[DXBC] ROV: Force late Z write with kill instructions

src/xenia/gpu/dxbc_shader_translator.h

@@ -189,12 +189,12 @@ class DxbcShaderTranslator : public ShaderTranslator {
     kSysFlag_ROVStencilTest_Shift,
     // If the depth/stencil test has failed, but resulted in a stencil value
     // that is different than the one currently in the depth buffer, write it
-    // anyway and don't run the shader (to check if the sample may be discarded
-    // some way). This, however, also results in depth/stencil testing done
-    // entirely early even when it passes to prevent writing in divergent places
-    // in the shader. When the shader can kill, this must be set only for
-    // RB_DEPTHCONTROL EARLY_Z_ENABLE, not for alpha test/alpha to coverage
-    // disabled.
+    // anyway and don't run the rest of the shader (to check if the sample may
+    // be discarded some way) - use when alpha test and alpha to coverage are
+    // disabled. Ignored by the shader if not applicable to it (like if it has
+    // kill instructions or writes the depth output).
+    // TODO(Triang3l): Replace with an alpha-to-mask flag, check if
+    // (flags & (alpha test | alpha to mask)) == (always | disabled).
     kSysFlag_ROVDepthStencilEarlyWrite_Shift,
 
     kSysFlag_Count,

src/xenia/gpu/dxbc_shader_translator_om.cc

@@ -1024,6 +1024,14 @@ void DxbcShaderTranslator::ROV_DepthStencilTest() {
     // temp.z = viewport maximum depth if not writing to oDepth
     // temp.w = whether depth/stencil has been modified
     DxbcOpINE(temp_w_dest, sample_depth_stencil_src, temp_w_src);
+    if (depth_stencil_early && !CanWriteZEarly()) {
+      // Set the sample bit in bits 4:7 of system_temp_rov_params_.x - always
+      // need to write late in this shader, as it may do something like
+      // explicitly killing pixels.
+      DxbcOpBFI(DxbcDest::R(system_temp_rov_params_, 0b0001), DxbcSrc::LU(1),
+                DxbcSrc::LU(4 + i), temp_w_src,
+                DxbcSrc::R(system_temp_rov_params_, DxbcSrc::kXXXX));
+    } else {
       // Check if need to write.
       // temp.x? = resulting sample depth/stencil
       // temp.y = polygon offset if not writing to oDepth
@@ -1057,7 +1065,7 @@ void DxbcShaderTranslator::ROV_DepthStencilTest() {
           // Need to still run the shader to know whether to write the
           // depth/stencil value.
           DxbcOpElse();
-        // Set sample bit out of bits 4:7 of system_temp_rov_params_.x if need
+          // Set the sample bit in bits 4:7 of system_temp_rov_params_.x if need
           // to write later (after checking if the sample is not discarded by a
           // kill instruction, alphatest or alpha-to-coverage).
           DxbcOpOr(DxbcDest::R(system_temp_rov_params_, 0b0001),
@@ -1069,6 +1077,7 @@ void DxbcShaderTranslator::ROV_DepthStencilTest() {
       }
       // Close the write check.
       DxbcOpEndIf();
+    }
 
     // Release sample_temp.
     PopSystemTemp();