Merge branch 'master' of https://github.com/xenia-project/xenia into canary_experimental

src/xenia/gpu/d3d12/pipeline_cache.cc

@@ -2394,6 +2394,10 @@ void PipelineCache::CreateDxbcGeometryShader(
   // to again.
   // Also, FXC generates only movs (from statically or dynamically indexed
   // v[#][#], from r#, or from a literal) to o# for some reason.
+  // emit_then_cut_stream must not be used - it crashes the shader compiler of
+  // AMD Software: Adrenalin Edition 23.3.2 on RDNA 3 if it's conditional (after
+  // a `retc` or inside an `if`), and it doesn't seem to be generated by FXC or
+  // DXC at all.
 
   // Discard the whole primitive if any vertex has a NaN position (may also be
   // set to NaN for emulation of vertex killing with the OR operator).
@@ -2539,11 +2543,9 @@ void PipelineCache::CreateDxbcGeometryShader(
               dxbc::Src::V2D(
                   0, input_register_clip_and_cull_distances + (j >> 2)));
         }
-        if (i < 3) {
-          a.OpEmitStream(stream);
-        }
+        a.OpEmitStream(stream);
       }
-      a.OpEmitThenCutStream(stream);
+      a.OpCutStream(stream);
     } break;
 
     case PipelineGeometryShader::kRectangleList: {
@@ -2689,7 +2691,8 @@ void PipelineCache::CreateDxbcGeometryShader(
                               clip_distance_mask),
                 dxbc::Src::R(1));
       }
-      a.OpEmitThenCutStream(stream);
+      a.OpEmitStream(stream);
+      a.OpCutStream(stream);
     } break;
 
     case PipelineGeometryShader::kQuadList: {
@@ -2720,11 +2723,9 @@ void PipelineCache::CreateDxbcGeometryShader(
                   input_vertex_index,
                   input_register_clip_and_cull_distances + (j >> 2)));
         }
-        if (i < 3) {
-          a.OpEmitStream(stream);
-        }
+        a.OpEmitStream(stream);
       }
-      a.OpEmitThenCutStream(stream);
+      a.OpCutStream(stream);
     } break;
 
     default:

src/xenia/gpu/dxbc.h

@@ -2197,6 +2197,8 @@ class Assembler {
     ++stat_.instruction_count;
     ++stat_.cut_instruction_count;
   }
+  // Don't use emit_then_cut_stream - crashes AMD Software: Adrenalin Edition
+  // 23.3.2 shader compiler on RDNA 3 if used conditionally.
   void OpEmitThenCutStream(const Dest& stream) {
     uint32_t operands_length = stream.GetLength();
     code_.reserve(code_.size() + 1 + operands_length);

src/xenia/gpu/vulkan/vulkan_render_target_cache.cc

@@ -3174,7 +3174,9 @@ VkShaderModule VulkanRenderTargetCache::GetTransferShader(
       source_stencil[0] != spv::NoResult) {
     // For the depth -> depth case, write the stencil directly to the output.
     assert_true(mode.output == TransferOutput::kDepth);
-    builder.createStore(source_stencil[0], output_fragment_stencil_ref);
+    builder.createStore(
+        builder.createUnaryOp(spv::OpBitcast, type_int, source_stencil[0]),
+        output_fragment_stencil_ref);
   }
 
   if (dest_is_64bpp) {
@@ -3518,13 +3520,15 @@ VkShaderModule VulkanRenderTargetCache::GetTransferShader(
               if (output_fragment_stencil_ref != spv::NoResult) {
                 builder.createStore(
                     builder.createUnaryOp(
-                        spv::OpConvertFToU, type_uint,
-                        builder.createBinOp(
-                            spv::OpFAdd, type_float,
-                            builder.createBinOp(spv::OpFMul, type_float,
-                                                source_color[0][0],
-                                                unorm_scale),
-                            unorm_round_offset)),
+                        spv::OpBitcast, type_int,
+                        builder.createUnaryOp(
+                            spv::OpConvertFToU, type_uint,
+                            builder.createBinOp(
+                                spv::OpFAdd, type_float,
+                                builder.createBinOp(spv::OpFMul, type_float,
+                                                    source_color[0][0],
+                                                    unorm_scale),
+                                unorm_round_offset))),
                     output_fragment_stencil_ref);
               }
             }
@@ -4331,6 +4335,17 @@ VkShaderModule VulkanRenderTargetCache::GetTransferShader(
                 builder.createOp(spv::OpPhi, type_float, id_vector_temp);
           }
           builder.createStore(fragment_depth32, output_fragment_depth);
+          // Unpack the stencil into the stencil reference output if needed and
+          // not already written.
+          if (!packed_only_depth &&
+              output_fragment_stencil_ref != spv::NoResult) {
+            builder.createStore(
+                builder.createUnaryOp(
+                    spv::OpBitcast, type_int,
+                    builder.createBinOp(spv::OpBitwiseAnd, type_uint, packed,
+                                        builder.makeUintConstant(UINT8_MAX))),
+                output_fragment_stencil_ref);
+          }
         }
       } break;
       case TransferOutput::kStencilBit: {

src/xenia/ui/imgui_drawer.h

@@ -13,6 +13,7 @@
 #include <cstddef>
 #include <cstdint>
 #include <memory>
+#include <optional>
 #include <vector>
 
 #include "third_party/imgui/imgui.h"