Implement Relaxed SIMD proposal (#1994)

This adds support for the new opcodes from the Relaxed SIMD proposal
(https://github.com/WebAssembly/relaxed-simd) behind the
"--enable-relaxed-simd" flag.

The exception is the f32x4.relaxed_dot_bf16x8_add_f32x4 instruction
which is not yet implemented.

8 files changed, 92 insertions, 8 deletions

diff --git a/include/wabt/feature.def b/include/wabt/feature.def
index 26b1db79..00a4e7f2 100644
--- a/include/wabt/feature.def
+++ b/include/wabt/feature.def
@@ -39,3 +39,4 @@ WABT_FEATURE(gc,                  "gc",                      false,   "Garbage c
 WABT_FEATURE(memory64,            "memory64",                false,   "64-bit memory")
 WABT_FEATURE(multi_memory,        "multi-memory",            false,   "Multi-memory")
 WABT_FEATURE(extended_const,      "extended-const",          false,   "Extended constant expressions")
+WABT_FEATURE(relaxed_simd,        "relaxed-simd",            false,   "Relaxed SIMD")
diff --git a/include/wabt/interp/interp.h b/include/wabt/interp/interp.h
index dc405a8e..6af605ac 100644
--- a/include/wabt/interp/interp.h
+++ b/include/wabt/interp/interp.h
@@ -1217,6 +1217,12 @@ class Thread {
   template <typename S, typename T>
   RunResult DoSimdDot();
   template <typename S, typename T>
+  RunResult DoSimdDotAdd();
+  template <typename S>
+  RunResult DoSimdRelaxedMadd();
+  template <typename S>
+  RunResult DoSimdRelaxedNmadd();
+  template <typename S, typename T>
   RunResult DoSimdLoadExtend(Instr, Trap::Ptr* out_trap);
   template <typename S, typename T>
   RunResult DoSimdExtaddPairwise();
diff --git a/include/wabt/ir.h b/include/wabt/ir.h
index bf52a602..68846bc0 100644
--- a/include/wabt/ir.h
+++ b/include/wabt/ir.h
@@ -202,6 +202,54 @@ struct Const {
 };
 using ConstVector = std::vector<Const>;
 
+enum class ExpectationType {
+  Values,
+  Either,
+};
+
+class Expectation {
+ public:
+  Expectation() = delete;
+  virtual ~Expectation() = default;
+  ExpectationType type() const { return type_; }
+
+  Location loc;
+
+  ConstVector expected;
+
+ protected:
+  explicit Expectation(ExpectationType type, const Location& loc = Location())
+      : loc(loc), type_(type) {}
+
+ private:
+  ExpectationType type_;
+};
+
+template <ExpectationType TypeEnum>
+class ExpectationMixin : public Expectation {
+ public:
+  static bool classof(const Expectation* expectation) {
+    return expectation->type() == TypeEnum;
+  }
+
+  explicit ExpectationMixin(const Location& loc = Location())
+      : Expectation(TypeEnum, loc) {}
+};
+
+class ValueExpectation : public ExpectationMixin<ExpectationType::Values> {
+ public:
+  explicit ValueExpectation(const Location& loc = Location())
+      : ExpectationMixin<ExpectationType::Values>(loc) {}
+};
+
+struct EitherExpectation : public ExpectationMixin<ExpectationType::Either> {
+ public:
+  explicit EitherExpectation(const Location& loc = Location())
+      : ExpectationMixin<ExpectationType::Either>(loc) {}
+};
+
+typedef std::unique_ptr<Expectation> ExpectationPtr;
+
 struct FuncSignature {
   TypeVector param_types;
   TypeVector result_types;
@@ -1375,7 +1423,7 @@ class RegisterCommand : public CommandMixin<CommandType::Register> {
 class AssertReturnCommand : public CommandMixin<CommandType::AssertReturn> {
  public:
   ActionPtr action;
-  ConstVector expected;
+  ExpectationPtr expected;
 };
 
 template <CommandType TypeEnum>
diff --git a/include/wabt/opcode-code-table.h b/include/wabt/opcode-code-table.h
index 3adb478d..13c5838e 100644
--- a/include/wabt/opcode-code-table.h
+++ b/include/wabt/opcode-code-table.h
@@ -24,7 +24,12 @@
 extern "C" {
 #endif
 
-#define WABT_OPCODE_CODE_TABLE_SIZE 65536
+#define WABT_OPCODE_CODE_TABLE_SIZE 131072
+
+/*
+ * Number of bits required to store an opcode
+ */
+#define MAX_OPCODE_BITS 9
 
 /* This structure is defined in C because C++ doesn't (yet) allow you to use
  * designated array initializers, i.e. [10] = {foo}.
diff --git a/include/wabt/opcode.def b/include/wabt/opcode.def
index 9005418e..e10f9cba 100644
--- a/include/wabt/opcode.def
+++ b/include/wabt/opcode.def
@@ -501,6 +501,28 @@ WABT_OPCODE(V128, V128,  ___, ___,  0,  0xfd, 0xfd, I32X4TruncSatF64X2UZero, "i3
 WABT_OPCODE(V128, V128,  ___, ___,  0,  0xfd, 0xfe, F64X2ConvertLowI32X4S, "f64x2.convert_low_i32x4_s", "")
 WABT_OPCODE(V128, V128,  ___, ___,  0,  0xfd, 0xff, F64X2ConvertLowI32X4U, "f64x2.convert_low_i32x4_u", "")
 
+/* Relaxed-SIMD opcodes */
+WABT_OPCODE(V128, V128, V128, ___,  0,  0xfd, 0x100, I8X16RelaxedSwizzle, "i8x16.relaxed_swizzle", "")
+WABT_OPCODE(V128, V128, ___,  ___,  0,  0xfd, 0x101, I32X4RelaxedTruncF32X4S, "i32x4.relaxed_trunc_f32x4_s", "")
+WABT_OPCODE(V128, V128, ___,  ___,  0,  0xfd, 0x102, I32X4RelaxedTruncF32X4U, "i32x4.relaxed_trunc_f32x4_u", "")
+WABT_OPCODE(V128, V128, ___,  ___,  0,  0xfd, 0x103, I32X4RelaxedTruncF64X2SZero, "i32x4.relaxed_trunc_f64x2_s_zero", "")
+WABT_OPCODE(V128, V128, ___,  ___,  0,  0xfd, 0x104, I32X4RelaxedTruncF64X2UZero, "i32x4.relaxed_trunc_f64x2_u_zero", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x105, F32X4RelaxedMadd, "f32x4.relaxed_madd", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x106, F32X4RelaxedNmadd, "f32x4.relaxed_nmadd", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x107, F64X2RelaxedMadd, "f64x2.relaxed_madd", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x108, F64X2RelaxedNmadd, "f64x2.relaxed_nmadd", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x109, I8X16RelaxedLaneSelect, "i8x16.relaxed_laneselect", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x10a, I16X8RelaxedLaneSelect, "i16x8.relaxed_laneselect", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x10b, I32X4RelaxedLaneSelect, "i32x4.relaxed_laneselect", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x10c, I64X2RelaxedLaneSelect, "i64x2.relaxed_laneselect", "")
+WABT_OPCODE(V128, V128, V128, ___,  0,  0xfd, 0x10d, F32X4RelaxedMin, "f32x4.relaxed_min", "")
+WABT_OPCODE(V128, V128, V128, ___,  0,  0xfd, 0x10e, F32X4RelaxedMax, "f32x4.relaxed_max", "")
+WABT_OPCODE(V128, V128, V128, ___,  0,  0xfd, 0x10f, F64X2RelaxedMin, "f64x2.relaxed_min", "")
+WABT_OPCODE(V128, V128, V128, ___,  0,  0xfd, 0x110, F64X2RelaxedMax, "f64x2.relaxed_max", "")
+WABT_OPCODE(V128, V128, V128, ___,  0,  0xfd, 0x111, I16X8RelaxedQ15mulrS, "i16x8.relaxed_q15mulr_s", "")
+WABT_OPCODE(V128, V128, V128, ___,  0,  0xfd, 0x112, I16X8DotI8X16I7X16S, "i16x8.dot_i8x16_i7x16_s", "")
+WABT_OPCODE(V128, V128, V128, V128, 0,  0xfd, 0x113, I32X4DotI8X16I7X16AddS, "i32x4.dot_i8x16_i7x16_add_s", "")
+
 /* Thread opcodes (--enable-threads) */
 WABT_OPCODE(I32,  I32,  I32,  ___,  4,  0xfe, 0x00, MemoryAtomicNotify, "memory.atomic.notify", "")
 WABT_OPCODE(I32,  I32,  I32,  I64,  4,  0xfe, 0x01, MemoryAtomicWait32, "memory.atomic.wait32", "")
diff --git a/include/wabt/opcode.h b/include/wabt/opcode.h
index 47e27f9b..a7c7b249 100644
--- a/include/wabt/opcode.h
+++ b/include/wabt/opcode.h
@@ -107,12 +107,11 @@ struct Opcode {
   };
 
   static uint32_t PrefixCode(uint8_t prefix, uint32_t code) {
-    // For now, 8 bits is enough for all codes.
-    if (code >= 0x100) {
-      // Clamp to 0xff, since we know that it is an invalid code.
-      code = 0xff;
+    if (code >= (1 << MAX_OPCODE_BITS)) {
+      // Clamp to (2^bits - 1), since we know that it is an invalid code.
+      code = (1 << MAX_OPCODE_BITS) - 1;
     }
-    return (prefix << 8) | code;
+    return (prefix << MAX_OPCODE_BITS) | code;
   }
 
   // The Opcode struct only stores an enumeration (Opcode::Enum) of all valid
@@ -137,7 +136,7 @@ struct Opcode {
                                   uint8_t* out_prefix,
                                   uint32_t* out_code) {
     uint32_t prefix_code = ~static_cast<uint32_t>(e) + 1;
-    *out_prefix = prefix_code >> 8;
+    *out_prefix = prefix_code >> MAX_OPCODE_BITS;
     *out_code = prefix_code & 0xff;
   }
 
diff --git a/include/wabt/token.def b/include/wabt/token.def
index b8869abd..1e95da62 100644
--- a/include/wabt/token.def
+++ b/include/wabt/token.def
@@ -34,6 +34,7 @@ WABT_TOKEN(Data, "data")
 WABT_TOKEN(Declare, "declare")
 WABT_TOKEN(Delegate, "delegate")
 WABT_TOKEN(Do, "do")
+WABT_TOKEN(Either, "either")
 WABT_TOKEN(Elem, "elem")
 WABT_TOKEN(Eof, "EOF")
 WABT_TOKEN(Tag, "tag")
diff --git a/include/wabt/wast-parser.h b/include/wabt/wast-parser.h
index 23194a7e..2aeb14bf 100644
--- a/include/wabt/wast-parser.h
+++ b/include/wabt/wast-parser.h
@@ -186,6 +186,8 @@ class WastParser {
   Result ParseF32(Const*, ConstType type);
   Result ParseF64(Const*, ConstType type);
   Result ParseConst(Const*, ConstType type);
+  Result ParseExpectedValues(ExpectationPtr*);
+  Result ParseEither(ConstVector*);
   Result ParseExternref(Const*);
   Result ParseExpectedNan(ExpectedNan* expected);
   Result ParseConstList(ConstVector*, ConstType type);