shader: Implement VertexA stage

4 files changed, 128 insertions, 0 deletions

diff --git a/src/shader_recompiler/ir_opt/dual_vertex_pass.cpp b/src/shader_recompiler/ir_opt/dual_vertex_pass.cpp
new file mode 100644
index 000000000..f35c6478a
--- /dev/null
+++ b/src/shader_recompiler/ir_opt/dual_vertex_pass.cpp
@@ -0,0 +1,74 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <algorithm>
+#include <ranges>
+#include <tuple>
+#include <type_traits>
+
+#include "common/bit_cast.h"
+#include "common/bit_util.h"
+#include "shader_recompiler/exception.h"
+#include "shader_recompiler/frontend/ir/ir_emitter.h"
+#include "shader_recompiler/ir_opt/passes.h"
+
+namespace Shader::Optimization {
+
+void VertexATransformPass(IR::Program& program) {
+    bool replaced_join{};
+    bool eliminated_epilogue{};
+    for (IR::Block* const block : program.post_order_blocks) {
+        for (IR::Inst& inst : block->Instructions()) {
+            switch (inst.GetOpcode()) {
+            case IR::Opcode::Return:
+                inst.ReplaceOpcode(IR::Opcode::Join);
+                replaced_join = true;
+                break;
+            case IR::Opcode::Epilogue:
+                inst.Invalidate();
+                eliminated_epilogue = true;
+                break;
+            default:
+                break;
+            }
+            if (replaced_join && eliminated_epilogue) {
+                return;
+            }
+        }
+    }
+}
+
+void VertexBTransformPass(IR::Program& program) {
+    for (IR::Block* const block : program.post_order_blocks | std::views::reverse) {
+        for (IR::Inst& inst : block->Instructions()) {
+            if (inst.GetOpcode() == IR::Opcode::Prologue) {
+                return inst.Invalidate();
+            }
+        }
+    }
+}
+
+void DualVertexJoinPass(IR::Program& program) {
+    const auto& blocks = program.blocks;
+    s64 s = static_cast<s64>(blocks.size()) - 1;
+    if (s < 1) {
+        throw NotImplementedException("Dual Vertex Join pass failed, expected atleast 2 blocks!");
+    }
+    for (s64 index = 0; index < s; index++) {
+        IR::Block* const current_block = blocks[index];
+        IR::Block* const next_block = blocks[index + 1];
+        for (IR::Inst& inst : current_block->Instructions()) {
+            if (inst.GetOpcode() == IR::Opcode::Join) {
+                IR::IREmitter ir{*current_block, IR::Block::InstructionList::s_iterator_to(inst)};
+                ir.Branch(next_block);
+                inst.Invalidate();
+                // only 1 join should exist
+                return;
+            }
+        }
+    }
+    throw NotImplementedException("Dual Vertex Join pass failed, no join present!");
+}
+
+} // namespace Shader::Optimization
diff --git a/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp b/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp
index 87eca2a0d..1d11a00d8 100644
--- a/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp
+++ b/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp
@@ -499,4 +499,30 @@ void GlobalMemoryToStorageBufferPass(IR::Program& program) {
     }
 }
 
+template <typename Descriptors, typename Descriptor, typename Func>
+static u32 Add(Descriptors& descriptors, const Descriptor& desc, Func&& pred) {
+    // TODO: Handle arrays
+    const auto it{std::ranges::find_if(descriptors, pred)};
+    if (it != descriptors.end()) {
+        return static_cast<u32>(std::distance(descriptors.begin(), it));
+    }
+    descriptors.push_back(desc);
+    return static_cast<u32>(descriptors.size()) - 1;
+}
+
+void JoinStorageInfo(Info& base, Info& source) {
+    auto& descriptors = base.storage_buffers_descriptors;
+    for (auto& desc : source.storage_buffers_descriptors) {
+        auto it{std::ranges::find_if(descriptors, [&desc](const auto& existing) {
+            return desc.cbuf_index == existing.cbuf_index &&
+                   desc.cbuf_offset == existing.cbuf_offset && desc.count == existing.count;
+        })};
+        if (it != descriptors.end()) {
+            it->is_written |= desc.is_written;
+            continue;
+        }
+        descriptors.push_back(desc);
+    }
+}
+
 } // namespace Shader::Optimization
diff --git a/src/shader_recompiler/ir_opt/passes.h b/src/shader_recompiler/ir_opt/passes.h
index 186104713..e9cb8546a 100644
--- a/src/shader_recompiler/ir_opt/passes.h
+++ b/src/shader_recompiler/ir_opt/passes.h
@@ -22,4 +22,11 @@ void SsaRewritePass(IR::Program& program);
 void TexturePass(Environment& env, IR::Program& program);
 void VerificationPass(const IR::Program& program);
 
+// Dual Vertex
+void VertexATransformPass(IR::Program& program);
+void VertexBTransformPass(IR::Program& program);
+void DualVertexJoinPass(IR::Program& program);
+void JoinTextureInfo(Info& base, Info& source);
+void JoinStorageInfo(Info& base, Info& source);
+
 } // namespace Shader::Optimization
diff --git a/src/shader_recompiler/ir_opt/texture_pass.cpp b/src/shader_recompiler/ir_opt/texture_pass.cpp
index cfa6b34b9..2b38bcf42 100644
--- a/src/shader_recompiler/ir_opt/texture_pass.cpp
+++ b/src/shader_recompiler/ir_opt/texture_pass.cpp
@@ -426,4 +426,25 @@ void TexturePass(Environment& env, IR::Program& program) {
     }
 }
 
+void JoinTextureInfo(Info& base, Info& source) {
+    Descriptors descriptors{
+        base.texture_buffer_descriptors,
+        base.image_buffer_descriptors,
+        base.texture_descriptors,
+        base.image_descriptors,
+    };
+    for (auto& desc : source.texture_buffer_descriptors) {
+        descriptors.Add(desc);
+    }
+    for (auto& desc : source.image_buffer_descriptors) {
+        descriptors.Add(desc);
+    }
+    for (auto& desc : source.texture_descriptors) {
+        descriptors.Add(desc);
+    }
+    for (auto& desc : source.image_descriptors) {
+        descriptors.Add(desc);
+    }
+}
+
 } // namespace Shader::Optimization