[D3D12] Move interpolators above position because it's more important for pixel shaders

src/xenia/gpu/d3d12/shaders/primitive_rectangle_list.gs.hlsl

@@ -1,6 +1,6 @@
 struct XeVertex {
-  float4 position : SV_Position;
   float4 interpolators[16] : TEXCOORD;
+  float4 position : SV_Position;
   float point_size : PSIZE;
 };
 
@@ -8,16 +8,16 @@ struct XeVertex {
 void main(triangle XeVertex xe_in[3], inout TriangleStream<XeVertex> xe_stream) {
   XeVertex xe_out;
 
-  xe_out.position = xe_in[0].position;
   xe_out.interpolators = xe_in[0].interpolators;
+  xe_out.position = xe_in[0].position;
   xe_out.point_size = xe_in[0].point_size;
   xe_stream.Append(xe_out);
-  xe_out.position = xe_in[1].position;
   xe_out.interpolators = xe_in[1].interpolators;
+  xe_out.position = xe_in[1].position;
   xe_out.point_size = xe_in[1].point_size;
   xe_stream.Append(xe_out);
-  xe_out.position = xe_in[2].position;
   xe_out.interpolators = xe_in[2].interpolators;
+  xe_out.position = xe_in[2].position;
   xe_out.point_size = xe_in[2].point_size;
   xe_stream.Append(xe_out);
   xe_stream.RestartStrip();
@@ -38,46 +38,46 @@ void main(triangle XeVertex xe_in[3], inout TriangleStream<XeVertex> xe_stream)
     //  |   //   |   2:   1,-1
     //  | -      |   3: [ 1, 1 ]
     //  1 ------[3]
-    xe_out.position = xe_in[2].position;
     xe_out.interpolators = xe_in[2].interpolators;
+    xe_out.position = xe_in[2].position;
     xe_out.point_size = xe_in[2].point_size;
     xe_stream.Append(xe_out);
-    xe_out.position = xe_in[1].position;
     xe_out.interpolators = xe_in[1].interpolators;
+    xe_out.position = xe_in[1].position;
     xe_out.point_size = xe_in[1].point_size;
     xe_stream.Append(xe_out);
-    xe_out.position = float4(xe_in[1].position.xy -
-                             xe_in[0].position.xy +
-                             xe_in[2].position.xy,
-                             xe_in[2].position.zw);
     [unroll] for (int i = 0; i < 16; ++i) {
       xe_out.interpolators[i] = xe_in[1].interpolators[i] -
                                 xe_in[0].interpolators[i] +
                                 xe_in[2].interpolators[i];
     }
+    xe_out.position = float4(xe_in[1].position.xy -
+                             xe_in[0].position.xy +
+                             xe_in[2].position.xy,
+                             xe_in[2].position.zw);
   } else {
     //  0 ------ 1   0:  -1,-1
     //  | -      |   1:   1,-1
     //  |   \\   |   2:   1, 1
     //  |      - |   3: [-1, 1 ]
     // [3] ----- 2
-    xe_out.position = xe_in[0].position;
     xe_out.interpolators = xe_in[0].interpolators;
+    xe_out.position = xe_in[0].position;
     xe_out.point_size = xe_in[0].point_size;
     xe_stream.Append(xe_out);
-    xe_out.position = xe_in[2].position;
     xe_out.interpolators = xe_in[2].interpolators;
+    xe_out.position = xe_in[2].position;
     xe_out.point_size = xe_in[2].point_size;
     xe_stream.Append(xe_out);
-    xe_out.position = float4(xe_in[0].position.xy -
-                             xe_in[1].position.xy +
-                             xe_in[2].position.xy,
-                             xe_in[2].position.zw);
     [unroll] for (int i = 0; i < 16; ++i) {
       xe_out.interpolators[i] = xe_in[0].interpolators[i] -
                                 xe_in[1].interpolators[i] +
                                 xe_in[2].interpolators[i];
     }
+    xe_out.position = float4(xe_in[0].position.xy -
+                             xe_in[1].position.xy +
+                             xe_in[2].position.xy,
+                             xe_in[2].position.zw);
   }
   xe_out.point_size = xe_in[2].point_size;
   xe_stream.Append(xe_out);

src/xenia/gpu/hlsl_shader_translator.cc

@@ -258,8 +258,8 @@ std::vector<uint8_t> HlslShaderTranslator::CompleteTranslation() {
         "XE_BYTE_SWAP_OVERLOAD(uint4)\n"
         "\n"
         "struct XeVertexShaderOutput {\n"
-        "  float4 position : SV_Position;\n"
         "  float4 interpolators[%u] : TEXCOORD;\n"
+        "  float4 position : SV_Position;\n"
         "  float point_size : PSIZE;\n"
         "};\n"
         "\n"
@@ -283,8 +283,8 @@ std::vector<uint8_t> HlslShaderTranslator::CompleteTranslation() {
     // XE_PIXEL_SHADER_WRITES_DEPTH in the beginning of the final output.
     source.AppendFormat(
         "struct XePixelShaderInput {\n"
-        "  float4 position : SV_Position;\n"
         "  float4 interpolators[%u] : TEXCOORD;\n"
+        "  float4 position : SV_Position;\n"
         "};\n"
         "\n"
         "struct XePixelShaderOutput {\n"