec, idx);
+  }
+
+  createSIMDReplace(
+    op: SIMDReplaceOp,
+    vec: ExpressionRef,
+    idx: u8,
+    value: ExpressionRef
+  ): ExpressionRef {
+    return _BinaryenSIMDReplace(this.ref, op, vec, idx, value);
+  }
+
+  createSIMDShuffle(
+    vec1: ExpressionRef,
+    vec2: ExpressionRef,
+    mask: Uint8Array
+  ): ExpressionRef {
+    assert(mask.length == 16);
+    var cArr = allocU8Array(mask);
+    try {
+      return _BinaryenSIMDShuffle(this.ref, vec1, vec2, cArr);
+    } finally {
+      memory.free(cArr);
+    }
+  }
+
+  createSIMDBitselect(
+    vec1: ExpressionRef,
+    vec2: ExpressionRef,
+    cond: ExpressionRef
+  ): ExpressionRef {
+    return _BinaryenSIMDBitselect(this.ref, vec1, vec2, cond);
+  }
+
+  createSIMDShift(
+    op: SIMDShiftOp,
+    vec: ExpressionRef,
+    shift: ExpressionRef
+  ): ExpressionRef {
+    return _BinaryenSIMDShift(this.ref, op, vec, shift);
+  }
+
   // meta
 
   addGlobal(

src/program.ts

@@ -523,6 +523,8 @@ export class Program extends DiagnosticEmitter {
       i64_new(options.hasFeature(Feature.BULK_MEMORY) ? 1 : 0, 0));
     this.registerConstantInteger(LibrarySymbols.ASC_FEATURE_SIMD, Type.bool,
       i64_new(options.hasFeature(Feature.SIMD) ? 1 : 0, 0));
+    this.registerConstantInteger(LibrarySymbols.ASC_FEATURE_THREADS, Type.bool,
+      i64_new(options.hasFeature(Feature.THREADS) ? 1 : 0, 0));
 
     // remember deferred elements
     var queuedImports = new Array<QueuedImport>();
@@ -2564,6 +2566,7 @@ export class Function extends TypedElement {
   tempI64s: Local[] | null = null;
   tempF32s: Local[] | null = null;
   tempF64s: Local[] | null = null;
+  tempV128s: Local[] | null = null;
 
   // used by flows to keep track of break labels
   nextBreakId: i32 = 0;

src/resolver.ts

@@ -74,7 +74,8 @@ import {
 } from "./common";
 
 import {
-  makeMap
+  makeMap,
+  isPowerOf2
 } from "./util";
 
 import {
@@ -1572,13 +1573,9 @@ export class Resolver extends DiagnosticEmitter {
             }
             if (!fieldType) break; // did report above
             let fieldInstance = new Field(<FieldPrototype>member, instance, fieldType);
-            switch (fieldType.byteSize) { // align
-              case 1: break;
-              case 2: { if (memoryOffset & 1) ++memoryOffset; break; }
-              case 4: { if (memoryOffset & 3) memoryOffset = (memoryOffset | 3) + 1; break; }
-              case 8: { if (memoryOffset & 7) memoryOffset = (memoryOffset | 7) + 1; break; }
-              default: assert(false);
-            }
+            assert(isPowerOf2(fieldType.byteSize));
+            let mask = fieldType.byteSize - 1;
+            if (memoryOffset & mask) memoryOffset = (memoryOffset | mask) + 1;
             fieldInstance.memoryOffset = memoryOffset;
             memoryOffset += fieldType.byteSize;
             instance.add(member.name, fieldInstance); // reports

src/types.ts

@@ -55,6 +55,8 @@ export const enum TypeKind {
   F64,
 
   // vectors
+
+  /** A 128-bit vector. */
   V128,
 
   // other

src/util/index.ts

@@ -10,3 +10,8 @@ export * from "./collections";
 export * from "./path";
 export * from "./text";
 export * from "./binary";
+
+/** Tests if `x` is a power of two. */
+export function isPowerOf2(x: i32): bool {
+  return x != 0 && (x & (x - 1)) == 0;
+}