SIMD saturating sub implementation. (#747)

WebAssembly · Feb 7, 2018 · a74072b · a74072b
1 parent f0b0677
commit a74072b
Show file tree

Hide file tree

Showing 7 changed files with 1,596 additions and 1,377 deletions.
diff --git a/src/binary-reader.cc b/src/binary-reader.cc
@@ -838,6 +838,10 @@ Result BinaryReader::ReadFunctionBody(Offset end_offset) {
       case Opcode::I8X16AddSaturateU:
       case Opcode::I16X8AddSaturateS:
       case Opcode::I16X8AddSaturateU:
+      case Opcode::I8X16SubSaturateS:
+      case Opcode::I8X16SubSaturateU:
+      case Opcode::I16X8SubSaturateS:
+      case Opcode::I16X8SubSaturateU:
         ERROR_UNLESS_OPCODE_ENABLED(opcode);
         CALLBACK(OnBinaryExpr, opcode);
         CALLBACK0(OnOpcodeBare);

diff --git a/src/interp.cc b/src/interp.cc
@@ -927,6 +927,21 @@ ValueTypeRep<T> AddSaturate(ValueTypeRep<T> lhs_rep, ValueTypeRep<T> rhs_rep) {
   }
 }
 
+template <typename T, typename R>
+ValueTypeRep<T> SubSaturate(ValueTypeRep<T> lhs_rep, ValueTypeRep<T> rhs_rep) {
+  T max = std::numeric_limits<R>::max();
+  T min = std::numeric_limits<R>::min(); 
+  T result = static_cast<T>(lhs_rep) - static_cast<T>(rhs_rep);
+
+  if(result < min) {
+    return ToRep(min);
+  } else if (result > max) {
+    return ToRep(max);
+  } else {
+    return ToRep(result);
+  }
+}
+
 // {i,f}{32,64}.sub
 template <typename T>
 ValueTypeRep<T> Sub(ValueTypeRep<T> lhs_rep, ValueTypeRep<T> rhs_rep) {
@@ -2424,6 +2439,22 @@ Result Thread::Run(int num_instructions) {
       case Opcode::I16X8AddSaturateU:
         CHECK_TRAP(SimdBinop<v128, uint16_t>(AddSaturate<uint32_t, uint16_t>));
         break;
+
+      case Opcode::I8X16SubSaturateS:
+        CHECK_TRAP(SimdBinop<v128, int8_t>(SubSaturate<int32_t, int8_t>));
+        break;
+
+      case Opcode::I8X16SubSaturateU:
+        CHECK_TRAP(SimdBinop<v128, uint8_t>(SubSaturate<int32_t, uint8_t>));
+        break;
+
+      case Opcode::I16X8SubSaturateS:
+        CHECK_TRAP(SimdBinop<v128, int16_t>(SubSaturate<int32_t, int16_t>));
+        break;
+
+      case Opcode::I16X8SubSaturateU:
+        CHECK_TRAP(SimdBinop<v128, uint16_t>(SubSaturate<int32_t, uint16_t>));
+        break;
       // The following opcodes are either never generated or should never be
       // executed.
       case Opcode::Block:
@@ -2914,7 +2945,11 @@ void Thread::Trace(Stream* stream) {
     case Opcode::I8X16AddSaturateS: 
     case Opcode::I8X16AddSaturateU: 
     case Opcode::I16X8AddSaturateS: 
-    case Opcode::I16X8AddSaturateU: { 
+    case Opcode::I16X8AddSaturateU: 
+    case Opcode::I8X16SubSaturateS: 
+    case Opcode::I8X16SubSaturateU: 
+    case Opcode::I16X8SubSaturateS: 
+    case Opcode::I16X8SubSaturateU: { 
       stream->Writef("%s $0x%08x %08x %08x %08x  $0x%08x %08x %08x %08x\n", opcode.GetName(), Pick(2).v128_bits.v[0],
                        Pick(2).v128_bits.v[1], Pick(2).v128_bits.v[2], Pick(2).v128_bits.v[3],Pick(1).v128_bits.v[0],
                        Pick(1).v128_bits.v[1], Pick(1).v128_bits.v[2], Pick(1).v128_bits.v[3]);
@@ -3233,6 +3268,10 @@ void Environment::Disassemble(Stream* stream,
       case Opcode::I8X16AddSaturateU:
       case Opcode::I16X8AddSaturateS:
       case Opcode::I16X8AddSaturateU:
+      case Opcode::I8X16SubSaturateS:
+      case Opcode::I8X16SubSaturateU:
+      case Opcode::I16X8SubSaturateS:
+      case Opcode::I16X8SubSaturateU:
         stream->Writef("%s %%[-2], %%[-1]\n", opcode.GetName());
         break;
 

diff --git a/src/opcode.cc b/src/opcode.cc
@@ -208,6 +208,10 @@ bool Opcode::IsEnabled(const Features& features) const {
     case Opcode::I8X16AddSaturateU:
     case Opcode::I16X8AddSaturateS:
     case Opcode::I16X8AddSaturateU:
+    case Opcode::I8X16SubSaturateS:
+    case Opcode::I8X16SubSaturateU:
+    case Opcode::I16X8SubSaturateS:
+    case Opcode::I16X8SubSaturateU:
       return features.simd_enabled();
 
     // Interpreter opcodes are never "enabled".

diff --git a/src/opcode.def b/src/opcode.def
@@ -262,6 +262,10 @@ WABT_OPCODE(V128, V128, V128, ___, 0, 0xfd, 0x27,  I8X16AddSaturateS, "i8x16.add
 WABT_OPCODE(V128, V128, V128, ___, 0, 0xfd, 0x28,  I8X16AddSaturateU, "i8x16.add_saturate_u")
 WABT_OPCODE(V128, V128, V128, ___, 0, 0xfd, 0x29,  I16X8AddSaturateS, "i16x8.add_saturate_s")
 WABT_OPCODE(V128, V128, V128, ___, 0, 0xfd, 0x2a,  I16X8AddSaturateU, "i16x8.add_saturate_u")
+WABT_OPCODE(V128, V128, V128, ___, 0, 0xfd, 0x2b,  I8X16SubSaturateS, "i8x16.sub_saturate_s")
+WABT_OPCODE(V128, V128, V128, ___, 0, 0xfd, 0x2c,  I8X16SubSaturateU, "i8x16.sub_saturate_u")
+WABT_OPCODE(V128, V128, V128, ___, 0, 0xfd, 0x2d,  I16X8SubSaturateS, "i16x8.sub_saturate_s")
+WABT_OPCODE(V128, V128, V128, ___, 0, 0xfd, 0x2e,  I16X8SubSaturateU, "i16x8.sub_saturate_u")
 
 
 WABT_OPCODE(I32, I32, I32, ___, 4, 0xfe,  0x00, AtomicWake, "atomic.wake")