[DXBC] Fix and cleanup depth-only PS register usage

src/xenia/gpu/dxbc_shader_translator.cc

@@ -833,15 +833,17 @@ void DxbcShaderTranslator::StartTranslation() {
     if (edram_rov_used_) {
       // Will be initialized unconditionally.
       system_temp_rov_params_ = PushSystemTemp();
+      if (ROV_IsDepthStencilEarly() || writes_depth()) {
         // If the shader doesn't write to oDepth, each component will be written
         // to if depth/stencil is enabled and the respective sample is covered -
-      // so need to initialize now because the first writes will be conditional.
-      // If the shader writes to oDepth, this is oDepth of the shader, written
-      // by the guest code, so initialize because assumptions can't be made
-      // about the integrity of the guest code.
+        // so need to initialize now because the first writes will be
+        // conditional. If the shader writes to oDepth, this is oDepth of the
+        // shader, written by the guest code, so initialize because assumptions
+        // can't be made about the integrity of the guest code.
         system_temp_rov_depth_stencil_ =
             PushSystemTemp(writes_depth() ? 0b0001 : 0b1111);
       }
+    }
     for (uint32_t i = 0; i < 4; ++i) {
       if (writes_color_target(i)) {
         system_temps_color_[i] = PushSystemTemp(0b1111);
@@ -887,16 +889,16 @@ void DxbcShaderTranslator::StartTranslation() {
     system_temp_loop_count_ = PushSystemTemp(0b1111);
     system_temp_grad_h_lod_ = PushSystemTemp(0b1111);
     system_temp_grad_v_ = PushSystemTemp(0b0111);
-  }
 
-  // Zero general-purpose registers to prevent crashes when the game references
-  // them after only initializing them conditionally.
+    // Zero general-purpose registers to prevent crashes when the game
+    // references them after only initializing them conditionally.
     for (uint32_t i = IsDxbcPixelShader() ? kInterpolatorCount : 0;
          i < register_count(); ++i) {
-    DxbcOpMov(
-        uses_register_dynamic_addressing() ? DxbcDest::X(0, i) : DxbcDest::R(i),
+      DxbcOpMov(uses_register_dynamic_addressing() ? DxbcDest::X(0, i)
+                                                   : DxbcDest::R(i),
                 DxbcSrc::LF(0.0f));
     }
+  }
 
   // Write stage-specific prologue.
   if (IsDxbcVertexOrDomainShader()) {
@@ -1163,8 +1165,12 @@ void DxbcShaderTranslator::CompleteShaderCode() {
       }
     }
     if (edram_rov_used_) {
-      // Release system_temp_rov_params_ and system_temp_rov_depth_stencil_.
-      PopSystemTemp(2);
+      if (ROV_IsDepthStencilEarly() || writes_depth()) {
+        // Release system_temp_rov_depth_stencil_.
+        PopSystemTemp();
+      }
+      // Release system_temp_rov_params_.
+      PopSystemTemp();
     }
   }
 

src/xenia/gpu/dxbc_shader_translator.h

@@ -2522,7 +2522,9 @@ class DxbcShaderTranslator : public ShaderTranslator {
   //     dwords.
   uint32_t system_temp_rov_params_;
   // ROV only - new depth/stencil data. 4 VGPRs when not writing to oDepth, 1
-  // VGPR when writing to oDepth.
+  // VGPR when writing to oDepth. Not used in the depth-only pixel shader (or,
+  // more formally, if neither early depth-stencil nor oDepth are used) because
+  // it always calculates and writes in the same place.
   // When not writing to oDepth: New per-sample depth/stencil values, generated
   // during early depth/stencil test (actual writing checks coverage bits).
   // When writing to oDepth: X also used to hold the depth written by the

src/xenia/gpu/dxbc_shader_translator_om.cc

@@ -581,10 +581,14 @@ void DxbcShaderTranslator::ROV_DepthStencilTest() {
     // Perform depth/stencil test for the sample, get the result in bits 4
     // (passed) and 8 (new depth/stencil buffer value is different).
     DxbcOpCall(DxbcSrc::Label(label_rov_depth_stencil_sample_));
-    // Write the resulting depth/stencil value in system_temps_subroutine_[0].x
-    // to the sample's depth in system_temp_rov_depth_stencil_.
+    if (ROV_IsDepthStencilEarly()) {
+      // Write the resulting depth/stencil value in
+      // system_temps_subroutine_[0].x to the sample's depth in
+      // system_temp_rov_depth_stencil_.
       DxbcOpMov(DxbcDest::R(system_temp_rov_depth_stencil_, 1 << i),
                 DxbcSrc::R(system_temps_subroutine_, DxbcSrc::kXXXX));
+    }
+    if (!is_depth_only_pixel_shader_) {
       if (i) {
         // Shift the result bits to the correct position.
         DxbcOpIShL(DxbcDest::R(system_temps_subroutine_, 0b0010),
@@ -593,11 +597,12 @@ void DxbcShaderTranslator::ROV_DepthStencilTest() {
       }
       // Add the result in system_temps_subroutine_[0].y to
       // system_temp_rov_params_.x. Bits 0:3 will be cleared in case of test
-    // failure (only doing this for covered samples), bits 4:7 will be added if
-    // need to defer writing.
+      // failure (only doing this for covered samples), bits 4:7 will be added
+      // if need to defer writing.
       DxbcOpXOr(DxbcDest::R(system_temp_rov_params_, 0b0001),
                 DxbcSrc::R(system_temp_rov_params_, DxbcSrc::kXXXX),
                 DxbcSrc::R(system_temps_subroutine_, DxbcSrc::kYYYY));
+    }
 
     // Close the sample conditional.
     DxbcOpEndIf();
@@ -1595,11 +1600,13 @@ void DxbcShaderTranslator::CompletePixelShader_WriteToROV() {
     ROV_DepthStencilTest();
   }
 
-  // Check if any sample is still covered after depth testing and writing, skip
-  // color writing completely in this case.
+  if (!is_depth_only_pixel_shader_) {
+    // Check if any sample is still covered after depth testing and writing,
+    // skip color writing completely in this case.
     DxbcOpAnd(temp_x_dest, DxbcSrc::R(system_temp_rov_params_, DxbcSrc::kXXXX),
               DxbcSrc::LU(0b1111));
     DxbcOpRetC(false, temp_x_src);
+  }
 
   // Write color values.
   for (uint32_t i = 0; i < 4; ++i) {