Detect write to psize and include conditionally.

src/xenia/gpu/d3d11/d3d11_geometry_shader.cc

@@ -136,9 +136,11 @@ int D3D11GeometryShader::Generate(D3D11VertexShader* vertex_shader,
       "  float4 o[%d] : XE_O;\n",
       D3D11Shader::MAX_INTERPOLATORS);
   }
+  if (alloc_counts.point_size) {
+    output->Append(
+      "  float4 oPointSize : PSIZE;\n");
+  }
   output->Append(
-    // TODO(benvanik): only pull in point size if required.
-    "  float4 oPointSize : PSIZE;\n"
     "};\n");
 
   output->Append(
@@ -163,6 +165,8 @@ int D3D11PointSpriteGeometryShader::Generate(D3D11VertexShader* vertex_shader,
     return 1;
   }
 
+  auto alloc_counts = vertex_shader->alloc_counts();
+
   // TODO(benvanik): fetch default point size from register and use that if
   //     the VS doesn't write oPointSize.
   // TODO(benvanik): clamp to min/max.
@@ -176,8 +180,19 @@ int D3D11PointSpriteGeometryShader::Generate(D3D11VertexShader* vertex_shader,
     "   float2( 1.0,  1.0),\n"
     "   float2(-1.0, -1.0),\n"
     "   float2( 1.0, -1.0),\n"
-    "  };\n"
+    "  };\n");
-    "  float psize = max(input[0].oPointSize.x, 1.0);\n"
+  if (alloc_counts.point_size) {
+    // Point size specified in input.
+    // TODO(benvanik): pull in psize min/max.
+    output->Append(
+      "  float psize = max(input[0].oPointSize.x, 1.0);\n");
+  } else {
+    // Point size from register.
+    // TODO(benvanik): pull in psize.
+    output->Append(
+      "  float psize = 1.0;\n");
+  }
+  output->Append(
     "  for (uint n = 0; n < 4; n++) {\n"
     "    VERTEX v = input[0];\n"
     "    v.oPos.xy += offsets[n] * psize;\n"
@@ -204,6 +219,8 @@ int D3D11RectListGeometryShader::Generate(D3D11VertexShader* vertex_shader,
     return 1;
   }
 
+  auto alloc_counts = vertex_shader->alloc_counts();
+
   output->Append(
     "[maxvertexcount(4)]\n"
     "void main(triangle VERTEX input[3], inout TriangleStream<VERTEX> output) {\n"
@@ -212,10 +229,11 @@ int D3D11RectListGeometryShader::Generate(D3D11VertexShader* vertex_shader,
     "    output.Append(v);\n"
     "  }\n"
     "  VERTEX v = input[2];\n"
-    "  v.oPos += input[1].oPos - input[0].oPos;\n"
+    "  v.oPos += input[1].oPos - input[0].oPos;\n");
-    // TODO(benvanik): only if needed?
+  if (alloc_counts.point_size) {
+    output->Append(
       "  v.oPointSize += input[1].oPointSize - input[0].oPointSize;\n");
-  auto alloc_counts = vertex_shader->alloc_counts();
+  }
   for (uint32_t n = 0; n < alloc_counts.params; n++) {
     // TODO(benvanik): this may be wrong - the count is a bad metric.
     output->Append(

src/xenia/gpu/d3d11/d3d11_shader.cc

@@ -449,8 +449,10 @@ const char* D3D11VertexShader::Translate(xe_gpu_program_cntl_t* program_cntl) {
       "  float4 o[%d] : XE_O;\n",
       MAX_INTERPOLATORS);
   }
+  if (alloc_counts_.point_size) {
     output->append(
       "  float4 oPointSize : PSIZE;\n");
+  }
   output->append(
     "};\n");
 
@@ -461,8 +463,11 @@ const char* D3D11VertexShader::Translate(xe_gpu_program_cntl_t* program_cntl) {
 
   // Always write position, as some shaders seem to only write certain values.
   output->append(
-    "  o.oPos = float4(0.0, 0.0, 0.0, 0.0);\n"
+    "  o.oPos = float4(0.0, 0.0, 0.0, 0.0);\n");
+  if (alloc_counts_.point_size) {
+    output->append(
       "  o.oPointSize = float4(1.0, 0.0, 0.0, 0.0);\n");
+  }
 
   // TODO(benvanik): remove this, if possible (though the compiler may be smart
   //     enough to do it for us).

src/xenia/gpu/shader.cc

@@ -127,8 +127,18 @@ void Shader::GatherExec(const instr_cf_exec_t* cf) {
       }
     } else {
       // TODO(benvanik): gather registers used, predicate bits used, etc.
-      /*const instr_alu_t* alu =
+      const instr_alu_t* alu =
-          (const instr_alu_t*)(dwords_ + alu_off * 3);*/
+          (const instr_alu_t*)(dwords_ + alu_off * 3);
+      if (alu->vector_write_mask) {
+        if (alu->export_data && alu->vector_dest == 63) {
+          alloc_counts_.point_size = true;
+        }
+      }
+      if (alu->scalar_write_mask || !alu->vector_write_mask) {
+        if (alu->export_data && alu->scalar_dest == 63) {
+          alloc_counts_.point_size = true;
+        }
+      }
     }
     sequence >>= 2;
   }

src/xenia/gpu/shader.h

@@ -57,6 +57,7 @@ public:
     uint32_t  positions;
     uint32_t  params;
     uint32_t  memories;
+    bool      point_size;
   } alloc_counts_t;
   const alloc_counts_t& alloc_counts() const { return alloc_counts_; }