Merge pull request #1960 from ReinUsesLisp/shader-ir-ldg

video_core: Implement LDG through heuristics based on IR

src/video_core/CMakeLists.txt

@@ -87,6 +87,7 @@ add_library(video_core STATIC
     shader/decode.cpp
     shader/shader_ir.cpp
     shader/shader_ir.h
+    shader/track.cpp
     surface.cpp
     surface.h
     textures/astc.cpp

src/video_core/engines/shader_bytecode.h

@@ -208,6 +208,8 @@ enum class UniformType : u64 {
     SignedShort = 3,
     Single = 4,
     Double = 5,
+    Quad = 6,
+    UnsignedQuad = 7,
 };
 
 enum class StoreType : u64 {
@@ -784,6 +786,12 @@ union Instruction {
         BitField<44, 2, u64> unknown;
     } st_l;
 
+    union {
+        BitField<48, 3, UniformType> type;
+        BitField<46, 2, u64> cache_mode;
+        BitField<20, 24, s64> immediate_offset;
+    } ldg;
+
     union {
         BitField<0, 3, u64> pred0;
         BitField<3, 3, u64> pred3;

src/video_core/renderer_opengl/gl_global_cache.cpp

@@ -4,8 +4,13 @@
 
 #include <glad/glad.h>
 
+#include "common/assert.h"
+#include "common/logging/log.h"
+#include "core/core.h"
+#include "core/memory.h"
 #include "video_core/renderer_opengl/gl_global_cache.h"
 #include "video_core/renderer_opengl/gl_rasterizer.h"
+#include "video_core/renderer_opengl/gl_shader_decompiler.h"
 #include "video_core/renderer_opengl/utils.h"
 
 namespace OpenGL {
@@ -18,7 +23,72 @@ CachedGlobalRegion::CachedGlobalRegion(VAddr addr, u32 size) : addr{addr}, size{
     LabelGLObject(GL_BUFFER, buffer.handle, addr, "GlobalMemory");
 }
 
+void CachedGlobalRegion::Reload(u32 size_) {
+    constexpr auto max_size = static_cast<u32>(RasterizerOpenGL::MaxGlobalMemorySize);
+
+    size = size_;
+    if (size > max_size) {
+        size = max_size;
+        LOG_CRITICAL(HW_GPU, "Global region size {} exceeded the expected size {}!", size_,
+                     max_size);
+    }
+
+    // TODO(Rodrigo): Get rid of Memory::GetPointer with a staging buffer
+    glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer.handle);
+    glBufferData(GL_SHADER_STORAGE_BUFFER, size, Memory::GetPointer(addr), GL_DYNAMIC_DRAW);
+}
+
+GlobalRegion GlobalRegionCacheOpenGL::TryGetReservedGlobalRegion(VAddr addr, u32 size) const {
+    const auto search{reserve.find(addr)};
+    if (search == reserve.end()) {
+        return {};
+    }
+    return search->second;
+}
+
+GlobalRegion GlobalRegionCacheOpenGL::GetUncachedGlobalRegion(VAddr addr, u32 size) {
+    GlobalRegion region{TryGetReservedGlobalRegion(addr, size)};
+    if (!region) {
+        // No reserved surface available, create a new one and reserve it
+        region = std::make_shared<CachedGlobalRegion>(addr, size);
+        ReserveGlobalRegion(region);
+    }
+    region->Reload(size);
+    return region;
+}
+
+void GlobalRegionCacheOpenGL::ReserveGlobalRegion(const GlobalRegion& region) {
+    reserve[region->GetAddr()] = region;
+}
+
 GlobalRegionCacheOpenGL::GlobalRegionCacheOpenGL(RasterizerOpenGL& rasterizer)
     : RasterizerCache{rasterizer} {}
 
+GlobalRegion GlobalRegionCacheOpenGL::GetGlobalRegion(
+    const GLShader::GlobalMemoryEntry& global_region,
+    Tegra::Engines::Maxwell3D::Regs::ShaderStage stage) {
+
+    auto& gpu{Core::System::GetInstance().GPU()};
+    const auto cbufs = gpu.Maxwell3D().state.shader_stages[static_cast<u64>(stage)];
+    const auto cbuf_addr = gpu.MemoryManager().GpuToCpuAddress(
+        cbufs.const_buffers[global_region.GetCbufIndex()].address + global_region.GetCbufOffset());
+    ASSERT(cbuf_addr);
+
+    const auto actual_addr_gpu = Memory::Read64(*cbuf_addr);
+    const auto size = Memory::Read32(*cbuf_addr + 8);
+    const auto actual_addr = gpu.MemoryManager().GpuToCpuAddress(actual_addr_gpu);
+    ASSERT(actual_addr);
+
+    // Look up global region in the cache based on address
+    GlobalRegion region = TryGet(*actual_addr);
+
+    if (!region) {
+        // No global region found - create a new one
+        region = GetUncachedGlobalRegion(*actual_addr, size);
+        Register(region);
+    }
+
+    return region;
+}
+
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_global_cache.h

@@ -5,9 +5,13 @@
 #pragma once
 
 #include <memory>
+#include <unordered_map>
+
 #include <glad/glad.h>
 
+#include "common/assert.h"
 #include "common/common_types.h"
+#include "video_core/engines/maxwell_3d.h"
 #include "video_core/rasterizer_cache.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 
@@ -40,6 +44,9 @@ public:
         return buffer.handle;
     }
 
+    /// Reloads the global region from guest memory
+    void Reload(u32 size_);
+
     // TODO(Rodrigo): When global memory is written (STG), implement flushing
     void Flush() override {
         UNIMPLEMENTED();
@@ -55,6 +62,17 @@ private:
 class GlobalRegionCacheOpenGL final : public RasterizerCache<GlobalRegion> {
 public:
     explicit GlobalRegionCacheOpenGL(RasterizerOpenGL& rasterizer);
+
+    /// Gets the current specified shader stage program
+    GlobalRegion GetGlobalRegion(const GLShader::GlobalMemoryEntry& descriptor,
+                                 Tegra::Engines::Maxwell3D::Regs::ShaderStage stage);
+
+private:
+    GlobalRegion TryGetReservedGlobalRegion(VAddr addr, u32 size) const;
+    GlobalRegion GetUncachedGlobalRegion(VAddr addr, u32 size);
+    void ReserveGlobalRegion(const GlobalRegion& region);
+
+    std::unordered_map<VAddr, GlobalRegion> reserve;
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -300,6 +300,7 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
     // Next available bindpoints to use when uploading the const buffers and textures to the GLSL
     // shaders. The constbuffer bindpoint starts after the shader stage configuration bind points.
     u32 current_constbuffer_bindpoint = Tegra::Engines::Maxwell3D::Regs::MaxShaderStage;
+    u32 current_gmem_bindpoint = 0;
     u32 current_texture_bindpoint = 0;
     std::array<bool, Maxwell::NumClipDistances> clip_distances{};
 
@@ -358,6 +359,10 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
             SetupConstBuffers(static_cast<Maxwell::ShaderStage>(stage), shader, primitive_mode,
                               current_constbuffer_bindpoint);
 
+        // Configure global memory regions for this shader stage.
+        current_gmem_bindpoint = SetupGlobalRegions(static_cast<Maxwell::ShaderStage>(stage),
+                                                    shader, primitive_mode, current_gmem_bindpoint);
+
         // Configure the textures for this shader stage.
         current_texture_bindpoint = SetupTextures(static_cast<Maxwell::ShaderStage>(stage), shader,
                                                   primitive_mode, current_texture_bindpoint);
@@ -993,6 +998,23 @@ u32 RasterizerOpenGL::SetupConstBuffers(Maxwell::ShaderStage stage, Shader& shad
     return current_bindpoint + static_cast<u32>(entries.size());
 }
 
+u32 RasterizerOpenGL::SetupGlobalRegions(Maxwell::ShaderStage stage, Shader& shader,
+                                         GLenum primitive_mode, u32 current_bindpoint) {
+    for (const auto& global_region : shader->GetShaderEntries().global_memory_entries) {
+        const auto& region =
+            global_cache.GetGlobalRegion(global_region, static_cast<Maxwell::ShaderStage>(stage));
+        const GLuint block_index{shader->GetProgramResourceIndex(global_region)};
+        ASSERT(block_index != GL_INVALID_INDEX);
+
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, current_bindpoint, region->GetBufferHandle());
+        glShaderStorageBlockBinding(shader->GetProgramHandle(primitive_mode), block_index,
+                                    current_bindpoint);
+        ++current_bindpoint;
+    }
+
+    return current_bindpoint;
+}
+
 u32 RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, Shader& shader,
                                     GLenum primitive_mode, u32 current_unit) {
     MICROPROFILE_SCOPE(OpenGL_Texture);

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -137,6 +137,16 @@ private:
     u32 SetupConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, Shader& shader,
                           GLenum primitive_mode, u32 current_bindpoint);
 
+    /**
+     * Configures the current global memory regions to use for the draw command.
+     * @param stage The shader stage to configure buffers for.
+     * @param shader The shader object that contains the specified stage.
+     * @param current_bindpoint The offset at which to start counting new buffer bindpoints.
+     * @returns The next available bindpoint for use in the next shader stage.
+     */
+    u32 SetupGlobalRegions(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, Shader& shader,
+                           GLenum primitive_mode, u32 current_bindpoint);
+
     /**
      * Configures the current textures to use for the draw command.
      * @param stage The shader stage to configure textures for.

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -108,11 +108,23 @@ CachedShader::CachedShader(VAddr addr, Maxwell::ShaderProgram program_type)
 }
 
 GLuint CachedShader::GetProgramResourceIndex(const GLShader::ConstBufferEntry& buffer) {
-    const auto search{resource_cache.find(buffer.GetHash())};
+    const auto search{cbuf_resource_cache.find(buffer.GetHash())};
-    if (search == resource_cache.end()) {
+    if (search == cbuf_resource_cache.end()) {
         const GLuint index{
             glGetProgramResourceIndex(program.handle, GL_UNIFORM_BLOCK, buffer.GetName().c_str())};
-        resource_cache[buffer.GetHash()] = index;
+        cbuf_resource_cache[buffer.GetHash()] = index;
+        return index;
+    }
+
+    return search->second;
+}
+
+GLuint CachedShader::GetProgramResourceIndex(const GLShader::GlobalMemoryEntry& global_mem) {
+    const auto search{gmem_resource_cache.find(global_mem.GetHash())};
+    if (search == gmem_resource_cache.end()) {
+        const GLuint index{glGetProgramResourceIndex(program.handle, GL_SHADER_STORAGE_BLOCK,
+                                                     global_mem.GetName().c_str())};
+        gmem_resource_cache[global_mem.GetHash()] = index;
         return index;
     }
 

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -76,6 +76,9 @@ public:
     /// Gets the GL program resource location for the specified resource, caching as needed
     GLuint GetProgramResourceIndex(const GLShader::ConstBufferEntry& buffer);
 
+    /// Gets the GL program resource location for the specified resource, caching as needed
+    GLuint GetProgramResourceIndex(const GLShader::GlobalMemoryEntry& global_mem);
+
     /// Gets the GL uniform location for the specified resource, caching as needed
     GLint GetUniformLocation(const GLShader::SamplerEntry& sampler);
 
@@ -107,7 +110,8 @@ private:
         OGLProgram triangles_adjacency;
     } geometry_programs;
 
-    std::map<u32, GLuint> resource_cache;
+    std::map<u32, GLuint> cbuf_resource_cache;
+    std::map<u32, GLuint> gmem_resource_cache;
     std::map<u32, GLint> uniform_cache;
 };
 

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -34,6 +34,8 @@ using Operation = const OperationNode&;
 enum : u32 { POSITION_VARYING_LOCATION = 0, GENERIC_VARYING_START_LOCATION = 1 };
 constexpr u32 MAX_CONSTBUFFER_ELEMENTS =
     static_cast<u32>(RasterizerOpenGL::MaxConstbufferSize) / (4 * sizeof(float));
+constexpr u32 MAX_GLOBALMEMORY_ELEMENTS =
+    static_cast<u32>(RasterizerOpenGL::MaxGlobalMemorySize) / sizeof(float);
 
 enum class Type { Bool, Bool2, Float, Int, Uint, HalfFloat };
 
@@ -143,6 +145,7 @@ public:
         DeclareInputAttributes();
         DeclareOutputAttributes();
         DeclareConstantBuffers();
+        DeclareGlobalMemory();
         DeclareSamplers();
 
         code.AddLine("void execute_" + suffix + "() {");
@@ -190,12 +193,15 @@ public:
     ShaderEntries GetShaderEntries() const {
         ShaderEntries entries;
         for (const auto& cbuf : ir.GetConstantBuffers()) {
-            ConstBufferEntry desc(cbuf.second, stage, GetConstBufferBlock(cbuf.first), cbuf.first);
+            entries.const_buffers.emplace_back(cbuf.second, stage, GetConstBufferBlock(cbuf.first),
-            entries.const_buffers.push_back(desc);
+                                               cbuf.first);
         }
         for (const auto& sampler : ir.GetSamplers()) {
-            SamplerEntry desc(sampler, stage, GetSampler(sampler));
+            entries.samplers.emplace_back(sampler, stage, GetSampler(sampler));
-            entries.samplers.push_back(desc);
+        }
+        for (const auto& gmem : ir.GetGlobalMemoryBases()) {
+            entries.global_memory_entries.emplace_back(gmem.cbuf_index, gmem.cbuf_offset, stage,
+                                                       GetGlobalMemoryBlock(gmem));
         }
         entries.clip_distances = ir.GetClipDistances();
         entries.shader_length = ir.GetLength();
@@ -375,6 +381,15 @@ private:
         }
     }
 
+    void DeclareGlobalMemory() {
+        for (const auto& entry : ir.GetGlobalMemoryBases()) {
+            code.AddLine("layout (std430) buffer " + GetGlobalMemoryBlock(entry) + " {");
+            code.AddLine("    float " + GetGlobalMemory(entry) + "[MAX_GLOBALMEMORY_ELEMENTS];");
+            code.AddLine("};");
+            code.AddNewLine();
+        }
+    }
+
     void DeclareSamplers() {
         const auto& samplers = ir.GetSamplers();
         for (const auto& sampler : samplers) {
@@ -538,6 +553,12 @@ private:
                 UNREACHABLE_MSG("Unmanaged offset node type");
             }
 
+        } else if (const auto gmem = std::get_if<GmemNode>(node)) {
+            const std::string real = Visit(gmem->GetRealAddress());
+            const std::string base = Visit(gmem->GetBaseAddress());
+            const std::string final_offset = "(ftou(" + real + ") - ftou(" + base + ")) / 4";
+            return fmt::format("{}[{}]", GetGlobalMemory(gmem->GetDescriptor()), final_offset);
+
         } else if (const auto lmem = std::get_if<LmemNode>(node)) {
             return fmt::format("{}[ftou({}) / 4]", GetLocalMemory(), Visit(lmem->GetAddress()));
 
@@ -1471,6 +1492,15 @@ private:
         return GetDeclarationWithSuffix(index, "cbuf");
     }
 
+    std::string GetGlobalMemory(const GlobalMemoryBase& descriptor) const {
+        return fmt::format("gmem_{}_{}_{}", descriptor.cbuf_index, descriptor.cbuf_offset, suffix);
+    }
+
+    std::string GetGlobalMemoryBlock(const GlobalMemoryBase& descriptor) const {
+        return fmt::format("gmem_block_{}_{}_{}", descriptor.cbuf_index, descriptor.cbuf_offset,
+                           suffix);
+    }
+
     std::string GetConstBufferBlock(u32 index) const {
         return GetDeclarationWithSuffix(index, "cbuf_block");
     }
@@ -1505,8 +1535,10 @@ private:
 };
 
 std::string GetCommonDeclarations() {
-    return "#define MAX_CONSTBUFFER_ELEMENTS " + std::to_string(MAX_CONSTBUFFER_ELEMENTS) +
+    const auto cbuf = std::to_string(MAX_CONSTBUFFER_ELEMENTS);
-           "\n"
+    const auto gmem = std::to_string(MAX_GLOBALMEMORY_ELEMENTS);
+    return "#define MAX_CONSTBUFFER_ELEMENTS " + cbuf + "\n" +
+           "#define MAX_GLOBALMEMORY_ELEMENTS " + gmem + "\n" +
            "#define ftoi floatBitsToInt\n"
            "#define ftou floatBitsToUint\n"
            "#define itof intBitsToFloat\n"

src/video_core/renderer_opengl/gl_shader_decompiler.h

@@ -71,9 +71,43 @@ private:
     Maxwell::ShaderStage stage{};
 };
 
+class GlobalMemoryEntry {
+public:
+    explicit GlobalMemoryEntry(u32 cbuf_index, u32 cbuf_offset, Maxwell::ShaderStage stage,
+                               std::string name)
+        : cbuf_index{cbuf_index}, cbuf_offset{cbuf_offset}, stage{stage}, name{std::move(name)} {}
+
+    u32 GetCbufIndex() const {
+        return cbuf_index;
+    }
+
+    u32 GetCbufOffset() const {
+        return cbuf_offset;
+    }
+
+    const std::string& GetName() const {
+        return name;
+    }
+
+    Maxwell::ShaderStage GetStage() const {
+        return stage;
+    }
+
+    u32 GetHash() const {
+        return (static_cast<u32>(stage) << 24) | (cbuf_index << 16) | cbuf_offset;
+    }
+
+private:
+    u32 cbuf_index{};
+    u32 cbuf_offset{};
+    Maxwell::ShaderStage stage{};
+    std::string name;
+};
+
 struct ShaderEntries {
     std::vector<ConstBufferEntry> const_buffers;
     std::vector<SamplerEntry> samplers;
+    std::vector<GlobalMemoryEntry> global_memory_entries;
     std::array<bool, Maxwell::NumClipDistances> clip_distances{};
     std::size_t shader_length{};
 };

src/video_core/shader/decode/memory.cpp

@@ -4,6 +4,7 @@
 
 #include <algorithm>
 #include <vector>
+#include <fmt/format.h>
 
 #include "common/assert.h"
 #include "common/common_types.h"
@@ -119,6 +120,54 @@ u32 ShaderIR::DecodeMemory(BasicBlock& bb, const BasicBlock& code, u32 pc) {
         }
         break;
     }
+    case OpCode::Id::LDG: {
+        const u32 count = [&]() {
+            switch (instr.ldg.type) {
+            case Tegra::Shader::UniformType::Single:
+                return 1;
+            case Tegra::Shader::UniformType::Double:
+                return 2;
+            case Tegra::Shader::UniformType::Quad:
+            case Tegra::Shader::UniformType::UnsignedQuad:
+                return 4;
+            default:
+                UNIMPLEMENTED_MSG("Unimplemented LDG size!");
+                return 1;
+            }
+        }();
+
+        const Node addr_register = GetRegister(instr.gpr8);
+        const Node base_address = TrackCbuf(addr_register, code, static_cast<s64>(code.size()));
+        const auto cbuf = std::get_if<CbufNode>(base_address);
+        ASSERT(cbuf != nullptr);
+        const auto cbuf_offset_imm = std::get_if<ImmediateNode>(cbuf->GetOffset());
+        ASSERT(cbuf_offset_imm != nullptr);
+        const auto cbuf_offset = cbuf_offset_imm->GetValue() * 4;
+
+        bb.push_back(Comment(
+            fmt::format("Base address is c[0x{:x}][0x{:x}]", cbuf->GetIndex(), cbuf_offset)));
+
+        const GlobalMemoryBase descriptor{cbuf->GetIndex(), cbuf_offset};
+        used_global_memory_bases.insert(descriptor);
+
+        const Node immediate_offset =
+            Immediate(static_cast<u32>(instr.ldg.immediate_offset.Value()));
+        const Node base_real_address =
+            Operation(OperationCode::UAdd, NO_PRECISE, immediate_offset, addr_register);
+
+        for (u32 i = 0; i < count; ++i) {
+            const Node it_offset = Immediate(i * 4);
+            const Node real_address =
+                Operation(OperationCode::UAdd, NO_PRECISE, base_real_address, it_offset);
+            const Node gmem = StoreNode(GmemNode(real_address, base_address, descriptor));
+
+            SetTemporal(bb, i, gmem);
+        }
+        for (u32 i = 0; i < count; ++i) {
+            SetRegister(bb, instr.gpr0.Value() + i, GetTemporal(i));
+        }
+        break;
+    }
     case OpCode::Id::ST_A: {
         UNIMPLEMENTED_IF_MSG(instr.gpr8.Value() != Register::ZeroIndex,
                              "Indirect attribute loads are not supported");

src/video_core/shader/shader_ir.h

@@ -257,6 +257,15 @@ private:
     bool is_indirect{};
 };
 
+struct GlobalMemoryBase {
+    u32 cbuf_index{};
+    u32 cbuf_offset{};
+
+    bool operator<(const GlobalMemoryBase& rhs) const {
+        return std::tie(cbuf_index, cbuf_offset) < std::tie(rhs.cbuf_index, rhs.cbuf_offset);
+    }
+};
+
 struct MetaArithmetic {
     bool precise{};
 };
@@ -478,14 +487,26 @@ private:
 /// Global memory node
 class GmemNode final {
 public:
-    explicit constexpr GmemNode(Node address) : address{address} {}
+    explicit constexpr GmemNode(Node real_address, Node base_address,
+                                const GlobalMemoryBase& descriptor)
+        : real_address{real_address}, base_address{base_address}, descriptor{descriptor} {}
 
-    Node GetAddress() const {
+    Node GetRealAddress() const {
-        return address;
+        return real_address;
+    }
+
+    Node GetBaseAddress() const {
+        return base_address;
+    }
+
+    const GlobalMemoryBase& GetDescriptor() const {
+        return descriptor;
     }
 
 private:
-    const Node address;
+    const Node real_address;
+    const Node base_address;
+    const GlobalMemoryBase descriptor;
 };
 
 /// Commentary, can be dropped
@@ -543,6 +564,10 @@ public:
         return used_clip_distances;
     }
 
+    const std::set<GlobalMemoryBase>& GetGlobalMemoryBases() const {
+        return used_global_memory_bases;
+    }
+
     std::size_t GetLength() const {
         return static_cast<std::size_t>(coverage_end * sizeof(u64));
     }
@@ -734,6 +759,10 @@ private:
     void WriteLop3Instruction(BasicBlock& bb, Tegra::Shader::Register dest, Node op_a, Node op_b,
                               Node op_c, Node imm_lut, bool sets_cc);
 
+    Node TrackCbuf(Node tracked, const BasicBlock& code, s64 cursor);
+
+    std::pair<Node, s64> TrackRegister(const GprNode* tracked, const BasicBlock& code, s64 cursor);
+
     template <typename... T>
     Node Operation(OperationCode code, const T*... operands) {
         return StoreNode(OperationNode(code, operands...));
@@ -786,6 +815,7 @@ private:
     std::map<u32, ConstBuffer> used_cbufs;
     std::set<Sampler> used_samplers;
     std::array<bool, Tegra::Engines::Maxwell3D::Regs::NumClipDistances> used_clip_distances{};
+    std::set<GlobalMemoryBase> used_global_memory_bases;
 
     Tegra::Shader::Header header;
 };

src/video_core/shader/track.cpp

@@ -0,0 +1,76 @@
+// Copyright 2018 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <algorithm>
+#include <utility>
+#include <variant>
+
+#include "video_core/shader/shader_ir.h"
+
+namespace VideoCommon::Shader {
+
+namespace {
+std::pair<Node, s64> FindOperation(const BasicBlock& code, s64 cursor,
+                                   OperationCode operation_code) {
+    for (; cursor >= 0; --cursor) {
+        const Node node = code[cursor];
+        if (const auto operation = std::get_if<OperationNode>(node)) {
+            if (operation->GetCode() == operation_code)
+                return {node, cursor};
+        }
+    }
+    return {};
+}
+} // namespace
+
+Node ShaderIR::TrackCbuf(Node tracked, const BasicBlock& code, s64 cursor) {
+    if (const auto cbuf = std::get_if<CbufNode>(tracked)) {
+        // Cbuf found, but it has to be immediate
+        return std::holds_alternative<ImmediateNode>(*cbuf->GetOffset()) ? tracked : nullptr;
+    }
+    if (const auto gpr = std::get_if<GprNode>(tracked)) {
+        if (gpr->GetIndex() == Tegra::Shader::Register::ZeroIndex) {
+            return nullptr;
+        }
+        // Reduce the cursor in one to avoid infinite loops when the instruction sets the same
+        // register that it uses as operand
+        const auto [source, new_cursor] = TrackRegister(gpr, code, cursor - 1);
+        if (!source) {
+            return nullptr;
+        }
+        return TrackCbuf(source, code, new_cursor);
+    }
+    if (const auto operation = std::get_if<OperationNode>(tracked)) {
+        for (std::size_t i = 0; i < operation->GetOperandsCount(); ++i) {
+            if (const auto found = TrackCbuf((*operation)[i], code, cursor)) {
+                // Cbuf found in operand
+                return found;
+            }
+        }
+        return nullptr;
+    }
+    return nullptr;
+}
+
+std::pair<Node, s64> ShaderIR::TrackRegister(const GprNode* tracked, const BasicBlock& code,
+                                             s64 cursor) {
+    for (; cursor >= 0; --cursor) {
+        const auto [found_node, new_cursor] = FindOperation(code, cursor, OperationCode::Assign);
+        if (!found_node) {
+            return {};
+        }
+        const auto operation = std::get_if<OperationNode>(found_node);
+        ASSERT(operation);
+
+        const auto& target = (*operation)[0];
+        if (const auto gpr_target = std::get_if<GprNode>(target)) {
+            if (gpr_target->GetIndex() == tracked->GetIndex()) {
+                return {(*operation)[1], new_cursor};
+            }
+        }
+    }
+    return {};
+}
+
+} // namespace VideoCommon::Shader