OPCODE_DOT_PRODUCT_4 constant propagation

src/xenia/cpu/compiler/passes/constant_propagation_pass.cc

@@ -657,6 +657,14 @@ bool ConstantPropagationPass::Run(HIRBuilder* builder) {
           }
           break;
 
+        case OPCODE_DOT_PRODUCT_4:
+          if (i->src1.value->IsConstant() && i->src2.value->IsConstant()) {
+            v->set_from(i->src1.value);
+            v->DotProduct4(i->src2.value);
+            i->Remove();
+          }
+          break;
+
         default:
           // Ignored.
           break;

src/xenia/cpu/hir/value.cc

@@ -1206,6 +1206,25 @@ void Value::DotProduct3(Value* other) {
   }
 }
 
+void Value::DotProduct4(Value* other) {
+  assert_true(this->type == VEC128_TYPE && other->type == VEC128_TYPE);
+  switch (type) {
+    case VEC128_TYPE:
+      alignas(16) float result[4];
+      __m128 src1 = _mm_load_ps(constant.v128.f32);
+      __m128 src2 = _mm_load_ps(other->constant.v128.f32);
+      __m128 dest = _mm_dp_ps(src1, src2, 0b11110001);
+      _mm_store_ps(result, dest);
+      // TODO(rick): is this sane?
+      type = FLOAT32_TYPE;
+      constant.f32 = result[0];
+      break;
+    default:
+      assert_unhandled_case(type);
+      break;
+  }
+}
+
 void Value::ByteSwap() {
   switch (type) {
     case INT8_TYPE:

src/xenia/cpu/hir/value.h

@@ -508,6 +508,7 @@ class Value {
   void VectorAdd(Value* other, TypeName type, bool is_unsigned, bool saturate);
   void VectorSub(Value* other, TypeName type, bool is_unsigned, bool saturate);
   void DotProduct3(Value* other);
+  void DotProduct4(Value* other);
   void ByteSwap();
   void CountLeadingZeros(const Value* other);
   bool Compare(Opcode opcode, Value* other);