pipeline: print a console message before the shader cache warm-up

launcher: close the terminal after a clean exit and run several games at once

Co-authored-by: nmzik <Nmzik@mail.ru>

CMakeLists.txt

@@ -337,7 +337,7 @@ if (KYTY_CLANG_CL)
 	set_source_files_properties(${kyty_emulator_src} PROPERTIES COMPILE_FLAGS "-Wno-pragma-pack -Wno-deprecated-declarations -D_TIMESPEC_DEFINED")
 endif()
 
-set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus OpenXR::openxr_loader)
+set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus OpenXR::openxr_loader libzstd_static)
 
 # Linux system libraries.
 if(LINUX)
@@ -451,6 +451,11 @@ add_executable(lru_cache_tests EXCLUDE_FROM_ALL
 )
 target_include_directories(lru_cache_tests PRIVATE ${inc_headers})
 
+add_executable(slot_vector_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/SlotVectorTests.cpp"
+)
+target_include_directories(slot_vector_tests PRIVATE ${inc_headers})
+
 add_executable(ime_dialog_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/ImeDialogTests.cpp"
 	"${KYTY_SOURCE_DIR}/libs/imeCommon.cpp"

src/common/additionalContent.cpp

@@ -109,4 +109,23 @@ std::vector<AddContentEntry> FindAddContent(const std::filesystem::path& dir,
 	return entries;
 }
 
+std::optional<AddContentEntry> ParseAddContent(std::string_view value) {
+	static const std::regex label_pattern("^[A-Z0-9]{16}$");
+	const auto              eq    = value.find('=');
+	const auto              label = value.substr(0, eq);
+	if (!std::regex_match(label.begin(), label.end(), label_pattern)) {
+		return std::nullopt;
+	}
+
+	AddContentEntry entry {{}, std::string(label), {}, {}};
+	if (eq != std::string_view::npos) {
+		entry.mount_dir = PathFromUtf8(FixFilenameSlash(value.substr(eq + 1)));
+		std::error_code ec;
+		if (!std::filesystem::is_directory(entry.mount_dir, ec)) {
+			return std::nullopt;
+		}
+	}
+	return entry;
+}
+
 } // namespace Common

src/common/additionalContent.h

@@ -2,7 +2,9 @@
 #define KYTY_COMMON_ADDITIONAL_CONTENT_H_
 
 #include <filesystem>
+#include <optional>
 #include <string>
+#include <string_view>
 #include <vector>
 
 namespace Common {
@@ -10,7 +12,7 @@ namespace Common {
 struct AddContentEntry {
 	std::string           content_id;
 	std::string           entitlement_label;
-	std::filesystem::path mount_dir;
+	std::filesystem::path mount_dir; // Empty for an add-on without extra data.
 	std::string           title;
 };
 
@@ -18,6 +20,9 @@ struct AddContentEntry {
 std::vector<AddContentEntry> FindAddContent(const std::filesystem::path& dir,
                                             const std::string&           title_id);
 
+// Parses a user add-on: "LABEL" without extra data, or "LABEL=DIR" with its data in DIR.
+std::optional<AddContentEntry> ParseAddContent(std::string_view value);
+
 } // namespace Common
 
 #endif /* KYTY_COMMON_ADDITIONAL_CONTENT_H_ */

src/common/emulatorConfig.cpp

@@ -133,6 +133,10 @@ bool ShaderValidationEnabled() {
 	return g_config->shader_validation_enabled;
 }
 
+bool ShaderCacheEnabled() {
+	return g_config->shader_cache_enabled;
+}
+
 ShaderOptimizationType GetShaderOptimizationType() {
 	return g_config->shader_optimization_type;
 }

src/common/emulatorConfig.h

@@ -72,6 +72,7 @@ struct ConfigOptions {
 	uint32_t               console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
 	bool                   shader_validation_enabled   = false;
+	bool                   shader_cache_enabled        = false;
 	ShaderOptimizationType shader_optimization_type    = ShaderOptimizationType::None;
 	LogDirection           shader_log_direction        = LogDirection::Silent;
 	std::filesystem::path  shader_log_folder           = "_Shaders";
@@ -132,6 +133,7 @@ uint32_t GetConsoleLanguage();
 bool     VulkanValidationEnabled();
 
 bool                   ShaderValidationEnabled();
+bool                   ShaderCacheEnabled();
 ShaderOptimizationType GetShaderOptimizationType();
 LogDirection           GetShaderLogDirection();
 std::filesystem::path  GetShaderLogFolder();

src/common/slotVector.h

@@ -1,15 +1,15 @@
 #ifndef EMULATOR_SRC_COMMON_SLOTVECTOR_H_
 #define EMULATOR_SRC_COMMON_SLOTVECTOR_H_
 
-#include "common/assert.h"
-#include "common/abi.h"
+#include "common/common.h"
 
 #include <compare>
 #include <cstdint>
-#include <deque>
 #include <functional>
 #include <limits>
-#include <optional>
+#include <memory>
+#include <new>
+#include <numeric>
 #include <utility>
 #include <vector>
 
@@ -19,105 +19,105 @@ struct SlotId {
 	static constexpr uint32_t INVALID_INDEX = std::numeric_limits<uint32_t>::max();
 
 	constexpr SlotId() noexcept = default;
-	constexpr SlotId(uint32_t value) noexcept: index(value), generation(1) {}
-	constexpr SlotId(uint32_t value, uint32_t slot_generation) noexcept
-	    : index(value), generation(slot_generation) {}
+	constexpr SlotId(uint32_t index) noexcept: index(index) {}
 
-	[[nodiscard]] constexpr explicit operator bool() const noexcept { return index != INVALID_INDEX; }
+	[[nodiscard]] constexpr explicit operator bool() const noexcept {
+		return index != INVALID_INDEX;
+	}
 	constexpr auto operator<=>(const SlotId&) const noexcept = default;
 
 	uint32_t index = INVALID_INDEX;
-	uint32_t generation = 0;
 };
 
-// Stable-address slot storage for cache resources that are intentionally non-movable.
+// Fixed-size chunks keep the stored, non-movable resources in place. Occupancy is a byte per
+// slot, not a bitset: a deferred erase runs while fault threads read other slots.
 template <typename T>
 class SlotVector {
+	static constexpr uint32_t ChunkSize = 256;
+
 public:
 	SlotVector() = default;
 	KYTY_CLASS_NO_COPY(SlotVector);
 
+	~SlotVector() noexcept {
+		ForEach([](SlotId, T& value) { value.~T(); });
+	}
+
 	[[nodiscard]] T& operator[](SlotId id) noexcept {
-		EXIT_IF(!is_allocated(id));
-		return *m_values[id.index].value;
+		return m_chunks[id.index / ChunkSize][id.index % ChunkSize].object;
 	}
 
 	[[nodiscard]] const T& operator[](SlotId id) const noexcept {
-		EXIT_IF(!is_allocated(id));
-		return *m_values[id.index].value;
+		return m_chunks[id.index / ChunkSize][id.index % ChunkSize].object;
 	}
 
 	[[nodiscard]] T* try_get(SlotId id) noexcept {
-		return is_allocated(id) ? &*m_values[id.index].value : nullptr;
+		return is_allocated(id) ? &(*this)[id] : nullptr;
 	}
 
 	[[nodiscard]] const T* try_get(SlotId id) const noexcept {
-		return is_allocated(id) ? &*m_values[id.index].value : nullptr;
+		return is_allocated(id) ? &(*this)[id] : nullptr;
 	}
 
 	[[nodiscard]] bool is_allocated(SlotId id) const noexcept {
-		return id && id.index < m_values.size() &&
-		       m_values[id.index].generation == id.generation && m_values[id.index].value.has_value();
+		return id.index < m_stored.size() && m_stored[id.index] != 0;
 	}
 
 	template <typename... Args>
 	[[nodiscard]] SlotId insert(Args&&... args) {
-		uint32_t index = 0;
 		if (m_free_list.empty()) {
-			index = static_cast<uint32_t>(m_values.size());
-			m_values.emplace_back();
-			m_values.back().value.emplace(std::forward<Args>(args)...);
-		} else {
-			index = m_free_list.back();
-			m_free_list.pop_back();
-			EXIT_IF(m_values[index].value.has_value());
-			m_values[index].value.emplace(std::forward<Args>(args)...);
+			Reserve();
 		}
-		++m_size;
-		return SlotId {index, m_values[index].generation};
+		const uint32_t index = m_free_list.back();
+		m_free_list.pop_back();
+		new (&(*this)[index]) T(std::forward<Args>(args)...);
+		m_stored[index] = 1;
+		return SlotId {index};
 	}
 
 	void erase(SlotId id) noexcept {
-		EXIT_IF(!is_allocated(id));
-		auto& slot = m_values[id.index];
-		slot.value.reset();
-		if (++slot.generation == 0) {
-			slot.generation = 1;
-		}
+		(*this)[id].~T();
+		m_stored[id.index] = 0;
 		m_free_list.push_back(id.index);
-		--m_size;
 	}
 
-	[[nodiscard]] size_t size() const noexcept { return m_size; }
-	[[nodiscard]] size_t capacity() const noexcept { return m_values.size(); }
-
 	template <typename F>
 	void ForEach(F&& fn) {
-		for (uint32_t index = 0; index < m_values.size(); ++index) {
-			if (m_values[index].value) {
-				fn(SlotId {index, m_values[index].generation}, *m_values[index].value);
+		for (uint32_t index = 0; index < m_stored.size(); ++index) {
+			if (m_stored[index] != 0) {
+				fn(SlotId {index}, (*this)[index]);
 			}
 		}
 	}
 
 	template <typename F>
 	void ForEach(F&& fn) const {
-		for (uint32_t index = 0; index < m_values.size(); ++index) {
-			if (m_values[index].value) {
-				fn(SlotId {index, m_values[index].generation}, *m_values[index].value);
+		for (uint32_t index = 0; index < m_stored.size(); ++index) {
+			if (m_stored[index] != 0) {
+				fn(SlotId {index}, (*this)[index]);
 			}
 		}
 	}
 
 private:
-	struct Slot {
-		std::optional<T> value;
-		uint32_t         generation = 1;
+	union Slot {
+		Slot() noexcept {}
+		~Slot() noexcept {}
+
+		T object;
 	};
 
-	std::deque<Slot>             m_values;
-	std::vector<uint32_t>        m_free_list;
-	size_t                       m_size = 0;
+	void Reserve() {
+		const auto first = static_cast<uint32_t>(m_stored.size());
+		m_chunks.push_back(std::make_unique<Slot[]>(ChunkSize));
+		m_stored.resize(first + ChunkSize);
+		m_free_list.resize(ChunkSize);
+		std::iota(m_free_list.rbegin(), m_free_list.rend(), first);
+	}
+
+	std::vector<std::unique_ptr<Slot[]>> m_chunks;
+	std::vector<uint8_t>                 m_stored;
+	std::vector<uint32_t>                m_free_list;
 };
 
 } // namespace Common
@@ -125,7 +125,7 @@ private:
 template <>
 struct std::hash<Common::SlotId> {
 	[[nodiscard]] size_t operator()(Common::SlotId id) const noexcept {
-		return std::hash<uint64_t> {}((static_cast<uint64_t>(id.generation) << 32u) | id.index);
+		return std::hash<uint32_t> {}(id.index);
 	}
 };
 

src/emulator.cpp

@@ -188,7 +188,7 @@ void Run(const RunOptions& options) {
 
 	MountSandboxDirs();
 
-	Loader::AddContentLoad(options.addcont_dir, title_id);
+	Loader::AddContentLoad(options.addcont_dir, options.addcont, title_id);
 	auto* rt = Common::Singleton<Loader::RuntimeLinker>::Instance();
 	Libs::InitAll(rt->Symbols());
 

src/emulator.h

@@ -1,6 +1,7 @@
 #ifndef EMULATOR_INCLUDE_EMULATOR_EMULATOR_H_
 #define EMULATOR_INCLUDE_EMULATOR_EMULATOR_H_
 
+#include "common/additionalContent.h"
 #include "common/emulatorConfig.h"
 #include "common/stringUtils.h"
 
@@ -9,11 +10,12 @@
 namespace Emulator {
 
 struct RunOptions {
-	Config::ConfigOptions config;
-	std::filesystem::path app0_dir;
-	std::filesystem::path elf;
-	std::filesystem::path game_patch;
-	std::filesystem::path addcont_dir;
+	Config::ConfigOptions                config;
+	std::filesystem::path                app0_dir;
+	std::filesystem::path                elf;
+	std::filesystem::path                game_patch;
+	std::filesystem::path                addcont_dir;
+	std::vector<Common::AddContentEntry> addcont;
 };
 
 void Run(const RunOptions& options);

src/graphics/host_gpu/memoryTracker.cpp

@@ -66,18 +66,6 @@ RegionManager* MemoryTracker::GetOrCreateRegion(uint64_t index) {
 	return ptr;
 }
 
-void MemoryTracker::QueueBdaSync(uint64_t vaddr, uint64_t size) {
-	const auto begin = Common::AlignDown(vaddr, TRACKER_PAGE_SIZE);
-	const auto end   = Common::AlignUp(vaddr + size, TRACKER_PAGE_SIZE);
-	std::lock_guard lock(m_bda_sync_mutex);
-	m_bda_sync.Add(begin, end - begin);
-}
-
-RangeSet MemoryTracker::TakeBdaSync() {
-	std::lock_guard lock(m_bda_sync_mutex);
-	return std::exchange(m_bda_sync, RangeSet {});
-}
-
 bool MemoryTracker::IsRegionCpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	return Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
@@ -100,7 +88,6 @@ void MemoryTracker::MarkRegionAsCpuModified(uint64_t vaddr, uint64_t size) {
 		std::scoped_lock lock(manager->lock);
 		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
-	QueueBdaSync(vaddr, size);
 }
 
 void MemoryTracker::MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size) {
@@ -141,8 +128,6 @@ void MemoryTracker::UntrackMemory(uint64_t vaddr, uint64_t size) {
 	Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
-	locks.clear();
-	QueueBdaSync(vaddr, size);
 }
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/memoryTracker.h

@@ -29,10 +29,6 @@ public:
 	void               MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UnmarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UntrackMemory(uint64_t vaddr, uint64_t size);
-	// Queue after changing ownership; a concurrent drain must not consume an event
-	// before the dirty pages it describes become visible.
-	void                   QueueBdaSync(uint64_t vaddr, uint64_t size);
-	[[nodiscard]] RangeSet TakeBdaSync();
 	// Removes protection from a range and flushes GPU-owned data when required.
 	template <typename Flush>
 	void InvalidateRegion(uint64_t vaddr, uint64_t size, Flush&& on_flush) noexcept {
@@ -55,7 +51,6 @@ public:
 				on_flush();
 			}
 		});
-		QueueBdaSync(vaddr, size);
 	}
 #if KYTY_BUILD == KYTY_BUILD_DEBUG
 	void ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
@@ -154,8 +149,6 @@ private:
 	std::unique_ptr<std::atomic<RegionManager*>[]> m_regions;
 	std::vector<std::unique_ptr<RegionManager>>    m_region_storage;
 	std::mutex                                     m_region_mutex;
-	std::mutex                                     m_bda_sync_mutex;
-	RangeSet                                       m_bda_sync;
 	PageManager&                                   m_page_manager;
 };
 

src/graphics/host_gpu/renderer/cache/bufferCache.cpp

@@ -127,7 +127,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		}
 		WriteDataBuffer(m_bda_pagetable_buffer, table_offset,
 		                addresses.data(), addresses.size() * sizeof(vk::DeviceAddress));
-		m_memory_tracker.QueueBdaSync(buffer.CpuAddress(), buffer.Size());
+		QueueBdaSync(buffer.CpuAddress(), buffer.Size());
 	} else {
 		const auto found = m_buffers.find(buffer.CpuAddress());
 		EXIT_IF(found == m_buffers.end() || found->second != id);
@@ -299,6 +299,7 @@ void BufferCache::InvalidateMemory(uint64_t vaddr, uint64_t size) {
 	}
 	m_memory_tracker.InvalidateRegion(vaddr, size,
 	                                  [this, vaddr, size] { ReadMemory(vaddr, size, true); });
+	QueueBdaSync(vaddr, size);
 }
 
 void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
@@ -318,6 +319,7 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
 		}
 		if (is_write) {
 			m_memory_tracker.MarkRegionAsCpuModified(vaddr, size);
+			QueueBdaSync(vaddr, size);
 		}
 	});
 }
@@ -719,14 +721,16 @@ void BufferCache::ProcessFaultBuffer() {
 	m_fault_manager.ProcessFaultBuffer();
 }
 
-void BufferCache::SynchronizeBda(const RangeSet& mapped) {
-	// Take before walking: writes queued during synchronization remain pending.
-	const auto pending = m_memory_tracker.TakeBdaSync();
-	pending.ForEach([&](uint64_t begin, uint64_t end) {
-		mapped.ForEachInRange(begin, end - begin, [this](uint64_t start, uint64_t finish) {
-			SynchronizeBuffersInRange(start, finish - start);
-		});
-	});
+void BufferCache::QueueBdaSync(uint64_t vaddr, uint64_t size) {
+	const auto begin = Common::AlignDown(vaddr, TRACKER_PAGE_SIZE);
+	const auto end   = Common::AlignUp(vaddr + size, TRACKER_PAGE_SIZE);
+	std::lock_guard lock(m_bda_sync_mutex);
+	m_bda_sync.Add(begin, end - begin);
+}
+
+RangeSet BufferCache::TakeBdaSync() {
+	std::lock_guard lock(m_bda_sync_mutex);
+	return std::exchange(m_bda_sync, {});
 }
 
 void BufferCache::SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size) {

src/graphics/host_gpu/renderer/cache/bufferCache.h

@@ -13,6 +13,7 @@
 
 #include <array>
 #include <map>
+#include <mutex>
 #include <span>
 #include <utility>
 #include <vector>
@@ -85,19 +86,16 @@ public:
 	void               MarkColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys);
 	[[nodiscard]] bool IsColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys) const;
 	void               ProcessFaultBuffer();
-	void               SynchronizeBda(const RangeSet& mapped);
-	void               QueueBdaSync(uint64_t vaddr, uint64_t size) {
-		m_memory_tracker.QueueBdaSync(vaddr, size);
-	}
+	void               SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
+	// Queue after the range's state change becomes visible, or a drain could consume it first.
+	void                   QueueBdaSync(uint64_t vaddr, uint64_t size);
+	[[nodiscard]] RangeSet TakeBdaSync();
 	void               RunGarbageCollector();
 
 private:
 	friend struct BufferCacheTestAccess;
 
-	bool IsBufferInvalid(BufferId id) const {
-		const auto* buffer = m_slot_buffers.try_get(id);
-		return buffer == nullptr || buffer->is_deleted;
-	}
+	bool IsBufferInvalid(BufferId id) const { return !id || m_slot_buffers[id].is_deleted; }
 
 	using BufferMap = std::map<uint64_t, BufferId>;
 	struct OverlapResult {
@@ -122,7 +120,6 @@ private:
 	void DeleteBuffer(BufferId id);
 	[[nodiscard]] bool SynchronizeBuffer(Buffer& buffer, uint64_t vaddr, uint64_t size,
 	                                     bool is_written, bool is_texel_buffer);
-	void SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] vk::Buffer UploadCopies(Buffer& buffer, std::span<vk::BufferCopy> copies,
 	                                      uint64_t total_size);
 	[[nodiscard]] bool SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size);
@@ -144,6 +141,8 @@ private:
 	// One per TileManager::ColorClearSlots key slot.
 	std::array<RangeSet, 8>                           m_checked_color_metadata;
 	MemoryTracker                                     m_memory_tracker;
+	std::mutex                                        m_bda_sync_mutex;
+	RangeSet                                          m_bda_sync;
 	StreamBuffer                                      m_staging_buffer;
 	StreamBuffer                                      m_stream_buffer;
 	StreamBuffer                                      m_download_buffer;

src/graphics/host_gpu/renderer/cache/textureCache.cpp

@@ -251,24 +251,24 @@ void TextureCache::UnregisterImage(ImageId id) {
 }
 
 void TextureCache::DeleteImage(ImageId id) {
-	auto* image = m_slot_images.try_get(id);
-	if (image == nullptr || !image->registered) {
+	auto& image = m_slot_images[id];
+	if (!image.registered) {
 		return;
 	}
-	if (image->depth_id) {
-		std::erase(m_slot_images[image->depth_id].stencil_ids, id);
+	if (image.depth_id) {
+		std::erase(m_slot_images[image.depth_id].stencil_ids, id);
 	} else {
 		// Detach first: deleting each association also unlinks it from its depth owner.
-		for (const auto association: std::exchange(image->stencil_ids, {})) {
+		for (const auto association: std::exchange(image.stencil_ids, {})) {
 			FreeImage(association);
 		}
 	}
-	if (image->IsGpuModified()) {
+	if (image.IsGpuModified()) {
 		EXIT("TextureCache: deleting a GPU-modified image without resolving its contents\n");
 	}
 	m_download_images.erase(id);
-	if (image->info.metadata.kind == ImageMetadataKind::Htile) {
-		m_surface_metas.erase(image->info.metadata.range.address);
+	if (image.info.metadata.kind == ImageMetadataKind::Htile) {
+		m_surface_metas.erase(image.info.metadata.range.address);
 	}
 	UnregisterImage(id);
 	if (m_scheduler.Active()) {
@@ -429,12 +429,6 @@ TextureCache::ImageIds TextureCache::FindImagesInRegion(uint64_t address, uint64
 		return {};
 	}
 
-	uint32_t query_epoch = ++m_image_query_epoch;
-	if (query_epoch == 0) {
-		m_slot_images.ForEach([](ImageId, const Image& image) { image.query_epoch = 0; });
-		query_epoch = ++m_image_query_epoch;
-	}
-
 	ImageIds result;
 	ForEachPage(address, size, [&](uint64_t page) {
 		const auto* owners = m_image_page_table.Find(page);
@@ -442,19 +436,17 @@ TextureCache::ImageIds TextureCache::FindImagesInRegion(uint64_t address, uint64
 			return;
 		}
 		owners->ForEach([&](ImageId id) {
-			auto* image = m_slot_images.try_get(id);
-			if (image == nullptr) {
+			const auto& image = m_slot_images[id];
+			if (image.picked || !image.Overlaps(address, size, page_overlap)) {
 				return;
 			}
-			if (image->query_epoch == query_epoch) {
-				return;
-			}
-			image->query_epoch = query_epoch;
-			if (image->Overlaps(address, size, page_overlap)) {
-				result.push_back(id);
-			}
+			image.picked = true;
+			result.push_back(id);
 		});
 	});
+	for (const auto id: result) {
+		m_slot_images[id].picked = false;
+	}
 	return result;
 }
 
@@ -787,15 +779,14 @@ void Resolve(TextureCache& cache, ImageId cached_id, ImageId merged_id) {
 TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& requested,
                                                          BindingType binding, ImageId cached_id,
                                                          ImageId merged_id) {
-	auto owner = m_slot_images.try_get(cached_id);
-	if (owner == nullptr || !owner->registered) {
+	auto& cached = m_slot_images[cached_id];
+	if (!cached.registered) {
 		return {merged_id};
 	}
-	if (BcAlias::Matches(requested, owner->info) && !owner->IsCpuDirty()) {
+	if (BcAlias::Matches(requested, cached.info) && !cached.IsCpuDirty()) {
 		BcAlias::Resolve(*this, cached_id, merged_id);
 		return {merged_id};
 	}
-	auto&      cached       = *owner;
 	const auto current_tick = m_scheduler.CurrentTick();
 	const bool safe_to_delete =
 	    current_tick - std::min(current_tick, cached.tick_accessed_last) > NumFramesBeforeRemoval &&
@@ -869,9 +860,8 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		if (requested.pixel_format != cached.info.pixel_format ||
 		    requested.data.size <= cached.info.data.size) {
 			const auto result_id = merged_id ? merged_id : cached_id;
-			const auto result    = m_slot_images.try_get(result_id);
-			return {result != nullptr && ImageViewOps::FormatsCompatible(result->info.pixel_format,
-			                                                             requested.pixel_format)
+			return {ImageViewOps::FormatsCompatible(m_slot_images[result_id].info.pixel_format,
+			                                        requested.pixel_format)
 			            ? result_id
 			            : ImageId {}};
 		}
@@ -1218,9 +1208,8 @@ ImageId TextureCache::AssociateStencil(ImageId depth_id, GuestRange stencil) {
 
 	ImageId association {};
 	for (const auto id: FindImagesInRegion(stencil.address, stencil.size, false)) {
-		const auto owner = m_slot_images.try_get(id);
-		if (owner != nullptr && owner->info.data == stencil &&
-		    owner->info.extent == depth.info.extent) {
+		const auto& owner = m_slot_images[id];
+		if (owner.info.data == stencil && owner.info.extent == depth.info.extent) {
 			association = id;
 		}
 	}
@@ -1315,10 +1304,9 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 			}
 		} else if (block_alias) {
 			for (const auto candidate: candidates) {
-				const auto* owner = m_slot_images.try_get(candidate);
-				if (candidate != result && owner != nullptr && owner->registered &&
-				    !owner->IsCpuDirty() &&
-				    BcAlias::Matches(m_slot_images[result].info, owner->info)) {
+				const auto& owner = m_slot_images[candidate];
+				if (candidate != result && owner.registered && !owner.IsCpuDirty() &&
+				    BcAlias::Matches(m_slot_images[result].info, owner.info)) {
 					BcAlias::Resolve(*this, candidate, result);
 				}
 			}
@@ -1381,14 +1369,14 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
 	std::scoped_lock lock {m_lock};
 	ImageIds         matches;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr || owner->info.data.address != address) {
+		const auto* owner = &m_slot_images[id];
+		if (owner->info.data.address != address) {
 			continue;
 		}
 		if (ensure_valid && owner->depth_id) {
-			owner = m_slot_images.try_get(owner->depth_id);
+			owner = &m_slot_images[owner->depth_id];
 		}
-		if (owner == nullptr || (ensure_valid && !owner->SafeToDownload())) {
+		if (ensure_valid && !owner->SafeToDownload()) {
 			continue;
 		}
 		matches.push_back(id);
@@ -1405,11 +1393,8 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
 			}
 		}
 	}
-	if (selected && ensure_valid) {
-		const auto owner = m_slot_images.try_get(selected);
-		if (owner != nullptr && owner->depth_id) {
-			selected = owner->depth_id;
-		}
+	if (selected && ensure_valid && m_slot_images[selected].depth_id) {
+		selected = m_slot_images[selected].depth_id;
 	}
 	return selected;
 }
@@ -1430,9 +1415,8 @@ vk::ImageView TextureCache::FindTexture(ImageId id, const ImageDesc& desc) {
 	    desc.info.data.End() <= image.info.stencil.End()) {
 		for (const auto stencil_id:
 		     FindImagesInRegion(image.info.stencil.address, image.info.stencil.size, false)) {
-			const auto* stencil = m_slot_images.try_get(stencil_id);
-			if (stencil != nullptr && stencil->depth_id == id &&
-			    stencil->info.data == image.info.stencil) {
+			const auto& stencil = m_slot_images[stencil_id];
+			if (stencil.depth_id == id && stencil.info.data == image.info.stencil) {
 				RefreshImage(stencil_id);
 				break;
 			}
@@ -1543,27 +1527,24 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	vk::ImageAspectFlags  aspect {};
 	ImageSubresourceRange subresources;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr) {
-			continue;
-		}
+		const auto&          owner = m_slot_images[id];
 		vk::ImageAspectFlags candidate {};
 		ImageId              candidate_id = id;
 		ImageSubresourceRange candidate_subresources {
-		    0, owner->info.resources.levels, 0, owner->info.TransferLayers()};
-		if (owner->depth_id && owner->info.data.address == address &&
-		    owner->info.data.size == size) {
+		    0, owner.info.resources.levels, 0, owner.info.TransferLayers()};
+		if (owner.depth_id && owner.info.data.address == address &&
+		    owner.info.data.size == size) {
 			candidate              = vk::ImageAspectFlagBits::eStencil;
-			candidate_id           = owner->depth_id;
-			candidate_subresources = owner->stencil_subresources;
-			owner        = m_slot_images.try_get(candidate_id);
-			if (owner == nullptr || owner->backing.image == nullptr || !owner->info.HasStencil()) {
+			candidate_id           = owner.depth_id;
+			candidate_subresources = owner.stencil_subresources;
+			const auto& depth      = m_slot_images[candidate_id];
+			if (depth.backing.image == nullptr || !depth.info.HasStencil()) {
 				continue;
 			}
-		} else if (!owner->depth_id && owner->info.data.address == address &&
-		           owner->info.data.size == size) {
-			candidate = owner->info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
-			                                  : vk::ImageAspectFlagBits::eColor;
+		} else if (!owner.depth_id && owner.info.data.address == address &&
+		           owner.info.data.size == size) {
+			candidate = owner.info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
+			                                 : vk::ImageAspectFlagBits::eColor;
 		}
 		if (!candidate) {
 			continue;
@@ -1904,23 +1885,20 @@ void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
 	const auto page_begin = Common::AlignDown(address, TRACKER_PAGE_SIZE);
 	const auto page_end   = Common::AlignUp(address + size, TRACKER_PAGE_SIZE);
 	for (const auto id: FindImagesInRegion(address, size, true)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr) {
-			continue;
-		}
-		if (owner->Overlaps(address, size)) {
-			owner->MarkCpuDirty();
+		auto& owner = m_slot_images[id];
+		if (owner.Overlaps(address, size)) {
+			owner.MarkCpuDirty();
 			UntrackImage(id);
 			continue;
 		}
-		const auto image_begin = owner->info.data.address;
-		const auto image_end   = owner->info.data.End();
+		const auto image_begin = owner.info.data.address;
+		const auto image_end   = owner.info.data.End();
 		if (page_end < image_end) {
 			UntrackImageHead(id);
 		} else if (image_begin < page_begin) {
 			UntrackImageTail(id);
 		} else {
-			MarkAsMaybeDirty(id, *owner);
+			MarkAsMaybeDirty(id, owner);
 		}
 	}
 }
@@ -2060,8 +2038,7 @@ void TextureCache::ProcessDownloadImages() {
 	}
 	std::scoped_lock lock {m_lock};
 	for (const auto id: m_download_images) {
-		const auto owner = m_slot_images.try_get(id);
-		if (owner != nullptr && owner->registered && owner->IsGpuModified()) {
+		if (m_slot_images[id].IsGpuModified()) {
 			(void)DownloadImageMemory(id);
 		}
 	}

src/graphics/host_gpu/renderer/cache/textureCache.h

@@ -184,7 +184,6 @@ private:
 	uint64_t                                          m_pressure_gc_memory = 1536ull * 1024 * 1024;
 	uint64_t         m_critical_gc_memory     = 3ull * 1024 * 1024 * 1024;
 	uint64_t         m_gc_tick                = 0;
-	mutable uint32_t m_image_query_epoch      = 0;
 	bool             m_readback_linear_images = false;
 	bool             m_readback_tiled_images  = false;
 

src/graphics/host_gpu/renderer/debug.cpp

@@ -595,13 +595,6 @@ void LogDrawPhase(const char* draw_name, const char* phase) {
 	}
 }
 
-void LogPipelineTrace(const char* phase, uint64_t vertex_program_id, uint64_t pixel_program_id) {
-	if (graphics_debug_dump_enabled()) {
-		LOGF("PipelineTrace: %s VS=%" PRIu64 " PS=%" PRIu64 "\n", phase, vertex_program_id,
-		     pixel_program_id);
-	}
-}
-
 static void VpPrint(const char* func, const HW::ScreenViewport& vp,
                     const HW::ScanModeControl& smc) {
 	LOGF("%s\n", func);

src/graphics/host_gpu/renderer/image/image.h

@@ -139,14 +139,17 @@ public:
 	}
 
 	static inline std::atomic<uint64_t> heap_epoch {0};
+	static inline std::atomic<uint64_t> serials {0};
 
+	// Unique per image, unlike ImageId: freed slots are reused.
+	const uint64_t   serial = serials.fetch_add(1u, std::memory_order_relaxed) + 1u;
 	ImageInfo        info;
 	VulkanImage      backing;
 	std::vector<CachedImageView> views;
 	ImageUsage       usage;
 	ImageBinding     binding;
 	bool             registered     = false;
-	mutable uint32_t query_epoch    = 0;
+	mutable bool     picked         = false; // Set only while FindImagesInRegion collects it.
 	uint64_t         track_addr     = 0;
 	uint64_t         track_addr_end = 0;
 	ImageId          depth_id {};

src/graphics/host_gpu/renderer/pipeline/descriptors.cpp

@@ -672,9 +672,7 @@ void RenderExecutor::BindRenderTarget(ImageId id) {
 
 void RenderExecutor::ResetBindings() {
 	for (const auto id: m_bound_images) {
-		if (auto* image = m_context.GetTextureCache().m_slot_images.try_get(id); image != nullptr) {
-			image->binding = {};
-		}
+		m_context.GetTextureCache().m_slot_images[id].binding = {};
 	}
 	m_bound_images.clear();
 }
@@ -683,7 +681,7 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
                                      PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(!runtime);
-	const auto& program  = *runtime.program;
+	const auto& program  = runtime.module->info;
 	const auto& snapshot = *runtime.resources;
 	prepared.runtime = &runtime;
 	prepared.gds = {nullptr, 0, VK_WHOLE_SIZE};
@@ -725,7 +723,7 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 	for (auto* stage: stages) {
 		auto& prepared = *stage;
 		EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-		const auto& program  = *prepared.runtime->program;
+		const auto& program  = prepared.runtime->module->info;
 		const auto& snapshot = *prepared.runtime->resources;
 		const auto& layout = program.bindings;
 		prepared.buffer_sources.resize(layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
@@ -782,7 +780,7 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-	const auto& program   = *prepared.runtime->program;
+	const auto& program   = prepared.runtime->module->info;
 	const auto& snapshot  = *prepared.runtime->resources;
 	const auto& layout    = program.bindings;
 	EXIT_IF(prepared.buffer_sources.size() != layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
@@ -815,7 +813,7 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-	const auto& program  = *prepared.runtime->program;
+	const auto& program  = prepared.runtime->module->info;
 	const auto& snapshot = *prepared.runtime->resources;
 	auto&       images   = prepared.images;
 	EXIT_IF(images.size() != program.info.images.size());
@@ -823,12 +821,9 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	// Acquire each view before a later overlapping descriptor can replace its image.
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		if (!images[i].image_id) continue;
-		const auto old_image = texture_cache.m_slot_images.try_get(images[i].image_id);
-		if (old_image == nullptr || !old_image->registered ||
-		    old_image->binding.needs_rebind) {
-			if (old_image != nullptr) {
-				old_image->binding = {};
-			}
+		auto& old_image = texture_cache.m_slot_images[images[i].image_id];
+		if (!old_image.registered || old_image.binding.needs_rebind) {
+			old_image.binding = {};
 			images[i] = ResolveTexture(program.info.images[i], snapshot.images[i]);
 			BindImage(images[i].image_id,
 			          images[i].desc.type == TextureCache::BindingType::Storage);
@@ -864,8 +859,8 @@ void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const>
 	bool uses_image_heap = false;
 	FindBuffers(stages);
 	for (auto* stage: stages) {
-		uses_dma |= stage->runtime->program->info.uses_dma;
-		uses_image_heap |= stage->runtime->program->info.uses_image_heap;
+		uses_dma |= stage->runtime->module->info.info.uses_dma;
+		uses_image_heap |= stage->runtime->module->info.info.uses_image_heap;
 	}
 	if (uses_dma) {
 		m_context.PrepareBda();
@@ -879,12 +874,9 @@ void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const>
 	auto& cache = m_context.GetTextureCache();
 	for (auto& target: colors) {
 		EXIT_IF(!target.image_id);
-		const auto old_image = cache.m_slot_images.try_get(target.image_id);
-		if (old_image == nullptr || !old_image->registered ||
-		    old_image->binding.needs_rebind) {
-			if (old_image != nullptr) {
-				old_image->binding = {};
-			}
+		auto& old_image = cache.m_slot_images[target.image_id];
+		if (!old_image.registered || old_image.binding.needs_rebind) {
+			old_image.binding = {};
 			target.desc.view_info.base_level = target.guest_mip_level;
 			target.desc.view_info.base_layer = target.guest_array_layer;
 			target.image_id = cache.FindImage(target.desc);
@@ -917,7 +909,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	                                       : vk::ShaderStageFlags {};
 	for (const auto* prepared: prepared_bindings) {
 		EXIT_IF(prepared == nullptr || prepared->runtime == nullptr || !*prepared->runtime);
-		const auto& program = *prepared->runtime->program;
+		const auto& program = prepared->runtime->module->info;
 		write_count += std::popcount(program.bindings.descriptor_mask);
 		for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
 			const auto index = std::countr_zero(mask);
@@ -938,17 +930,16 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 		const auto& reads = reader->runtime->resources->specialization_reads;
 		if (reads.empty()) continue;
 		for (const auto* writer: prepared_bindings) {
-			if (writer->runtime->program->has_address_writes) {
+			if (writer->runtime->module->info.has_address_writes) {
 				EXIT("scalar resource reads cannot be proven disjoint from shader address writes\n");
 			}
 		}
 		for (const auto [address, size]: reads) {
 			for (const auto id: m_bound_images) {
-				const auto* image = m_context.GetTextureCache().m_slot_images.try_get(id);
-				if (image == nullptr ||
-				    (!image->binding.shader_write && !image->binding.is_target)) continue;
-				for (const auto written: {image->info.data, image->info.stencil,
-				                          image->info.metadata.range}) {
+				const auto& image = m_context.GetTextureCache().m_slot_images[id];
+				if (!image.binding.shader_write && !image.binding.is_target) continue;
+				for (const auto written: {image.info.data, image.info.stencil,
+				                          image.info.metadata.range}) {
 					if (written.size != 0 && ImageRangeOverlaps(address, size,
 					                                          written.address, written.size)) {
 						EXIT("scalar resource reads overlap an image or attachment write\n");
@@ -956,7 +947,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 				}
 			}
 			for (const auto* writer: prepared_bindings) {
-				const auto& program = *writer->runtime->program;
+				const auto& program = writer->runtime->module->info;
 				for (const auto& resource: program.info.buffers) {
 					if (!resource.written || resource.descriptor_index == UINT32_MAX) continue;
 					const auto& written = writer->buffer_sources[resource.descriptor_index];
@@ -982,7 +973,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	std::array<vk::DescriptorBufferInfo, 2> address_buffers;
 
 	for (auto* prepared: prepared_bindings) {
-		const auto& program       = *prepared->runtime->program;
+		const auto& program       = prepared->runtime->module->info;
 		auto&       descriptors   = *prepared;
 		const auto  shader_stage  = NativeShaderStage(program.stage);
 		const auto  shader_stages = ShaderPipelineStages(shader_stage);

src/graphics/host_gpu/renderer/pipeline/imageHeap.cpp

@@ -350,12 +350,17 @@ void ImageHeap::Release(View& view, uint32_t entry) {
 }
 
 void ImageHeap::Unreference(Element& element) {
-	if (auto* image = m_context.GetTextureCache().m_slot_images.try_get(element.binding.image_id)) {
+	if (auto* image = ImageOf(element)) {
 		image->heap_refs.fetch_sub(1u, std::memory_order_relaxed);
 		image->tick_accessed_last = m_context.GetCommandScheduler().CurrentTick();
 	}
 }
 
+Image* ImageHeap::ImageOf(const Element& element) const {
+	auto* image = m_context.GetTextureCache().m_slot_images.try_get(element.binding.image_id);
+	return image != nullptr && image->serial == element.serial ? image : nullptr;
+}
+
 uint32_t ImageHeap::Allocate() {
 	auto& scheduler = m_context.GetCommandScheduler();
 	while (!m_retired.empty() && scheduler.IsFree(m_retired.front().first)) {
@@ -393,7 +398,7 @@ void ImageHeap::Rebind(RenderExecutor& executor) {
 			for (const auto& [value, index]: view->elements) {
 				const auto& binding = m_elements[index].binding;
 				if (!binding.image_id) continue;
-				const auto* image = textures.m_slot_images.try_get(binding.image_id);
+				const auto* image = ImageOf(m_elements[index]);
 				if (image == nullptr || !image->registered || image->depth_id ||
 				    image->binding.needs_rebind) {
 					mark(index);
@@ -434,6 +439,7 @@ void ImageHeap::Rebind(RenderExecutor& executor) {
 			if (binding.image_id) {
 				auto& image = textures.GetImage(binding.image_id);
 				image.heap_refs.fetch_add(1u, std::memory_order_relaxed);
+				m_elements[index].serial = image.serial;
 				binding.layout     = image.ReadLayout();
 				binding.image_view = textures.FindTexture(binding.image_id, binding.desc);
 				if (binding.desc.info.metadata.kind == ImageMetadataKind::Dcc) {
@@ -471,8 +477,7 @@ void ImageHeap::WriteDescriptors() {
 void ImageHeap::Commit(CommandBuffer& buffer, vk::PipelineBindPoint bind_point,
                        vk::PipelineLayout layout) {
 	KYTY_PROFILER_FUNCTION();
-	auto&      textures = m_context.GetTextureCache();
-	const auto epoch    = Image::heap_epoch.load(std::memory_order_relaxed);
+	const auto epoch = Image::heap_epoch.load(std::memory_order_relaxed);
 	for (auto* view: m_active) {
 		if (view->dirty_begin != view->dirty_end) {
 			const auto slots = std::span(view->slots).subspan(view->dirty_begin,
@@ -493,7 +498,7 @@ void ImageHeap::Commit(CommandBuffer& buffer, vk::PipelineBindPoint bind_point,
 		// this draw's discovery replaced lives until the tick ends, and its slots still select it.
 		std::erase_if(view->away, [&](uint32_t index) {
 			const auto& element = m_elements[index];
-			auto*       image   = textures.m_slot_images.try_get(element.binding.image_id);
+			auto*       image   = ImageOf(element);
 			if (element.refs == 0u || image == nullptr) return true;
 			if (image->binding.is_target || image->binding.shader_write ||
 			    image->binding.force_general) {

src/graphics/host_gpu/renderer/pipeline/imageHeap.h

@@ -57,9 +57,10 @@ private:
 	struct Element {
 		DescriptorValue value;
 		TextureBinding  binding;
-		uint32_t        kind  = 0;
-		uint32_t        refs  = 0; // Slots of the owning view that select it.
-		bool            stale = false;
+		uint32_t        kind   = 0;
+		uint32_t        refs   = 0; // Slots of the owning view that select it.
+		bool            stale  = false;
+		uint64_t        serial = 0; // Image::serial of binding.image_id.
 	};
 
 	struct Memory;
@@ -109,6 +110,7 @@ private:
 	uint32_t Intern(View& view, uint32_t kind, const DescriptorValue& value);
 	void     Release(View& view, uint32_t entry);
 	void     Unreference(Element& element);
+	Image*   ImageOf(const Element& element) const;
 	uint32_t Allocate();
 	void     Retire(View& view);
 	void     WriteDescriptors();

src/graphics/host_gpu/renderer/pipeline/pipelineCache.h

@@ -4,42 +4,57 @@
 #include "common/abi.h"
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/file.h"
+#include "common/uniqueFunction.h"
+#include "graphics/host_gpu/renderer/colorRenderTarget.h"
+#include "graphics/host_gpu/renderer/depthRenderTarget.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/rectListShader.h"
 #include "graphics/shader/shader.h"
 
 #include <chrono>
+#include <condition_variable>
 #include <cstddef>
 #include <filesystem>
 #include <memory>
+#include <mutex>
+#include <queue>
 #include <span>
+#include <thread>
 #include <type_traits>
 #include <unordered_map>
+#include <vector>
 
 namespace Libs::Graphics {
 
 struct GraphicContext;
-struct RenderColorInfo;
-struct RenderDepthInfo;
+struct ShaderParams;
 class CommandBuffer;
 
-namespace HW {
-class Context;
-class Shader;
-class UserConfig;
-struct ComputeShaderInfo;
-} // namespace HW
+namespace ShaderRecompiler {
+struct CompileOptions;
+struct ShaderSource;
+struct TranslateResult;
+} // namespace ShaderRecompiler
 
 namespace ShaderRecompiler::IR {
 class ImageHeapSource;
 } // namespace ShaderRecompiler::IR
 
+namespace ShaderRecompiler::Spirv {
+struct Profile;
+} // namespace ShaderRecompiler::Spirv
+
 #pragma pack(push, 1)
 
 struct PipelineStaticParameters {
 	bool                       negative_one_to_one      = false;
 	bool                       depth_clip_enable        = true;
 	vk::PrimitiveTopology      topology                 = vk::PrimitiveTopology::ePointList;
+	// 0 without tessellation, 3 for a rectangle list, else the input control points.
+	uint8_t                    patch_control_points     = 0;
 	bool                       primitive_restart_enable = false;
 	uint32_t                   samples                  = 1;
 	bool                       sample_shading_enable    = false;
@@ -58,57 +73,108 @@ struct PipelineStaticParameters {
 	uint8_t                    alpha_destblend[RENDER_COLOR_ATTACHMENTS_MAX]      = {};
 	bool                       separate_alpha_blend[RENDER_COLOR_ATTACHMENTS_MAX] = {};
 	bool                       blend_enable[RENDER_COLOR_ATTACHMENTS_MAX]         = {};
-
-	bool operator==(const PipelineStaticParameters& other) const noexcept;
 };
 
-#pragma pack(pop)
-
-static_assert(std::is_trivially_copyable_v<PipelineStaticParameters>);
-static_assert(std::is_standard_layout_v<PipelineStaticParameters>);
-static_assert(alignof(PipelineStaticParameters) == 1);
-static_assert(sizeof(PipelineStaticParameters) == 117);
-
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};
 	vk::Format                                           depth_format   = vk::Format::eUndefined;
 	vk::Format                                           stencil_format = vk::Format::eUndefined;
 	uint32_t                                             color_count    = 0;
-
-	bool operator==(const PipelineRenderingState&) const = default;
 };
 
 struct PipelineVertexInputState {
 	struct Binding {
-		uint32_t stride                           = 0;
-		bool     instance                         = false;
-		bool     operator==(const Binding&) const = default;
+		uint32_t stride   = 0;
+		uint8_t  instance = 0;
 	};
 	struct Attribute {
-		uint32_t offset                             = 0;
-		uint8_t  binding                            = 0;
-		bool     operator==(const Attribute&) const = default;
+		uint32_t offset  = 0;
+		uint8_t  binding = 0;
 	};
 
 	std::array<Binding, ShaderVertexInputInfo::RES_MAX>   bindings {};
 	std::array<Attribute, ShaderVertexInputInfo::RES_MAX> attributes {};
 	uint8_t                                               binding_count   = 0;
 	uint8_t                                               attribute_count = 0;
-
-	bool operator==(const PipelineVertexInputState&) const = default;
 };
 
-struct ShaderProgram {
-	uint64_t         id     = 0;
-	vk::ShaderModule module = nullptr;
+// Everything a graphics pipeline is made of besides the stage modules its ids name. Aligned so
+// that rendering.color_formats can be passed to Vulkan.
+struct alignas(4) GraphicsPipelineKey {
+	// VS, LS or mesh; HS; TES; PS. 0 marks an absent stage.
+	std::array<uint64_t, 4>  stage_ids {};
+	PipelineRenderingState   rendering;
+	PipelineVertexInputState vertex_input;
+	PipelineStaticParameters static_params;
 
-	explicit operator bool() const { return id != 0 && module != nullptr; }
+	bool operator==(const GraphicsPipelineKey& other) const;
+};
+
+#pragma pack(pop)
+
+// Keys are compared and hashed as bytes.
+static_assert(std::has_unique_object_representations_v<GraphicsPipelineKey>);
+
+// A compiled stage with what pipeline creation needs from it besides the key.
+struct ShaderModule {
+	uint64_t                                     id     = 0;
+	vk::ShaderModule                             module = nullptr;
+	ShaderRecompiler::IR::CompiledShaderInfo     info;
+	ShaderRecompiler::IR::ResourceSpecialization specialization;
+	std::vector<uint32_t>                        function_code;
+	// VS or LS: the host format of each vertex input attribute.
+	std::vector<vk::Format>                      vertex_formats;
+	// Rectangle-list PS: the parameters it reads.
+	std::vector<RectListParameter>               rect_parameters;
+};
+
+// A draw's topology, shader stages (from PipelineCache::GetGraphicsPrograms) and targets.
+struct DrawRenderState {
+	RenderDepthInfo                      depth_info;
+	RenderColorInfo                      color_info[RENDER_COLOR_ATTACHMENTS_MAX] = {};
+	uint32_t                             color_count                              = 0;
+	// Color targets the pixel program writes.
+	uint32_t                             mrt_mask                                 = 0;
+	vk::PrimitiveTopology                topology = vk::PrimitiveTopology::ePointList;
+	bool                                 primitive_restart_enable                 = false;
+	bool                                 ps_active                                = false;
+	std::array<ShaderVertexInputInfo, 3> vertex_info;
+	ShaderPixelInputInfo                 ps_input_info;
+
+	[[nodiscard]] std::span<const ShaderVertexInputInfo> VertexStages() const {
+		return {vertex_info.data(), vertex_info[1].stage ? 3u : 1u};
+	}
+};
+
+// Writes the shader store on its own thread, running jobs in the order they were posted.
+class ShaderStoreWriter {
+public:
+	// Keeps the first `kept` bytes of the store at `path`, or starts it over with `header`. The
+	// thread runs only if that succeeded.
+	void Open(const std::filesystem::path& path, std::span<const uint8_t> header, uint64_t kept);
+	explicit operator bool() const { return m_thread.joinable(); }
+	// Compresses `record` and appends it on the store thread, in call order.
+	void Append(std::vector<uint8_t>&& record);
+
+private:
+	friend struct PipelineCacheTestAccess;
+
+	void Post(Common::UniqueFunction<void>&& job);
+
+	Common::File                             m_file;
+	std::mutex                               m_mutex;
+	std::condition_variable_any              m_ready;
+	std::queue<Common::UniqueFunction<void>> m_jobs;
+	// Declared last, so it drains the queue and joins before the rest is destroyed.
+	std::jthread m_thread;
 };
 
 // The owning renderer serializes access, including saves while the GPU is running.
 class PipelineCache {
 public:
-	PipelineCache(GraphicContext& graphics, ShaderRecompiler::IR::ImageHeapSource* image_heap);
+	// Without a shader store path nothing is kept across runs.
+	PipelineCache(GraphicContext& graphics, ShaderRecompiler::IR::ImageHeapSource* image_heap,
+	              const std::filesystem::path& shader_store = {});
 	~PipelineCache();
 	KYTY_CLASS_NO_COPY(PipelineCache);
 	void Save();
@@ -121,87 +187,106 @@ public:
 		bool                    uses_image_heap       = false;
 	};
 
-	struct GraphicsPrograms {
-		std::array<ShaderProgram, 3> vertex;
-		ShaderProgram pixel;
+	// Reads the shaders and registers bound in `command` into the state's shader stages.
+	void GetGraphicsPrograms(const CommandBuffer& command, DrawRenderState& state);
+	// Null skips the dispatch.
+	const ShaderModule* GetComputeProgram(const HW::ComputeShaderInfo& regs,
+	                                      ShaderComputeInputInfo&      input_info);
 
-		[[nodiscard]] uint32_t VertexStageCount() const { return vertex[1] ? 3u : 1u; }
-	};
-
-	GraphicsPrograms
-	GetGraphicsPrograms(const HW::VertexShaderInfo& vertex_regs,
-	                    const HW::PixelShaderInfo& pixel_regs, const HW::ShaderRegisters& sh,
-	                    const HW::Context& context, const HW::UserConfig& user_config,
-	                    std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
-	                    bool pixel_active, std::array<ShaderVertexInputInfo, 3>& vertex_info,
-	                    ShaderPixelInputInfo& pixel_info);
-	ShaderProgram GetComputeProgram(const HW::ComputeShaderInfo& regs,
-	                                const HW::ShaderRegisters&   sh,
-	                                ShaderComputeInputInfo&      input_info);
-
-	Pipeline& GetGraphicsPipeline(std::span<const RenderColorInfo>       colors,
-	                              const RenderDepthInfo&                 depth,
-	                              std::span<const ShaderVertexInputInfo> vertex_info,
-	                              CommandBuffer& command, const ShaderPixelInputInfo* ps_input_info,
-	                              vk::PrimitiveTopology topology, bool primitive_restart_enable,
-	                              const GraphicsPrograms& programs);
-	Pipeline& GetComputePipeline(const ShaderComputeInputInfo& input_info,
-	                             const ShaderProgram&          compute_program);
+	Pipeline& GetGraphicsPipeline(const CommandBuffer& command, const DrawRenderState& state);
+	Pipeline& GetComputePipeline(const ShaderComputeInputInfo& input_info);
 
 private:
-	struct ProgramCache;
+	friend struct PipelineCacheTestAccess;
 
-	struct GraphicsPipelineKey {
-		PipelineRenderingState   rendering;
-		std::array<uint64_t, 3>  vertex_shader_ids {};
-		uint64_t                 ps_shader_id = 0;
-		PipelineVertexInputState vertex_input;
-		PipelineStaticParameters static_params;
+	struct Program {
+		std::unique_ptr<ShaderRecompiler::ShaderSource> call_source;
+		ShaderRecompiler::IR::ResourcePlan           resource_plan;
+		ShaderRecompiler::IR::ResourceSnapshot       resources;
+		ShaderRecompiler::IR::ResourceSpecialization specialization;
+		std::vector<ShaderModule>                    modules;
+		bool                                         skip_dispatch = false;
 
-		bool operator==(const GraphicsPipelineKey& other) const {
-			return rendering == other.rendering && vertex_shader_ids == other.vertex_shader_ids &&
-			       ps_shader_id == other.ps_shader_id && vertex_input == other.vertex_input &&
-			       static_params == other.static_params;
-		}
+		[[nodiscard]] std::span<const uint32_t> FunctionCode() const;
+		bool RefreshCallSource(const ShaderRecompiler::IR::SrtRuntime& runtime);
+		template <bool HasCalls>
+		ShaderModule* FindModule(const ShaderRecompiler::IR::SrtRuntime& runtime,
+		                         uint32_t                                push_data_cursor);
 	};
 
-	struct PipelineKeyHash {
-		static void Mix(std::size_t& hash, std::size_t value) {
-			hash ^= value + static_cast<std::size_t>(0x9e3779b97f4a7c15ull) + (hash << 6u) +
-			        (hash >> 2u);
-		}
+	struct ProgramKey {
+		ShaderType            stage           = ShaderType::Unknown;
+		uint64_t              hash            = 0;
+		uint32_t              user_data_count = 0;
+		std::vector<uint32_t> static_state;
+
+		bool operator==(const ProgramKey&) const = default;
+	};
+
+	// Variants of one source share a bucket; ProgramKey equality compares the state.
+	struct ProgramKeyHash {
+		std::size_t operator()(const ProgramKey& key) const { return key.hash; }
 	};
 
 	struct GraphicsPipelineKeyHash {
 		std::size_t operator()(const GraphicsPipelineKey& key) const;
 	};
 
-	GraphicContext&               m_graphics;
-	std::unique_ptr<ProgramCache> m_program_cache;
+	using ModuleMap = std::unordered_map<uint64_t, const ShaderModule*>;
+
+	GraphicContext&                        m_graphics;
+	ShaderRecompiler::IR::ImageHeapSource* m_image_heap;
+	std::unordered_map<ProgramKey, Program, ProgramKeyHash> m_program_cache;
+	// Modules compiled per stage.
+	std::array<uint32_t, static_cast<size_t>(ShaderType::TessellationEvaluation) + 1> m_compiled {};
+	ProgramKey                             m_lookup_key;
 	vk::PipelineCache             m_driver_cache = nullptr;
 	std::filesystem::path         m_driver_cache_path;
 	std::chrono::steady_clock::time_point m_next_driver_cache_save;
 	bool                          m_driver_cache_dirty = false;
-	std::unordered_map<GraphicsPipelineKey, std::unique_ptr<Pipeline>, GraphicsPipelineKeyHash>
-	                                                        m_graphics_pipelines;
-	std::unordered_map<uint64_t, std::unique_ptr<Pipeline>> m_compute_pipelines;
+	std::unordered_map<GraphicsPipelineKey, Pipeline, GraphicsPipelineKeyHash> m_graphics_pipelines;
+	std::unordered_map<uint64_t, Pipeline>                                     m_compute_pipelines;
+	std::unordered_map<uint64_t, ShaderTessellationInputInfo>                  m_tess_layouts;
+	// Records everything first created; closed without a shader store.
+	ShaderStoreWriter m_store;
 
+	template <typename InputInfo>
+	const ShaderModule* GetProgram(const ShaderParams& params, InputInfo& input_info,
+	                               uint32_t& push_data_cursor);
+	ShaderModule& CompileModule(Program& program, const ProgramKey& key,
+	                            ShaderRecompiler::CompileOptions  options,
+	                            ShaderRecompiler::TranslateResult translated,
+	                            uint32_t                          push_data_cursor);
 	void InitializeDriverCache();
 	void CheckpointDriverCache();
+
+	// pipelineSerialization.cpp
+	void WarmUp(const std::filesystem::path& path);
+	bool LoadRecord(std::span<const uint8_t>& records, std::vector<GraphicsPipelineKey>& pipelines);
+	void StoreModule(const ProgramKey& key, const Program& program, const ShaderModule& module,
+	                 const std::vector<uint32_t>& spirv);
+	void StorePipeline(const GraphicsPipelineKey& key);
+	void StoreLayout(uint64_t key, const ShaderTessellationInputInfo& layout);
+	// Every module by id, with id 0 naming an absent stage.
+	[[nodiscard]] ModuleMap MapModules() const;
+	// False when a stage id of the key is not in `modules`.
+	bool CreateGraphicsPipeline(Pipeline& pipeline, const GraphicsPipelineKey& key,
+	                            const ModuleMap& modules);
 };
 
-void LogPipelineTrace(const char* phase, uint64_t vertex_program_id, uint64_t pixel_program_id);
+// The shader store of the running title; empty unless --shader-cache is on in a Release build
+// without debug dumps, which change the SPIR-V.
+std::filesystem::path ShaderStorePath();
+// Host capabilities that change the emitted SPIR-V.
+ShaderRecompiler::Spirv::Profile SpirvProfile(const GraphicContext& graphics);
+
+// Stages are indexed like GraphicsPipelineKey::stage_ids.
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const PipelineRenderingState&          rendering,
-                            const PipelineVertexInputState&        vertex_input,
-                            std::span<const ShaderVertexInputInfo> vertex_info,
-                            const ShaderPixelInputInfo*            ps_input_info,
-                            const PipelineCache::GraphicsPrograms& programs,
-                            const PipelineStaticParameters&        static_params,
-                            vk::PipelineCache                      driver_cache);
+                            const GraphicsPipelineKey&                 key,
+                            const std::array<const ShaderModule*, 4>& stages,
+                            vk::PipelineCache                          driver_cache);
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const ShaderComputeInputInfo& input_info,
-                            vk::ShaderModule compute_module, vk::PipelineCache driver_cache);
+                            const ShaderModule& module, vk::PipelineCache driver_cache);
 
 } // namespace Libs::Graphics
 

src/graphics/host_gpu/renderer/pipeline/pipelineSerialization.cpp

@@ -0,0 +1,329 @@
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
+
+#include "common/logging/log.h"
+#include "common/stringUtils.h"
+#include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/shader/recompiler/ShaderRecompiler.h"
+#include "graphics/shader/recompiler/ir/Serialize.h"
+
+#include <algorithm>
+#include <chrono>
+#include <cstring>
+#include <fmt/format.h>
+#include <ranges>
+#include <xxhash.h>
+#include <zstd.h>
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#elif defined(__APPLE__)
+#include <mach-o/dyld.h>
+#endif
+
+// The shader store: a header, then zstd-compressed records of everything a run created first.
+// Records are only appended, so a store cut short by a crash keeps every complete record.
+namespace Libs::Graphics {
+
+namespace {
+
+using ShaderRecompiler::IR::IoAll;
+using ShaderRecompiler::IR::Reader;
+using ShaderRecompiler::IR::Writer;
+
+enum class Record : uint8_t { Module, GraphicsPipeline, TessellationLayout };
+
+// Precedes each record's zstd frame.
+struct Frame {
+	uint32_t compressed_size = 0;
+	uint32_t raw_size        = 0;
+	uint64_t checksum        = 0; // XXH3 of the compressed bytes.
+};
+
+// Changes with every build, so records never meet code that encodes them differently. Zero when
+// the executable cannot be found.
+uint64_t BuildStamp() {
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+	std::array<wchar_t, 32768> name {};
+	GetModuleFileNameW(nullptr, name.data(), static_cast<DWORD>(name.size()));
+	const std::filesystem::path path = name.data();
+#elif defined(__APPLE__)
+	std::array<char, 4096> name {};
+	auto size = static_cast<uint32_t>(name.size());
+	_NSGetExecutablePath(name.data(), &size);
+	const std::filesystem::path path = name.data();
+#else
+	const std::filesystem::path path = "/proc/self/exe";
+#endif
+	std::error_code error;
+	const auto      time = std::filesystem::last_write_time(path, error).time_since_epoch().count();
+	return error ? 0 : static_cast<uint64_t>(time);
+}
+
+// What every record depends on besides its own fields; empty when the build is unknown.
+std::vector<uint8_t> StoreHeader(const GraphicContext& graphics) {
+	const auto stamp = BuildStamp();
+	if (stamp == 0) {
+		return {};
+	}
+	Writer header;
+	IoAll(header, stamp, SpirvProfile(graphics),
+	      graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize);
+	return std::move(header.bytes);
+}
+
+// Record fields, shared by writing and reading.
+template <class S, class Key> void IoKey(S& io, Key& key) {
+	auto& [stage, hash, user_data_count, static_state] = key;
+	IoAll(io, stage, hash, user_data_count, static_state);
+}
+
+template <class S, class Module> void IoModule(S& io, Module& stage) {
+	// The handle comes from the SPIR-V; stored programs have no function code.
+	auto& [id, handle, info, specialization, function_code, vertex_formats, rect_parameters] = stage;
+	IoAll(io, id, info, specialization, vertex_formats, rect_parameters);
+}
+
+} // namespace
+
+void ShaderStoreWriter::Open(const std::filesystem::path& path, std::span<const uint8_t> header,
+                             uint64_t kept) {
+	Common::File::CreateDirectories(path.parent_path());
+	if (kept == 0) {
+		if (m_file.Create(path)) {
+			m_file.Write(header.data(), static_cast<uint32_t>(header.size()));
+		}
+	} else if (m_file.Open(path, Common::File::Mode::ReadWrite)) {
+		m_file.Truncate(kept);
+		m_file.Seek(kept);
+	}
+	if (m_file.IsInvalid()) {
+		return;
+	}
+	m_thread = std::jthread([this](std::stop_token stop) {
+		std::unique_lock lock(m_mutex);
+		// Once stop is requested the wait returns false only when the queue is empty.
+		while (m_ready.wait(lock, stop, [this] { return !m_jobs.empty(); })) {
+			const auto job = std::move(m_jobs.front());
+			m_jobs.pop();
+			lock.unlock();
+			job();
+			lock.lock();
+		}
+	});
+}
+
+void ShaderStoreWriter::Append(std::vector<uint8_t>&& record) {
+	Post([this, record = std::move(record)] {
+		std::vector<uint8_t> bytes(sizeof(Frame) + ZSTD_compressBound(record.size()));
+		const auto size = ZSTD_compress(bytes.data() + sizeof(Frame), bytes.size() - sizeof(Frame),
+		                                record.data(), record.size(), 3);
+		EXIT_IF(ZSTD_isError(size));
+		const Frame frame {.compressed_size = static_cast<uint32_t>(size),
+		                   .raw_size        = static_cast<uint32_t>(record.size()),
+		                   .checksum        = XXH3_64bits(bytes.data() + sizeof(Frame), size)};
+		std::memcpy(bytes.data(), &frame, sizeof(frame));
+		// A failed write leaves a partial record; the next run drops it and everything after it.
+		m_file.Write(bytes.data(), static_cast<uint32_t>(sizeof(Frame) + size));
+	});
+}
+
+void ShaderStoreWriter::Post(Common::UniqueFunction<void>&& job) {
+	{
+		std::scoped_lock lock(m_mutex);
+		m_jobs.push(std::move(job));
+	}
+	m_ready.notify_one();
+}
+
+void PipelineCache::WarmUp(const std::filesystem::path& path) {
+	if (path.empty()) {
+		return;
+	}
+	const auto header = StoreHeader(m_graphics);
+	if (header.empty()) {
+		return;
+	}
+	const auto           start = std::chrono::steady_clock::now();
+	std::vector<uint8_t> bytes;
+	if (Common::File file;
+	    Common::File::IsFileExisting(path) && file.Open(path, Common::File::Mode::Read)) {
+		bytes.resize(file.Size());
+		uint32_t read = 0;
+		file.Read(bytes.data(), static_cast<uint32_t>(bytes.size()), &read);
+		bytes.resize(read);
+	}
+	Log::WriteToConsoleAndLog(
+	    fmt::format("Shader cache: compiling shaders and pipelines from {} ({:.1f} MB)\n",
+	                Common::PathToString(path), static_cast<double>(bytes.size()) / 1'000'000.0));
+	std::vector<GraphicsPipelineKey> pipelines;
+	auto                             records = std::span<const uint8_t>(bytes);
+	const bool reusable =
+	    records.size() >= header.size() && std::ranges::equal(records.first(header.size()), header);
+	if (reusable) {
+		records = records.subspan(header.size());
+		while (LoadRecord(records, pipelines)) {
+		}
+	}
+	// A damaged record ends the store; the records before it stay.
+	m_store.Open(path, header, reusable ? bytes.size() - records.size() : 0);
+
+	const auto modules = MapModules();
+	for (const auto& key: pipelines) {
+		Pipeline pipeline;
+		if (CreateGraphicsPipeline(pipeline, key, modules)) {
+			m_graphics_pipelines.emplace(key, pipeline);
+		}
+	}
+	// A compute pipeline depends only on its module.
+	for (const auto* module: modules | std::views::values) {
+		if (module != nullptr && module->info.stage == ShaderType::Compute) {
+			CreatePipelineInternal(m_graphics, m_compute_pipelines[module->id], *module,
+			                       m_driver_cache);
+		}
+	}
+	if (!m_graphics_pipelines.empty() || !m_compute_pipelines.empty()) {
+		CheckpointDriverCache();
+	}
+	Log::WriteToConsoleAndLog(fmt::format(
+	    "Shader cache: loaded {} programs, {} modules, {} pipelines in {} ms\n",
+	    m_program_cache.size(), modules.size() - 1,
+	    m_graphics_pipelines.size() + m_compute_pipelines.size(),
+	    std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() -
+	                                                          start)
+	        .count()));
+}
+
+bool PipelineCache::LoadRecord(std::span<const uint8_t>&        records,
+                               std::vector<GraphicsPipelineKey>& pipelines) {
+	Frame frame;
+	if (records.size() < sizeof(frame)) {
+		return false;
+	}
+	std::memcpy(&frame, records.data(), sizeof(frame));
+	if (records.size() - sizeof(frame) < frame.compressed_size) {
+		return false;
+	}
+	const auto compressed = records.subspan(sizeof(frame), frame.compressed_size);
+	if (XXH3_64bits(compressed.data(), compressed.size()) != frame.checksum ||
+	    ZSTD_getFrameContentSize(compressed.data(), compressed.size()) != frame.raw_size) {
+		return false;
+	}
+	std::vector<uint8_t> raw(frame.raw_size);
+	if (ZSTD_decompress(raw.data(), raw.size(), compressed.data(), compressed.size()) != raw.size()) {
+		return false;
+	}
+	Reader     reader {raw};
+	auto       type     = Record::Module;
+	const auto complete = [&] { return reader.ok && reader.bytes.empty(); };
+	IoAll(reader, type);
+	switch (type) {
+		case Record::Module: {
+			ProgramKey            key;
+			std::vector<uint8_t>  plan;
+			ShaderModule          module;
+			std::vector<uint32_t> spirv;
+			IoKey(reader, key);
+			IoAll(reader, plan);
+			IoModule(reader, module);
+			IoAll(reader, spirv);
+			if (!complete()) {
+				return false;
+			}
+			const auto [program, inserted] = m_program_cache.try_emplace(std::move(key));
+			if (inserted &&
+			    !ShaderRecompiler::IR::LoadResourcePlan(plan, program->second.resource_plan)) {
+				m_program_cache.erase(program);
+				return false;
+			}
+			module.module = CompileSPV(spirv, m_graphics.device);
+			if (module.module == nullptr) {
+				return false;
+			}
+			program->second.modules.push_back(std::move(module));
+			break;
+		}
+		case Record::GraphicsPipeline: {
+			GraphicsPipelineKey key;
+			reader.Raw(&key, sizeof(key));
+			if (!complete()) {
+				return false;
+			}
+			pipelines.push_back(key);
+			break;
+		}
+		case Record::TessellationLayout: {
+			uint64_t                    key = 0;
+			ShaderTessellationInputInfo layout;
+			IoAll(reader, key, layout);
+			if (!complete()) {
+				return false;
+			}
+			m_tess_layouts.try_emplace(key, layout);
+			break;
+		}
+		default: return false;
+	}
+	records = records.subspan(sizeof(frame) + compressed.size());
+	return true;
+}
+
+// Records are encoded here, since programs and modules change later.
+void PipelineCache::StoreModule(const ProgramKey& key, const Program& program,
+                                const ShaderModule& module, const std::vector<uint32_t>& spirv) {
+	if (!m_store) {
+		return;
+	}
+	Writer record;
+	IoAll(record, Record::Module);
+	IoKey(record, key);
+	IoAll(record, ShaderRecompiler::IR::SaveResourcePlan(program.resource_plan));
+	IoModule(record, module);
+	IoAll(record, spirv);
+	m_store.Append(std::move(record.bytes));
+}
+
+void PipelineCache::StorePipeline(const GraphicsPipelineKey& key) {
+	if (!m_store) {
+		return;
+	}
+	Writer record;
+	IoAll(record, Record::GraphicsPipeline);
+	record.Raw(&key, sizeof(key));
+	m_store.Append(std::move(record.bytes));
+}
+
+void PipelineCache::StoreLayout(uint64_t key, const ShaderTessellationInputInfo& layout) {
+	if (!m_store) {
+		return;
+	}
+	Writer record;
+	IoAll(record, Record::TessellationLayout, key, layout);
+	m_store.Append(std::move(record.bytes));
+}
+
+PipelineCache::ModuleMap PipelineCache::MapModules() const {
+	ModuleMap modules {{0, nullptr}};
+	for (const auto& program: m_program_cache | std::views::values) {
+		for (const auto& module: program.modules) {
+			modules.emplace(module.id, &module);
+		}
+	}
+	return modules;
+}
+
+bool PipelineCache::CreateGraphicsPipeline(Pipeline& pipeline, const GraphicsPipelineKey& key,
+                                           const ModuleMap& modules) {
+	std::array<const ShaderModule*, 4> stages {};
+	for (size_t i = 0; i < stages.size(); i++) {
+		const auto found = modules.find(key.stage_ids[i]);
+		if (found == modules.end()) {
+			return false;
+		}
+		stages[i] = found->second;
+	}
+	CreatePipelineInternal(m_graphics, pipeline, key, stages, m_driver_cache);
+	return true;
+}
+
+} // namespace Libs::Graphics

src/graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.cpp

@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 
 #include "common/assert.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "graphics/shader/shader.h"
 #include "graphics/shader/shaderBindings.h"
 
@@ -90,7 +91,7 @@ vk::BufferMemoryBarrier MakeGdsDependency(vk::Buffer buffer) {
 
 bool HasShaderBufferWrites(const ShaderStageRuntime& runtime) {
 	EXIT_IF(!runtime);
-	const auto& program   = *runtime.program;
+	const auto& program   = runtime.module->info;
 	const auto& resources = *runtime.resources;
 	EXIT_IF(resources.buffers.size() != program.info.buffers.size());
 	// Address writes reach memory outside the bound buffer descriptors.

src/graphics/host_gpu/renderer/pipeline/shaders.cpp

@@ -13,7 +13,6 @@
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
-#include "graphics/shader/recompiler/BufferFormat.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/rectListShader.h"
 #include "graphics/shader/shader.h"
@@ -21,124 +20,11 @@
 #include <algorithm>
 #include <array>
 #include <bit>
-#include <limits>
 #include <span>
 #include <vector>
 
 namespace Libs::Graphics {
 
-// IDK: maybe we can remove it?
-constexpr uint8_t kTemporaryVertexAttribFormat113 =
-    static_cast<uint8_t>(Prospero::VertexAttribFormat::k16_16SInt);
-constexpr uint32_t kTemporaryPs5BufferFormat121 = 121u;
-
-static bool NarrowInputFormat(vk::Format& format, uint32_t& size, uint32_t used_components) {
-	if (used_components == 0 || used_components >= size) {
-		return false;
-	}
-
-	switch (format) {
-		case vk::Format::eR32G32B32A32Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR32Sfloat; break;
-				case 2: format = vk::Format::eR32G32Sfloat; break;
-				case 3: format = vk::Format::eR32G32B32Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR32G32B32Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR32Sfloat; break;
-				case 2: format = vk::Format::eR32G32Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR16G16B16A16Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR16Sfloat; break;
-				case 2: format = vk::Format::eR16G16Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR8G8B8A8Unorm:
-			switch (used_components) {
-				case 1: format = vk::Format::eR8Unorm; break;
-				case 2: format = vk::Format::eR8G8Unorm; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR8G8B8A8Snorm:
-			if (used_components != 2) {
-				return false;
-			}
-			format = vk::Format::eR8G8Snorm;
-			size   = 2;
-			return true;
-		case vk::Format::eR8G8B8A8Uint:
-			switch (used_components) {
-				case 1: format = vk::Format::eR8Uint; break;
-				case 2: format = vk::Format::eR8G8Uint; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		default: break;
-	}
-
-	return false;
-}
-
-static void GetInputFormat(const ShaderBufferResource& res, vk::Format& format, uint32_t& size,
-                           uint32_t used_components) {
-	const auto fmt        = res.Format();
-	const auto raw_format = res.RawFormat();
-	if (raw_format == kTemporaryVertexAttribFormat113) {
-		static bool logged_113 = false;
-		if (!logged_113) {
-			LOGF("InputFormat: temporary: accepting invalid PS5 buffer format 113 as "
-			     "vk::Format::eR32G32B32A32Sfloat\n");
-			logged_113 = true;
-		}
-		format = vk::Format::eR32G32B32A32Sfloat;
-		size   = 4;
-		if (NarrowInputFormat(format, size, used_components)) {
-			LOGF("InputFormat: narrowing fmt=%u to %s for used_components=%u\n", raw_format,
-			     vk::to_string(format).c_str(), used_components);
-		}
-		return;
-	}
-	if (raw_format == kTemporaryPs5BufferFormat121) {
-		static bool logged_121 = false;
-		if (!logged_121) {
-			LOGF("InputFormat: accepting PS5 buffer format 121 as vk::Format::eR16G16Sfloat\n");
-			logged_121 = true;
-		}
-		format = vk::Format::eR16G16Sfloat;
-		size   = 2;
-		return;
-	}
-
-	format = VulkanFormat(fmt);
-	size   = ShaderRecompiler::Format::GetFormatInfo(fmt).component_count;
-	if (format == vk::Format::eUndefined || size == 0) {
-		EXIT("unknown vertex format: fmt = %u\n", raw_format);
-	}
-
-	if (NarrowInputFormat(format, size, used_components)) {
-		static std::atomic<uint64_t> log_count = 0;
-		auto                         log_id    = log_count.fetch_add(1);
-		if (log_id < 32) {
-			LOGF("VertexInput: narrowed vertex format to %" PRIu32
-			     " component(s) for shader fetch\n",
-			     used_components);
-		}
-	}
-}
-
 static vk::BlendFactor GetBlendFactor(uint32_t factor) {
 	switch (static_cast<Prospero::BlendFactor>(factor)) {
 		case Prospero::BlendFactor::kZero: return vk::BlendFactor::eZero;
@@ -242,33 +128,27 @@ static void CreateDescriptorLayout(GraphicContext& graphics, PipelineCache::Pipe
 
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const PipelineRenderingState&          rendering,
-                            const PipelineVertexInputState&        vertex_input,
-                            std::span<const ShaderVertexInputInfo> vertex_info,
-                            const ShaderPixelInputInfo*            ps_input_info,
-                            const PipelineCache::GraphicsPrograms& programs,
-                            const PipelineStaticParameters&        static_params,
-                            vk::PipelineCache                      driver_cache) {
-	const auto& vs_input_info  = vertex_info.front();
-	const auto& vertex_program = programs.vertex[0];
-	const auto& pixel_program  = programs.pixel;
-	const bool  tessellation   = vertex_info.size() == 3;
-	const bool ps_active = ps_input_info != nullptr;
-	EXIT_IF(!vertex_program || (ps_active && !pixel_program));
-	const bool with_depth = rendering.depth_format != vk::Format::eUndefined ||
-	                        rendering.stencil_format != vk::Format::eUndefined;
-	EXIT_IF(!vs_input_info.stage);
-	const bool mesh = vs_input_info.stage.program->stage == ShaderType::Mesh;
+                            const GraphicsPipelineKey&                 key,
+                            const std::array<const ShaderModule*, 4>& stages,
+                            vk::PipelineCache                          driver_cache) {
+	const auto& rendering     = key.rendering;
+	const auto& vertex_input  = key.vertex_input;
+	const auto& static_params = key.static_params;
+	const auto& vertex        = *stages[0];
+	const auto* pixel         = stages[3];
+	const bool  with_depth    = rendering.depth_format != vk::Format::eUndefined ||
+	                            rendering.stencil_format != vk::Format::eUndefined;
+	const bool  mesh          = vertex.info.stage == ShaderType::Mesh;
 	EXIT_NOT_IMPLEMENTED(mesh && !graphics.mesh_shader_enabled);
-	const bool rect_list =
-	    !mesh && !tessellation && static_params.topology == vk::PrimitiveTopology::ePatchList;
+	const bool rect_list = static_params.patch_control_points != 0 && stages[1] == nullptr;
 
 	vk::ShaderModule tess_control_shader_module = nullptr;
 	vk::ShaderModule tess_eval_shader_module    = nullptr;
 
 	if (rect_list) {
-		const auto shaders =
-		    BuildRectListShaders(vs_input_info, ps_active ? ps_input_info : nullptr);
+		std::span<const RectListParameter> parameters;
+		if (pixel != nullptr) parameters = pixel->rect_parameters;
+		const auto shaders = BuildRectListShaders(parameters, vertex.info.param_export_mask);
 		tess_control_shader_module = CompileSPV(shaders.control, graphics.device);
 		if (graphics_debug_dump_enabled()) {
 			LOGF("PipelineTrace: vkCreateShaderModule RectList TCS done module=%p\n",
@@ -287,11 +167,12 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	vk::PipelineShaderStageCreateInfo shader_stages[4] {};
 	uint32_t                          shader_stage_count = 0;
-	for (uint32_t i = 0; i < vertex_info.size(); i++) {
-		shader_stages[shader_stage_count++] = {.stage =
-		                                           NativeShaderStage(vertex_info[i].logical_stage),
-		                                       .module = programs.vertex[i].module,
-		                                       .pName  = "main"};
+	for (const auto* stage: stages) {
+		if (stage != nullptr) {
+			shader_stages[shader_stage_count++] = {.stage  = NativeShaderStage(stage->info.stage),
+			                                       .module = stage->module,
+			                                       .pName  = "main"};
+		}
 	}
 	if (rect_list) {
 		shader_stages[shader_stage_count++] = {.stage =
@@ -303,11 +184,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		                                       .module = tess_eval_shader_module,
 		                                       .pName  = "main"};
 	}
-	if (ps_active) {
-		shader_stages[shader_stage_count++] = {.stage  = vk::ShaderStageFlagBits::eFragment,
-		                                       .module = pixel_program.module,
-		                                       .pName  = "main"};
-	}
 
 	vk::VertexInputAttributeDescription input_attr[ShaderVertexInputInfo::RES_MAX] {};
 	vk::VertexInputBindingDescription   input_desc[ShaderVertexInputInfo::RES_MAX] {};
@@ -323,46 +199,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		input_attr[index].binding  = vertex_input.attributes[index].binding;
 		input_attr[index].location = index;
 		input_attr[index].offset   = vertex_input.attributes[index].offset;
-
-		uint32_t   attr_size     = 4;
-		const auto registers_num = vs_input_info.resources_dst[index].registers_num;
-		const auto compiled_components =
-		    vs_input_info.stage.program->info.vertex_fetch_components[index];
-		const auto used_components =
-		    compiled_components > 0 ? static_cast<int>(compiled_components) : registers_num;
-		GetInputFormat(vs_input_info.resources[index], input_attr[index].format, attr_size,
-		               static_cast<uint32_t>(used_components));
-
-		if (graphics_debug_dump_enabled()) {
-			static std::atomic_uint log_count = 0;
-			const auto              log_id    = log_count.fetch_add(1, std::memory_order_relaxed);
-			if (log_id < 128) {
-				LOGF("VertexInputState[%u]: attr=%u binding=%u offset=%u stride=%u fmt=%d "
-				     "src_fmt=%u dst=v%u regs=%u"
-				     " fetched_components=%u attr_size=%u swizzle=%u,%u,%u,%u\n",
-				     log_id, index, input_attr[index].binding, input_attr[index].offset,
-				     input_desc[input_attr[index].binding].stride,
-				     static_cast<int>(input_attr[index].format),
-				     static_cast<uint32_t>(vs_input_info.resources[index].Format()),
-				     static_cast<uint32_t>(vs_input_info.resources_dst[index].register_start),
-				     static_cast<uint32_t>(registers_num), static_cast<uint32_t>(used_components),
-				     attr_size, static_cast<uint32_t>(vs_input_info.resources[index].DstSelX()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelY()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelZ()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelW()));
-			}
-		}
-
-		if (vs_input_info.resources[index].OutOfBounds() != 0) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("VertexInput: temporary: accepting PS5 out-of-bounds behavior %" PRIu8 "\n",
-				     vs_input_info.resources[index].OutOfBounds());
-				logged = true;
-			}
-		}
-
-		EXIT_IF(registers_num < 1 || registers_num > 4);
+		input_attr[index].format   = vertex.vertex_formats[index];
 	}
 
 	vk::PipelineVertexInputStateCreateInfo vertex_input_info {};
@@ -470,17 +307,13 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	std::vector<vk::DescriptorSetLayoutBinding> descriptor_bindings;
 	vk::ShaderStageFlags graphics_stages = vk::ShaderStageFlagBits::eFragment;
-	for (const auto& stage: vertex_info) {
-		const auto native_stage = NativeShaderStage(stage.logical_stage);
-		AddLayoutBindings(descriptor_bindings, *stage.stage.program, native_stage);
-		pipeline.uses_image_heap |= stage.stage.program->info.uses_image_heap;
-		graphics_stages |= native_stage;
-	}
-	if (ps_active) {
-		EXIT_IF(!ps_input_info->stage);
-		AddLayoutBindings(descriptor_bindings, *ps_input_info->stage.program,
-		                  vk::ShaderStageFlagBits::eFragment);
-		pipeline.uses_image_heap |= ps_input_info->stage.program->info.uses_image_heap;
+	for (const auto* stage: stages) {
+		if (stage != nullptr) {
+			const auto native_stage = NativeShaderStage(stage->info.stage);
+			AddLayoutBindings(descriptor_bindings, stage->info, native_stage);
+			pipeline.uses_image_heap |= stage->info.info.uses_image_heap;
+			graphics_stages |= native_stage;
+		}
 	}
 	CreateDescriptorLayout(graphics, pipeline, descriptor_bindings);
 	const vk::PushConstantRange push_constants {graphics_stages, 0,
@@ -498,7 +331,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	if (graphics_debug_dump_enabled()) {
 		LOGF("PipelineTrace: vkCreatePipelineLayout begin VS=%" PRIu64 " PS=%" PRIu64
 		     " set_layouts=1 push_constants=%" PRIu32 "\n",
-		     vertex_program.id, ps_active ? pixel_program.id : 0, 1u);
+		     key.stage_ids[0], key.stage_ids[3], 1u);
 	}
 	auto result = graphics.device.createPipelineLayout(&pipeline_layout_info, nullptr,
 	                                                   &pipeline.pipeline_layout);
@@ -508,8 +341,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
-
 	vk::PipelineDepthStencilStateCreateInfo depth_stencil_info {};
 
 	std::vector<vk::DynamicState> dynamic_states {
@@ -555,9 +386,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	pipeline_info.pVertexInputState        = mesh ? nullptr : &vertex_input_info;
 	pipeline_info.pInputAssemblyState      = mesh ? nullptr : &input_assembly;
 	vk::PipelineTessellationStateCreateInfo tessellation_state {};
-	tessellation_state.patchControlPoints =
-	    tessellation ? vs_input_info.tess.input_control_points : 3u;
-	pipeline_info.pTessellationState = (rect_list || tessellation) ? &tessellation_state : nullptr;
+	tessellation_state.patchControlPoints = static_params.patch_control_points;
+	pipeline_info.pTessellationState =
+	    static_params.patch_control_points != 0 ? &tessellation_state : nullptr;
 	pipeline_info.pViewportState          = &viewport_state;
 	pipeline_info.pRasterizationState     = &rasterizer;
 	pipeline_info.pMultisampleState       = &multisampling;
@@ -573,7 +404,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		LOGF("PipelineTrace: vkCreateGraphicsPipelines begin VS=%" PRIu64 " PS=%" PRIu64
 		     " topology=%" PRIu32 " color_mask=0x%08" PRIx32
 		     " depth=%s blend=%s dyn_states=%" PRIu32 "\n",
-		     vertex_program.id, ps_active ? pixel_program.id : 0,
+		     key.stage_ids[0], key.stage_ids[3],
 		     static_cast<uint32_t>(static_params.topology), static_params.color_mask[0],
 		     (with_depth ? "true" : "false"), (static_params.blend_enable[0] ? "true" : "false"),
 		     dynamic_state.dynamicStateCount);
@@ -586,29 +417,18 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline == nullptr);
-
-	if (tess_control_shader_module != nullptr) {
-		graphics.device.destroyShaderModule(tess_control_shader_module, nullptr);
-	}
-	if (tess_eval_shader_module != nullptr) {
-		graphics.device.destroyShaderModule(tess_eval_shader_module, nullptr);
-	}
+	graphics.device.destroyShaderModule(tess_control_shader_module, nullptr);
+	graphics.device.destroyShaderModule(tess_eval_shader_module, nullptr);
 }
 
-// NOLINTNEXTLINE(readability-function-cognitive-complexity)
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const ShaderComputeInputInfo& input_info,
-                            vk::ShaderModule compute_module, vk::PipelineCache driver_cache) {
-	EXIT_IF(compute_module == nullptr);
-
+                            const ShaderModule& module, vk::PipelineCache driver_cache) {
 	vk::PipelineShaderStageCreateInfo                     comp_shader_stage_info {};
 	vk::PipelineShaderStageRequiredSubgroupSizeCreateInfo comp_subgroup_size {};
 	comp_shader_stage_info.stage  = vk::ShaderStageFlagBits::eCompute;
-	comp_shader_stage_info.module = compute_module;
+	comp_shader_stage_info.module = module.module;
 	comp_shader_stage_info.pName  = "main";
-	EXIT_IF(!input_info.stage);
-	const auto wave_size = input_info.stage.program->wave_size;
+	const auto wave_size = module.info.wave_size;
 	if (graphics.compute_subgroup_size_control_enabled &&
 	    wave_size >= graphics.min_subgroup_size && wave_size <= graphics.max_subgroup_size) {
 		comp_subgroup_size.requiredSubgroupSize = wave_size;
@@ -616,9 +436,8 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 
 	std::vector<vk::DescriptorSetLayoutBinding> descriptor_bindings;
-	AddLayoutBindings(descriptor_bindings, *input_info.stage.program,
-	                  vk::ShaderStageFlagBits::eCompute);
-	pipeline.uses_image_heap = input_info.stage.program->info.uses_image_heap;
+	AddLayoutBindings(descriptor_bindings, module.info, vk::ShaderStageFlagBits::eCompute);
+	pipeline.uses_image_heap = module.info.info.uses_image_heap;
 	CreateDescriptorLayout(graphics, pipeline, descriptor_bindings);
 	const vk::PushConstantRange push_constants {vk::ShaderStageFlagBits::eCompute, 0,
 	                                            ShaderRecompiler::IR::NativePushConstantSize};
@@ -632,18 +451,12 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	EXIT_IF(pipeline.pipeline_layout != nullptr);
 
-	LOGF("PipelineTrace: vkCreatePipelineLayout CS begin set_layouts=1 push_constants=%u\n",
-	     1u);
 	auto result = graphics.device.createPipelineLayout(&pipeline_layout_info, nullptr,
 	                                                  &pipeline.pipeline_layout);
-	LOGF("PipelineTrace: vkCreatePipelineLayout CS done result=%s layout=%p\n",
-	     vk::to_string(result).c_str(), static_cast<void*>(pipeline.pipeline_layout));
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
-
 	vk::ComputePipelineCreateInfo info {};
-	if (input_info.stage.program->info.uses_gds) {
+	if (module.info.info.uses_gds) {
 		info.flags = vk::PipelineCreateFlagBits::eDispatchBase;
 	}
 	info.stage             = comp_shader_stage_info;
@@ -652,15 +465,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	EXIT_IF(pipeline.pipeline != nullptr);
 
-	LOGF("PipelineTrace: vkCreateComputePipelines begin layout=%p\n",
-	     static_cast<void*>(pipeline.pipeline_layout));
 	result = graphics.device.createComputePipelines(driver_cache, 1, &info, nullptr,
 	                                                &pipeline.pipeline);
-	LOGF("PipelineTrace: vkCreateComputePipelines done result=%s pipeline=%p\n",
-	     vk::to_string(result).c_str(), static_cast<void*>(pipeline.pipeline));
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
-
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline == nullptr);
 }
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/renderer/render.h

@@ -23,14 +23,9 @@ class Shader;
 } // namespace HW
 
 struct GraphicContext;
-struct ShaderBufferResource;
-struct ShaderComputeInputInfo;
-struct RenderDepthInfo;
-struct RenderColorInfo;
 struct DrawCallInfo;
 struct DrawEmitInfo;
 struct DrawIndexBufferSource;
-struct DrawRenderState;
 class RenderContext;
 class CommandScheduler;
 struct RenderExecutorTestAccess;
@@ -198,9 +193,8 @@ private:
 	                                          uint32_t            render_target_slice_offset,
 	                                          DrawRenderState& state);
 	void ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buffer, const DrawCallInfo& draw,
-	                         DrawRenderState& state, vk::PrimitiveTopology topology,
-	                         const DrawEmitInfo& emit, const DrawIndexBufferSource& index_source,
-	                         bool primitive_restart_enable);
+	                         DrawRenderState& state, const DrawEmitInfo& emit,
+	                         const DrawIndexBufferSource& index_source);
 	[[nodiscard]] RenderState AcquireRenderTargets(CommandBuffer& buffer, RenderColorInfo* colors,
 	                                               uint32_t color_count, RenderDepthInfo& depth,
 	                                               vk::ImageAspectFlags& feedback_aspects,

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -53,7 +53,7 @@ static bool FillSourcesDisjoint(std::span<const ShaderRecompiler::IR::Descriptor
 
 bool RenderExecutor::TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
                                                 const CommandBuffer&          buffer) {
-	const auto& program   = *input.stage.program;
+	const auto& program   = input.stage.module->info;
 	const auto& resources = *input.stage.resources;
 	if (resources.buffers.size() != program.info.buffers.size()) {
 		EXIT("compute runtime buffer count does not match shader metadata\n");
@@ -135,7 +135,7 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
                                           CommandBuffer& command, uint32_t group_x,
                                           uint32_t group_y, uint32_t group_z, uint32_t mode) {
-	const auto& program   = *input.stage.program;
+	const auto& program   = input.stage.module->info;
 	const auto& resources = *input.stage.resources;
 	const auto& fill      = resources.uniform_fill;
 	auto&       cache     = command.GetContext().GetTextureCache();
@@ -208,7 +208,7 @@ bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
 	if (logged_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
 		LOGF("GraphicsRenderDispatchDirect: compute %s shader=0x%016" PRIx64
 		     " addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
-		     image_clear ? "image clear" : "buffer fill", input.stage.program->shader_hash,
+		     image_clear ? "image clear" : "buffer fill", input.stage.module->info.shader_hash,
 		     descriptor.Base48(), size, packed_clear);
 	}
 	return true;
@@ -216,7 +216,7 @@ bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
 
 static void BindSharedMemory(RenderContext& context, ShaderComputeInputInfo& input,
                              PreparedBindings& bindings, uint64_t indirect_args = 0) {
-	if (input.stage.program->bindings.descriptor_counts[
+	if (input.stage.module->info.bindings.descriptor_counts[
 	        static_cast<size_t>(ShaderRecompiler::IR::DescriptorBindingKind::SharedMemory)] == 0) {
 		return;
 	}
@@ -248,7 +248,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
                                     uint32_t thread_group_z, uint32_t mode, bool async_compute) {
 	EXIT_IF(buffer.IsInvalid());
 	m_context.GetCommandScheduler().PopPendingOperations();
-	auto& ctx    = buffer.GetRegisters();
 	auto& sh_ctx = buffer.GetShaders();
 
 	buffer.SetDebugInfo(static_cast<uint32_t>(CommandBufferDebugOp::DispatchDirect), submit_id,
@@ -293,7 +292,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 
 	const auto& cs_regs = sh_ctx.GetCs();
-	const auto& sh_regs = ctx.GetShaderRegisters();
 
 	ShaderComputeInputInfo input_info {};
 	input_info.async_compute = async_compute;
@@ -312,16 +310,14 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		}
 	}
 	m_context.GetImageHeap().BeginDraw();
-	const auto compute_program =
-	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
-	if (!compute_program) return;
+	if (!m_context.GetPipelineCache().GetComputeProgram(cs_regs, input_info)) return;
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
 		input_info.dispatch_threads_num[2]    = thread_group_z;
 	}
 
-	const auto& program   = *input_info.stage.program;
+	const auto& program   = input_info.stage.module->info;
 	const auto& resources = *input_info.stage.resources;
 	if (resources.specialization_reads.empty() &&
 	    (TryConsumeComputeMetaClear(input_info, buffer) ||
@@ -351,8 +347,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 
 	buffer.EndRendering();
-	auto& pipeline =
-	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info);
 	auto& bindings = m_compute_bindings;
 	PrepareBindings(input_info.stage, bindings);
 	if (program.bindings.dispatch_thread_dword != ShaderRecompiler::IR::PushData::NoStart) {
@@ -502,16 +497,14 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.async_compute = async_compute;
 	input_info.indirect_args_address = args_addr;
 	m_context.GetImageHeap().BeginDraw();
-	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
-	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
-	if (!compute_program) return;
+	if (!m_context.GetPipelineCache().GetComputeProgram(cs_regs, input_info)) return;
 	buffer.EndRendering();
-	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info);
 	auto& bindings = m_compute_bindings;
 	PrepareBindings(input_info.stage, bindings);
 	PreparedBindings* descriptor_stage = &bindings;
 	FindBuffers(std::span {&descriptor_stage, 1u});
-	const auto& program = *input_info.stage.program;
+	const auto& program = input_info.stage.module->info;
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -10,10 +10,10 @@
 
 namespace Libs::Graphics {
 
-RenderContext::RenderContext(GraphicContext& graphics)
+RenderContext::RenderContext(GraphicContext& graphics, const std::filesystem::path& shader_store)
     : m_graphics(graphics), m_render_executor(*this), m_command_scheduler(*this, graphics),
       m_descriptor_heap(graphics, m_command_scheduler.GetMasterSemaphore()),
-      m_pipeline_cache(graphics, &m_image_heap), m_sampler_cache(graphics),
+      m_pipeline_cache(graphics, &m_image_heap, shader_store), m_sampler_cache(graphics),
       m_buffer_cache(graphics, m_command_scheduler, m_page_manager, m_texture_cache),
       m_texture_cache(graphics, m_command_scheduler, m_page_manager, m_buffer_cache),
       m_image_heap(*this, m_page_manager), m_occlusion_queries(graphics, m_command_scheduler) {
@@ -134,7 +134,11 @@ void RenderContext::PrepareBda() {
 		m_bda_logged = true;
 	}
 	std::shared_lock lock(m_mapped_ranges_mutex);
-	m_buffer_cache.SynchronizeBda(m_mapped_ranges);
+	m_buffer_cache.TakeBdaSync().ForEach([this](uint64_t begin, uint64_t end) {
+		m_mapped_ranges.ForEachInRange(begin, end - begin, [this](uint64_t start, uint64_t finish) {
+			m_buffer_cache.SynchronizeBuffersInRange(start, finish - start);
+		});
+	});
 	m_fault_process_pending = true;
 }
 

src/graphics/host_gpu/renderer/renderContext.h

@@ -31,7 +31,8 @@ class GuestGpu;
 
 class RenderContext {
 public:
-	explicit RenderContext(GraphicContext& graphics);
+	explicit RenderContext(GraphicContext&              graphics,
+	                       const std::filesystem::path& shader_store = ShaderStorePath());
 	~RenderContext();
 	KYTY_CLASS_NO_COPY(RenderContext);
 

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -17,7 +17,6 @@
 #include "graphics/host_gpu/renderer/colorRenderTarget.h"
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/depthRenderTarget.h"
-#include "graphics/host_gpu/renderer/image/textureCommon.h"
 #include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 #include "graphics/host_gpu/renderer/render.h"
@@ -58,7 +57,7 @@ std::pair<int32_t, uint32_t> ResolveDrawOffsets(uint32_t index_offset,
 	}
 
 	EXIT_IF(!vs_input_info.stage);
-	const auto& program   = *vs_input_info.stage.program;
+	const auto& program   = vs_input_info.stage.module->info;
 	const auto& resources = *vs_input_info.stage.resources;
 	if (index_offset == 0 &&
 	    program.info.vertex_offset_sgpr >= static_cast<int32_t>(program.user_data_base)) {
@@ -198,7 +197,7 @@ static void LogDrawTargetState(const char* draw_name, const RenderColorInfo& col
 	const auto& dc             = ctx.GetDepthControl();
 	const auto& vp             = ctx.GetScreenViewport();
 	const auto& vp0            = vp.viewports[0];
-	const auto& ps_resources   = ps_input_info.stage.program->info;
+	const auto& ps_resources   = ps_input_info.stage.module->info.info;
 	const auto  sampled_images = std::count_if(
 	    ps_resources.images.begin(), ps_resources.images.end(), [](const auto& image) {
 		    return image.resource_class == ShaderRecompiler::IR::ImageResourceClass::Sampled;
@@ -321,7 +320,7 @@ static void UpdateGraphicsDynamicState(const CommandBuffer& buffer, DynamicState
 	const auto& ctx = buffer.GetRegisters();
 	const auto&        vp  = ctx.GetScreenViewport();
 	const vk::Extent2D framebuffer_extent {rendering.width, rendering.height};
-	const auto& outputs = vs_input_info.stage.program->info.outputs;
+	const auto& outputs = vs_input_info.stage.module->info.info.outputs;
 	const bool  indexed_viewports =
 	    std::any_of(outputs.begin(), outputs.end(), [](const auto& output) {
 		    return output.kind == ShaderRecompiler::IR::StageOutputKind::ViewportIndex;
@@ -433,22 +432,6 @@ static bool DrawHasValidVertexShader(const HW::Shader& sh_ctx) {
 	return vs.es_regs.data_addr != 0;
 }
 
-static bool PixelShaderHasDepthOrCoverageSideEffects(const HW::ShaderRegisters& sh_regs) {
-	const auto& db = sh_regs.db_shader_control;
-	return db.shader_kill_enable || db.shader_z_export_enable || db.shader_mask_export_enable ||
-	       db.shader_dual_export_enable || db.shader_execute_on_noop || sh_regs.AlphaToMask();
-}
-
-struct DrawRenderState {
-	RenderDepthInfo       depth_info;
-	RenderColorInfo       color_info[RENDER_COLOR_ATTACHMENTS_MAX] = {};
-	uint32_t              color_count                              = 0;
-	bool                  ps_active                                = true;
-	std::array<ShaderVertexInputInfo, 3> vertex_info;
-	ShaderPixelInputInfo  ps_input_info;
-	PipelineCache::GraphicsPrograms programs;
-};
-
 struct DrawCallInfo {
 	CommandBufferDebugOp debug_op       = CommandBufferDebugOp::DrawIndex;
 	uint32_t             index_count    = 0;
@@ -474,9 +457,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	for (uint32_t i = 0; i < color_count; i++) {
 		auto& target = colors[i];
 		EXIT_IF(!target.image_id);
-		const auto owner = cache.m_slot_images.try_get(target.image_id);
-		if (owner == nullptr || (!owner->registered && !owner->info.data.Empty()) ||
-		    owner->binding.needs_rebind) {
+		const auto& owner = cache.m_slot_images[target.image_id];
+		if ((!owner.registered && !owner.info.data.Empty()) || owner.binding.needs_rebind) {
 			EXIT("color target changed after render-state discovery\n");
 		}
 		const auto image_view = cache.FindRenderTarget(target.image_id, target.desc);
@@ -502,8 +484,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		attachment.image_layout = layout;
 	}
 	if (depth.image_id) {
-		const auto owner = cache.m_slot_images.try_get(depth.image_id);
-		if (owner == nullptr || !owner->registered || owner->binding.needs_rebind) {
+		const auto& owner = cache.m_slot_images[depth.image_id];
+		if (!owner.registered || owner.binding.needs_rebind) {
 			EXIT("depth target changed after render-state discovery\n");
 		}
 		const auto  image_view = cache.FindDepthTarget(depth.image_id, depth.desc);
@@ -589,23 +571,6 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	return state;
 }
 
-static uint32_t DrawColorOutputMask(const HW::Context& ctx) {
-	const auto& sh_regs     = ctx.GetShaderRegisters();
-	const auto  write_mask  = ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask;
-	uint32_t    output_mask = ctx.HasObjectPrimitiveIdOutput() ? 1u << 7u : 0;
-	for (uint32_t index = 0; index < RENDER_COLOR_ATTACHMENTS_MAX; index++) {
-		const auto slot = ShaderPixelExportTarget(sh_regs.m_cbShaderMask, index);
-		if (slot >= RENDER_COLOR_ATTACHMENTS_MAX) {
-			break;
-		}
-		if (sh_regs.target_output_mode[index] != 0 &&
-		    render_target_mask_slot(write_mask, slot) != 0) {
-			output_mask |= 1u << slot;
-		}
-	}
-	return output_mask;
-}
-
 enum class CbColorMode : uint8_t {
 	Disable            = 0,
 	Normal             = 1,
@@ -870,62 +835,19 @@ static bool ResolvePrimitiveRestart(const CommandBuffer& buffer,
 	return false;
 }
 
-static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
-                           uint32_t color_output_mask, DrawRenderState& state) {
-	auto& ctx    = buffer.GetRegisters();
-	auto& sh_ctx = buffer.GetShaders();
-
-	const auto& vertex_shader_info = sh_ctx.GetVs();
-	const auto& pixel_shader_info  = sh_ctx.GetPs();
-	const auto& shader_regs        = ctx.GetShaderRegisters();
-
-	state.programs      = {};
-	state.ps_input_info = {};
-	std::array<Prospero::ColorComponentMapping, RENDER_COLOR_ATTACHMENTS_MAX>
-	    target_export_mapping {};
-	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		const auto& rt = ctx.GetRenderTarget(slot);
-		if ((color_output_mask & (1u << slot)) != 0 && rt.base.addr != 0) {
-			target_export_mapping[slot] =
-			    TextureGetRenderTargetFormat(rt.info.format, rt.info.channel_type,
-			                                 rt.info.channel_order)
-			        .export_mapping;
-		}
-	}
-	auto& pipeline_cache = buffer.GetContext().GetPipelineCache();
-	buffer.GetContext().GetImageHeap().BeginDraw();
-	if (draw.IsIndexed()) {
-		LogDrawPhase(draw.Name(), "GetGraphicsPrograms");
-	}
-	state.programs = pipeline_cache.GetGraphicsPrograms(
-	    vertex_shader_info, pixel_shader_info, shader_regs, ctx, buffer.GetUserConfig(),
-	    target_export_mapping, state.ps_active, state.vertex_info, state.ps_input_info);
-}
-
 bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCallInfo& draw,
                                             uint32_t            render_target_slice_offset,
 	                                        DrawRenderState& state) {
-	const auto& shader_regs       = buffer.GetRegisters().GetShaderRegisters();
-	const auto  color_output_mask = DrawColorOutputMask(buffer.GetRegisters());
-	state.ps_active = buffer.GetShaders().GetPs().ps_regs.data_addr != 0 &&
-	                  (color_output_mask != 0 ||
-	                   PixelShaderHasDepthOrCoverageSideEffects(shader_regs));
-	RefreshShaders(buffer, draw, color_output_mask, state);
-	uint32_t mrt_mask = 0;
-	if (state.ps_active) {
-		for (const auto& output: state.ps_input_info.stage.program->info.outputs) {
-			if (output.kind == ShaderRecompiler::IR::StageOutputKind::Mrt ||
-			    output.kind == ShaderRecompiler::IR::StageOutputKind::ObjectPrimitiveId) {
-				mrt_mask |= 1u << output.location;
-			}
-		}
+	m_context.GetImageHeap().BeginDraw();
+	if (draw.IsIndexed()) {
+		LogDrawPhase(draw.Name(), "GetGraphicsPrograms");
 	}
-	mrt_mask &= color_output_mask;
+	m_context.GetPipelineCache().GetGraphicsPrograms(buffer, state);
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "ResolveRenderColorTarget");
 	}
 	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		if ((mrt_mask & (1u << slot)) != 0) {
+		if ((state.mrt_mask & (1u << slot)) != 0) {
 			ResolveRenderColorTarget(buffer, state.color_info[state.color_count],
 			                         render_target_slice_offset, slot);
 			if (state.color_info[state.color_count].image_id) {
@@ -1124,16 +1046,15 @@ bool RenderExecutor::CopyDepthToColor(CommandBuffer& buffer, const DrawCallInfo&
 
 void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buffer,
                                          const DrawCallInfo& draw, DrawRenderState& state,
-                                         vk::PrimitiveTopology topology, const DrawEmitInfo& emit,
-                                         const DrawIndexBufferSource& index_source,
-	                                     bool primitive_restart_enable) {
-	auto& ucfg = buffer.GetUserConfig();
+                                         const DrawEmitInfo&          emit,
+                                         const DrawIndexBufferSource& index_source) {
+	auto&      ucfg                     = buffer.GetUserConfig();
+	const bool primitive_restart_enable = state.primitive_restart_enable;
 	if (CopyDepthToColor(buffer, draw, emit, state)) {
 		return;
 	}
-	const auto vertex_stages =
-	    std::span {state.vertex_info.data(), state.programs.VertexStageCount()};
-	const bool mesh_active = state.vertex_info[0].stage.program->stage == ShaderType::Mesh;
+	const auto vertex_stages = state.VertexStages();
+	const bool mesh_active = state.vertex_info[0].stage.module->info.stage == ShaderType::Mesh;
 	uint32_t   mesh_groups = 0;
 	if (mesh_active) {
 		const auto& mesh = state.vertex_info[0].mesh;
@@ -1203,10 +1124,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "CreatePipeline");
 	}
-	auto& pipeline = m_context.GetPipelineCache().GetGraphicsPipeline(
-	    std::span {state.color_info, state.color_count}, state.depth_info, vertex_stages, buffer,
-	    state.ps_active ? &state.ps_input_info : nullptr, topology, primitive_restart_enable,
-	    state.programs);
+	auto& pipeline = m_context.GetPipelineCache().GetGraphicsPipeline(buffer, state);
 	vk::ImageAspectFlags feedback_aspects;
 	const auto rendering =
 	    AcquireRenderTargets(buffer, state.color_info, state.color_count, state.depth_info,
@@ -1323,7 +1241,7 @@ static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&
                                                    uint32_t                     first_instance,
                                                    const DrawIndirectArgs&      indirect) {
 	if (indirect.address != 0) {
-		if (vs.stage.program->stage == ShaderType::Mesh ||
+		if (vs.stage.module->info.stage == ShaderType::Mesh ||
 		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy) {
 			return std::nullopt;
 		}
@@ -1388,8 +1306,8 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	hw_check(buffer);
 
-	vk::PrimitiveTopology topology = vk::PrimitiveTopology::ePointList;
-	if (!GetDrawTopology(ucfg, topology)) {
+	DrawRenderState state {};
+	if (!GetDrawTopology(ucfg, state.topology)) {
 		return true;
 	}
 
@@ -1410,7 +1328,7 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 		default: EXIT("unknown index_type_and_size: %u\n", args.index_type_and_size);
 	}
 	index_source.size = static_cast<uint64_t>(args.index_count) * index_source.guest_element_size;
-	const bool primitive_restart = ResolvePrimitiveRestart(buffer, index_source);
+	state.primitive_restart_enable = ResolvePrimitiveRestart(buffer, index_source);
 
 	std::vector<uint16_t> expanded_indices;
 	if (index_source.guest_element_size == 1) {
@@ -1418,7 +1336,8 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 		const auto* src = static_cast<const uint8_t*>(args.index_addr);
 		expanded_indices.resize(args.index_count);
 		for (uint32_t i = 0; i < args.index_count; i++) {
-			expanded_indices[i] = primitive_restart && src[i] == 0xffu ? 0xffffu : src[i];
+			expanded_indices[i] =
+			    state.primitive_restart_enable && src[i] == 0xffu ? 0xffffu : src[i];
 		}
 		index_source.host_data = expanded_indices.data();
 		index_source.size      = expanded_indices.size() * sizeof(uint16_t);
@@ -1426,7 +1345,6 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndex, args.index_count,
 	                        args.instance_count, args.first_instance};
-	DrawRenderState state {};
 	if (!PrepareDrawRenderState(buffer, draw, args.render_target_slice_offset, state)) {
 		ResetBindings();
 		return true;
@@ -1437,8 +1355,7 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 	const auto emit = ResolveDrawEmit(ucfg, state.vertex_info[0], args.offset_source,
 	                                  args.base_vertex, args.first_instance, args.indirect);
 	if (emit) {
-		ExecutePreparedDraw(submit_id, buffer, draw, state, topology, *emit, index_source,
-		                    primitive_restart);
+		ExecutePreparedDraw(submit_id, buffer, draw, state, *emit, index_source);
 	}
 	ResetBindings();
 	return emit.has_value();
@@ -1494,12 +1411,11 @@ bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndexAuto,
 	                         args.vertex_count, args.instance_count, args.first_instance};
 
-	vk::PrimitiveTopology topology = vk::PrimitiveTopology::ePointList;
-	if (!GetDrawTopology(ucfg, topology)) {
+	DrawRenderState state {};
+	if (!GetDrawTopology(ucfg, state.topology)) {
 		ResetBindings();
 		return true;
 	}
-	DrawRenderState state {};
 	if (!PrepareDrawRenderState(buffer, draw, args.render_target_slice_offset, state)) {
 		ResetBindings();
 		return true;
@@ -1507,7 +1423,7 @@ bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 
 	const bool rect_list = Prospero::IsRectList(ucfg.GetPrimType());
 	if (rect_list && state.vertex_info[0].buffers_num == 0 &&
-	    state.vertex_info[0].stage.program->param_export_mask == 0 &&
+	    state.vertex_info[0].stage.module->info.param_export_mask == 0 &&
 	    state.ps_input_info.input_num != 0) {
 		if (graphics_debug_dump_enabled()) {
 			LOGF("DrawIndexAuto: skipping rect-list draw with no VS param exports and PS inputs: "
@@ -1524,7 +1440,7 @@ bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	                                  static_cast<int32_t>(args.first_vertex), args.first_instance,
 	                                  args.indirect);
 	if (emit) {
-		ExecutePreparedDraw(submit_id, buffer, draw, state, topology, *emit, {}, false);
+		ExecutePreparedDraw(submit_id, buffer, draw, state, *emit, {});
 	}
 	ResetBindings();
 	return emit.has_value();

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -30,10 +30,6 @@ namespace Libs::Graphics::ShaderRecompiler {
 
 namespace {
 
-const char* GetDumpLabel(const CompileOptions& options) {
-	return options.dump_label != nullptr ? options.dump_label : "ShaderRecompiler";
-}
-
 std::string MakeIrDump(std::string_view cfg, const IR::Program& ir) {
 	std::string dump = "CFG:\n";
 	dump += cfg;
@@ -44,30 +40,17 @@ std::string MakeIrDump(std::string_view cfg, const IR::Program& ir) {
 	return dump;
 }
 
-const char* StageName(ShaderType stage) {
-	switch (stage) {
-		case ShaderType::Compute: return "CS";
-		case ShaderType::Vertex: return "VS";
-		case ShaderType::Local: return "LS";
-		case ShaderType::TessellationControl: return "HS";
-		case ShaderType::TessellationEvaluation: return "TES";
-		case ShaderType::Mesh: return "MS";
-		case ShaderType::Pixel: return "PS";
-		default: return "unknown";
-	}
-}
-
 void LogDispatcherFallback(const CompileOptions& options, const CFG::Graph& cfg, const char* phase) {
 	const auto* block        = cfg.FindBlock(cfg.failure_block);
 	const auto  start        = block != nullptr ? block->start_pc : UINT32_MAX;
 	const auto  end          = block != nullptr ? block->end_pc : UINT32_MAX;
 	const auto  predecessors = block != nullptr ? block->predecessors.size() : 0u;
 	const auto  successors   = block != nullptr ? block->successors.size() : 0u;
-	LOGF("%s CFG dispatcher fallback: stage=%s hash=0x%016" PRIx64
+	LOGF("ShaderRecompiler %s CFG dispatcher fallback: hash=0x%016" PRIx64
 	     " phase=%s failure=%s block=%" PRIu32 " pc=0x%08" PRIx32 "..0x%08" PRIx32 " preds=%" PRIu64
 	     " succs=%" PRIu64 " blocks=%" PRIu64 " loops=%" PRIu64 " back_edges=%" PRIu64
 	     " reason=%s\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash, phase,
+	     StageName(options.stage), options.shader_hash, phase,
 	     CFG::FailureKindToString(cfg.failure_kind).c_str(), cfg.failure_block, start, end,
 	     static_cast<uint64_t>(predecessors), static_cast<uint64_t>(successors),
 	     static_cast<uint64_t>(cfg.blocks.size()), static_cast<uint64_t>(cfg.natural_loops.size()),
@@ -584,14 +567,30 @@ void PrepareCallTargets(ShaderSource& source, const CompileOptions& options) {
 	for (auto& inst: block) {
 		inst.SetParent(nullptr);
 		if (inst.GetOpcode() == ValueOpcode::LoadAddressU32 || inst.GetOpcode() == ValueOpcode::ReadConstBuffer)
-			inst.SetFlags(SrtReadFlags {.index = inst.Flags<MemoryFlags>().index});
+			inst.SetFlags(MemoryFlags {.index = inst.Flags<MemoryFlags>().index});
 	}
 	query.value_storage.splice(query.value_storage.end(), block.Instructions());
 	source.call_targets = std::move(static_cast<ResourcePlan&>(query));
+	std::vector<Value> targets;
+	for (const auto& site: source.calls) targets.insert(targets.end(), site.target.begin(), site.target.end());
+	CompileSrtWalker(source.call_targets, targets);
 }
 
 } // namespace
 
+const char* StageName(ShaderType stage) {
+	switch (stage) {
+		case ShaderType::Compute: return "CS";
+		case ShaderType::Vertex: return "VS";
+		case ShaderType::Local: return "LS";
+		case ShaderType::TessellationControl: return "HS";
+		case ShaderType::TessellationEvaluation: return "TES";
+		case ShaderType::Mesh: return "MS";
+		case ShaderType::Pixel: return "PS";
+		default: return "unknown";
+	}
+}
+
 ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOptions& options) {
 	EXIT_IF(code.empty());
 	ShaderSource source;
@@ -617,12 +616,18 @@ const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtR
 	EXIT_IF(source.unsupported_call_pc != UINT32_MAX);
 	auto& reads = source.reads;
 	reads.clear();
-	IR::SrtReadCapture capture(runtime, reads);
-	const auto clean = IR::CleanRuntime(capture.ObservedRuntime());
-	IR::SrtWalker walker(source.call_targets, clean);
+	std::vector<uint32_t> flat;
+	const bool walked = IR::WalkSrt(source.call_targets, runtime, flat, &reads);
+	const auto read_code = [&](uint64_t address, std::span<uint32_t> words) {
+		if (runtime.read_specialization_memory == nullptr ||
+		    !runtime.read_specialization_memory(runtime.userdata, address, words)) return false;
+		reads.emplace_back(address, words.size_bytes());
+		return true;
+	};
 	for (auto& site: source.calls) {
 		uint32_t low = 0, high = 0;
-		if (!walker.Evaluate(site.target[0], low) || !walker.Evaluate(site.target[1], high))
+		if (!walked || !IR::WalkedValue(source.call_targets, site.target[0], low) ||
+		    !IR::WalkedValue(source.call_targets, site.target[1], high))
 			EXIT("shader call at pc 0x%08x has an unavailable scalar target",
 			     source.decoded.instructions[site.instruction].pc);
 		// SWAPPC reads the old aliased pair and ignores the target's low two bits.
@@ -646,7 +651,7 @@ const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtR
 				break;
 			}
 			linked.observed_function.resize(function.size);
-			if (!clean.read_memory(clean.userdata, source.calls[i].address, linked.observed_function) ||
+			if (!read_code(source.calls[i].address, linked.observed_function) ||
 			    !std::ranges::equal(expected, linked.observed_function)) {
 				unchanged = false;
 				break;
@@ -676,8 +681,7 @@ const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtR
 				uint32_t count = 1;
 				Decoder::Instruction inst;
 				for (;;) {
-					if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
-					                       std::span(words).first(count)))
+					if (!read_code(address + uint64_t {index} * sizeof(uint32_t), std::span(words).first(count)))
 						EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
 					inst = {};
 					Decoder::DecodeInstruction(words, 0u, inst);
@@ -756,6 +760,7 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 		     static_cast<unsigned>(options.stage));
 	}
 
+	const auto* stage        = StageName(options.stage);
 	const auto compile_begin = std::chrono::steady_clock::now();
 	const auto phase_ms = [&compile_begin]() {
 		return static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::milliseconds>(
@@ -765,27 +770,25 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	std::string decoded_dump;
 	if (options.dump_ir) {
 		decoded_dump = Decoder::ProgramToString(decoded);
-		if (options.early_dump) {
-			LOGF("%s decoded RDNA2 (early):\n%s", GetDumpLabel(options), decoded_dump.c_str());
-		}
+		LOGF("ShaderRecompiler %s decoded RDNA2:\n%s", stage, decoded_dump.c_str());
 	}
 
-	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " CFG BuildGraph\n", GetDumpLabel(options),
-	     StageName(options.stage), options.shader_hash);
+	LOGF("ShaderRecompiler %s phase begin: hash=0x%016" PRIx64 " CFG BuildGraph\n", stage,
+	     options.shader_hash);
 	auto native_cfg = CFG::BuildGraph(decoded);
 	CFG::Graph structured_cfg;
 	auto* selected_cfg = &native_cfg;
-	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " CFG BuildGraph blocks=%" PRIu64
+	LOGF("ShaderRecompiler %s phase end: hash=0x%016" PRIx64 " CFG BuildGraph blocks=%" PRIu64
 	     " loops=%" PRIu64 " back_edges=%" PRIu64 " elapsed_ms=%" PRIu64 "\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
+	     stage, options.shader_hash,
 	     static_cast<uint64_t>(native_cfg.blocks.size()),
 	     static_cast<uint64_t>(native_cfg.natural_loops.size()),
 	     static_cast<uint64_t>(native_cfg.back_edges.size()), phase_ms());
 	if (native_cfg.irreducible) {
 		LogDispatcherFallback(options, native_cfg, "build");
 	} else {
-		LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " CFG Structurize\n",
-		     GetDumpLabel(options), StageName(options.stage), options.shader_hash);
+		LOGF("ShaderRecompiler %s phase begin: hash=0x%016" PRIx64 " CFG Structurize\n", stage,
+		     options.shader_hash);
 		structured_cfg = CFG::Structurize(native_cfg);
 		if (structured_cfg.unsupported) {
 			native_cfg.unsupported = true;
@@ -795,12 +798,12 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 			LogDispatcherFallback(options, native_cfg, "structurize");
 		} else {
 			selected_cfg = &structured_cfg;
-			LOGF("%s structured CFG success: blocks=%" PRIu64 "\n", GetDumpLabel(options),
+			LOGF("ShaderRecompiler %s structured CFG success: blocks=%" PRIu64 "\n", stage,
 			     static_cast<uint64_t>(selected_cfg->blocks.size()));
 		}
-		LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " CFG Structurize blocks=%" PRIu64
+		LOGF("ShaderRecompiler %s phase end: hash=0x%016" PRIx64 " CFG Structurize blocks=%" PRIu64
 		     " loops=%" PRIu64 " elapsed_ms=%" PRIu64 "\n",
-		     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
+		     stage, options.shader_hash,
 		     static_cast<uint64_t>(selected_cfg->blocks.size()),
 		     static_cast<uint64_t>(selected_cfg->natural_loops.size()), phase_ms());
 	}
@@ -813,7 +816,7 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 		    decoded, *options.input_info.vertex, options.user_data_base,
 		    static_cast<uint32_t>(options.user_data.size()), options.wave_size);
 		if (!embedded_fetch.loads.empty()) {
-			LOGF("%s embedded vertex fetch plan: detected=%" PRIu64 "\n", GetDumpLabel(options),
+			LOGF("ShaderRecompiler %s embedded vertex fetch plan: detected=%" PRIu64 "\n", stage,
 			     static_cast<uint64_t>(embedded_fetch.loads.size()));
 		}
 	}
@@ -826,12 +829,12 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	    .input_info       = options.input_info,
 	    .embedded_fetch   = embedded_fetch.loads.empty() ? nullptr : &embedded_fetch,
 	};
-	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " IR TranslateProgram\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash);
+	LOGF("ShaderRecompiler %s phase begin: hash=0x%016" PRIx64 " IR TranslateProgram\n", stage,
+	     options.shader_hash);
 	auto ir = Frontend::TranslateProgram(decoded, cfg, translate_options);
-	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " IR TranslateProgram blocks=%" PRIu64
+	LOGF("ShaderRecompiler %s phase end: hash=0x%016" PRIx64 " IR TranslateProgram blocks=%" PRIu64
 	     " elapsed_ms=%" PRIu64 "\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
+	     stage, options.shader_hash,
 	     static_cast<uint64_t>(ir.blocks.size()), phase_ms());
 	IR::RewriteToSsa(ir.blocks);
 	const auto simplify = [&ir] {
@@ -842,8 +845,9 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	simplify();
 	const auto read_lane_stats = IR::EliminateReadLane(ir, ir.wave_size);
 	if (read_lane_stats.rewritten_reads != 0 || read_lane_stats.rewritten_reductions != 0) {
-		LOGF("%s read-lane elimination: reads=%" PRIu32 " reductions=%" PRIu32 "\n",
-		     GetDumpLabel(options), read_lane_stats.rewritten_reads, read_lane_stats.rewritten_reductions);
+		LOGF("ShaderRecompiler %s read-lane elimination: reads=%" PRIu32
+		     " reductions=%" PRIu32 "\n",
+		     stage, read_lane_stats.rewritten_reads, read_lane_stats.rewritten_reductions);
 	}
 	// Cleanup also exposes profitable folds by removing dead predicate consumers.
 	simplify();
@@ -857,10 +861,8 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	std::string cfg_dump;
 	if (options.dump_ir) {
 		cfg_dump = CFG::GraphToString(cfg);
-		if (options.early_dump) {
-			LOGF("%s native IR before resource tracking:\n%s", GetDumpLabel(options),
-			     MakeIrDump(cfg_dump, ir).c_str());
-		}
+		LOGF("ShaderRecompiler %s native IR before resource tracking:\n%s", stage,
+		     MakeIrDump(cfg_dump, ir).c_str());
 	}
 	IR::TrackResources(ir, decoded, native_cfg, options.input_info);
 	TranslateResult result;
@@ -877,6 +879,7 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
                              uint32_t push_data_start_dword) {
 	const auto emit_begin = std::chrono::steady_clock::now();
 	auto& ir = translated.program;
+	const auto* stage = StageName(ir.stage);
 	IR::ApplyResourceSpecialization(ir, specialization);
 	// The resource plan owns host descriptor evaluation now. Keep only dependencies consumed
 	// by GPU memory operations; bound descriptor dwords must not retain shader instructions.
@@ -923,19 +926,16 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 	std::string ir_dump;
 	if (options.dump_ir) {
 		ir_dump = MakeIrDump(translated.cfg_dump, ir);
-		if (options.early_dump) {
-			LOGF("%s native IR (early):\n%s", GetDumpLabel(options), ir_dump.c_str());
-		}
+		LOGF("ShaderRecompiler %s native IR:\n%s", stage, ir_dump.c_str());
 	}
 
-	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " SPIR-V EmitProgram\n",
-	     GetDumpLabel(options), StageName(ir.stage), ir.shader_hash);
+	LOGF("ShaderRecompiler %s phase begin: hash=0x%016" PRIx64 " SPIR-V EmitProgram\n", stage,
+	     ir.shader_hash);
 	auto spirv = Spirv::EmitProgram(ir, options.input_info, push_data_start_dword, *options.profile,
 	                                options.emit_debug_info);
-	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " SPIR-V EmitProgram words=%" PRIu64
-	     " elapsed_ms=%" PRIu64 "\n",
-	     GetDumpLabel(options), StageName(ir.stage), ir.shader_hash,
-	     static_cast<uint64_t>(spirv.size()),
+	LOGF("ShaderRecompiler %s phase end: hash=0x%016" PRIx64 " SPIR-V EmitProgram words=%" PRIu64
+	     " wave_size=%u elapsed_ms=%" PRIu64 "\n",
+	     stage, ir.shader_hash, static_cast<uint64_t>(spirv.size()), ir.wave_size,
 	     static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::milliseconds>(
 	                               std::chrono::steady_clock::now() - emit_begin)
 	                               .count()));

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -22,10 +22,8 @@ struct CompileOptions {
 	uint32_t                    user_data_base  = 0;
 	uint64_t                    shader_hash     = 0;
 	bool                        dump_ir                    = true;
-	bool                        early_dump                 = false;
 	bool                        emit_debug_info            = false;
 	const Spirv::Profile*       profile                    = &Spirv::DefaultProfile;
-	const char*                 dump_label                 = nullptr;
 	std::span<const uint32_t>   user_data;
 	std::span<const uint32_t>   back_code;
 	ShaderStageInputInfo        input_info;
@@ -69,6 +67,7 @@ struct ShaderSource {
 	std::vector<std::pair<uint64_t, uint64_t>> reads;
 };
 
+[[nodiscard]] const char* StageName(ShaderType stage);
 [[nodiscard]] ShaderSource PrepareShaderSource(std::span<const uint32_t> code,
 	                                           const CompileOptions& options);
 const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime);

src/graphics/shader/recompiler/backend/spirv/spirvEmitterAlu.cpp

@@ -54,9 +54,30 @@ uint32_t EmitExt(EmitterState& state, uint32_t type, uint32_t opcode,
 	return result;
 }
 
-uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t value) {
+uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
+	// Clamping first keeps the truncating conversion defined; NaN and values past the largest
+	// float below the integer limit are selected afterwards.
+	const auto clamped =
+	    EmitExt(state, TypeF32(state), GLSLstd450NClamp,
+	            {src, ConstantF32(state, signed_value ? 0xcf000000u : 0u),
+	             ConstantF32(state, signed_value ? 0x4effffffu : 0x4f7fffffu)});
+	const auto converted = signed_value
+	                           ? Unary(state, spv::OpBitcast, TypeU32(state),
+	                                   Unary(state, spv::OpConvertFToS, TypeI32(state), clamped))
+	                           : Unary(state, spv::OpConvertFToU, TypeU32(state), clamped);
+	const auto above     = Binary(state, spv::OpFOrdGreaterThanEqual, TypeBool(state), src,
+	                              ConstantF32(state, signed_value ? 0x4f000000u : 0x4f800000u));
+	const auto high =
+	    Select(state, TypeU32(state), above,
+	           ConstantU32(state, signed_value ? 0x7fffffffu : 0xffffffffu), converted);
+	return Select(state, TypeU32(state), Unary(state, spv::OpIsNan, TypeBool(state), src),
+	              ConstantU32(state, 0), high);
+}
+
+} // namespace
+
+uint32_t EmitF32BitsToF16RtzBits(EmitterState& state, uint32_t bits) {
 	// Native half conversion causes regressions; use compact integer RTZ conversion.
-	const auto bits = EmitBitcastF32ToU32(state, value);
 	const auto abs  = EmitAndConstant(state, bits, 0x7fffffffu);
 	const auto sign = EmitAndConstant(
 	    state, Binary(state, spv::OpShiftRightLogical, TypeU32(state), bits, ConstantU32(state, 16)),
@@ -92,28 +113,6 @@ uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t value) {
 	return EmitOrU32(state, sign, magnitude);
 }
 
-uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
-	// Clamping first keeps the truncating conversion defined; NaN and values past the largest
-	// float below the integer limit are selected afterwards.
-	const auto clamped =
-	    EmitExt(state, TypeF32(state), GLSLstd450NClamp,
-	            {src, ConstantF32(state, signed_value ? 0xcf000000u : 0u),
-	             ConstantF32(state, signed_value ? 0x4effffffu : 0x4f7fffffu)});
-	const auto converted = signed_value
-	                           ? Unary(state, spv::OpBitcast, TypeU32(state),
-	                                   Unary(state, spv::OpConvertFToS, TypeI32(state), clamped))
-	                           : Unary(state, spv::OpConvertFToU, TypeU32(state), clamped);
-	const auto above     = Binary(state, spv::OpFOrdGreaterThanEqual, TypeBool(state), src,
-	                              ConstantF32(state, signed_value ? 0x4f000000u : 0x4f800000u));
-	const auto high =
-	    Select(state, TypeU32(state), above,
-	           ConstantU32(state, signed_value ? 0x7fffffffu : 0xffffffffu), converted);
-	return Select(state, TypeU32(state), Unary(state, spv::OpIsNan, TypeBool(state), src),
-	              ConstantU32(state, 0), high);
-}
-
-} // namespace
-
 uint32_t EmitFPFma32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
 	const auto result = EmitGlsl<GLSLstd450Fma, IR::Type::F32>(state, a, b, c);
 	state.builder.AddAnnotation(spv::OpDecorate, result, spv::DecorationNoContraction);
@@ -199,7 +198,8 @@ uint32_t EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value
 
 uint32_t EmitPackFloat2x16Rtz(EmitterState& state, uint32_t arg0, uint32_t arg1) {
 	return EmitNative<spv::OpBitFieldInsert, IR::Type::U32>(
-	    state, EmitF32ToF16RtzBits(state, arg0), EmitF32ToF16RtzBits(state, arg1),
+	    state, EmitF32BitsToF16RtzBits(state, EmitBitcastF32ToU32(state, arg0)),
+	    EmitF32BitsToF16RtzBits(state, EmitBitcastF32ToU32(state, arg1)),
 	    ConstantU32(state, 16), ConstantU32(state, 16));
 }
 

src/graphics/shader/recompiler/backend/spirv/spirvEmitterFlow.cpp

@@ -114,6 +114,18 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 	return EmitInputComponentU32(state, kind, component);
 }
 
+// Whether EXEC enables a source lane within the caller's 32-lane half.
+uint32_t EmitSourceLaneActive(ValueEmitContext& ctx, uint32_t exec, uint32_t lane) {
+	auto&      state  = ctx.state;
+	const auto scope  = ConstantU32(state, spv::ScopeSubgroup);
+	const auto ballot = EmitNative<spv::OpGroupNonUniformBallot, IR::Type::U32x4>(state, scope, exec);
+	if (ctx.half != 0) {
+		lane = EmitAndConstant(state, lane, 31);
+	}
+	return EmitNative<spv::OpGroupNonUniformBallotBitExtract, IR::Type::U1>(state, scope, ballot,
+	                                                                        lane);
+}
+
 uint32_t EmitDppWriteCondition(ValueEmitContext& ctx, const IR::DppMoveFlags& flags,
                                uint32_t exec) {
 	auto&    state    = ctx.state;
@@ -135,9 +147,24 @@ uint32_t EmitDppWriteCondition(ValueEmitContext& ctx, const IR::DppMoveFlags& fl
 	// Only row shifts (0x101-0x10f, 0x111-0x11f) can vacate a lane; other controls always read one.
 	const bool row_shift = flags.control > 0x100u && flags.control < 0x120u &&
 	                       flags.control != 0x110u;
-	if (!flags.bound_control && row_shift) {
+	if (flags.bound_control || (flags.fetch_inactive && !row_shift)) {
+		return writable;
+	}
+	// Without BOUND_CTRL, a vacated source or, without FI, an EXEC-disabled one disables the write.
+	const auto target = EmitDppTargetLane(state, flags);
+	if (row_shift) {
+		writable = Binary(state, spv::OpLogicalAnd, TypeBool(state), writable, target.valid);
+	}
+	if (!flags.fetch_inactive) {
+		// Vacated lanes target outside their row; wrapping keeps the ballot index in range.
+		auto source = target.lane;
+		if (row_shift) {
+			lane   = lane != 0 ? lane : EmitSubgroupLocalInvocationId(state);
+			source = EmitNative<spv::OpBitFieldInsert, IR::Type::U32>(
+			    state, lane, source, ConstantU32(state, 0), ConstantU32(state, 4));
+		}
 		writable = Binary(state, spv::OpLogicalAnd, TypeBool(state), writable,
-		                  EmitDppTargetLane(state, flags).valid);
+		                  EmitSourceLaneActive(ctx, exec, source));
 	}
 	return writable;
 }
@@ -816,27 +843,23 @@ uint32_t EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 }
 
 uint32_t EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst) {
-	auto&      state     = ctx.state;
-	const auto flags     = inst.Flags<IR::PermlaneFlags>();
-	const auto subid     = EmitSubgroupLocalInvocationId(state);
-	const auto row       = state.builder.AllocateId();
-	const auto row_value = state.builder.AllocateId();
-	const auto lane      = state.builder.AllocateId();
-	const auto lane8     = state.builder.AllocateId();
-	const auto shift     = state.builder.AllocateId();
-	const auto upper     = state.builder.AllocateId();
-	const auto selected  = state.builder.AllocateId();
-	const auto shifted   = state.builder.AllocateId();
-	const auto index     = state.builder.AllocateId();
-	const auto target    = state.builder.AllocateId();
+	auto&      state    = ctx.state;
+	const auto flags    = inst.Flags<IR::PermlaneFlags>();
+	const auto subid    = EmitSubgroupLocalInvocationId(state);
+	const auto row      = state.builder.AllocateId();
+	const auto lane     = state.builder.AllocateId();
+	const auto lane8    = state.builder.AllocateId();
+	const auto shift    = state.builder.AllocateId();
+	const auto upper    = state.builder.AllocateId();
+	const auto selected = state.builder.AllocateId();
+	const auto shifted  = state.builder.AllocateId();
+	const auto index    = state.builder.AllocateId();
+	const auto target   = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), row, subid,
 	                          ConstantU32(state, 0xfffffff0u));
-	if (flags.x16) {
-		state.builder.AddFunction(spv::OpBitwiseXor, TypeU32(state), row_value, row,
-		                          ConstantU32(state, 16));
-	} else {
-		state.builder.AddFunction(spv::OpCopyObject, TypeU32(state), row_value, row);
-	}
+	const auto row_value =
+	    flags.x16 ? Binary(state, spv::OpBitwiseXor, TypeU32(state), row, ConstantU32(state, 16))
+		          : row;
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), lane, subid,
 	                          ConstantU32(state, 15));
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), lane8, lane,
@@ -852,14 +875,11 @@ uint32_t EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	                          ConstantU32(state, 15));
 	state.builder.AddFunction(spv::OpBitwiseOr, TypeU32(state), target, row_value, index);
 	const auto shuffled = ctx.Shuffle(inst, 0, target);
-	uint32_t   result   = shuffled;
-	if (!flags.fetch_inactive) {
-		const auto source_exec = ctx.Shuffle(inst, 3, target);
-		result                 = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpSelect, TypeU32(state), result, source_exec, shuffled,
-		                          ConstantU32(state, 0));
+	if (flags.fetch_inactive) {
+		return shuffled;
 	}
-	return result;
+	return Select(state, TypeU32(state), EmitSourceLaneActive(ctx, ctx.Arg(inst, 3), target),
+	              shuffled, ctx.Arg(inst, 4));
 }
 
 uint32_t EmitGetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {

src/graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h

@@ -56,6 +56,7 @@ EMIT_NATIVE(BitFieldSExtract, OpBitFieldSExtract, U32, uint32_t, uint32_t, uint3
 EMIT_NATIVE(SelectU1, OpSelect, U1, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(SelectU32, OpSelect, U32, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(SelectF32, OpSelect, F32, uint32_t, uint32_t, uint32_t)
+EMIT_NATIVE(SelectF64, OpSelect, F64, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(IAdd32, OpIAdd, U32, uint32_t, uint32_t)
 EMIT_NATIVE(ISub32, OpISub, U32, uint32_t, uint32_t)
 EMIT_NATIVE(IMul32, OpIMul, U32, uint32_t, uint32_t)

src/graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h

@@ -531,6 +531,7 @@ inline constexpr auto EmitFNegateValue = EmitNative<spv::OpFNegate, IR::Type::F3
 inline constexpr auto EmitFAbsValue = EmitGlsl<GLSLstd450FAbs, IR::Type::F32, uint32_t>;
 
 uint32_t EmitF16BitsToF32(EmitterState& state, uint32_t bits);
+uint32_t EmitF32BitsToF16RtzBits(EmitterState& state, uint32_t bits);
 
 void EmitProgram(EmitterState& state);
 

src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemoryHelpers.cpp

@@ -472,6 +472,17 @@ uint32_t EncodeFormattedStoreComponent(EmitterState& state, const Format::Buffer
                                        UnormRounding rounding) {
 	EXIT_IF(rounding != UnormRounding::RoundEven && info.type != Format::ComponentType::Unorm);
 	const auto bits = info.component_bits[component];
+	if (info.type == Format::ComponentType::Float && bits < 16u) {
+		// Truncate to f16, whose bias the unsigned 11/10-bit floats share, and drop mantissa
+		// bits; negatives and NaN, whose f16 bits lie above +inf, store zero.
+		const auto half = data_bits == 16u ? EmitAndConstant(state, data, 0xffffu)
+		                                   : EmitF32BitsToF16RtzBits(state, data);
+		return Select(state, TypeU32(state),
+		              EmitCompareU32Constant(state, spv::OpUGreaterThan, half, 0x7c00u),
+		              ConstantU32(state, 0),
+		              EmitBinaryU32(state, spv::OpShiftRightLogical, half,
+		                            ConstantU32(state, 15u - bits)));
+	}
 	if (data_bits == 16u) {
 		if (info.type == Format::ComponentType::Uint) {
 			data = EmitBitFieldUExtract(state, data, ConstantU32(state, 0), ConstantU32(state, 16));

src/graphics/shader/recompiler/frontend/translate/Control.cpp

@@ -465,9 +465,13 @@ void Translator::V_PERMLANE16_B32(const Decoder::Instruction& inst, bool x16) {
 	    .fetch_inactive = inst.dst.op_sel,
 	    .bound_control  = inst.dst.op_sel_hi,
 	};
-	const auto result =
-	    ir.Emit(IR::ValueOpcode::Permlane16U32,
-	            {ReadU32(inst.src0), ReadU32(inst.src1), ReadU32(inst.src2), ir.GetExec()}, flags);
+	// An EXEC-disabled source lane without FI disables the write unless BOUND_CTRL reads zero.
+	const auto inactive = flags.fetch_inactive || flags.bound_control
+	                          ? IR::U32(IR::Value(0u))
+	                          : ir.GetVectorReg(static_cast<IR::VectorReg>(inst.dst.reg));
+	const auto result   = ir.Emit(
+	    IR::ValueOpcode::Permlane16U32,
+	    {ReadU32(inst.src0), ReadU32(inst.src1), ReadU32(inst.src2), ir.GetExec(), inactive}, flags);
 	auto dst      = DestinationOperand(inst);
 	dst.op_sel    = false;
 	dst.op_sel_hi = false;

src/graphics/shader/recompiler/frontend/translate/Convert.cpp

@@ -49,18 +49,11 @@ void Translator::V_CVT_F16_16(const Decoder::Instruction& inst, bool signed_valu
 }
 
 void Translator::V_CVT_16_F16(const Decoder::Instruction& inst, bool signed_value) {
-	const auto value = ReadF16AsF32(inst.src0);
 	// The 32-bit conversion already saturates and maps NaN to zero.
-	if (signed_value) {
-		const auto converted = ir.Emit(IR::ValueOpcode::ConvertS32F32, {value});
-		const auto low = ir.Emit(IR::ValueOpcode::SMax32, {converted, IR::Value(0xffff8000u)});
-		Write16Bits(DestinationOperand(inst),
-		            IR::U32(ir.Emit(IR::ValueOpcode::SMin32, {low, IR::Value(0x7fffu)})));
-		return;
-	}
-	const auto converted = ir.Emit(IR::ValueOpcode::ConvertU32F32, {value});
-	Write16Bits(DestinationOperand(inst),
-	            IR::U32(ir.Emit(IR::ValueOpcode::UMin32, {converted, IR::Value(0xffffu)})));
+	const auto converted = IR::U32(
+	    ir.Emit(signed_value ? IR::ValueOpcode::ConvertS32F32 : IR::ValueOpcode::ConvertU32F32,
+	            {ReadF16AsF32(inst.src0)}));
+	Write16Bits(DestinationOperand(inst), Saturate16(converted, signed_value));
 }
 
 void Translator::V_CVT_RPI_I32_F32(const Decoder::Instruction& inst) {
@@ -130,6 +123,12 @@ void Translator::V_CVT_PK_U8_F32(const Decoder::Instruction& inst) {
 	                     {ReadU32(SourceAt(inst, 2)), byte, shift, IR::Value(8u)}));
 }
 
+void Translator::V_CVT_PK_16_32(const Decoder::Instruction& inst, bool signed_value) {
+	WriteOperand(DestinationOperand(inst),
+	             PackU16Lanes(Saturate16(ReadU32(inst.src0), signed_value),
+	                          Saturate16(ReadU32(inst.src1), signed_value)));
+}
+
 void Translator::V_PACK_B32_F16(const Decoder::Instruction& inst) {
 	WriteOperand(DestinationOperand(inst),
 	             PackU16Lanes(Read16LaneBits(inst.src0, false), Read16LaneBits(inst.src1, false)));
@@ -140,4 +139,12 @@ IR::U32 Translator::PackU16Lanes(IR::U32 low, IR::U32 high) {
 	    ir.Emit(IR::ValueOpcode::BitFieldInsert, {low, high, IR::Value(16u), IR::Value(16u)}));
 }
 
+IR::U32 Translator::Saturate16(IR::U32 value, bool signed_value) {
+	if (!signed_value) {
+		return IR::U32(ir.Emit(IR::ValueOpcode::UMin32, {value, IR::Value(0xffffu)}));
+	}
+	const auto low = ir.Emit(IR::ValueOpcode::SMax32, {value, IR::Value(0xffff8000u)});
+	return IR::U32(ir.Emit(IR::ValueOpcode::SMin32, {low, IR::Value(0x7fffu)}));
+}
+
 } // namespace Libs::Graphics::ShaderRecompiler::Frontend

src/graphics/shader/recompiler/frontend/translate/Float.cpp

@@ -1,6 +1,7 @@
 #include "graphics/shader/recompiler/frontend/translate/Translator.h"
 
 #include <array>
+#include <bit>
 #include <utility>
 
 namespace Libs::Graphics::ShaderRecompiler::Frontend {
@@ -118,6 +119,33 @@ void Translator::FloatUnary(const Decoder::Instruction& inst, IR::ValueOpcode op
 	WriteOperand(DestinationOperand(inst), ir.Emit(opcode, {ReadOperand(inst.src0, type)}));
 }
 
+void Translator::V_FRACT(const Decoder::Instruction& inst, IR::Type type) {
+	// x - floor(x) rounds to 1.0 for a tiny negative x; v_fract returns the largest value below
+	// one instead. The ordered compare lets NaN through.
+	const auto destination = DestinationOperand(inst);
+	if (type == IR::Type::F64) {
+		const auto result    = ir.Emit(IR::ValueOpcode::FPFract64, {ReadOperand(inst.src0, type)});
+		const auto below_one = ir.Emit(IR::ValueOpcode::BitCastF64U64,
+		                               {IR::Value(uint64_t {0x3fefffffffffffff})});
+		const auto above     = ir.Emit(IR::ValueOpcode::FPOrdGreaterThan64, {result, below_one});
+		WriteOperand(destination, ir.Emit(IR::ValueOpcode::SelectF64, {above, below_one, result}));
+		return;
+	}
+	const bool half   = type == IR::Type::F16;
+	const auto source = half ? ReadF16AsF32(inst.src0) : IR::F32(ReadOperand(inst.src0, type));
+	const auto result = IR::F32(ir.Emit(IR::ValueOpcode::FPFract32, {source}));
+	// Clamp in the destination precision: 0x3f7fe000 is the half 0x3bff.
+	const auto below_one =
+	    IR::F32(IR::Value::F32(std::bit_cast<float>(half ? 0x3f7fe000u : 0x3f7fffffu)));
+	const auto above   = IR::U1(ir.Emit(IR::ValueOpcode::FPOrdGreaterThan32, {result, below_one}));
+	const auto clamped = SelectF32(above, below_one, result);
+	if (half) {
+		WriteF16(destination, clamped);
+	} else {
+		WriteOperand(destination, clamped);
+	}
+}
+
 void Translator::FloatBinary(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                              bool reverse) {
 	std::array<IR::Value, 2> args;

src/graphics/shader/recompiler/frontend/translate/Integer.cpp

@@ -25,8 +25,7 @@ void Translator::V_MAD_I16(const Decoder::Instruction& inst) {
 	const auto rhs    = ReadU16AsU32(inst.src1, true);
 	auto       result = ir.IAdd(ir.IMul(lhs, rhs), ReadU16AsU32(inst.src2, true));
 	if (inst.dst.clamp) {
-		result = IR::U32(ir.Emit(IR::ValueOpcode::SMax32, {result, IR::Value(0xffff8000u)}));
-		result = IR::U32(ir.Emit(IR::ValueOpcode::SMin32, {result, IR::Value(0x7fffu)}));
+		result = Saturate16(result, true);
 	}
 	Write16Bits(DestinationOperand(inst), result);
 }

src/graphics/shader/recompiler/frontend/translate/Translator.h

@@ -114,6 +114,7 @@ private:
 
 	IR::F32 SelectF32(IR::U1 condition, IR::F32 true_value, IR::F32 false_value);
 	IR::U32 PackU16Lanes(IR::U32 low, IR::U32 high);
+	IR::U32 Saturate16(IR::U32 value, bool signed_value);
 	void EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar, bool cmpx);
 	void EmitCompareConstant(const Decoder::Instruction& inst, bool value, bool scalar, bool cmpx);
 	void EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode, IR::Type type,
@@ -141,6 +142,7 @@ private:
 	void V_CVT_PKRTZ_F16_F32(const Decoder::Instruction& inst);
 	void V_CVT_PKNORM_F32(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
 	void V_CVT_PK_U8_F32(const Decoder::Instruction& inst);
+	void V_CVT_PK_16_32(const Decoder::Instruction& inst, bool signed_value);
 	void V_PACK_B32_F16(const Decoder::Instruction& inst);
 	void PackedFloat16(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                   bool quiet_snan);
@@ -151,6 +153,7 @@ private:
 	void Float16Ternary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                    bool mix);
 	void FloatUnary(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
+	void V_FRACT(const Decoder::Instruction& inst, IR::Type type);
 	void FloatBinary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool reverse);
 	void FloatTernary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                  bool mix);

src/graphics/shader/recompiler/frontend/translate/Vector.cpp

@@ -495,8 +495,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 			return;
 		case O::V_CVT_PK_U8_F32: V_CVT_PK_U8_F32(inst); return;
 		case O::V_PACK_B32_F16: V_PACK_B32_F16(inst); return;
-		case O::V_CVT_PK_U16_U32:
-		case O::V_CVT_PK_I16_I32: return PackB16(inst, false, false);
+		case O::V_CVT_PK_U16_U32: return V_CVT_PK_16_32(inst, false);
+		case O::V_CVT_PK_I16_I32: return V_CVT_PK_16_32(inst, true);
 
 		case O::V_LSHLREV_B16:
 			return Integer16Shift(inst, IR::ValueOpcode::ShiftLeftLogical32, false);
@@ -563,7 +563,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CEIL_F16: return Float16Unary(inst, IR::ValueOpcode::FPCeil32, false);
 		case O::V_TRUNC_F16: return Float16Unary(inst, IR::ValueOpcode::FPTrunc32, false);
 		case O::V_RNDNE_F16: return Float16Unary(inst, IR::ValueOpcode::FPRoundEven32, false);
-		case O::V_FRACT_F16: return Float16Unary(inst, IR::ValueOpcode::FPFract32, false);
+		case O::V_FRACT_F16: return V_FRACT(inst, IR::Type::F16);
 		case O::V_SIN_F16: return Float16Trig(inst, IR::ValueOpcode::FPSin);
 		case O::V_COS_F16: return Float16Trig(inst, IR::ValueOpcode::FPCos);
 		case O::V_MIN3_F16: return Float16Ternary(inst, IR::ValueOpcode::FPMinTri32, false, false);
@@ -574,8 +574,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_RCP_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecip32);
 		case O::V_RCP_F64: return FloatUnary(inst, IR::ValueOpcode::FPRecip64);
 		case O::V_RCP_IFLAG_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecipIFlag32);
-		case O::V_FRACT_F32: return FloatUnary(inst, IR::ValueOpcode::FPFract32);
-		case O::V_FRACT_F64: return FloatUnary(inst, IR::ValueOpcode::FPFract64);
+		case O::V_FRACT_F32: return V_FRACT(inst, IR::Type::F32);
+		case O::V_FRACT_F64: return V_FRACT(inst, IR::Type::F64);
 		case O::V_TRUNC_F32: return FloatUnary(inst, IR::ValueOpcode::FPTrunc32);
 		case O::V_TRUNC_F64: return FloatUnary(inst, IR::ValueOpcode::FPTrunc64);
 		case O::V_CEIL_F32: return FloatUnary(inst, IR::ValueOpcode::FPCeil32);

src/graphics/shader/recompiler/ir/Serialize.h

@@ -0,0 +1,181 @@
+#ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SERIALIZE_H_
+#define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SERIALIZE_H_
+
+#include "common/assert.h"
+#include "graphics/shader/recompiler/backend/spirv/SpirvEmitter.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/rectListShader.h"
+
+#include <algorithm>
+#include <bit>
+#include <optional>
+#include <span>
+#include <string>
+#include <tuple>
+#include <type_traits>
+#include <unordered_map>
+#include <vector>
+
+// Shader cache records. Io(io, value) appends value to a Writer or reads it back from a Reader
+// field by field, so padding never reaches the bytes; instructions travel as plan-local indices.
+namespace Libs::Graphics::ShaderRecompiler::IR {
+
+struct Writer {
+	std::vector<uint8_t>                      bytes;
+	std::unordered_map<const Inst*, uint32_t> index; // Instructions that values may refer to.
+
+	void Raw(const void* data, size_t size) {
+		const auto* begin = static_cast<const uint8_t*>(data);
+		bytes.insert(bytes.end(), begin, begin + size);
+	}
+};
+
+struct Reader {
+	std::span<const uint8_t> bytes;
+	std::vector<Inst*>       insts;     // Instructions by index.
+	bool                     ok = true; // False once a read ran past the end or was invalid.
+
+	void Raw(void* data, size_t size) {
+		ok = ok && size <= bytes.size();
+		if (!ok) return;
+		std::copy_n(bytes.data(), size, static_cast<uint8_t*>(data));
+		bytes = bytes.subspan(size);
+	}
+	// Every element takes at least one byte, which bounds counts read from damaged records.
+	uint32_t Count(uint32_t count) {
+		ok = ok && count <= bytes.size();
+		return ok ? count : 0u;
+	}
+};
+
+template <class S> constexpr bool Reading = std::is_same_v<S, Reader>;
+template <class T, template <class...> class Of> constexpr bool IsA = false;
+template <template <class...> class Of, class... A> constexpr bool IsA<Of<A...>, Of> = true;
+template <class T, class U> concept Is = std::is_same_v<std::remove_const_t<T>, U>;
+
+template <class S, class T> void Io(S& io, T& value);
+
+template <class S, class... T> void IoAll(S& io, T&&... values) {
+	(Io(io, values), ...);
+}
+
+// The value's Type, Opaque for an instruction, then the instruction index or the zero-extended
+// immediate.
+inline void IoValue(Writer& writer, const Value& value) {
+	auto     type = Type::Opaque;
+	uint64_t bits = 0;
+	if (const auto* inst = value.TryInstruction()) {
+		bits = writer.index.at(inst);
+	} else {
+		type = value.GetType();
+		switch (type) {
+			case Type::Void: break;
+			case Type::ScalarReg: bits = RegIndex(value.ScalarRegister()); break;
+			case Type::VectorReg: bits = RegIndex(value.VectorRegister()); break;
+			case Type::U1: bits = value.U1(); break;
+			case Type::U8: bits = value.U8(); break;
+			case Type::U16: bits = value.U16(); break;
+			case Type::U32: bits = value.U32(); break;
+			case Type::U64: bits = value.U64(); break;
+			case Type::F16: bits = value.F16Bits(); break;
+			case Type::F32: bits = std::bit_cast<uint32_t>(value.F32Value()); break;
+			default: EXIT("value type %u has no cache encoding\n", static_cast<uint32_t>(type));
+		}
+	}
+	IoAll(writer, type, bits);
+}
+
+inline void IoValue(Reader& reader, Value& value) {
+	auto     type = Type::Void;
+	uint64_t bits = 0;
+	IoAll(reader, type, bits);
+	switch (type) {
+		case Type::Void: value = {}; break;
+		case Type::Opaque:
+			reader.ok = reader.ok && bits < reader.insts.size();
+			if (reader.ok) value = Value(reader.insts[bits]);
+			break;
+		case Type::ScalarReg: value = Value(static_cast<ScalarReg>(bits)); break;
+		case Type::VectorReg: value = Value(static_cast<VectorReg>(bits)); break;
+		case Type::U1: value = Value(bits != 0u); break;
+		case Type::U8: value = Value(static_cast<uint8_t>(bits)); break;
+		case Type::U16: value = Value(static_cast<uint16_t>(bits)); break;
+		case Type::U32: value = Value(static_cast<uint32_t>(bits)); break;
+		case Type::U64: value = Value(bits); break;
+		case Type::F16: value = Value::F16(static_cast<uint16_t>(bits)); break;
+		case Type::F32: value = Value::F32(std::bit_cast<float>(static_cast<uint32_t>(bits))); break;
+		default: reader.ok = false; break;
+	}
+}
+
+// Scalars and standard containers; structs have their own overloads below.
+template <class S, class T> void Io(S& io, T& value) {
+	using U = std::remove_const_t<T>;
+	static_assert(!std::is_pointer_v<U>, "pointers do not persist");
+	if constexpr (std::is_arithmetic_v<U> || std::is_enum_v<U>) {
+		io.Raw(&value, sizeof(U));
+	} else if constexpr (std::is_same_v<U, Value>) {
+		IoValue(io, value);
+	} else if constexpr (IsA<U, std::vector> || IsA<U, std::basic_string>) {
+		auto count = static_cast<uint32_t>(value.size());
+		Io(io, count);
+		if constexpr (Reading<S>) value.resize(io.Count(count));
+		using E = typename U::value_type;
+		if constexpr (std::is_arithmetic_v<E> || std::is_enum_v<E>) {
+			io.Raw(value.data(), value.size() * sizeof(E));
+		} else {
+			for (auto& element: value) Io(io, element);
+		}
+	} else if constexpr (IsA<U, std::optional>) {
+		bool has = value.has_value();
+		Io(io, has);
+		if constexpr (Reading<S>) {
+			if (has) value = typename U::value_type {}; // Not emplace(): clang rejects nested aggregates.
+			else value.reset();
+		}
+		if (has) Io(io, *value);
+	} else {
+		static_assert(requires { std::tuple_size<U>::value; }, "struct without an Io overload");
+		std::apply([&](auto&... elements) { IoAll(io, elements...); }, value); // std::array, std::pair
+	}
+}
+
+// Binds every field, so that a field added later fails to compile until it is listed.
+#define KYTY_IO_FIELDS(Struct, ...)                                                                \
+	template <class S, Is<Struct> T> void Io(S& io, T& object) {                                   \
+		auto& [__VA_ARGS__] = object;                                                              \
+		IoAll(io, __VA_ARGS__);                                                                    \
+	}
+KYTY_IO_FIELDS(MemoryInfo, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w, x, y)
+KYTY_IO_FIELDS(BufferResource, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s)
+KYTY_IO_FIELDS(ImageResource, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w)
+KYTY_IO_FIELDS(SamplerResource, a, b, c, d, e, f, g, h, i, j, k)
+KYTY_IO_FIELDS(SampledResourcePair, a, b, c)
+KYTY_IO_FIELDS(StageInput, a, b, c, d, e)
+KYTY_IO_FIELDS(StageOutput, a, b, c, d)
+KYTY_IO_FIELDS(ShaderInfo, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w, x, y, z,
+               A, B, C, D, E, F, G, H, I, J)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::CountedRows, a, b)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead, a, b, c, d, e, f, g, h, i, j)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::PointerTable, a, b, c, d)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p)
+KYTY_IO_FIELDS(DescriptorSource, a, b, c)
+KYTY_IO_FIELDS(SrtRead, a, b)
+KYTY_IO_FIELDS(ResourceBlock, a, b, c, d)
+KYTY_IO_FIELDS(UniformFill, a, b, c, d, e, f)
+KYTY_IO_FIELDS(UniformFillPlan, a, b, c, d)
+KYTY_IO_FIELDS(ResourceSpecialization::Buffer, a, b, c, d, e, f, g, h)
+KYTY_IO_FIELDS(ResourceSpecialization::Image, a, b, c, d, e, f, g, h, i, j, k)
+KYTY_IO_FIELDS(ResourceSpecialization::Sampler, a, b, c)
+KYTY_IO_FIELDS(ResourceSpecialization, a, b, c)
+KYTY_IO_FIELDS(Bindings, a, b, c, d, e)
+KYTY_IO_FIELDS(CompiledShaderInfo, a, b, c, d, e, f, g, h, i, j)
+KYTY_IO_FIELDS(Spirv::Profile, a, b, c)
+KYTY_IO_FIELDS(ShaderTessellationInputInfo, a, b, c, d, e, f, g, h, i)
+KYTY_IO_FIELDS(RectListParameter, a, b, c)
+#undef KYTY_IO_FIELDS
+
+} // namespace Libs::Graphics::ShaderRecompiler::IR
+
+#endif /* EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SERIALIZE_H_ */

src/graphics/shader/recompiler/ir/ShaderIR.h

@@ -14,7 +14,6 @@
 
 #include <array>
 #include <bit>
-#include <deque>
 #include <list>
 #include <memory>
 #include <optional>
@@ -28,8 +27,6 @@ class CodeGenerator;
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-class SrtWalker;
-
 // Finite descriptor domains also bound the arithmetic used to prove their indices.
 constexpr uint32_t MaxIndirectDescriptorProbes = 65536u;
 
@@ -658,24 +655,10 @@ struct UniformFillPlan {
 // condition and fill values without translated blocks, plus reusable evaluation scratch.
 struct ResourcePlan {
 	struct SelectorKeys {
-		uint32_t              source = UINT32_MAX;
 		std::vector<uint32_t> values;
 		bool                  resolved = false;
 	};
 
-	struct EvaluationContext {
-		struct Entry {
-			uint64_t value      = 0;
-			uint64_t generation = 0;
-		};
-
-		std::vector<Entry> values;
-		uint64_t           generation = 0;
-		const SrtWalker*    session = nullptr;
-		Value              active_mask;
-		uint8_t            read_policy = 0;
-	};
-
 	ResourcePlan();
 	~ResourcePlan();
 
@@ -695,21 +678,18 @@ struct ResourcePlan {
 	std::vector<ResourceBlock>          control_flow;
 	std::unordered_map<const Inst*, std::vector<uint32_t>> control_phi_predecessors;
 	std::vector<SrtRead>                srt_reads;
-	bool                                requires_specialization_memory = false;
 	std::vector<std::pair<uint64_t, uint64_t>> source_reads;
 	bool                                capture_specialization_reads = false;
 	bool                                srt_plan_complete          = false;
 	bool                                resource_tracking_complete = false;
 	ShaderInfo                          info;
 	UniformFillPlan                     uniform_fill;
+	// x86 walker over the retained SRT, descriptor and condition values.
 	std::unique_ptr<Xbyak::CodeGenerator> srt_code;
-	// GPU-thread scratch for nested clean/EXEC memos, activity and material keys.
-	mutable std::deque<EvaluationContext> evaluation_contexts;
-	mutable uint32_t                       evaluation_value_count = 0;
-	mutable std::vector<uint8_t>            active_sources;
-	mutable std::vector<uint8_t>            visited_blocks;
-	mutable std::vector<uint8_t>            reachable_blocks;
-	mutable std::vector<uint32_t>           pending_blocks;
+	uint32_t                              srt_user_data = 0; // User-data DWORDs the walker reads.
+	uint32_t                              srt_captures  = 0; // Scratch offset of loaded addresses.
+	// GPU-thread scratch for walker results and material keys.
+	mutable std::vector<uint64_t>           srt_scratch;
 	mutable std::vector<uint32_t>           material_keys;
 	mutable std::vector<SelectorKeys>       selector_keys;
 };

src/graphics/shader/recompiler/ir/Value.h

@@ -123,13 +123,9 @@ public:
 	[[nodiscard]] Block*                  PhiBlock(size_t index) const;
 	[[nodiscard]] Block*                  Parent() const;
 	[[nodiscard]] const std::vector<Use>& Uses() const;
-	// Runtime indices belong to the resource plan that owns this instruction.
-	[[nodiscard]] uint32_t EvaluationIndex(uint32_t& count) const {
-		if (evaluation_index == UINT32_MAX) {
-			evaluation_index = count++;
-		}
-		return evaluation_index;
-	}
+	// SRT walker output index, assigned by the resource plan that owns this instruction.
+	[[nodiscard]] uint32_t WalkOutput() const { return walk_output; }
+	void SetWalkOutput(uint32_t index) const { walk_output = index; }
 	// Backend definitions belong to one emission. Wave64 emulation needs one per half.
 	[[nodiscard]] uint32_t Definition(uint32_t half) const { return definitions[half]; }
 	void SetDefinition(uint32_t half, uint32_t definition) const {
@@ -171,7 +167,7 @@ private:
 	ValueOpcode         opcode;
 	uint8_t             num_args;
 	bool                live = false;
-	mutable uint32_t    evaluation_index = UINT32_MAX;
+	mutable uint32_t    walk_output = UINT32_MAX;
 	mutable std::array<uint32_t, 2> definitions {};
 	uint64_t            flags;
 	Block*              parent = nullptr;

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc

@@ -96,6 +96,7 @@ VALUE_OPCODE(WqmU64, U64, U64)
 // Core value operations used to expand architectural side effects.
 VALUE_OPCODE(SelectU1, U1, U1, U1, U1)
 VALUE_OPCODE(SelectF32, F32, U1, F32, F32)
+VALUE_OPCODE(SelectF64, F64, U1, F64, F64)
 VALUE_OPCODE(IAdd32, U32, U32, U32)
 VALUE_OPCODE(IAdd64, U64, U64, U64)
 VALUE_OPCODE(ISub32, U32, U32, U32)
@@ -222,7 +223,7 @@ VALUE_OPCODE(ReadFirstLane, U32, U32, U1)
 VALUE_OPCODE(ReadLane, U32, U32, U32)
 VALUE_OPCODE(SubgroupUMinU32, U32, U32)
 VALUE_OPCODE(WriteLane, U32, U32, U32, U32)
-VALUE_OPCODE(Permlane16U32, U32, U32, U32, U32, U1)
+VALUE_OPCODE(Permlane16U32, U32, U32, U32, U32, U1, U32)
 VALUE_OPCODE(PermuteU32, U32, U32, U32, U1)
 VALUE_OPCODE(BpermuteU32, U32, U32, U32, U1)
 

src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp

@@ -3,6 +3,7 @@
 #include "common/assert.h"
 #include "graphics/guest_gpu/gpu_format.h"
 #include "graphics/shader/recompiler/BufferFormat.h"
+#include "graphics/shader/recompiler/ir/Serialize.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shaderBindings.h"
 
@@ -186,35 +187,6 @@ const DescriptorSource* Source(const ResourcePlan& program, uint32_t source) {
 	return &program.descriptor_sources[source];
 }
 
-void MarkCleanReads(const std::unordered_set<const Inst*>& planned_reads,
-                    const DescriptorSource* source, Value extra = {}) {
-	if (source == nullptr && extra.IsEmpty()) {
-		return;
-	}
-	std::vector<Value> pending;
-	if (source != nullptr) {
-		pending.assign(source->dwords.begin(), source->dwords.begin() + source->dword_count);
-	}
-	if (!extra.IsEmpty()) pending.push_back(extra);
-	std::vector<const Inst*> visited;
-	while (!pending.empty()) {
-		auto* inst = pending.back().Resolve().TryInstruction();
-		pending.pop_back();
-		if (inst == nullptr || std::ranges::find(visited, inst) != visited.end()) {
-			continue;
-		}
-		visited.push_back(inst);
-		if (planned_reads.contains(inst)) {
-			auto flags = inst->Flags<SrtReadFlags>();
-			flags.clean = 1u;
-			inst->SetFlags(flags);
-		}
-		for (size_t arg = 0; arg < inst->NumArgs(); arg++) {
-			pending.push_back(inst->Arg(arg));
-		}
-	}
-}
-
 bool ReadTable(uint64_t base, uint64_t size, uint64_t offset,
                const SrtRuntime& runtime, std::span<uint32_t> words) {
 	const auto count = std::min<uint64_t>(words.size(), offset < size ? (size - offset) / 4u : 0u);
@@ -309,22 +281,22 @@ bool CanReadIndexedSelector(const ShaderBufferResource& material, uint32_t offse
 	           uint64_t {UINT32_MAX} + 1u;
 }
 
-bool EvaluateIndirectCount(const DescriptorSource::IndirectDescriptor& indirect,
-                           const SrtRuntime& runtime, SrtWalker& clean, uint32_t& count) {
+bool EvaluateIndirectCount(const ResourcePlan& program, const DescriptorSource::IndirectDescriptor& indirect,
+                           const SrtRuntime& runtime, uint32_t& count) {
 	if (indirect.workgroup_axis != UINT32_MAX) {
 		if (runtime.workgroup_counts.size() != 3u || indirect.workgroup_axis >= 3u) return false;
 		count = std::ranges::find(runtime.workgroup_counts, 0u) != runtime.workgroup_counts.end()
 		    ? 0u : runtime.workgroup_counts[indirect.workgroup_axis];
 		return true;
 	}
-	if (!clean.Evaluate(indirect.key_count, count)) return false;
+	if (!WalkedValue(program, indirect.key_count, count)) return false;
 	if (indirect.signed_count && std::bit_cast<int32_t>(count) < 0) count = 0;
 	return true;
 }
 
-bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
+bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::IndirectDescriptor& indirect,
                          const DescriptorValue& material_value, const SrtRuntime& runtime,
-                         SrtWalker& clean, std::vector<uint32_t>& keys) {
+                         std::vector<uint32_t>& keys) {
 	const auto* selector = &*indirect.selector;
 	keys.clear();
 	const auto read_selector = [&](uint64_t base, uint64_t size, uint64_t offset,
@@ -363,7 +335,7 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 		const auto& rows = *selector->rows;
 		DescriptorValue count_value;
 		ShaderBufferResource counts, material;
-		if (!clean.EvaluateDescriptor(rows.source, count_value) ||
+		if (!WalkedDescriptor(program, rows.source, count_value) ||
 		    !DecodeBufferDescriptor(count_value, counts) ||
 		    !DecodeBufferDescriptor(material_value, material) || selector->stride != 4u ||
 		    selector->offset != 0u || rows.stride == 0u || (rows.stride & 3u) != 0u ||
@@ -392,7 +364,7 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 				if (candidate_end - first > MaxIndirectDescriptorProbes) break;
 				end = candidate_end;
 			}
-			// Coalesce small gaps, but interpret only selected prefixes. The strict
+			// Coalesce small gaps, but interpret only selected prefixes. The specialization
 			// reader captures the entire fetched span for writable-alias rejection.
 			batch.resize(end - first);
 			if (!ReadScalarTable(material.Base48(), material.GetSize(), first * 4u, runtime, batch))
@@ -408,8 +380,8 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 		std::array<uint32_t, 4> masks {};
 		uint32_t first = 0, count = 0;
 		if (!DecodeBufferDescriptor(material_value, material) ||
-		    !clean.Evaluate(indirect.selector_first, first) ||
-		    !EvaluateIndirectCount(indirect, runtime, clean, count) || count != masks.size() * 32u ||
+		    !WalkedValue(program, indirect.selector_first, first) ||
+		    !EvaluateIndirectCount(program, indirect, runtime, count) || count != masks.size() * 32u ||
 		    selector->stride != sizeof(masks) || selector->offset != 0u ||
 		    material.Type() != 0u || material.Stride() != sizeof(masks) ||
 		    material.SwizzleEnabled() || material.AddTid() || material.OutOfBounds() != 0u ||
@@ -462,8 +434,8 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 			base = material.Base48();
 			size = material.GetSize();
 		}
-		if (!clean.Evaluate(indirect.selector_first, first) ||
-		    !EvaluateIndirectCount(indirect, runtime, clean, count) || count > MaxIndirectDescriptorProbes)
+		if (!WalkedValue(program, indirect.selector_first, first) ||
+		    !EvaluateIndirectCount(program, indirect, runtime, count) || count > MaxIndirectDescriptorProbes)
 			return false;
 		// SMEM uses its instruction's byte stride; V# stride only supplies its bounds.
 		const auto step = selector->stride != 0u ? selector->stride : material.Stride();
@@ -483,8 +455,8 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 		uint32_t mask = 0;
 		uint32_t count = 0;
 		if (material_value.dword_count != 2u ||
-		    !clean.Evaluate(indirect.selector_mask, mask) ||
-		    !EvaluateIndirectCount(indirect, runtime, clean, count) || count == 0u || count > 32u) {
+		    !WalkedValue(program, indirect.selector_mask, mask) ||
+		    !EvaluateIndirectCount(program, indirect, runtime, count) || count == 0u || count > 32u) {
 			return false;
 		}
 		if (count < 32u) mask &= (1u << count) - 1u;
@@ -541,8 +513,7 @@ template <typename Specialization, typename Normalize>
 bool MaterializeIndirectDescriptor(const ResourcePlan&                         program,
                                    uint32_t source_index,
                                    uint32_t resource_index, uint32_t dword_count, uint32_t read_dwords,
-                                   const SrtRuntime& runtime, SrtWalker& clean,
-                                   ResourceSnapshot&             snapshot,
+                                   const SrtRuntime& runtime, ResourceSnapshot& snapshot,
                                    std::vector<DescriptorValue>& descriptors,
                                    std::vector<Specialization>&  specializations,
                                    Normalize&& normalize) {
@@ -605,8 +576,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		if (selector == nullptr) keys.clear();
 		DescriptorValue material_value;
 		if ((selector != nullptr && (byte_keys || !selector_keys->resolved) &&
-		     !clean.EvaluateDescriptor(selector->source, material_value)) ||
-		    !clean.EvaluateDescriptor(indirect.table_source, table_value)) {
+		     !WalkedDescriptor(program, selector->source, material_value)) ||
+		    !WalkedDescriptor(program, indirect.table_source, table_value)) {
 			return false;
 		}
 		if (table_value.dword_count == 2u) {
@@ -659,7 +630,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			                                       : uint64_t {1} << 32u;
 			if (!indirect.key_count.IsEmpty()) {
 				uint32_t key_count = 0u;
-				if (!EvaluateIndirectCount(indirect, runtime, clean, key_count)) return false;
+				if (!EvaluateIndirectCount(program, indirect, runtime, key_count)) return false;
 				domain = std::min(domain, uint64_t {key_count});
 			}
 			if (indexed_pointer &&
@@ -771,8 +742,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		if (selector == nullptr) {
 			uint32_t key_count = 0;
 			uint32_t first_key = 0;
-			if (!EvaluateIndirectCount(indirect, runtime, clean, key_count)) return false;
-			if (!indirect.selector_first.IsEmpty() && !clean.Evaluate(indirect.selector_first, first_key))
+			if (!EvaluateIndirectCount(program, indirect, runtime, key_count)) return false;
+			if (!indirect.selector_first.IsEmpty() && !WalkedValue(program, indirect.selector_first, first_key))
 				return false;
 			if (indirect.table_stride == 0u || key_count > MaxIndirectDescriptorProbes ||
 			    (indirect.selector_first.IsEmpty() && key_count != 0u &&
@@ -788,7 +759,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		} else {
 			if (!indirect.selector_mask.IsEmpty() && table_value.dword_count != 2u) return false;
 			if (!selector_keys->resolved) {
-				if (!ResolveSelectorKeys(indirect, material_value, runtime, clean, keys)) return false;
+				if (!ResolveSelectorKeys(program, indirect, material_value, runtime, keys)) return false;
 				selector_keys->resolved = true;
 			}
 		}
@@ -827,7 +798,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			                     std::span(candidate.dwords).first(read_dwords))) {
 				return read_failed("table descriptor read", key, 0u, table_base, table_offset, descriptor_bytes);
 			}
-		} else if (!clean.EvaluateDescriptor(sources[entry], candidate)) {
+		} else if (!WalkedDescriptor(program, sources[entry], candidate)) {
 			return false;
 		}
 		if (!normalize(candidate))
@@ -1341,6 +1312,17 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 	return result;
 }
 
+// State an extracted or loaded plan derives from its fields: one key set per selector group, which
+// extraction numbers densely in order of first use (false when they are not), and the walker.
+static bool DerivePlanState(ResourcePlan& plan) {
+	for (const auto group: plan.selector_groups) {
+		if (group == plan.selector_keys.size()) plan.selector_keys.emplace_back();
+		if (group != UINT32_MAX && group >= plan.selector_keys.size()) return false;
+	}
+	if (plan.srt_plan_complete && plan.resource_tracking_complete) CompileSrtWalker(plan);
+	return true;
+}
+
 ResourcePlan ExtractResourcePlan(const Program& program) {
 	ResourcePlan plan;
 	plan.stage                      = program.stage;
@@ -1380,16 +1362,13 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		if (const auto found = cloned.find(source); found != cloned.end()) {
 			return Value(found->second);
 		}
-		if (source->GetOpcode() == ValueOpcode::LoadBufferU32) {
-			plan.requires_specialization_memory = true;
-			plan.capture_specialization_reads   = true;
-		}
+		if (source->GetOpcode() == ValueOpcode::LoadBufferU32) plan.capture_specialization_reads = true;
 		auto& target =
 		    plan.value_storage.emplace_back(source->GetOpcode(), source->Flags<uint64_t>());
 		if (source->GetOpcode() == ValueOpcode::LoadAddressU32 ||
 		    source->GetOpcode() == ValueOpcode::ReadConstBuffer ||
 		    source->GetOpcode() == ValueOpcode::LoadBufferU32) {
-			target.SetFlags(SrtReadFlags {.index = source->Flags<MemoryFlags>().index});
+			target.SetFlags(MemoryFlags {.index = source->Flags<MemoryFlags>().index});
 		}
 		cloned.emplace(source, &target);
 		if (source->GetOpcode() == ValueOpcode::Phi) {
@@ -1411,7 +1390,6 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		target.dword_count    = source.dword_count;
 		target.indirect_descriptor = source.indirect_descriptor;
 		if (target.indirect_descriptor.has_value()) {
-			plan.requires_specialization_memory = true;
 			capture_indirect_reads = true;
 			target.indirect_descriptor->key_count = Clone(target.indirect_descriptor->key_count);
 			target.indirect_descriptor->selector_first =
@@ -1425,17 +1403,18 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	}
 	// Destination tables share keys when their complete selector domains agree.
 	// Resolve those domains lazily at dispatch time, without repeated guest reads.
+	std::vector<uint32_t> group_sources; // The first source of each group.
 	plan.selector_groups.resize(plan.descriptor_sources.size(), UINT32_MAX);
 	for (uint32_t source = 0; source < plan.descriptor_sources.size(); ++source) {
 		const auto& descriptor = program.descriptor_sources[source].indirect_descriptor;
 		if (!descriptor || !descriptor->selector) continue;
 		const auto& a = *descriptor;
 		uint32_t group = 0;
-		for (; group < plan.selector_keys.size(); ++group) {
-			const auto& b = *program.descriptor_sources[plan.selector_keys[group].source].indirect_descriptor;
+		for (; group < group_sources.size(); ++group) {
+			const auto& b = *program.descriptor_sources[group_sources[group]].indirect_descriptor;
 			if (EquivalentSelectorDomain(program, a, b)) break;
 		}
-		if (group == plan.selector_keys.size()) plan.selector_keys.push_back({.source = source});
+		if (group == group_sources.size()) group_sources.push_back(source);
 		plan.selector_groups[source] = group;
 	}
 	plan.srt_reads.reserve(program.srt_reads.size());
@@ -1523,57 +1502,16 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	plan.uniform_fill.element_count = Clone(plan.uniform_fill.element_count);
 	plan.capture_specialization_reads |=
 	    capture_indirect_reads || capture_control_reads || !plan.source_reads.empty();
-	if (plan.capture_specialization_reads) {
-		// Every former flat alias now targets its retained raw read. Mark those reads
-		// on the normalized graph, so discarded Phi/EXEC dependencies stay discarded.
-		std::unordered_set<const Inst*> planned_reads;
-		for (const auto& read: plan.srt_reads) {
-			auto* inst = read.value.ResolveInstruction();
-			EXIT_IF(inst == nullptr || (inst->GetOpcode() != ValueOpcode::LoadAddressU32 &&
-			                           inst->GetOpcode() != ValueOpcode::ReadConstBuffer));
-			planned_reads.insert(inst);
-		}
-		for (const auto& source: plan.descriptor_sources) {
-			if (!source.indirect_descriptor.has_value()) continue;
-			const auto& indirect = *source.indirect_descriptor;
-			if (indirect.selector) {
-				MarkCleanReads(planned_reads, Source(plan, indirect.selector->source));
-				if (indirect.selector->rows)
-					MarkCleanReads(planned_reads, Source(plan, indirect.selector->rows->source));
-			}
-			MarkCleanReads(planned_reads, nullptr, indirect.selector_mask);
-			MarkCleanReads(planned_reads, nullptr, indirect.selector_first);
-			MarkCleanReads(planned_reads, nullptr, indirect.key_count);
-			if (indirect.sources.empty()) {
-				MarkCleanReads(planned_reads, Source(plan, indirect.table_source));
-			} else {
-				for (const auto candidate: indirect.sources) {
-					MarkCleanReads(planned_reads, Source(plan, candidate));
-				}
-			}
-		}
-		// Clean writable addresses prove that shader writes cannot overlap captured reads.
-		for (const auto& buffer: plan.info.buffers) {
-			if (buffer.written) MarkCleanReads(planned_reads, Source(plan, buffer.source));
-		}
-		for (const auto& image: plan.info.images) {
-			if (image.written) MarkCleanReads(planned_reads, Source(plan, image.source));
-		}
-	}
 	if (capture_indirect_reads || !plan.source_reads.empty())
 		plan.resource_tracking_complete &= !program.has_address_writes;
-	CompileSrtWalker(plan);
+	DerivePlanState(plan);
 	return plan;
 }
 
 bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime,
                           ResourceSnapshot& snapshot, ResourceSpecialization& specialization) {
 	for (auto& keys: program.selector_keys) keys.resolved = false;
-	if (!program.resource_tracking_complete ||
-	    ((program.requires_specialization_memory || !program.source_reads.empty()) &&
-	     runtime.read_specialization_memory == nullptr)) {
-		return false;
-	}
+	if (!program.resource_tracking_complete) return false;
 	const auto uses_workgroup_count = [](const DescriptorSource& source) {
 		return source.indirect_descriptor && source.indirect_descriptor->workgroup_axis != UINT32_MAX;
 	};
@@ -1582,20 +1520,16 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	     std::ranges::any_of(program.descriptor_sources, uses_workgroup_count));
 	auto& reads = snapshot.specialization_reads;
 	reads = program.source_reads;
-	SrtReadCapture capture {runtime, reads};
-	auto observed = capture_reads ? capture.ObservedRuntime() : runtime;
-	SrtWalker clean(program, CleanRuntime(observed));
-	SrtWalker walker(program, observed,
-	                 capture_reads || program.requires_specialization_memory ? &clean : nullptr);
-	if (!walker.RefreshFlatBuffer(snapshot.flattened_srt)) {
+	if (!WalkSrt(program, runtime, snapshot.flattened_srt, capture_reads ? &reads : nullptr)) {
 		return SpecializationFail(fmt::format("SRT refresh complete={} reads={} user_data={}",
 		    program.srt_plan_complete, program.srt_reads.size(), runtime.user_data.size()));
 	}
-	const auto active = std::span<const uint8_t>(program.active_sources);
+	SrtReadCapture capture {runtime, reads};
+	auto observed = capture_reads ? capture.ObservedRuntime() : runtime;
 	std::array<uint32_t, 3> workgroup_counts {};
 	if (runtime.indirect_args_address != 0u) {
 		for (uint32_t source = 0; source < program.descriptor_sources.size(); ++source) {
-			if ((!active.empty() && !active[source]) ||
+			if (!WalkedActive(program, source) ||
 			    !uses_workgroup_count(program.descriptor_sources[source])) continue;
 			if ((runtime.indirect_args_address & 3u) != 0u ||
 			    !ReadTable(runtime.indirect_args_address, sizeof(workgroup_counts), 0u,
@@ -1610,24 +1544,24 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	std::array<uint32_t, 4> stored {};
 	bool uniform_fill = words != 0;
 	if (uniform_fill && !fill.pattern_size.IsEmpty())
-		uniform_fill = clean.Evaluate(fill.pattern_size, words) && words >= 1 && words <= 4;
+		uniform_fill = WalkedValue(program, fill.pattern_size, words) && words >= 1 && words <= 4;
 	uint32_t elements = 0;
 	if (uniform_fill && !fill.element_count.IsEmpty())
-		uniform_fill = clean.Evaluate(fill.element_count, elements) && elements != 0;
+		uniform_fill = WalkedValue(program, fill.element_count, elements) && elements != 0;
 	for (uint32_t i = 0; i < words && uniform_fill; ++i) {
-		uniform_fill = clean.Evaluate(fill.values[i], stored[i]) && stored[i] == stored[0];
+		uniform_fill = WalkedValue(program, fill.values[i], stored[i]) && stored[i] == stored[0];
 	}
 	if (uniform_fill) {
 		snapshot.uniform_fill = fill.fill;
 		snapshot.uniform_fill.value = stored[0];
 		snapshot.uniform_fill.bounded_elements = elements;
 	}
-	const auto evaluate = [&](uint32_t source, DescriptorValue& value, bool written = false) {
+	const auto evaluate = [&](uint32_t source, DescriptorValue& value) {
 		if (source >= program.descriptor_sources.size()) {
 			return false;
 		}
-		if (active.empty() || active[source]) {
-			return (capture_reads && written ? clean : walker).EvaluateDescriptor(source, value) ||
+		if (WalkedActive(program, source)) {
+			return WalkedDescriptor(program, source, value) ||
 			    SpecializationFail(fmt::format("descriptor evaluation source={}", source));
 		}
 		value = {};
@@ -1651,9 +1585,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			if (source == nullptr) return false;
 			if (source->indirect_descriptor.has_value()) {
 				snapshot.buffers[i] = {.dword_count = 4u};
-				if (active.empty() || active[base.source]) {
+				if (WalkedActive(program, base.source)) {
 					if (!MaterializeIndirectDescriptor(
-					        program, base.source, i, 4u, 4u, observed, clean, snapshot,
+					        program, base.source, i, 4u, 4u, observed, snapshot,
 					        snapshot.buffers, specialization.buffers,
 					        [&](DescriptorValue& value) {
 						        return NormalizeIndirectBuffer(
@@ -1661,7 +1595,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 					        }))
 						return resource_failed("buffer", i, base.source, snapshot.buffers.size());
 				}
-			} else if (!evaluate(base.source, snapshot.buffers[i], base.written)) {
+			} else if (!evaluate(base.source, snapshot.buffers[i])) {
 				return false;
 			}
 		}
@@ -1711,13 +1645,13 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		}
 		if (source->indirect_descriptor.has_value()) {
 			snapshot.images[i] = {.dword_count = 8u};
-			if (!active.empty() && !active[image.source]) {
+			if (!WalkedActive(program, image.source)) {
 				continue;
 			}
 			const auto& indirect = *source->indirect_descriptor;
 			const bool bounded_table = IsBoundedDescriptorTable(program, indirect);
 			if (!MaterializeIndirectDescriptor(
-			        program, image.source, i, 8u, image.r128 ? 4u : 8u, observed, clean, snapshot, snapshot.images,
+			        program, image.source, i, 8u, image.r128 ? 4u : 8u, observed, snapshot, snapshot.images,
 			        specialization.images, [&](DescriptorValue& value) {
 				        // A broad heap also contains resources for other typed image operations.
 				        if (NullImageDescriptor(value) || !ValidImageDescriptor(value, image.r128) ||
@@ -1728,7 +1662,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 				return resource_failed("image", i, image.source, snapshot.images.size());
 			}
 		} else {
-			if (!evaluate(image.source, snapshot.images[i], image.written)) {
+			if (!evaluate(image.source, snapshot.images[i])) {
 				return false;
 			}
 			if (!ValidImageDescriptor(snapshot.images[i], image.r128)) {
@@ -1749,8 +1683,8 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			const auto* indirect = descriptor->indirect_descriptor ? &*descriptor->indirect_descriptor : nullptr;
 			if (indirect != nullptr && indirect->sources.empty()) {
 				snapshot.samplers[i] = {.dword_count = 4u};
-				if ((active.empty() || active[source]) && !MaterializeIndirectDescriptor(
-				        program, source, i, 4u, 4u, observed, clean, snapshot, snapshot.samplers,
+				if (WalkedActive(program, source) && !MaterializeIndirectDescriptor(
+				        program, source, i, 4u, 4u, observed, snapshot, snapshot.samplers,
 				        specialization.samplers,
 				        [](DescriptorValue&) { return true; }))
 					return resource_failed("sampler", i, source, snapshot.samplers.size());
@@ -2065,4 +1999,86 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 	image_remap.Apply(images);
 }
 
+namespace {
+
+using ControlPhis = decltype(ResourcePlan::control_phi_predecessors);
+
+// Instructions come first, so that any value can refer to them by index, then their arguments.
+// Plan Phis have no blocks; a control Phi lists the plan blocks of its arguments after them.
+void IoValues(Writer& writer, const std::list<Inst>& values, const ControlPhis& phis) {
+	IoAll(writer, static_cast<uint32_t>(values.size()));
+	for (const auto& inst: values) {
+		writer.index.emplace(&inst, static_cast<uint32_t>(writer.index.size()));
+		IoAll(writer, inst.GetOpcode(), inst.Flags<uint64_t>());
+	}
+	for (const auto& inst: values) {
+		if (inst.GetOpcode() != ValueOpcode::Phi) {
+			for (size_t index = 0; index < inst.NumArgs(); ++index) IoAll(writer, inst.Arg(index));
+			continue;
+		}
+		std::vector<Value> args;
+		for (size_t index = 0; index < inst.NumArgs(); ++index) {
+			EXIT_IF(inst.PhiBlock(index) != nullptr);
+			args.push_back(inst.Arg(index));
+		}
+		const auto found = phis.find(&inst);
+		IoAll(writer, args, found != phis.end() ? found->second : std::vector<uint32_t> {});
+	}
+}
+
+void IoValues(Reader& reader, std::list<Inst>& values, ControlPhis& phis) {
+	uint32_t count = 0;
+	Io(reader, count);
+	count = reader.Count(count);
+	for (uint32_t index = 0; index < count && reader.ok; ++index) {
+		ValueOpcode opcode {};
+		uint64_t    flags = 0;
+		IoAll(reader, opcode, flags);
+		reader.ok = reader.ok && opcode < ValueOpcode::Count;
+		if (reader.ok) reader.insts.push_back(&values.emplace_back(opcode, flags));
+	}
+	for (auto* inst: reader.insts) {
+		if (inst->GetOpcode() != ValueOpcode::Phi) {
+			for (size_t index = 0; index < inst->NumArgs(); ++index) {
+				Value arg;
+				Io(reader, arg);
+				inst->SetArg(index, arg);
+			}
+			continue;
+		}
+		std::vector<Value>    args;
+		std::vector<uint32_t> predecessors;
+		IoAll(reader, args, predecessors);
+		for (const auto arg: args) inst->AddPhiOperand(nullptr, arg);
+		if (!predecessors.empty()) phis.emplace(inst, std::move(predecessors));
+	}
+}
+
+// Binds every plan field: one added later fails to compile here until it is stored or derived.
+template <class S, class P> void IoPlan(S& io, P& plan) {
+	auto& [stage, shader_hash, user_data_base, user_data_count, values, memory_info, descriptor_sources,
+	       selector_groups, control_flow, control_phis, srt_reads, source_reads, capture_specialization_reads,
+	       srt_plan_complete, resource_tracking_complete, info, uniform_fill,
+	       // Derived on load.
+	       srt_code, srt_user_data, srt_captures, srt_scratch, material_keys, selector_keys] = plan;
+	IoValues(io, values, control_phis);
+	IoAll(io, stage, shader_hash, user_data_base, user_data_count, memory_info, descriptor_sources,
+	      selector_groups, control_flow, srt_reads, source_reads, capture_specialization_reads,
+	      srt_plan_complete, resource_tracking_complete, info, uniform_fill);
+}
+
+} // namespace
+
+std::vector<uint8_t> SaveResourcePlan(const ResourcePlan& plan) {
+	Writer writer;
+	IoPlan(writer, plan);
+	return std::move(writer.bytes);
+}
+
+bool LoadResourcePlan(std::span<const uint8_t> bytes, ResourcePlan& plan) {
+	Reader reader {bytes};
+	IoPlan(reader, plan);
+	return reader.ok && reader.bytes.empty() && DerivePlanState(plan);
+}
+
 } // namespace Libs::Graphics::ShaderRecompiler::IR

src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.h

@@ -53,6 +53,11 @@ struct ResourceSpecialization {
 // its values and is independent of the translated shader CFG.
 ResourcePlan ExtractResourcePlan(const Program& program);
 
+// Shader cache record of a plan; loading it into a default-constructed plan rebuilds the walker
+// and scratch state like extraction. A failed load leaves a plan that must not be used.
+std::vector<uint8_t> SaveResourcePlan(const ResourcePlan& plan);
+bool                 LoadResourcePlan(std::span<const uint8_t> bytes, ResourcePlan& plan);
+
 // Refreshes cached resources and specialization in place. A failed refresh must not be used.
 bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime,
                           ResourceSnapshot& snapshot, ResourceSpecialization& specialization);

src/graphics/shader/recompiler/ir/passes/SrtWalker.h

@@ -8,8 +8,6 @@
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-class Value;
-
 using SrtMemoryReader = bool (*)(void* userdata, uint64_t address, std::span<uint32_t> values);
 
 // Image descriptors at base + first + k * step, k < count, within size bytes that a sampled
@@ -42,7 +40,6 @@ protected:
 struct SrtRuntime {
 	std::span<const uint32_t> user_data;
 	uint64_t                  shader_base                = 0;
-	SrtMemoryReader           read_memory                = nullptr;
 	void*                     userdata                   = nullptr;
 	SrtMemoryReader           read_specialization_memory = nullptr;
 	std::span<const uint32_t> workgroup_counts;
@@ -57,20 +54,12 @@ public:
 	SrtRuntime ObservedRuntime();
 
 private:
-	static bool ReadStrict(void* userdata, uint64_t address, std::span<uint32_t> values);
-	static bool ReadOrdinary(void* userdata, uint64_t address, std::span<uint32_t> values);
+	static bool Read(void* userdata, uint64_t address, std::span<uint32_t> values);
 
 	SrtRuntime                                  m_source;
 	std::vector<std::pair<uint64_t, uint64_t>>& m_ranges;
 };
 
-// Retained reads no longer need the guest instruction PC. A clean read evaluates
-// its address and bounds through the strict reader as well as the final DWORD.
-struct SrtReadFlags {
-	uint32_t index = 0;
-	uint32_t clean = 0;
-};
-
 enum class RuntimeValueType { Any, Integer, ImmutableInteger };
 
 bool ValidateRuntimeValue(const ResourcePlan& program, Value value,
@@ -78,9 +67,20 @@ bool ValidateRuntimeValue(const ResourcePlan& program, Value value,
 // Uniform across a compute wave, including values derived from its workgroup ID.
 // This does not imply that the host can evaluate the value.
 bool ValidateWaveUniformValue(const ResourcePlan& program, Value value);
-void CompileSrtWalker(ResourcePlan& program);
-// Uses the strict reader for values that affect shader specialization.
-SrtRuntime CleanRuntime(SrtRuntime runtime);
+// Compiles the plan's x86 walker. Without values it walks the plan's resources; otherwise it
+// only evaluates the values, with every block reachable.
+void CompileSrtWalker(ResourcePlan& plan, std::span<const Value> values = {});
+// Runs the walker. flat receives the scalar reads of reachable blocks and reads, when non-null,
+// one DWORD range per guest load. False when a reachable read is unknown.
+bool WalkSrt(const ResourcePlan& plan, const SrtRuntime& runtime, std::vector<uint32_t>& flat,
+             std::vector<std::pair<uint64_t, uint64_t>>* reads);
+// Results of the last walk; false when unknown.
+bool WalkedValue(const ResourcePlan& plan, Value value, uint32_t& result);
+bool WalkedDescriptor(const ResourcePlan& plan, uint32_t source, DescriptorValue& result);
+bool WalkedActive(const ResourcePlan& plan, uint32_t source);
+// Length of the guest load at rip if it belongs to the walker running on this thread, else 0.
+// The exception handler completes such a load as a zero read.
+uint32_t SrtWalkerLoadLength(uint64_t rip);
 
 // Nonnegative affine coefficients: element zero is constant, remaining elements are
 // caller-selected leaves. Runtime range proofs must additionally exclude U32 wrap.
@@ -124,43 +124,6 @@ std::optional<std::array<uint64_t, Size>> AffineIndex(
 	return left;
 }
 
-// One memoized evaluation session shared by the entire shader resource refresh.
-class SrtWalker {
-public:
-	SrtWalker(const ResourcePlan& program, const SrtRuntime& runtime,
-	          SrtWalker* clean_evaluator = nullptr);
-	~SrtWalker();
-	SrtWalker(const SrtWalker&)            = delete;
-	SrtWalker& operator=(const SrtWalker&) = delete;
-
-	bool Evaluate(Value value, uint32_t& result);
-	bool EvaluateDescriptor(uint32_t source, DescriptorValue& result);
-	// Refreshes reachable scalar reads and active descriptor sources in one walk.
-	bool RefreshFlatBuffer(std::vector<uint32_t>& flat);
-
-private:
-	SrtWalker(SrtWalker& parent, Value active_mask, bool clean = false,
-	          SrtWalker* clean_evaluator = nullptr);
-	ResourcePlan::EvaluationContext& AcquireContext();
-	void ResetUnknowns();
-	static float Float32(uint64_t bits);
-	bool EvaluateWide(Value value, uint64_t& result);
-	bool Arg(const Inst& inst, size_t index, uint64_t& result);
-	bool EvaluatePhi(const Inst& inst, uint64_t& result);
-	bool EvaluateExtract(const Inst& inst, uint64_t& result);
-	bool EvaluateRawRead(const Inst& inst, uint64_t& result);
-	bool EvaluateInst(const Inst& inst, uint64_t& result);
-
-	const ResourcePlan&              m_program;
-	SrtRuntime                      m_runtime;
-	SrtWalker*                      m_clean_evaluator = nullptr;
-	Value                           m_active_mask;
-	const std::vector<uint8_t>*      m_reachable_blocks = nullptr;
-	const SrtWalker*                m_session;
-	uint8_t                         m_read_policy;
-	ResourcePlan::EvaluationContext& m_context;
-};
-
 } // namespace Libs::Graphics::ShaderRecompiler::IR
 
 #endif /* EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SRTWALKER_H_ */

src/graphics/shader/rectListShader.cpp

@@ -18,45 +18,9 @@ using ShaderRecompiler::Spirv::Builder;
 
 constexpr uint32_t SpirvVersion15 = 0x00010500u;
 
-struct Parameter {
-	uint32_t input_location  = 0;
-	uint32_t output_location = 0;
-	bool     flat            = false;
-};
-
-std::vector<Parameter> GetParameters(const ShaderVertexInputInfo& vertex_info,
-                                     const ShaderPixelInputInfo*  pixel_info) {
-	if (pixel_info == nullptr) {
-		return {};
-	}
-	EXIT_IF(pixel_info->input_num > ShaderVertexInputInfo::RES_MAX);
-	EXIT_IF(pixel_info->stage.program == nullptr);
-
-	std::vector<uint32_t> active_inputs;
-	for (const auto& input: pixel_info->stage.program->info.inputs) {
-		if (input.kind == ShaderRecompiler::IR::StageInputKind::Parameter) {
-			active_inputs.push_back(input.location);
-		}
-	}
-
-	std::vector<Parameter> parameters;
-	std::array<bool, 32> output_locations {};
-	for (const auto input: active_inputs) {
-		const auto input_location = ShaderPixelParameterMappedLocation(*pixel_info, input);
-		const auto output_location = ShaderPixelParameterLocation(*pixel_info, active_inputs, input);
-		if ((vertex_info.stage.program->param_export_mask & (1u << input_location)) != 0 &&
-		    !output_locations[output_location]) {
-			parameters.push_back({input_location, output_location,
-			                      ShaderPixelParameterIsFlat(*pixel_info, input)});
-			output_locations[output_location] = true;
-		}
-	}
-	return parameters;
-}
-
 class RectListEmitter {
 public:
-	RectListEmitter(const std::vector<Parameter>& parameters_, spv::ExecutionModel model)
+	RectListEmitter(const std::vector<RectListParameter>& parameters_, spv::ExecutionModel model)
 	    : parameters(parameters_) {
 		builder.AddMemoryModel(spv::AddressingModelLogical, spv::MemoryModelGLSL450);
 
@@ -315,42 +279,68 @@ private:
 		              Result(spv::OpFAdd, vec4_float_type, p1, p2));
 	}
 
-	Builder                       builder {SpirvVersion15};
-	const std::vector<Parameter>& parameters;
-	std::vector<uint32_t>         interfaces;
-	std::vector<uint32_t>         inputs;
-	std::vector<uint32_t>         outputs;
-	uint32_t                      main                  = 0;
-	uint32_t                      void_type             = 0;
-	uint32_t                      bool_type             = 0;
-	uint32_t                      uint_type             = 0;
-	uint32_t                      int_type              = 0;
-	uint32_t                      float_type            = 0;
-	uint32_t                      vec2_bool_type        = 0;
-	uint32_t                      vec2_float_type       = 0;
-	uint32_t                      vec3_float_type       = 0;
-	uint32_t                      vec4_float_type       = 0;
-	uint32_t                      function_type         = 0;
-	uint32_t                      per_vertex_type       = 0;
-	uint32_t                      ptr_input_float       = 0;
-	uint32_t                      ptr_input_vec4_float  = 0;
-	uint32_t                      ptr_output_float      = 0;
-	uint32_t                      ptr_output_vec4_float = 0;
-	uint32_t                      gl_in                 = 0;
-	uint32_t                      gl_out                = 0;
-	uint32_t                      tess_inner            = 0;
-	uint32_t                      tess_outer            = 0;
-	uint32_t                      tess_coord            = 0;
-	uint32_t                      invocation_id         = 0;
+	Builder                               builder {SpirvVersion15};
+	const std::vector<RectListParameter>& parameters;
+	std::vector<uint32_t>                 interfaces;
+	std::vector<uint32_t>                 inputs;
+	std::vector<uint32_t>                 outputs;
+	uint32_t                              main                  = 0;
+	uint32_t                              void_type             = 0;
+	uint32_t                              bool_type             = 0;
+	uint32_t                              uint_type             = 0;
+	uint32_t                              int_type              = 0;
+	uint32_t                              float_type            = 0;
+	uint32_t                              vec2_bool_type        = 0;
+	uint32_t                              vec2_float_type       = 0;
+	uint32_t                              vec3_float_type       = 0;
+	uint32_t                              vec4_float_type       = 0;
+	uint32_t                              function_type         = 0;
+	uint32_t                              per_vertex_type       = 0;
+	uint32_t                              ptr_input_float       = 0;
+	uint32_t                              ptr_input_vec4_float  = 0;
+	uint32_t                              ptr_output_float      = 0;
+	uint32_t                              ptr_output_vec4_float = 0;
+	uint32_t                              gl_in                 = 0;
+	uint32_t                              gl_out                = 0;
+	uint32_t                              tess_inner            = 0;
+	uint32_t                              tess_outer            = 0;
+	uint32_t                              tess_coord            = 0;
+	uint32_t                              invocation_id         = 0;
 };
 
 } // namespace
 
-RectListShaders BuildRectListShaders(const ShaderVertexInputInfo& vertex_info,
-                                     const ShaderPixelInputInfo*  pixel_info) {
-	const auto      parameters = GetParameters(vertex_info, pixel_info);
-	RectListEmitter control(parameters, spv::ExecutionModelTessellationControl);
-	RectListEmitter evaluation(parameters, spv::ExecutionModelTessellationEvaluation);
+std::vector<RectListParameter> RectListParameters(
+    const ShaderPixelInputInfo& pixel_info, const ShaderRecompiler::IR::CompiledShaderInfo& program) {
+	EXIT_IF(pixel_info.input_num > ShaderVertexInputInfo::RES_MAX);
+	std::vector<uint32_t> active_inputs;
+	for (const auto& input: program.info.inputs) {
+		if (input.kind == ShaderRecompiler::IR::StageInputKind::Parameter) {
+			active_inputs.push_back(input.location);
+		}
+	}
+	std::vector<RectListParameter> parameters;
+	for (const auto input: active_inputs) {
+		parameters.push_back({ShaderPixelParameterMappedLocation(pixel_info, input),
+		                      ShaderPixelParameterLocation(pixel_info, active_inputs, input),
+		                      ShaderPixelParameterIsFlat(pixel_info, input)});
+	}
+	return parameters;
+}
+
+RectListShaders BuildRectListShaders(std::span<const RectListParameter> parameters,
+                                     uint32_t                           param_export_mask) {
+	std::vector<RectListParameter> exported;
+	std::array<bool, 32>           output_locations {};
+	for (const auto& parameter: parameters) {
+		if ((param_export_mask & (1u << parameter.input_location)) != 0 &&
+		    !output_locations[parameter.output_location]) {
+			exported.push_back(parameter);
+			output_locations[parameter.output_location] = true;
+		}
+	}
+	RectListEmitter control(exported, spv::ExecutionModelTessellationControl);
+	RectListEmitter evaluation(exported, spv::ExecutionModelTessellationEvaluation);
 	return {control.EmitControl(), evaluation.EmitEvaluation()};
 }
 

src/graphics/shader/rectListShader.h

@@ -2,20 +2,35 @@
 #define EMULATOR_SRC_GRAPHICS_SHADER_RECTLISTSHADER_H_
 
 #include <cstdint>
+#include <span>
 #include <vector>
 
 namespace Libs::Graphics {
 
 struct ShaderPixelInputInfo;
-struct ShaderVertexInputInfo;
+
+namespace ShaderRecompiler::IR {
+struct CompiledShaderInfo;
+} // namespace ShaderRecompiler::IR
+
+// A pixel parameter the generated stages forward from vertex export input_location.
+struct RectListParameter {
+	uint32_t input_location  = 0;
+	uint32_t output_location = 0;
+	bool     flat            = false;
+};
 
 struct RectListShaders {
 	std::vector<uint32_t> control;
 	std::vector<uint32_t> evaluation;
 };
 
-RectListShaders BuildRectListShaders(const ShaderVertexInputInfo& vertex_info,
-                                     const ShaderPixelInputInfo*  pixel_info);
+// Every parameter a rectangle-list pixel program reads.
+std::vector<RectListParameter> RectListParameters(
+    const ShaderPixelInputInfo& pixel_info, const ShaderRecompiler::IR::CompiledShaderInfo& program);
+// Forwards the parameters the vertex stage exports, one per output location.
+RectListShaders BuildRectListShaders(std::span<const RectListParameter> parameters,
+                                     uint32_t                           param_export_mask);
 
 } // namespace Libs::Graphics
 

src/graphics/shader/shader.cpp

@@ -11,7 +11,6 @@
 #include "graphics/guest_gpu/hardwareContext.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
-#include "graphics/shader/recompiler/Tessellation.h"
 #include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 #include "graphics/shader/shaderCompiler.h"
 #include "graphics/shader/shaderVertexMetadata.h"
@@ -40,15 +39,19 @@
 
 namespace Libs::Graphics {
 
+// A mapped shader and its vertex metadata, parsed once when it is mapped.
+struct ShaderMapData: ShaderMappedData {
+	ShaderVertexMetadata vertex;
+	const char*          vertex_error = nullptr;
+};
+
 struct ShaderMapEntry {
-	ShaderMappedData data;
-	uint64_t hash;
+	ShaderMapData data;
+	uint64_t      hash;
 };
 
 static std::unique_ptr<std::unordered_map<uint64_t, ShaderMapEntry>> g_shader_map;
 static std::mutex                                                      g_shader_map_mutex;
-static std::unordered_map<uint64_t, ShaderTessellationInputInfo>       g_tess_layouts;
-static std::mutex                                                      g_tess_layouts_mutex;
 
 void ShaderInit() {
 	EXIT_IF(g_shader_map != nullptr);
@@ -64,8 +67,11 @@ void ShaderMapUserData(uint64_t addr, const ShaderMappedData& data) {
 		     addr, data.code_size_bytes);
 	}
 	const auto hash = XXH3_64bits(reinterpret_cast<const void*>(addr), data.code_size_bytes);
+	ShaderMapData mapped {data};
+	mapped.vertex_error =
+	    ShaderReadVertexMetadata(data, HW::UserSgprInfo::SGPRS_MAX, mapped.vertex);
 	std::scoped_lock lock(g_shader_map_mutex);
-	(*g_shader_map)[addr] = {data, hash};
+	(*g_shader_map)[addr] = {mapped, hash};
 }
 
 static ShaderMapEntry ShaderGetMappedData(uint64_t addr, const char* label) {
@@ -261,8 +267,8 @@ static uint32_t ShaderCalcPsSystemInputBase(const HW::ShaderRegisters& regs) {
 
 static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserSgprInfo& user_sgpr,
                                            uint32_t user_sgpr_num, const HW::ShaderRegisters& sh,
-                                           const ShaderMappedData& data,
-                                           ShaderVertexInputInfo&  info) {
+                                           const ShaderMapData&   data,
+                                           ShaderVertexInputInfo& info) {
 	KYTY_PROFILER_FUNCTION();
 
 	info = {};
@@ -277,13 +283,12 @@ static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserS
 		     " user_sgpr_num=%u\n",
 		     shader_addr, user_sgpr_num);
 	}
-	ShaderVertexMetadata metadata;
-	std::string          metadata_error;
-	if (!ShaderReadVertexMetadata(data, HW::UserSgprInfo::SGPRS_MAX, metadata, &metadata_error)) {
+	if (data.vertex_error != nullptr) {
 		LOGF("ShaderGetVertexInputInfo(): invalid AGC metadata shader=0x%016" PRIx64 ": %s\n",
-		     shader_addr, metadata_error.c_str());
+		     shader_addr, data.vertex_error);
 		return false;
 	}
+	const auto& metadata = data.vertex;
 
 	if (metadata.vertex_attrib_reg >= 0) {
 		info.fetch_external   = false;
@@ -377,7 +382,6 @@ static void ShaderGetStaticInputInfoPS(
 }
 
 static void ShaderGetStaticInputInfoCS(const HW::ComputeShaderInfo& regs,
-                                       const HW::ShaderRegisters& /*sh*/,
                                        const ShaderMappedData& data, ShaderComputeInputInfo& info) {
 	const bool async_compute             = info.async_compute;
 	const bool dispatch_thread_dimensions = info.dispatch_thread_dimensions;
@@ -655,18 +659,6 @@ PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context&
 	};
 	EXIT_IF(tess.input_control_points == 0 || tess.input_control_points > 32 ||
 	        tess.output_control_points == 0 || tess.output_control_points > 32);
-	// Reflection decodes both programs; the key also covers the VGT fields cached in tess.
-	const uint64_t key[] = {local_hash, control_hash,
-	                        uint64_t {sh.m_vgtLsHsConfig & 0xfff00u} << 32u | sh.m_vgtTfParam};
-	{
-		std::scoped_lock lock(g_tess_layouts_mutex);
-		auto [entry, inserted] = g_tess_layouts.try_emplace(XXH3_64bits(key, sizeof(key)), tess);
-		if (inserted) {
-			ShaderRecompiler::AnalyzeTessellationPrograms(params[0].code, params[1].code,
-			                                              entry->second);
-		}
-		tess = entry->second;
-	}
 	for (auto& stage: input_info) {
 		stage.tess = tess;
 	}
@@ -684,10 +676,9 @@ ShaderParams PrepareProgram(
 	    std::span<const uint32_t>(regs.ps_user_sgpr.value, regs.ps_regs.rsrc2.user_sgpr), data);
 }
 
-ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, const HW::ShaderRegisters& sh,
-                            ShaderComputeInputInfo& info) {
+ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, ShaderComputeInputInfo& info) {
 	const auto [data, hash] = ShaderGetMappedData(regs.cs_regs.data_addr, "ShaderGetInputInfoCS():");
-	ShaderGetStaticInputInfoCS(regs, sh, data, info);
+	ShaderGetStaticInputInfoCS(regs, data, info);
 	return GetShaderParams(
 	    regs.cs_regs.data_addr, hash,
 	    std::span<const uint32_t>(regs.cs_user_sgpr.value, regs.cs_regs.user_sgpr), data);

src/graphics/shader/shader.h

@@ -15,17 +15,6 @@
 
 namespace Libs::Graphics {
 
-namespace ShaderError {
-
-[[nodiscard]] inline bool Fail(std::string* error, std::string_view message) {
-	if (error != nullptr) {
-		error->assign(message.data(), message.size());
-	}
-	return false;
-}
-
-} // namespace ShaderError
-
 namespace HW {
 struct VertexShaderInfo;
 struct PixelShaderInfo;
@@ -45,16 +34,14 @@ enum class ShaderType {
 	TessellationEvaluation
 };
 
-namespace ShaderRecompiler::IR {
-struct CompiledShaderInfo;
-} // namespace ShaderRecompiler::IR
+struct ShaderModule;
 
 struct ShaderStageRuntime {
-	const ShaderRecompiler::IR::CompiledShaderInfo* program = nullptr;
-	const ShaderRecompiler::IR::ResourceSnapshot*   resources = nullptr;
+	const ShaderModule*                           module    = nullptr;
+	const ShaderRecompiler::IR::ResourceSnapshot* resources = nullptr;
 
 	[[nodiscard]] explicit operator bool() const {
-		return program != nullptr && resources != nullptr;
+		return module != nullptr && resources != nullptr;
 	}
 };
 

src/graphics/shader/shaderCompiler.h

@@ -32,6 +32,7 @@ void BuildStageStaticKey(const ShaderComputeInputInfo& input_info, std::vector<u
 
 ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context& context,
                             const HW::UserConfig& user_config, ShaderVertexInputInfo& input_info);
+// Each stage's tess holds the VGT fields; the caller adds the layout reflected from the programs.
 std::array<ShaderParams, 3>
 PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context& context,
                             std::array<ShaderVertexInputInfo, 3>& input_info);
@@ -39,8 +40,7 @@ ShaderParams PrepareProgram(
     const HW::PixelShaderInfo& regs, const HW::ShaderRegisters& sh,
     std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
     ShaderPixelInputInfo&                               input_info);
-ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, const HW::ShaderRegisters& sh,
-                            ShaderComputeInputInfo& input_info);
+ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, ShaderComputeInputInfo& input_info);
 
 } // namespace Libs::Graphics
 

src/graphics/shader/shaderVertexMetadata.cpp

@@ -5,10 +5,10 @@
 
 namespace Libs::Graphics {
 
-bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sgprs,
-                              ShaderVertexMetadata& metadata, std::string* error) {
+const char* ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sgprs,
+                                     ShaderVertexMetadata& metadata) {
 	if (data.user_data == nullptr) {
-		return ShaderError::Fail(error, "missing AGC user-data header");
+		return "missing AGC user-data header";
 	}
 
 	ShaderUserData user_data {};
@@ -17,8 +17,7 @@ bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sg
 	constexpr uint32_t DirectResourceCount =
 	    static_cast<uint32_t>(AgcDirectResourceType::Last) + 1u;
 	if (user_data.direct_resource_count > DirectResourceCount) {
-		return ShaderError::Fail(error,
-		                         "AGC direct-resource count exceeds the known resource domain");
+		return "AGC direct-resource count exceeds the known resource domain";
 	}
 
 	std::array<uint16_t, DirectResourceCount> direct_offsets {};
@@ -26,7 +25,7 @@ bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sg
 	    static_cast<uint64_t>(user_data.direct_resource_count) * sizeof(uint16_t);
 	if (direct_size != 0) {
 		if (user_data.direct_resource_offset == nullptr) {
-			return ShaderError::Fail(error, "missing AGC direct-resource offsets");
+			return "missing AGC direct-resource offsets";
 		}
 		std::memcpy(direct_offsets.data(), user_data.direct_resource_offset, direct_size);
 	}
@@ -47,33 +46,33 @@ bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sg
 	}
 
 	if (next.vertex_attrib_reg >= 0 && next.vertex_buffer_reg < 0) {
-		return ShaderError::Fail(error, "vertex attribute table requires a vertex buffer table");
+		return "vertex attribute table requires a vertex buffer table";
 	}
 	if (next.vertex_buffer_reg < 0) {
 		metadata = next;
-		return true;
+		return nullptr;
 	}
 	if (static_cast<uint32_t>(next.vertex_buffer_reg) + 1u >= max_user_sgprs) {
-		return ShaderError::Fail(error, "vertex table pointer exceeds the user-SGPR domain");
+		return "vertex table pointer exceeds the user-SGPR domain";
 	}
 	if (next.vertex_attrib_reg < 0) {
 		metadata = next;
-		return true;
+		return nullptr;
 	}
 	if (static_cast<uint32_t>(next.vertex_attrib_reg) + 1u >= max_user_sgprs) {
-		return ShaderError::Fail(error, "vertex table pointer exceeds the user-SGPR domain");
+		return "vertex table pointer exceeds the user-SGPR domain";
 	}
 	if (data.num_input_semantics == 0 ||
 	    data.num_input_semantics > ShaderVertexInputInfo::RES_MAX) {
-		return ShaderError::Fail(error, "vertex semantic count is outside the supported domain");
+		return "vertex semantic count is outside the supported domain";
 	}
 
 	if (data.input_semantics == nullptr) {
-		return ShaderError::Fail(error, "missing vertex input semantics");
+		return "missing vertex input semantics";
 	}
 	next.input_semantics = {data.input_semantics, data.num_input_semantics};
 	metadata                   = next;
-	return true;
+	return nullptr;
 }
 
 } // namespace Libs::Graphics

src/graphics/shader/shaderVertexMetadata.h

@@ -4,7 +4,6 @@
 #include "graphics/shader/shader.h"
 
 #include <span>
-#include <string>
 
 namespace Libs::Graphics {
 
@@ -14,9 +13,10 @@ struct ShaderVertexMetadata {
 	std::span<const ShaderSemantic> input_semantics;
 };
 
-// Validates the vertex metadata and returns a view into the shader header.
-bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sgprs,
-                              ShaderVertexMetadata& metadata, std::string* error);
+// Reads the vertex metadata as a view into the shader header. Returns nullptr if it is
+// valid, otherwise why it was rejected.
+const char* ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sgprs,
+                                     ShaderVertexMetadata& metadata);
 
 } // namespace Libs::Graphics
 

src/kernel/memoryAddressSpace.inc

@@ -820,12 +820,12 @@ public:
 		return ProtectMappedUnlocked(vaddr, size, mode);
 	}
 
+	// Tracker protections only read m_mapped; the page tracker orders them per page.
 	bool ProtectTransient(uint64_t vaddr, uint64_t size, VirtualMemory::Mode mode) {
 		if (size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
-
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return ProtectMappedUnlocked(vaddr, size, mode);
 	}
 

src/launcher/forms/configuration_edit_dialog.ui

@@ -141,6 +141,7 @@
               <property name="value"><number>60</number></property>
              </widget>
             </item>
+            <!--
             <item row="3" column="0">
              <widget class="QLabel" name="label_shader_optimization">
               <property name="text"><string>Shader optimization:</string></property>
@@ -151,6 +152,7 @@
               <property name="toolTip"><string>Optimize shaders for code size or performance</string></property>
              </widget>
             </item>
+            -->
             <item row="4" column="0">
              <widget class="QLabel" name="label_gpu">
               <property name="text"><string>GPU:</string></property>
@@ -191,6 +193,12 @@
               <property name="text"><string>Enable HDR display support</string></property>
              </widget>
             </item>
+            <item row="10" column="1">
+             <widget class="QCheckBox" name="checkBox_shader_cache">
+              <property name="toolTip"><string>Save compiled shaders and pipelines per game and load them at startup.</string></property>
+              <property name="text"><string>Cache shaders</string></property>
+             </widget>
+            </item>
            </layout>
           </widget>
          </item>

src/launcher/forms/main_dialog.ui

@@ -112,6 +112,13 @@
          </property>
         </widget>
        </item>
+       <item>
+        <widget class="QCheckBox" name="close_terminal">
+         <property name="text">
+          <string>Close terminal automatically</string>
+         </property>
+        </widget>
+       </item>
        <item>
         <layout class="QHBoxLayout" name="themeLayout">
          <item>

src/launcher/include/additionalContentDialog.h

@@ -4,26 +4,39 @@
 #include <QDialog>
 #include <QIcon>
 #include <QString>
+#include <QStringList>
+
+#include <functional>
 
 class Configuration;
 class QLabel;
+class QLineEdit;
 class QListWidget;
 class QPushButton;
 
 class AdditionalContentDialog final: public QDialog {
 public:
-	AdditionalContentDialog(const Configuration& game, const QString& folder,
-	                        QWidget* parent = nullptr);
+	// entries are the game's user add-ons, "LABEL" or "LABEL=DIR"; save receives every edit.
+	AdditionalContentDialog(const Configuration& game, const QString& folder, QStringList entries,
+	                        std::function<void(const QStringList&)> save,
+	                        QWidget*                                parent = nullptr);
 
 private:
 	void Scan();
+	void Add();
+	void Remove();
 
-	QString      m_title_id;
-	QString      m_folder;
-	QIcon        m_game_icon;
-	QLabel*      m_status = nullptr;
-	QListWidget* m_list   = nullptr;
-	QPushButton* m_open   = nullptr;
+	QString                                 m_title_id;
+	QString                                 m_folder;
+	QStringList                             m_entries;
+	std::function<void(const QStringList&)> m_save;
+	QIcon                                   m_game_icon;
+	QLabel*                                 m_status     = nullptr;
+	QListWidget*                            m_list       = nullptr;
+	QPushButton*                            m_open       = nullptr;
+	QPushButton*                            m_remove     = nullptr;
+	QLineEdit*                              m_label_edit = nullptr;
+	QLineEdit*                              m_dir_edit   = nullptr;
 };
 
 #endif // ADDITIONAL_CONTENT_DIALOG_H

src/launcher/include/configuration.h

@@ -102,16 +102,17 @@ public:
 
 	Configuration() = default;
 
-	QString    name;
-	QString    title_id;    /* Serial / title id from sce_sys/param.json */
-	QString    gameVersion; /* appVersion / contentVersion from sce_sys/param.json */
-	QString    firmwareVer; /* requiredSystemSoftwareVersion from sce_sys/param.json */
-	QString    basedir;     /* Game base directory */
-	QString    game_path;   /* Launcher-unique game path */
-	QString    addons_dir;  /* Global setting; not saved per game. */
-	bool       custom_settings = false;
-	GameStatus game_status     = GameStatus::Unknown;
-	QString    game_comment;
+	QString     name;
+	QString     title_id;    /* Serial / title id from sce_sys/param.json */
+	QString     gameVersion; /* appVersion / contentVersion from sce_sys/param.json */
+	QString     firmwareVer; /* requiredSystemSoftwareVersion from sce_sys/param.json */
+	QString     basedir;     /* Game base directory */
+	QString     game_path;   /* Launcher-unique game path */
+	QString     addons_dir;  /* Global setting; not saved per game. */
+	QStringList addcont;     /* User add-ons ("LABEL" or "LABEL=DIR"), saved by title ID. */
+	bool        custom_settings = false;
+	GameStatus  game_status     = GameStatus::Unknown;
+	QString     game_comment;
 
 	// Controller preferences always come from the global configuration.
 	ControllerSettings controller;
@@ -130,6 +131,7 @@ public:
 	bool                   sync_raw_image_buffers      = false;
 	bool                   tessellation_enabled        = false;
 	bool                   occlusion_queries_enabled   = false;
+	bool                   shader_cache_enabled        = true;
 	bool                   trophy_enabled              = true;
 	int                    trophy_duration_seconds     = 6;
 	bool                   perf_overlay_enabled        = false;
@@ -171,6 +173,7 @@ public:
 		sync_raw_image_buffers      = other.sync_raw_image_buffers;
 		tessellation_enabled        = other.tessellation_enabled;
 		occlusion_queries_enabled   = other.occlusion_queries_enabled;
+		shader_cache_enabled        = other.shader_cache_enabled;
 		trophy_enabled              = other.trophy_enabled;
 		trophy_duration_seconds     = other.trophy_duration_seconds;
 		perf_overlay_enabled        = other.perf_overlay_enabled;

src/launcher/include/configurationListWidget.h

@@ -87,6 +87,7 @@ private:
 	QString                       m_runtime_directory;
 	QStringList                   m_game_dirs;
 	QString                       m_addons_dir;
+	QMap<QString, QStringList>    m_addcont; // User add-ons by title ID.
 	Configuration                 m_global_info;
 	QMap<QString, Configuration*> m_custom_infos;
 	CompatibilityDatabase*        m_compatibility = nullptr;

src/launcher/include/mainDialog.h

@@ -22,7 +22,7 @@ public:
 	explicit MainDialog(QWidget* parent = nullptr);
 	~MainDialog() override = default;
 
-	void RunInterpreter(QProcess* process, const Configuration& info);
+	void RunInterpreter(QProcess* process, const Configuration& info, bool close_terminal);
 
 	static void WriteSettings(QSettings& s);
 	static void ReadSettings(QSettings& s);

src/launcher/src/additionalContentDialog.cpp

@@ -7,7 +7,10 @@
 #include <QDesktopServices>
 #include <QDialogButtonBox>
 #include <QDir>
+#include <QFileDialog>
+#include <QHBoxLayout>
 #include <QLabel>
+#include <QLineEdit>
 #include <QListWidget>
 #include <QPixmap>
 #include <QPushButton>
@@ -16,8 +19,11 @@
 #include <QVBoxLayout>
 
 AdditionalContentDialog::AdditionalContentDialog(const Configuration& game, const QString& folder,
-                                                 QWidget* parent)
-    : QDialog(parent), m_title_id(game.title_id.trimmed().toUpper()), m_folder(folder) {
+                                                 QStringList                             entries,
+                                                 std::function<void(const QStringList&)> save,
+                                                 QWidget*                                parent)
+    : QDialog(parent), m_title_id(game.title_id.trimmed().toUpper()), m_folder(folder),
+      m_entries(std::move(entries)), m_save(std::move(save)) {
 	setAttribute(Qt::WA_DeleteOnClose);
 	setWindowTitle(tr("Additional Content - %1").arg(game.name));
 	resize(720, 480);
@@ -39,6 +45,28 @@ AdditionalContentDialog::AdditionalContentDialog(const Configuration& game, cons
 	m_list->setVerticalScrollMode(QAbstractItemView::ScrollPerPixel);
 	layout->addWidget(m_list, 1);
 
+	auto* add_layout = new QHBoxLayout;
+	m_label_edit     = new QLineEdit(this);
+	m_label_edit->setPlaceholderText(tr("Label or content ID"));
+	m_dir_edit = new QLineEdit(this);
+	m_dir_edit->setPlaceholderText(tr("Add-on data folder (empty for a license only)"));
+	auto* browse = new QPushButton(tr("Browse..."), this);
+	auto* add    = new QPushButton(tr("Add"), this);
+	add->setDefault(true);
+	add_layout->addWidget(m_label_edit, 1);
+	add_layout->addWidget(m_dir_edit, 2);
+	add_layout->addWidget(browse);
+	add_layout->addWidget(add);
+	layout->addLayout(add_layout);
+	connect(browse, &QPushButton::clicked, this, [this]() {
+		const auto dir =
+		    QFileDialog::getExistingDirectory(this, tr("Add-on data folder"), m_dir_edit->text());
+		if (!dir.isEmpty()) {
+			m_dir_edit->setText(QDir::toNativeSeparators(dir));
+		}
+	});
+	connect(add, &QPushButton::clicked, this, &AdditionalContentDialog::Add);
+
 	m_status = new QLabel(this);
 	m_status->setTextFormat(Qt::PlainText);
 	m_status->setWordWrap(true);
@@ -47,10 +75,15 @@ AdditionalContentDialog::AdditionalContentDialog(const Configuration& game, cons
 	auto* buttons = new QDialogButtonBox(QDialogButtonBox::Close, this);
 	auto* refresh = buttons->addButton(tr("Refresh"), QDialogButtonBox::ActionRole);
 	m_open        = buttons->addButton(tr("Open folder"), QDialogButtonBox::ActionRole);
+	m_remove      = buttons->addButton(tr("Remove"), QDialogButtonBox::ActionRole);
 	layout->addWidget(buttons);
 	connect(refresh, &QPushButton::clicked, this, &AdditionalContentDialog::Scan);
 	connect(m_open, &QPushButton::clicked, this,
 	        [this]() { QDesktopServices::openUrl(QUrl::fromLocalFile(m_folder)); });
+	connect(m_remove, &QPushButton::clicked, this, &AdditionalContentDialog::Remove);
+	connect(m_list, &QListWidget::currentItemChanged, this, [this](const QListWidgetItem* item) {
+		m_remove->setEnabled(item != nullptr && item->data(Qt::UserRole).isValid());
+	});
 	connect(buttons, &QDialogButtonBox::rejected, this, &QDialog::close);
 
 	Scan();
@@ -73,18 +106,54 @@ void AdditionalContentDialog::Scan() {
 		                                 tr("%1\nInstalled").arg(title), m_list);
 		item->setToolTip(QString::fromStdString(entry.content_id));
 	}
+	for (const auto& entry: m_entries) {
+		const auto label = entry.section('=', 0, 0);
+		const auto dir   = entry.section('=', 1);
+		auto       text  = tr("%1\nLicense only").arg(label);
+		if (!dir.isEmpty()) {
+			text = (QDir(dir).exists() ? tr("%1\nData: %2") : tr("%1\nData folder not found: %2"))
+			           .arg(label, QDir::toNativeSeparators(dir));
+		}
+		auto* item = new QListWidgetItem(m_game_icon, text, m_list);
+		item->setData(Qt::UserRole, entry);
+	}
+	m_remove->setEnabled(false);
 
-	if (m_folder.isEmpty()) {
+	if (m_list->count() != 0) {
+		m_status->setText(tr("%n add-on(s)", nullptr, m_list->count()));
+	} else if (m_folder.isEmpty()) {
 		m_status->setText(tr("No add-ons folder is set. Open Global settings and set the "
 		                     "Add-ons folder to the location of your extracted add-on packages."));
 	} else if (!folder_exists) {
 		m_status->setText(tr("The add-ons folder does not exist:\n%1\nCheck the path in Global "
 		                     "settings.")
 		                      .arg(QDir::toNativeSeparators(m_folder)));
-	} else if (m_list->count() == 0) {
+	} else {
 		m_status->setText(
 		    tr("No add-ons found for this game in:\n%1").arg(QDir::toNativeSeparators(m_folder)));
-	} else {
-		m_status->setText(tr("%n installed add-on(s)", nullptr, m_list->count()));
 	}
 }
+
+void AdditionalContentDialog::Add() {
+	// A content ID ends with its entitlement label.
+	const auto label = m_label_edit->text().trimmed().section('-', -1);
+	const auto dir   = m_dir_edit->text().trimmed();
+	const auto entry = dir.isEmpty() ? label : label + '=' + QDir(dir).absolutePath();
+	if (!Common::ParseAddContent(entry.toStdString())) {
+		m_status->setText(tr("Enter a 16-character entitlement label (A-Z, 0-9) and an existing "
+		                     "data folder, or no folder for a license only."));
+		return;
+	}
+	m_entries.removeIf([&](const QString& e) { return e.section('=', 0, 0) == label; });
+	m_entries.append(entry);
+	m_save(m_entries);
+	m_label_edit->clear();
+	m_dir_edit->clear();
+	Scan();
+}
+
+void AdditionalContentDialog::Remove() {
+	m_entries.removeOne(m_list->currentItem()->data(Qt::UserRole).toString());
+	m_save(m_entries);
+	Scan();
+}

src/launcher/src/configuration.cpp

@@ -17,6 +17,7 @@ QVariantMap Configuration::GameSettings() const {
 	KYTY_CFG_SET(sync_raw_image_buffers);
 	KYTY_CFG_SET(tessellation_enabled);
 	KYTY_CFG_SET(occlusion_queries_enabled);
+	KYTY_CFG_SET(shader_cache_enabled);
 	KYTY_CFG_SET(trophy_enabled);
 	KYTY_CFG_SET(trophy_duration_seconds);
 	KYTY_CFG_SET(perf_overlay_enabled);
@@ -68,7 +69,8 @@ void Configuration::ReadGameSettingsValues(const Settings& s) {
 	KYTY_CFG_GET(sync_raw_image_buffers);
 	KYTY_CFG_GET(tessellation_enabled);
 	KYTY_CFG_GET(occlusion_queries_enabled);
-	trophy_enabled   = s.value("trophy_enabled", trophy_enabled).toBool();
+	shader_cache_enabled = s.value("shader_cache_enabled", shader_cache_enabled).toBool();
+	trophy_enabled       = s.value("trophy_enabled", trophy_enabled).toBool();
 	KYTY_CFG_GET(trophy_duration_seconds);
 	KYTY_CFG_GET(perf_overlay_enabled);
 	KYTY_CFG_GET(skip_notice_screen);

src/launcher/src/configurationEditDialog.cpp

@@ -290,6 +290,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_sync_raw_image_buffers->setChecked(info.sync_raw_image_buffers);
 	m_ui->checkBox_tessellation->setChecked(info.tessellation_enabled);
 	m_ui->checkBox_occlusion_queries->setChecked(info.occlusion_queries_enabled);
+	m_ui->checkBox_shader_cache->setChecked(info.shader_cache_enabled);
 	m_ui->checkBox_skip_notice_screen->setChecked(info.skip_notice_screen);
 	m_ui->checkBox_network_connected->setChecked(info.network_connected);
 	m_ui->checkBox_trophy_notifications->setChecked(info.trophy_enabled);
@@ -314,7 +315,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 #else
 	m_ui->checkBox_red_zone_protection->setVisible(false);
 #endif
-	ListInit(m_ui->comboBox_shader_optimization_type, info.shader_optimization_type);
+	// ListInit(m_ui->comboBox_shader_optimization_type, info.shader_optimization_type);
 	ListInit(m_ui->comboBox_shader_log_direction, info.shader_log_direction);
 	m_ui->lineEdit_shader_log_folder->setText(info.shader_log_folder);
 	m_ui->lineEdit_shader_log_folder->setEnabled(info.shader_log_direction ==
@@ -485,6 +486,7 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 	info.sync_raw_image_buffers    = ui.checkBox_sync_raw_image_buffers->isChecked();
 	info.tessellation_enabled      = ui.checkBox_tessellation->isChecked();
 	info.occlusion_queries_enabled = ui.checkBox_occlusion_queries->isChecked();
+	info.shader_cache_enabled      = ui.checkBox_shader_cache->isChecked();
 	info.skip_notice_screen        = ui.checkBox_skip_notice_screen->isChecked();
 	info.network_connected         = ui.checkBox_network_connected->isChecked();
 	info.trophy_enabled            = ui.checkBox_trophy_notifications->isChecked();
@@ -499,8 +501,8 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 #if defined(_WIN32)
 	info.red_zone_protection_enabled = ui.checkBox_red_zone_protection->isChecked();
 #endif
-	info.shader_optimization_type = TextToEnum<Configuration::ShaderOptimizationType>(
-	    ui.comboBox_shader_optimization_type->currentText());
+	// info.shader_optimization_type = TextToEnum<Configuration::ShaderOptimizationType>(
+	//     ui.comboBox_shader_optimization_type->currentText());
 	info.shader_log_direction = TextToEnum<Configuration::LogDirection>(
 	    ui.comboBox_shader_log_direction->currentText());
 	info.shader_log_folder           = ui.lineEdit_shader_log_folder->text();

src/launcher/src/configurationListWidget.cpp

@@ -79,6 +79,7 @@ constexpr char CONF_LAUNCHER[]     = "Launcher";
 constexpr char CONF_GAME_DIR[]     = "game_dir";
 constexpr char CONF_GAME_DIRS[]    = "game_dirs";
 constexpr char CONF_ADDONS_DIR[]   = "addons_dir";
+constexpr char CONF_ADDCONT[]      = "AdditionalContent";
 constexpr char CONF_GLOBAL[]       = "GlobalConfiguration";
 constexpr char SAVE_DATA_DIR[]     = "_SaveData";
 
@@ -348,6 +349,13 @@ void ConfigurationListWidget::WriteSettings() {
 	s->remove(CONF_GAME_DIR);
 	s->endGroup();
 
+	s->remove(CONF_ADDCONT);
+	s->beginGroup(CONF_ADDCONT);
+	for (auto it = m_addcont.constBegin(); it != m_addcont.constEnd(); ++it) {
+		s->setValue(it.key(), it.value());
+	}
+	s->endGroup();
+
 	s->remove(CONF_GLOBAL);
 	s->beginGroup(CONF_GLOBAL);
 	m_global_info.WriteSettings(s.get());
@@ -395,6 +403,13 @@ void ConfigurationListWidget::ReadSettings() {
 	m_addons_dir = NormalizeGameDirectory(s->value(CONF_ADDONS_DIR).toString());
 	s->endGroup();
 
+	m_addcont.clear();
+	s->beginGroup(CONF_ADDCONT);
+	for (const auto& title_id: s->childKeys()) {
+		m_addcont.insert(title_id, SettingsStringList(s->value(title_id)));
+	}
+	s->endGroup();
+
 	s->beginGroup(CONF_GLOBAL);
 	if (!s->childKeys().isEmpty()) {
 		m_global_info.ReadSettings(s.get());
@@ -458,6 +473,7 @@ ConfigurationListWidget::CreateConfiguration(const ConfigurationItem& item) cons
 	}
 	info->host_input_mapping = m_global_info.host_input_mapping;
 	info->addons_dir         = m_addons_dir;
+	info->addcont            = m_addcont.value(item.GetInfo().title_id.trimmed().toUpper());
 	return info;
 }
 
@@ -1273,11 +1289,22 @@ void ConfigurationListWidget::ShowContextMenu(QTreeWidgetItem* witem) {
 	connect(action_addcont, &QAction::triggered, this,
 	        [this, item = QPointer<ConfigurationItem>(item)]() {
 		        if (item != nullptr) {
-			        auto* dialog = new AdditionalContentDialog(item->GetInfo(), m_addons_dir, this);
+			        const auto title_id = item->GetInfo().title_id.trimmed().toUpper();
+			        auto*      dialog   = new AdditionalContentDialog(
+			            item->GetInfo(), m_addons_dir, m_addcont.value(title_id),
+			            [this, title_id](const QStringList& entries) {
+				            if (entries.isEmpty()) {
+					            m_addcont.remove(title_id);
+				            } else {
+					            m_addcont.insert(title_id, entries);
+				            }
+				            WriteSettings();
+			            },
+			            this);
 			        dialog->show();
 		        }
 	        });
-	action_addcont->setEnabled(item != nullptr);
+	action_addcont->setEnabled(item != nullptr && !item->GetInfo().title_id.trimmed().isEmpty());
 	QAction* action_compress =
 	    menu.addAction(tr("Compress to ZAR..."), this, SLOT(compress_game()));
 	QAction* action_remove_save_data =

src/launcher/src/mainDialog.cpp

@@ -17,7 +17,6 @@
 #include <QDir>
 #include <QFile>
 #include <QFileInfo>
-#include <QIODevice>
 #include <QLabel>
 #include <QMessageBox>
 #include <QObject>
@@ -28,7 +27,6 @@
 #include <QSettings>
 #include <QStringList>
 #include <QStyleHints>
-#include <QTextStream>
 #include <QVariant>
 #include <QtCore>
 
@@ -47,23 +45,17 @@
 class QWidget;
 
 #if defined(_WIN32)
-constexpr char EMULATOR_EXE[] = "kyty_emulator.exe";
+constexpr char  EMULATOR_EXE[] = "kyty_emulator.exe";
+constexpr char  CMD_EXE[]      = "cmd.exe";
+constexpr DWORD CMD_X_CHARS    = 175;
+constexpr DWORD CMD_Y_CHARS    = 1000;
 #else
 constexpr char EMULATOR_EXE[] = "kyty_emulator";
 #endif
-
-#if defined(_WIN32)
-constexpr char CMD_EXE[] = "cmd.exe";
-#elif defined(__linux__)
-constexpr char KYTY_BASH_FILE[] = "kyty_run.sh";
-#endif
-#if defined(_WIN32)
-constexpr DWORD CMD_X_CHARS = 175;
-constexpr DWORD CMD_Y_CHARS = 1000;
-#endif
 constexpr char SETTINGS_MAIN_DIALOG[]        = "MainDialog";
 constexpr char SETTINGS_MAIN_LAST_GEOMETRY[] = "geometry";
 constexpr char SETTINGS_CHECK_UPDATES[]      = "check_updates_on_startup";
+constexpr char SETTINGS_CLOSE_TERMINAL[]     = "close_terminal";
 constexpr char SETTINGS_THEME[]              = "theme";
 
 class MainDialogPrivate: public QObject {
@@ -89,6 +81,7 @@ public:
 private:
 	static QByteArray g_last_geometry;
 	static bool       g_check_updates_on_startup;
+	static bool       g_close_terminal;
 	static int        g_theme;
 
 	static void ApplyTheme(int theme) {
@@ -108,13 +101,12 @@ private:
 	TitleSound      m_title_sound;
 	QString         m_interpreter;
 
-	QProcess m_process;
-
-	QPointer<ConfigurationItem> m_running_item;
+	int m_running_games = 0;
 };
 
 QByteArray MainDialogPrivate::g_last_geometry;
 bool       MainDialogPrivate::g_check_updates_on_startup = true;
+bool       MainDialogPrivate::g_close_terminal           = false;
 int        MainDialogPrivate::g_theme                    = 0;
 
 MainDialog::MainDialog(QWidget* parent): QDialog(parent), m_p(new MainDialogPrivate(this)) {
@@ -122,7 +114,6 @@ MainDialog::MainDialog(QWidget* parent): QDialog(parent), m_p(new MainDialogPriv
 }
 
 MainDialogPrivate::~MainDialogPrivate() {
-	m_process.disconnect(this);
 	delete m_ui;
 }
 
@@ -134,10 +125,14 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 	m_main_dialog = main_dialog;
 	m_update_checker = new UpdateChecker(main_dialog);
 	m_ui->check_updates_on_startup->setChecked(g_check_updates_on_startup);
+	m_ui->close_terminal->setChecked(g_close_terminal);
 	m_ui->theme_combo->setCurrentIndex(g_theme);
 	ApplyTheme(g_theme);
 	m_ui->check_updates_link->setVisible(UpdateChecker::IsSupported());
 	m_ui->check_updates_on_startup->setVisible(UpdateChecker::IsSupported());
+#if defined(__APPLE__)
+	m_ui->close_terminal->setVisible(false); // The emulator runs without a terminal.
+#endif
 
 	main_dialog->setWindowFlags(Qt::Dialog /*| Qt::MSWindowsFixedSizeDialogHint*/);
 
@@ -146,13 +141,12 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 	connect(m_ui->widget, &ConfigurationListWidget::Select, this, [this]() {
 		Update();
 		const auto* item = m_ui->widget->GetSelectedItem();
-		m_title_sound.SetTitle(m_process.state() == QProcess::NotRunning && item != nullptr
-		                           ? item->GetInfo().basedir
-		                           : QString());
+		m_title_sound.SetTitle(m_running_games == 0 && item != nullptr ? item->GetInfo().basedir
+		                                                               : QString());
 	});
 	connect(m_ui->widget, &ConfigurationListWidget::PreviewControllerColor, this,
 	        [this](const QString& color) {
-		        if (m_process.state() == QProcess::NotRunning) {
+		        if (m_running_games == 0) {
 			        m_controller_preview.SetColor(color);
 		        }
 	        });
@@ -166,6 +160,10 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 		g_check_updates_on_startup = checked;
 		m_ui->widget->WriteSettings();
 	});
+	connect(m_ui->close_terminal, &QCheckBox::toggled, this, [this](bool checked) {
+		g_close_terminal = checked;
+		m_ui->widget->WriteSettings();
+	});
 	connect(m_ui->theme_combo, &QComboBox::currentIndexChanged, this, [this](int index) {
 		g_theme = index;
 		ApplyTheme(index);
@@ -176,16 +174,6 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 		m_ui->widget->WriteSettings();
 	});
 
-	connect(&m_process, &QProcess::stateChanged, this, [this](QProcess::ProcessState state) {
-		m_update_checker->SetGameRunning(state != QProcess::NotRunning);
-		if (state == QProcess::NotRunning) {
-			if (m_running_item != nullptr) {
-				m_running_item->SetRunning(false);
-			}
-			Update();
-		}
-	});
-
 	m_ui->label_settings_file->setText(tr("Settings file: ") + m_ui->widget->GetSettingsFile());
 
 	m_main_dialog->restoreGeometry(g_last_geometry);
@@ -288,6 +276,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	AddFlag(args, info.network_connected, "--network-connected");
 	AddFlag(args, info.tessellation_enabled, "--tessellation");
 	AddFlag(args, info.occlusion_queries_enabled, "--occlusion-queries");
+	AddFlag(args, info.shader_cache_enabled, "--shader-cache");
 	args << "--vblank-frequency" << QString::number(info.vblank_frequency);
 	args << "--console-language" << QString::number(info.console_language);
 	AddFlag(args, info.vulkan_validation_enabled, "--vulkan-validation");
@@ -317,6 +306,9 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	if (!info.addons_dir.isEmpty()) {
 		args << "--addcont-dir" << info.addons_dir;
 	}
+	for (const auto& entry: info.addcont) {
+		args << "--addcont" << entry;
+	}
 
 	const auto patch_plan = Cheats::PlanPath(info.title_id);
 	if (QFileInfo::exists(patch_plan)) {
@@ -327,32 +319,14 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 }
 
 #ifdef __linux__
-static QString BashQuote(QString value) {
-	value.replace('\'', "'\\''");
-	return QStringLiteral("'") + value + QStringLiteral("'");
-}
-
-static bool CreateBashScript(const QString& interpreter, const QStringList& args,
-                             const QString& file_name) {
-	QFile file(file_name);
-	if (file.open(QIODevice::WriteOnly | QIODevice::Text)) {
-		QTextStream s(&file);
-
-		s << "#!/bin/bash\n";
-		s << BashQuote(interpreter);
-		for (const auto& arg: args) {
-			s << " " << BashQuote(arg);
-		}
-		s << "\n";
-		s << "echo Press any key...\n";
-		s << "read -n1\n";
-
-		file.close();
-
-		return file.setPermissions(file.permissions() | QFile::ExeUser | QFile::ExeOwner |
-		                           QFile::ExeGroup);
-	}
-	return false;
+// Runs the command in bash, then waits for a key (with close_on_clean_exit, only after an error).
+// The command goes in as positional parameters ("$@"), so it needs no quoting.
+static QStringList BashCommand(const QString& interpreter, const QStringList& args,
+                               bool close_on_clean_exit) {
+	const QString run = close_on_clean_exit ? "\"$@\" && exit" : "\"$@\"";
+	return QStringList {"bash", "-c", run + "; echo Press any key...; read -n1", "kyty",
+	                    interpreter} +
+	       args;
 }
 
 // Find a terminal and its command separator.
@@ -416,28 +390,23 @@ static bool FindTerminal(QString* program, QStringList* prefix) {
 #endif
 
 #if defined(_WIN32)
-// Quote one token for cmd.exe so paths with spaces survive /K parsing.
-static QString WinCmdQuote(QString value) {
-	value.replace(QLatin1Char('"'), QStringLiteral("\\\""));
-	return QLatin1Char('"') + value + QLatin1Char('"');
-}
-
-static QString BuildWinCmdKCommand(const QString& interpreter, const QStringList& args) {
-	QString command = WinCmdQuote(QDir::toNativeSeparators(interpreter));
-	for (const auto& arg: args) {
-		command += QLatin1Char(' ');
-		command += WinCmdQuote(arg);
-	}
-	return command;
+// cmd /K arguments that keep the console open (with close_on_clean_exit, only after an error).
+// A quote is written as "" rather than \" so cmd's quote count stays even and operators inside
+// arguments stay quoted.
+static QString CmdCommand(const QString& interpreter, const QStringList& args,
+                          bool close_on_clean_exit) {
+	static const QRegularExpression quote_or_end(QStringLiteral(R"((\\*)("|\z))"));
+	// The UCRT reads 2n backslashes before a quote as n.
+	auto quoted = QStringList {interpreter} + args;
+	quoted.replaceInStrings(quote_or_end, QStringLiteral(R"(\1\1\2\2)"));
+	return QStringLiteral("/K \"\"") + quoted.join(QStringLiteral("\" \"")) +
+	       (close_on_clean_exit ? "\" && exit\"" : "\"\"");
 }
 #endif
 
-void MainDialog::RunInterpreter(QProcess* process, const Configuration& info) {
+void MainDialog::RunInterpreter(QProcess* process, const Configuration& info, bool close_terminal) {
 	const auto& interpreter = m_p->GetInterpreter();
 
-	QFileInfo f(interpreter);
-	auto      dir = f.absoluteDir();
-
 	auto args = CreateEmulatorArgs(info);
 	if (args.isEmpty()) {
 		QMessageBox::critical(this, tr("Error"), tr("Invalid emulator configuration"));
@@ -445,40 +414,24 @@ void MainDialog::RunInterpreter(QProcess* process, const Configuration& info) {
 		return;
 	}
 
-#ifdef __linux__
-	auto bash_file_name = dir.filePath(KYTY_BASH_FILE);
-	if (!CreateBashScript(interpreter, args, bash_file_name)) {
-		QMessageBox::critical(this, tr("Error"), tr("Can't create file:\n") + bash_file_name);
-		QApplication::quit();
-		return;
-	}
-
-	{
-		QString     terminal;
-		QStringList terminal_prefix;
-		// Pass the script as a file argument (not bash -c) so paths with spaces work.
-		if (FindTerminal(&terminal, &terminal_prefix)) {
-			process->setProgram(terminal);
-			process->setArguments(terminal_prefix + QStringList {"bash", bash_file_name});
-		} else {
-			// Run without a terminal as a fallback.
-			process->setProgram(QStringLiteral("bash"));
-			process->setArguments({bash_file_name});
-		}
-	}
-#elif defined(_WIN32)
-	{
-		// Use nativeArguments so Qt does not re-quote the /K command string.
-		process->setProgram(CMD_EXE);
-		process->setArguments({});
-		process->setNativeArguments(QStringLiteral("/K \"") +
-		                            BuildWinCmdKCommand(interpreter, args) + QLatin1Char('"'));
-	}
-#else
+	// Without a terminal, the emulator runs directly.
 	process->setProgram(interpreter);
 	process->setArguments(args);
+#ifdef __linux__
+	QString     terminal;
+	QStringList terminal_prefix;
+	if (FindTerminal(&terminal, &terminal_prefix)) {
+		process->setProgram(terminal);
+		process->setArguments(terminal_prefix + BashCommand(interpreter, args, close_terminal));
+	}
+#elif defined(_WIN32)
+	// Use nativeArguments so Qt does not re-quote the /K command string.
+	process->setProgram(CMD_EXE);
+	process->setArguments({});
+	process->setNativeArguments(
+	    CmdCommand(QDir::toNativeSeparators(interpreter), args, close_terminal));
 #endif
-	process->setWorkingDirectory(dir.path());
+	process->setWorkingDirectory(QFileInfo(interpreter).absolutePath());
 #if defined(_WIN32)
 	process->setCreateProcessArgumentsModifier([](QProcess::CreateProcessArguments* args) {
 		args->flags |= static_cast<uint32_t>(CREATE_NEW_CONSOLE);
@@ -525,6 +478,7 @@ void MainDialogPrivate::WriteSettings(QSettings& s) {
 		s.setValue(SETTINGS_MAIN_LAST_GEOMETRY, g_last_geometry);
 	}
 	s.setValue(SETTINGS_CHECK_UPDATES, g_check_updates_on_startup);
+	s.setValue(SETTINGS_CLOSE_TERMINAL, g_close_terminal);
 	s.setValue(SETTINGS_THEME, g_theme);
 
 	s.endGroup();
@@ -535,6 +489,7 @@ void MainDialogPrivate::ReadSettings(QSettings& s) {
 
 	g_last_geometry = s.value(SETTINGS_MAIN_LAST_GEOMETRY, g_last_geometry).toByteArray();
 	g_check_updates_on_startup = s.value(SETTINGS_CHECK_UPDATES, true).toBool();
+	g_close_terminal           = s.value(SETTINGS_CLOSE_TERMINAL, false).toBool();
 	g_theme                    = s.value(SETTINGS_THEME, 0).toInt();
 	if (g_theme < 0 || g_theme > 2) {
 		g_theme = 0;
@@ -547,17 +502,30 @@ void MainDialogPrivate::Run() {
 	if (m_update_checker->IsInstalling()) {
 		return;
 	}
-	m_running_item = m_ui->widget->GetSelectedItem();
-	if (m_running_item == nullptr) {
+	QPointer<ConfigurationItem> item = m_ui->widget->GetSelectedItem();
+	if (item == nullptr) {
 		return;
 	}
 
-	m_running_item->SetRunning(true);
+	item->SetRunning(true);
+	m_running_games++;
 	m_controller_preview.SetSuspended(true);
 	m_title_sound.SetTitle({});
 
-	auto info = m_ui->widget->CreateConfiguration(*m_running_item);
-	m_main_dialog->RunInterpreter(&m_process, *info);
+	auto* process = new QProcess(this);
+	connect(process, &QProcess::stateChanged, this,
+	        [this, process, item](QProcess::ProcessState state) {
+		        if (state == QProcess::NotRunning) {
+			        if (item != nullptr) {
+				        item->SetRunning(false);
+			        }
+			        m_running_games--;
+			        process->deleteLater();
+			        Update();
+		        }
+	        });
+	auto info = m_ui->widget->CreateConfiguration(*item);
+	m_main_dialog->RunInterpreter(process, *info, g_close_terminal);
 
 	Update();
 }
@@ -565,8 +533,7 @@ void MainDialogPrivate::Run() {
 void MainDialogPrivate::Update() {
 	const auto* item = m_ui->widget->GetSelectedItem();
 
-	bool run_enabled = (m_process.state() == QProcess::NotRunning && item != nullptr &&
-	                    !m_update_checker->IsInstalling());
+	bool run_enabled = item != nullptr && !item->IsRunning() && !m_update_checker->IsInstalling();
 
 	if (run_enabled) {
 		const auto& info = item->GetInfo();
@@ -575,7 +542,8 @@ void MainDialogPrivate::Update() {
 	}
 
 	m_ui->widget->SetRunEnabled(run_enabled);
-	if (m_process.state() != QProcess::NotRunning) {
+	m_update_checker->SetGameRunning(m_running_games > 0);
+	if (m_running_games > 0) {
 		m_controller_preview.SetSuspended(true);
 		return;
 	}

src/libs/libC.cpp

@@ -771,10 +771,10 @@ LIB_DEFINE(InitLibcInternal_1) {
 	LIB_FUNC("Q3VBxCXhUHs", LibcInternal::memcpy);
 	LIB_FUNC("+P6FRGH4LfA", LibcInternal::memmove);
 	LIB_FUNC("DfivPArhucg", LibcInternal::memcmp);
-	LIB_FUNC("aesyjrHVWy4", LibcInternal::strcmp);
-	LIB_FUNC("Ovb2dSJOAuE", LibcInternal::strncmp);
+	LIB_FUNC("Ovb2dSJOAuE", LibcInternal::strcmp);
+	LIB_FUNC("aesyjrHVWy4", LibcInternal::strncmp);
 	LIB_FUNC("j4ViWNHEgww", LibcInternal::strlen);
-	LIB_FUNC("5Xa2ACNECdo", LibcInternal::strcpy);
+	LIB_FUNC("kiZSXIWd9vg", LibcInternal::strcpy);
 	LIB_FUNC("6sJWiWSRuqk", LibcInternal::strncpy);
 	LIB_FUNC("Ls4tzzhimqQ", LibcInternal::strcat);
 	LIB_FUNC("ob5xAW4ln-0", LibcInternal::strchr);

src/libs/libKernel.cpp

@@ -1370,22 +1370,8 @@ static void KYTY_SYSV_ABI KernelRtldSetApplicationHeapAPI(void* api[]) {
 	[[maybe_unused]] auto* heap_posix_memalign = api[6];
 }
 
-static int64_t KYTY_SYSV_ABI write(int d, const char* str, int64_t size) {
-	// PRINT_NAME();
-
-	EXIT_NOT_IMPLEMENTED(d < 0);
-
-	if (Network::Net::IsSocket(d)) {
-		return Network::Net::Send(d, str, static_cast<uint64_t>(size), 0);
-	}
-
-	if (d > 2) {
-		return POSIX_N_CALL(FileSystem::KernelWrite(d, str, size));
-	}
-
-	int size_int = static_cast<int>(size);
-
-	return size_int;
+static int64_t KYTY_SYSV_ABI write(int d, const void* buf, size_t nbytes) {
+	return POSIX_N_CALL(FileSystem::KernelWrite(d, buf, nbytes));
 }
 
 static int KYTY_SYSV_ABI open(const char* path, int flags, int mode) {

src/libs/libNpEntitlementAccess.cpp

@@ -172,8 +172,9 @@ MakeAddcontInfo(const Loader::AddContentEntry& entry) {
 	NpEntitlementAccessAddcontEntitlementInfo info {};
 	std::memcpy(info.entitlement_label.data, entry.entitlement_label.data(),
 	            entry.entitlement_label.size());
-	info.package_type    = 2; // Additional content with extra data.
-	info.download_status = 4; // Installed.
+	const bool extra_data = !entry.mount_dir.empty();
+	info.package_type     = extra_data ? 2 : 3; // PSAC (with extra data) or PSAL (license only).
+	info.download_status  = extra_data ? 4 : 0; // INSTALLED or NO_EXTRA_DATA.
 	return info;
 }
 

src/loader/additionalContent.cpp

@@ -10,10 +10,24 @@ namespace Loader {
 
 static std::vector<AddContentEntry> g_add_content;
 
-void AddContentLoad(const std::filesystem::path& dir, const std::string& title_id) {
+static auto FindLabel(std::string_view label) {
+	return std::find_if(g_add_content.begin(), g_add_content.end(),
+	                    [&](const auto& entry) { return entry.entitlement_label == label; });
+}
+
+void AddContentLoad(const std::filesystem::path& dir, const std::vector<AddContentEntry>& user,
+                    const std::string& title_id) {
 	g_add_content = Common::FindAddContent(dir, title_id);
+	for (const auto& entry: user) {
+		const auto it = FindLabel(entry.entitlement_label);
+		if (it == g_add_content.end()) {
+			g_add_content.push_back(entry);
+		} else {
+			*it = entry;
+		}
+	}
 	for (const auto& entry: g_add_content) {
-		LOGF("AddContent: registered %s -> %s\n", entry.content_id.c_str(),
+		LOGF("AddContent: registered %s -> %s\n", entry.entitlement_label.c_str(),
 		     Common::PathToString(entry.mount_dir).c_str());
 	}
 }
@@ -26,10 +40,7 @@ const AddContentEntry* AddContentFind(const char* label, size_t label_size) {
 	if (label == nullptr || label_size == 0) {
 		return nullptr;
 	}
-	const std::string_view value(label, std::find(label, label + label_size, '\0') - label);
-	const auto             it =
-	    std::find_if(g_add_content.begin(), g_add_content.end(),
-	                 [&](const auto& entry) { return entry.entitlement_label == value; });
+	const auto it = FindLabel({label, std::find(label, label + label_size, '\0')});
 	return it == g_add_content.end() ? nullptr : &*it;
 }
 

src/loader/additionalContent.h

@@ -9,7 +9,9 @@ namespace Loader {
 
 using AddContentEntry = Common::AddContentEntry;
 
-void AddContentLoad(const std::filesystem::path& dir, const std::string& title_id);
+// User entries replace detected packages with the same label.
+void AddContentLoad(const std::filesystem::path& dir, const std::vector<AddContentEntry>& user,
+                    const std::string& title_id);
 const std::vector<AddContentEntry>& AddContentGetAll();
 const AddContentEntry*              AddContentFind(const char* label, size_t label_size);
 

src/loader/loadExec.cpp

@@ -149,7 +149,8 @@ int LoadExec(const char* path, const char* const argv[]) {
 	}
 	CloseHandle(process.hThread);
 	CloseHandle(process.hProcess);
-	std::quick_exit(0);
+	// Nonzero, or the launcher's "cmd /K ... && exit" closes the console the new process shares.
+	std::quick_exit(1);
 #else
 	std::ifstream            command("/proc/self/cmdline", std::ios::binary);
 	std::vector<std::string> host_arguments;

src/loader/runtimeLinker.cpp

@@ -14,6 +14,7 @@
 #include "common/threads.h"
 #include "common/virtualMemory.h"
 #include "graphics/host_gpu/pageManager.h"
+#include "graphics/shader/recompiler/ir/passes/SrtWalker.h"
 #include "kernel/fileSystem.h"
 #include "kernel/memory.h"
 #include "kernel/pthread.h"
@@ -672,6 +673,11 @@ static bool KytyExceptionHandler(const Common::HostException::ExceptionInfo& exc
 			Common::DebugCounters::Add(Common::DebugCounters::Counter::GpuFaults);
 			return true;
 		}
+		if (const auto length =
+		        Libs::Graphics::ShaderRecompiler::IR::SrtWalkerLoadLength(info->exception_address)) {
+			Loader::X64InstructionEmulator::SkipLoad(info->native_context, length);
+			return true;
+		}
 	}
 	// Report whatever guest context can be read safely before terminating: which guest thread
 	// faulted, the register file, the faulting code bytes and the top of its stack.

src/loader/x64InstructionEmulator.cpp

@@ -822,4 +822,10 @@ bool TryEmulate(void* native_context) {
 	       TryEmulateCpuExtensions(context);
 }
 
+void SkipLoad(void* native_context, size_t length) {
+	Context context {static_cast<decltype(Context::native)>(native_context)};
+	context.Gpr(0) = 0;
+	context.Advance(length);
+}
+
 } // namespace Loader::X64InstructionEmulator

src/loader/x64InstructionEmulator.h

@@ -1,9 +1,13 @@
 #ifndef KYTY_LOADER_X64_INSTRUCTION_EMULATOR_H_
 #define KYTY_LOADER_X64_INSTRUCTION_EMULATOR_H_
 
+#include <cstddef>
+
 namespace Loader::X64InstructionEmulator {
 
 [[nodiscard]] bool TryEmulate(void* native_context);
+// Completes a faulting load into EAX as a zero read and resumes after it.
+void SkipLoad(void* native_context, size_t length);
 
 } // namespace Loader::X64InstructionEmulator
 

src/main.cpp

@@ -48,6 +48,9 @@ static void PrintUsage() {
 	::printf("  --game <dir|elf|zar>                 Game directory, ELF, or ZArchive to load.\n");
 	::printf("  --game-patch <json>                  ETAHen cheat file.\n");
 	::printf("  --addcont-dir <dir>                  Add-ons (DLC) folder (default: none).\n");
+	::printf(
+	    "  --addcont <label>[=<dir>]            Add-on (DLC) by entitlement label with its data in\n"
+	    "                                       <dir>, or license only without <dir>. Repeatable.\n");
 	::printf("  --screen-width <num>                 Window width. Default: 1280.\n");
 	::printf("  --screen-height <num>                Window height. Default: 720.\n");
 	::printf(
@@ -77,6 +80,7 @@ static void PrintUsage() {
 	::printf("  --gpu-assisted-validation            Bounds-check shader accesses on the GPU.\n"
 	         "                                       Implies --vulkan-validation; very slow.\n");
 	::printf("  --shader-validation                  Enable shader validation.\n");
+	::printf("  --shader-cache                       Keep compiled shaders and pipelines between boots.\n");
 	::printf("  --tessellation                      Draw tessellation patches; skipped by default.\n");
 	::printf("  --occlusion-queries                  Count occlusion query samples on the GPU.\n");
 	::printf("  --shader-optimization-type <value>   None, Size, or Performance.\n");
@@ -202,6 +206,7 @@ static constexpr Flag FLAGS[] = {
     {"--vulkan-validation", &ConfigOptions::vulkan_validation_enabled},
     {"--gpu-assisted-validation", &ConfigOptions::gpu_assisted_validation_enabled},
     {"--shader-validation", &ConfigOptions::shader_validation_enabled},
+    {"--shader-cache", &ConfigOptions::shader_cache_enabled},
     {"--tessellation", &ConfigOptions::tessellation_enabled},
     {"--occlusion-queries", &ConfigOptions::occlusion_queries_enabled},
     {"--command-buffer-dump", &ConfigOptions::command_buffer_dump_enabled},
@@ -303,6 +308,14 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 				return false;
 			}
 			options.addcont_dir = Common::PathFromUtf8(Common::FixFilenameSlash(value));
+		} else if (arg == "--addcont") {
+			auto entry = Common::ParseAddContent(value);
+			if (!entry) {
+				::printf("--addcont needs <16-character A-Z/0-9 label>[=<existing folder>]: %s\n",
+				         value.c_str());
+				return false;
+			}
+			options.addcont.push_back(std::move(*entry));
 		} else if (arg == "--screen-width") {
 			if (!ParseUint32(value, options.config.screen_width) ||
 			    options.config.screen_width == 0) {

tests/KernelFileSystemTests.cpp

@@ -506,9 +506,9 @@ void TestAddContent(const std::filesystem::path& root) {
   write(folder / "duplicate" / "sce_sys" / "param.json", param("PPSA11111", label(1)));
   write(folder / label(0) / "payload.txt", "DLC payload");
 
-  Loader::AddContentLoad(folder, "");
+  Loader::AddContentLoad(folder, {}, "");
   Check(Loader::AddContentGetAll().empty(), "unknown title never exposes unrelated add-ons");
-  Loader::AddContentLoad(folder, "PPSA11111");
+  Loader::AddContentLoad(folder, {}, "PPSA11111");
   Check(Loader::AddContentGetAll().size() == 65,
         "scan filters title before deduplication and ignores invalid/non-package metadata");
 
@@ -603,7 +603,39 @@ void TestAddContent(const std::filesystem::path& root) {
   for (size_t i = 1; i < mounts.size(); ++i) {
     Check(unmount(&mounts[i]) == OK, "unmount fixture");
   }
-  Loader::AddContentLoad(folder / "missing", "PPSA11111");
+
+  const auto user_dir = root / "user-add-on";
+  write(user_dir / "payload.txt", "user payload");
+  const auto user_arg = label(0) + "=" + Common::PathToString(user_dir);
+  Check(!Common::ParseAddContent("SHORT") && !Common::ParseAddContent("lowercase0000000") &&
+            !Common::ParseAddContent("LICENSE0000000001") &&
+            !Common::ParseAddContent(user_arg + "/missing"),
+        "user add-on needs a 16-character A-Z/0-9 label and an existing folder");
+  const auto licensed = Common::ParseAddContent("LICENSE000000001");
+  const auto replaced = Common::ParseAddContent(user_arg);
+  const auto duplicate =
+      Common::ParseAddContent("LICENSE000000001=" + Common::PathToString(user_dir));
+  Check(licensed && replaced && duplicate, "user add-ons parse with and without extra data");
+  Loader::AddContentLoad(folder, {*duplicate, *licensed, *replaced}, "PPSA11111");
+  Check(list(0, nullptr, 0, &count) == OK && count == 69,
+        "user add-ons replace detected labels and append new ones");
+  Label user {};
+  std::memcpy(user.data, "LICENSE000000001", 17);
+  Check(get(0, &user, &single) == OK && single.package_type == 3 && single.download_status == 0 &&
+            mount(0, &user, &point) == static_cast<int>(0x80d90005u),
+        "last entry of a label wins and a license-only add-on has no extra data to mount");
+  Check(mount(0, &info[3].label, &point) == OK, "user add-on mounts its extra data");
+  const std::string user_file = std::string(point.data) + "/payload.txt";
+  const int user_fd = FileSystem::KernelOpen(user_file.c_str(), 0, 0);
+  std::array<char, 12> user_payload {};
+  Check(user_fd >= 3 &&
+            FileSystem::KernelRead(user_fd, user_payload.data(), user_payload.size()) ==
+                user_payload.size() &&
+            std::string_view(user_payload.data(), user_payload.size()) == "user payload" &&
+            FileSystem::KernelClose(user_fd) == OK && unmount(&point) == OK,
+        "user add-on replaces the detected folder of its label");
+
+  Loader::AddContentLoad(folder / "missing", {}, "PPSA11111");
   Check(list(0, info.data(), 4, &count) == OK && count == 3,
         "missing add-ons folder retains GTA V entitlements");
   for (size_t i = 0; i < existing.size(); ++i) {

tests/MemoryTrackerTests.cpp

@@ -213,12 +213,6 @@ struct ProtectionCall {
 std::vector<ProtectionCall> g_protection_log;
 std::mutex g_protection_log_mutex;
 
-struct ProtectionGate {
-  std::binary_semaphore entered{0};
-  std::binary_semaphore resume{0};
-};
-ProtectionGate *g_protection_gate = nullptr;
-
 void ResetProtectionLog() {
   g_protection_calls = 0;
   g_protection_log.clear();
@@ -226,10 +220,6 @@ void ResetProtectionLog() {
 
 bool ProtectAddressSpace(uint64_t vaddr, uint64_t size,
                          Common::VirtualMemory::Mode mode) {
-  if (g_protection_gate != nullptr && mode == Common::VirtualMemory::Mode::ReadWrite) {
-    g_protection_gate->entered.release();
-    g_protection_gate->resume.acquire();
-  }
   uint32_t protection = PAGE_NOACCESS;
   if (mode == Common::VirtualMemory::Mode::Read) {
     protection = PAGE_READONLY;
@@ -417,74 +407,6 @@ void TestCpuDirtyUpload() {
   Release(memory);
 }
 
-void TestBdaSyncDirtiness() {
-  constexpr auto page = Libs::Graphics::TRACKER_PAGE_SIZE;
-  TrackerHarness harness;
-  auto &tracker = harness.tracker;
-  auto *memory = Allocate(harness.page_manager, 3);
-  const auto address = reinterpret_cast<uint64_t>(memory);
-  Check(tracker.TakeBdaSync().Empty(), "new tracker queued unowned memory");
-  tracker.QueueBdaSync(address, 3 * page);
-  Check(tracker.IsRegionCpuModified(address, page), "new region was clean");
-  auto pending = tracker.TakeBdaSync();
-  Check(pending.Contains(address, 3 * page) && !pending.Intersects(address + 3 * page, page),
-        "mapped or registered range did not bound initial BDA synchronization");
-  tracker.ForEachUploadRange(address, 3 * page, false,
-                             [](uint64_t, uint64_t) noexcept {}, []() noexcept {});
-  Check(tracker.TakeBdaSync().Empty(), "clean upload queued redundant BDA work");
-
-  tracker.MarkRegionAsCpuModified(address + 17, 1);
-  pending = tracker.TakeBdaSync();
-  Check(pending.Contains(address, page) && !pending.Intersects(address + page, page) &&
-            tracker.TakeBdaSync().Empty(),
-        "byte dirtiness did not drain exactly its whole tracker page");
-  // A write after Take belongs to the next drain, even when its page was just taken.
-  tracker.MarkRegionAsCpuModified(address + 31, 1);
-  Check(tracker.TakeBdaSync().Contains(address, page), "re-dirty event was lost");
-  tracker.ForEachUploadRange(address, page, true,
-                             [](uint64_t, uint64_t) noexcept {}, []() noexcept {});
-  Check(tracker.TakeBdaSync().Empty(), "GPU ownership queued an unnecessary upload");
-  tracker.InvalidateRegion(address + 7, 1, [&] {
-    tracker.UnmarkRegionAsGpuModified(address, page);
-    tracker.MarkRegionAsCpuModified(address, page);
-  });
-  Check(tracker.TakeBdaSync().Contains(address, page),
-        "GPU-to-CPU invalidation did not queue BDA synchronization");
-  tracker.InvalidateRegion(address + page + 7, 1, [] {});
-  Check(tracker.TakeBdaSync().Contains(address + page, page),
-        "CPU invalidation did not queue BDA synchronization");
-  tracker.UntrackMemory(address, 3 * page);
-  Check(tracker.TakeBdaSync().Contains(address, 3 * page),
-        "untracked memory did not queue its newly dirty pages");
-  Release(memory);
-}
-
-void TestBdaSyncConcurrentDrain() {
-  constexpr auto page = Libs::Graphics::TRACKER_PAGE_SIZE;
-  TrackerHarness harness;
-  auto &tracker = harness.tracker;
-  auto *memory = Allocate(harness.page_manager, 1);
-  const auto address = reinterpret_cast<uint64_t>(memory);
-  tracker.ForEachUploadRange(address, page, false,
-                             [](uint64_t, uint64_t) noexcept {}, []() noexcept {});
-  (void)tracker.TakeBdaSync();
-  ProtectionGate gate;
-  g_protection_gate = &gate;
-  std::jthread writer([&] { tracker.MarkRegionAsCpuModified(address + 17, 1); });
-  gate.entered.acquire();
-  // Ownership is still changing. Publishing its event before this point would let
-  // a drain consume it before the writer has finished releasing page protection.
-  Check(tracker.TakeBdaSync().Empty(), "BDA event preceded its state transition");
-  gate.resume.release();
-  writer.join();
-  g_protection_gate = nullptr;
-  Check(tracker.TakeBdaSync().Contains(address, page) &&
-            tracker.IsRegionCpuModified(address, page),
-        "concurrent drain lost a completed CPU dirty transition");
-  tracker.UntrackMemory(address, page);
-  Release(memory);
-}
-
 void TestCleanUploadPreservesOwnership() {
   constexpr auto page_size = Libs::Graphics::TRACKER_PAGE_SIZE;
   constexpr auto region_size = Libs::Graphics::TRACKER_REGION_SIZE;
@@ -1280,8 +1202,6 @@ int main(int argc, char **argv) {
   TestQueriesDoNotRequireMappedOwnership();
   TestConcurrentRegionPublication();
   TestCpuDirtyUpload();
-  TestBdaSyncDirtiness();
-  TestBdaSyncConcurrentDrain();
   TestCleanUploadPreservesOwnership();
   TestRangeInvalidation();
   TestGpuReacquisitionAfterInvalidation();

tests/ResourceMaterializationTests.cpp

@@ -1,5 +1,7 @@
+#include "graphics/shader/recompiler/ir/Serialize.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "SrtWalkerTestMemory.h"
 
 #include <bit>
 #include <cstdio>
@@ -16,6 +18,22 @@ void Check(bool value, const char *text) {
   }
 }
 
+// Every extracted plan below is the one a restarted game loads from the shader cache: exactly
+// its bytes load, and the loaded plan saves back to them.
+Libs::Graphics::ShaderRecompiler::IR::ResourcePlan
+Extract(const Libs::Graphics::ShaderRecompiler::IR::Program &program) {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  const auto bytes = SaveResourcePlan(ExtractResourcePlan(program));
+  auto extended = bytes;
+  extended.push_back(0u);
+  ResourcePlan loaded, truncated, padded;
+  Check(LoadResourcePlan(bytes, loaded) && SaveResourcePlan(loaded) == bytes &&
+            !LoadResourcePlan(std::span(bytes).first(bytes.size() - 1u), truncated) &&
+            !LoadResourcePlan(extended, padded),
+        "resource plan did not round-trip through its exact cache bytes");
+  return loaded;
+}
+
 bool RejectSpecializationRead(void *userdata, uint64_t, std::span<uint32_t>) {
   ++*static_cast<uint32_t *>(userdata);
   return false;
@@ -32,6 +50,24 @@ AddValueBlock(Libs::Graphics::ShaderRecompiler::IR::Program &program) {
   return *result;
 }
 
+// Lists one extra source per block and walks the plan: each source's activity is its block's reachability.
+std::vector<uint8_t> ReachedBlocks(Libs::Graphics::ShaderRecompiler::IR::ResourcePlan &plan,
+                                   const Libs::Graphics::ShaderRecompiler::IR::SrtRuntime &runtime) {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  const auto first = static_cast<uint32_t>(plan.descriptor_sources.size());
+  for (uint32_t block = 0; block < plan.control_flow.size(); ++block) {
+    plan.descriptor_sources.emplace_back();
+    plan.control_flow[block].sources.push_back(first + block);
+  }
+  CompileSrtWalker(plan);
+  std::vector<uint32_t> flat;
+  Check(WalkSrt(plan, runtime, flat, nullptr), "reachability walk failed");
+  std::vector<uint8_t> reached;
+  for (uint32_t block = 0; block < plan.control_flow.size(); ++block)
+    reached.push_back(WalkedActive(plan, first + block));
+  return reached;
+}
+
 Libs::Graphics::ShaderRecompiler::IR::Program SrtProgram(uint64_t address) {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   Program program;
@@ -73,7 +109,7 @@ Libs::Graphics::ShaderRecompiler::IR::ResourcePlan UnbasedFlatPlan() {
   program.resource_tracking_complete = true;
   AddValueBlock(program);
   program.info.uses_dma = true;
-  return ExtractResourcePlan(program);
+  return Extract(program);
 }
 
 Libs::Graphics::ShaderRecompiler::IR::ResourcePlan UserDataBufferPlan() {
@@ -94,7 +130,7 @@ Libs::Graphics::ShaderRecompiler::IR::ResourcePlan UserDataBufferPlan() {
   source.dword_count = 4;
   program.descriptor_sources.push_back(source);
   program.info.buffers.push_back({.source = 0});
-  return ExtractResourcePlan(program);
+  return Extract(program);
 }
 
 Libs::Graphics::ShaderRecompiler::IR::Program MixedSamplerProgram() {
@@ -149,10 +185,10 @@ Libs::Graphics::ShaderRecompiler::IR::Program MixedSamplerProgram() {
   return program;
 }
 
-void TestMappedSrtUsesDirectReaderByDefault() {
+void TestWalkerLoadsBypassSpecializationReader() {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   const uint32_t dword = 0x12345678;
-  auto plan = ExtractResourcePlan(SrtProgram(reinterpret_cast<uint64_t>(&dword)));
+  auto plan = Extract(SrtProgram(reinterpret_cast<uint64_t>(&dword)));
   uint32_t specialization_reads = 0;
   const SrtRuntime runtime{.userdata = &specialization_reads,
                            .read_specialization_memory =
@@ -162,10 +198,10 @@ void TestMappedSrtUsesDirectReaderByDefault() {
   Check(MaterializeResources(plan, runtime, snapshot, specialization),
         "mapped SRT stage materialization failed");
   Check(specialization_reads == 0,
-        "ordinary SRT read used the specialization reader");
+        "walker load used the specialization reader");
   Check(snapshot.flattened_srt.size() == 1 &&
             snapshot.flattened_srt[0] == dword,
-        "cache rematerialization did not use the direct reader by default");
+        "walker did not load the mapped SRT DWORD");
 }
 
 void TestNativeNestedSrtRefresh() {
@@ -204,22 +240,11 @@ void TestNativeNestedSrtRefresh() {
   leaf->Arg(0).Instruction()->SetArg(1, pointer_high);
   leaf->SetArg(1, Value(7u));
   read(leaf->Arg(0), 11u);
-  auto plan = ExtractResourcePlan(program);
+  auto plan = Extract(program);
   Check(plan.srt_code != nullptr, "nested user-data tree did not compile an x86 walker");
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   const SrtRuntime runtime{.user_data = user_data};
-  {
-    SrtWalker walker(plan, runtime);
-    Check(walker.RefreshFlatBuffer(snapshot.flattened_srt),
-          "native SRT refresh failed");
-    const auto loaded = first[1];
-    first[1] ^= 0xffffffffu;
-    uint32_t value = 0;
-    Check(walker.Evaluate(plan.srt_reads[0].value, value) && value == loaded,
-          "descriptor evaluation reloaded a value already copied by the native walker");
-    first[1] = loaded;
-  }
   for (const auto *data : {&first, &second}) {
     table = pointer_words(data->data());
     Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
@@ -227,14 +252,20 @@ void TestNativeNestedSrtRefresh() {
                   std::vector<uint32_t>{(*data)[1], table[0], table[1], (*data)[2]},
           "native SRT refresh lost pointer masking, alignment, or current values");
   }
+  {
+    // Unmapped tables and the null pointer they yield read as zero.
+    FakeGuestMemory unmapped;
+    const std::array<uint32_t, 2> unmapped_table{0x1000u, 0u};
+    Check(MaterializeResources(plan, {.user_data = unmapped_table}, snapshot, specialization) &&
+              snapshot.flattened_srt == std::vector<uint32_t>(4, 0u),
+          "faulting walker loads of a table and a null pointer did not read zero");
+  }
   Check(!MaterializeResources(plan, {.user_data = std::span(user_data).first(1)},
                               snapshot, specialization),
         "native SRT walk accepted truncated user data");
   const std::array<uint32_t, 2> fake_user_data{0x1003u, 0xabcd0000u};
   uint32_t reads = 0;
-  const SrtRuntime callback_runtime{
-      .user_data = fake_user_data,
-      .read_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
         ++*static_cast<uint32_t *>(data);
         switch (address) {
           case 0x1000u: words[0] = 0x2003u; break;
@@ -244,18 +275,17 @@ void TestNativeNestedSrtRefresh() {
           default: return false;
         }
         return true;
-      },
-      .userdata = &reads};
-  Check(MaterializeResources(plan, callback_runtime, snapshot, specialization) && reads == 4 &&
+      }, &reads);
+  Check(MaterializeResources(plan, {.user_data = fake_user_data}, snapshot, specialization) && reads == 4 &&
             snapshot.flattened_srt ==
                 std::vector<uint32_t>{0x1234u, 0x2003u, 0xabcd0000u, 0x5678u},
-        "native-eligible SRT bypassed its callback or repeated nested pointer reads");
+        "native SRT walk repeated nested pointer reads");
   leaf->SetArg(0, root);
   leaf->SetArg(1, Value(0u));
   program.srt_reads.resize(1);
   for (const int32_t offset : {-1, 7}) {
     program.memory_info[0].offset = static_cast<uint32_t>(offset);
-    auto bounded = ExtractResourcePlan(program);
+    auto bounded = Extract(program);
     auto base = pointer_words(second.data() + 1);
     Check(bounded.srt_code != nullptr &&
               MaterializeResources(bounded, {.user_data = base}, snapshot, specialization) &&
@@ -266,17 +296,6 @@ void TestNativeNestedSrtRefresh() {
     Check(!MaterializeResources(bounded, {.user_data = base}, snapshot, specialization),
           "native SRT accepted address underflow or 48-bit overflow");
   }
-  high.Instruction()->SetArg(0, Value(ScalarReg(6)));
-  Check(ExtractResourcePlan(program).srt_code == nullptr,
-        "native SRT accepted nonadjacent user-data pointer words");
-  high.Instruction()->SetArg(0, Value(ScalarReg(5)));
-  program.memory_info[0].offset = 0;
-  program.srt_reads.push_back({pointer_low, 1});
-  program.srt_reads.push_back({pointer_high, 2});
-  leaf->SetArg(0, emit(ValueOpcode::GetAddressResource, {pointer_low, pointer_high}));
-  pointer_high.Instruction()->SetArg(1, Value(8u));
-  Check(ExtractResourcePlan(program).srt_code == nullptr,
-        "native SRT accepted nonadjacent nested pointer words");
 }
 
 void TestIntegerRuntimeValueFollowsSrtReads() {
@@ -328,7 +347,7 @@ void TestIntegerRuntimeValueFollowsSrtReads() {
         "cyclic SRT read-first-lane dependency was accepted");
 }
 
-void TestSrtAliasesRetainReadPolicy() {
+void TestSrtAliasesShareReads() {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   auto program = SrtProgram(0x1000u);
   auto &block = *program.blocks[0];
@@ -356,70 +375,33 @@ void TestSrtAliasesRetainReadPolicy() {
   DescriptorSource indirect;
   indirect.indirect_descriptor.emplace(DescriptorSource::IndirectDescriptor{}).sources = {0, 1};
   program.descriptor_sources.push_back(indirect);
-  auto plan = ExtractResourcePlan(program);
+  auto plan = Extract(program);
   for (const auto &inst : plan.value_storage) {
     Check(inst.GetOpcode() != ValueOpcode::ReadConst &&
               inst.GetOpcode() != ValueOpcode::GetSrtResource,
           "retained plan still contains flattened-read aliases");
   }
-  Check(plan.descriptor_sources[0].dwords[0] == plan.srt_reads[0].value &&
-            plan.srt_reads[0].value.Instruction()->Flags<SrtReadFlags>().clean == 1u &&
-            plan.srt_reads[1].value.Instruction()->Flags<SrtReadFlags>().clean == 0u &&
-            plan.srt_reads[2].value.Instruction()->Flags<SrtReadFlags>().clean == 1u,
-        "alias normalization lost read identity or retained discarded EXEC provenance");
-  Check(plan.srt_code == nullptr,
-        "captured strict reads compiled an unused x86 walker");
-  struct Reads { uint32_t strict = 0; uint32_t ordinary = 0; bool dirty = false; } reads;
-  const SrtRuntime runtime{
-      .read_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
-        ++static_cast<Reads *>(data)->ordinary;
-        if (address == 0x2000u) words[0] = 0x2222u;
-        else if (address == 0x3000u) words[0] = 0x8000u;
-        else if (address == 0x8000u) words[0] = 0xeeeeu;
-        else return false;
-        return true;
-      },
-      .userdata = &reads,
-      .read_specialization_memory = +[](void *data, uint64_t address,
-                                        std::span<uint32_t> words) {
+  Check(plan.descriptor_sources[0].dwords[0] == plan.srt_reads[0].value,
+        "alias normalization lost read identity");
+  struct Reads { uint32_t count = 0; bool dirty = false; } reads;
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
         auto &reads = *static_cast<Reads *>(data);
-        ++reads.strict;
-        if (reads.dirty) return false;
-        if (address == 0x3000u) words[0] = 0x1000u;
-        else if (address == 0x1000u) words[0] = 0x1111u;
+        ++reads.count;
+        if (address == 0x1000u) words[0] = 0x1111u;
+        else if (address == 0x2000u) words[0] = 0x2222u;
+        else if (address == 0x3000u && !reads.dirty) words[0] = 0x1000u;
         else return false;
         return true;
-      }};
+      }, &reads);
   std::vector<uint32_t> flat;
-  DescriptorValue descriptor;
-  {
-    SrtWalker clean(plan, CleanRuntime(runtime));
-    SrtWalker walker(plan, runtime, &clean);
-    Check(walker.RefreshFlatBuffer(flat) &&
-              flat == std::vector<uint32_t>{0x1111u, 0x2222u, 0x1000u} &&
-              walker.EvaluateDescriptor(0, descriptor) && descriptor.dwords[0] == 0x1111u &&
-              reads.strict == 2 && reads.ordinary == 1,
-          "normalized aliases changed nested strict reads or repeated a shared read");
-  }
-  Check(!SrtWalker(plan, runtime).RefreshFlatBuffer(flat),
-        "strict flat read accepted a missing clean evaluator");
-  auto no_reader = runtime;
-  no_reader.read_specialization_memory = nullptr;
-  {
-    SrtWalker clean(plan, CleanRuntime(no_reader));
-    Check(!SrtWalker(plan, no_reader, &clean).RefreshFlatBuffer(flat),
-          "strict flat read accepted a missing strict reader");
-  }
-  Check(SrtWalker(plan, runtime).EvaluateDescriptor(0, descriptor) &&
-            descriptor.dwords[0] == 0xeeeeu && reads.strict == 2 && reads.ordinary == 3,
-        "direct descriptor evaluation without a clean evaluator changed read domains");
-  reads.dirty = true;
-  {
-    SrtWalker clean(plan, CleanRuntime(runtime));
-    Check(!SrtWalker(plan, runtime, &clean).RefreshFlatBuffer(flat) &&
-              reads.strict == 3 && reads.ordinary == 3,
-          "dirty strict pointer fell back to an ordinary read");
-  }
+  Check(WalkSrt(plan, {}, flat, nullptr) && flat == std::vector<uint32_t>{0x1111u, 0x2222u, 0x1000u} &&
+            reads.count == 3,
+        "normalized aliases changed nested reads or repeated a shared read");
+  // An unreadable pointer reads as zero, and so does the null table it yields.
+  reads = {.dirty = true};
+  Check(WalkSrt(plan, {}, flat, nullptr) && flat == std::vector<uint32_t>{0u, 0x2222u, 0u} &&
+            reads.count == 3,
+        "unreadable pointer did not read as zero");
 }
 
 void TestUniformVectorDescriptorRead() {
@@ -453,59 +435,66 @@ void TestUniformVectorDescriptorRead() {
   program.descriptor_sources.push_back(source);
   program.info.buffers.push_back({.source = 0, .written = true});
   Check(ValidateRuntimeValue(program, Value(&count)), "uniform DWORD count was rejected");
-  struct Reads { uint32_t value = 72; uint32_t strict = 0; uint32_t ordinary = 0; bool clean = true; } reads;
-  const SrtRuntime runtime{
-      .read_memory = [](void *data, uint64_t, std::span<uint32_t> words) {
-        ++static_cast<Reads *>(data)->ordinary;
-        words[0] = 999;
-        return true;
-      },
-      .userdata = &reads,
-      .read_specialization_memory = [](void *data, uint64_t address, std::span<uint32_t> words) {
+  struct Reads { uint32_t value = 72; uint32_t count = 0; bool readable = true; } reads;
+  FakeGuestMemory fake([](void *data, uint64_t address, std::span<uint32_t> words) {
         auto &reads = *static_cast<Reads *>(data);
-        ++reads.strict;
-        if (!reads.clean || address != 0x1000u || words.size() != 1) return false;
+        ++reads.count;
+        if (!reads.readable || address != 0x1000u || words.size() != 1) return false;
         words[0] = reads.value;
         return true;
-      }};
+      }, &reads);
+  const SrtRuntime runtime;
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   for (const bool written : {false, true}) {
     program.info.buffers[0].written = written;
-    auto plan = ExtractResourcePlan(program);
-    Check(plan.control_flow.empty() && plan.requires_specialization_memory &&
-              plan.capture_specialization_reads,
+    auto plan = Extract(program);
+    Check(plan.control_flow.empty() && plan.capture_specialization_reads,
           "vector descriptor read depended on incidental control-flow capture");
     reads = {};
     Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
-              snapshot.buffers[0].dwords[2] == 145 && reads.strict == 1 && reads.ordinary == 0 &&
+              snapshot.buffers[0].dwords[2] == 145 && reads.count == 1 &&
               snapshot.specialization_reads ==
                   std::vector<std::pair<uint64_t, uint64_t>>{{0x1000u, 4u}},
           "vector descriptor input was not read and captured exactly once");
-    reads.clean = false;
-    reads.strict = 0;
-    Check(!MaterializeResources(plan, runtime, snapshot, specialization) &&
-              reads.strict == 1 && reads.ordinary == 0,
-          "dirty vector descriptor input fell back to an ordinary memory read");
+    // An unreadable input reads as zero, so the count selects 0 + 1.
+    reads = {.readable = false};
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              snapshot.buffers[0].dwords[2] == 1 && reads.count == 1,
+          "unreadable vector descriptor input did not read as zero");
+  }
+  // A read guarded by its EXEC mask depends on that mask, so it cannot reuse the unmasked read.
+  auto &guarded = block.AppendNewInst(ValueOpcode::LoadBufferU32,
+      {Value(&handle), Value(0u), Value(0u), Value(0u), Value(&active)});
+  guarded.SetFlags(MemoryFlags{.index = 0});
+  auto &guarded_lane = block.AppendNewInst(ValueOpcode::ReadFirstLane, {Value(&guarded), Value(&active)});
+  program.descriptor_sources.push_back({.dwords = {Value(&guarded_lane)}, .dword_count = 1});
+  {
+    auto plan = Extract(program);
+    uint32_t value = 0;
+    reads = {};
+    Check(EvaluateSrtValue(plan, runtime, plan.descriptor_sources.back().dwords[0], value) &&
+              value == 72u && reads.count == 1,
+          "EXEC-guarded vector read lost its mask");
   }
   count.SetArg(4, Value(false));
   auto &inactive = block.AppendNewInst(ValueOpcode::ReadFirstLane, {Value(&count), Value(false)});
   program.descriptor_sources.push_back({.dwords = {Value(&inactive)}, .dword_count = 1});
-  reads.strict = 0;
+  reads.count = 0;
   uint32_t result = 99;
   {
-    const auto plan = ExtractResourcePlan(program);
-    Check(SrtWalker(plan, runtime).Evaluate(plan.descriptor_sources.back().dwords[0], result) &&
-              result == 0 && reads.strict == 0 && reads.ordinary == 0,
+    auto plan = Extract(program);
+    Check(EvaluateSrtValue(plan, runtime, plan.descriptor_sources.back().dwords[0], result) &&
+              result == 0 && reads.count == 0,
           "literal false EXEC read vector memory");
   }
   count.SetArg(4, Value(true));
   handle.SetArg(3, Value(0x204u));
   program.descriptor_sources.back().dwords[0] = Value(&count);
   {
-    const auto plan = ExtractResourcePlan(program);
-    Check(SrtWalker(plan, runtime).Evaluate(plan.descriptor_sources.back().dwords[0], result) &&
-              result == 0 && reads.strict == 0 && reads.ordinary == 0,
+    auto plan = Extract(program);
+    Check(EvaluateSrtValue(plan, runtime, plan.descriptor_sources.back().dwords[0], result) &&
+              result == 0 && reads.count == 0,
           "invalid vector buffer format read memory");
   }
   handle.SetArg(3, Value(0x16204u));
@@ -523,7 +512,7 @@ void TestExactReciprocalDescriptorArithmetic() {
                               std::bit_cast<float>(253u << 23u)}) {
     auto &reciprocal = block.AppendNewInst(ValueOpcode::FPRecipIFlag32, {Value::F32(divisor)});
     uint32_t result = 0;
-    Check(SrtWalker(program, {}).Evaluate(Value(&reciprocal), result) &&
+    Check(EvaluateSrtValue(program, {}, Value(&reciprocal), result) &&
               result == std::bit_cast<uint32_t>(1.f / divisor),
           "power-of-two reciprocal was not exact");
   }
@@ -532,7 +521,7 @@ void TestExactReciprocalDescriptorArithmetic() {
                               std::bit_cast<float>(254u << 23u)}) {
     auto &reciprocal = block.AppendNewInst(ValueOpcode::FPRecipIFlag32, {Value::F32(divisor)});
     uint32_t result = 0;
-    Check(!SrtWalker(program, {}).Evaluate(Value(&reciprocal), result),
+    Check(!EvaluateSrtValue(program, {}, Value(&reciprocal), result),
           "unsupported reciprocal rounding or exceptional input was accepted");
   }
 }
@@ -569,7 +558,7 @@ void TestWrittenDescriptorPredicateReads(bool memory_condition) {
   source.dword_count = 4;
   program.descriptor_sources.push_back(source);
   program.info.buffers.push_back({.source = 0, .written = true});
-  // Only memory-derived host decisions need clean, disjoint descriptor reads.
+  // Only memory-derived host decisions capture their descriptor reads.
   auto &condition = block.AppendNewInst(ValueOpcode::IEqual32,
       {memory_condition ? Value(&read) : Value(&offset),
        Value(memory_condition ? 0x8000u : 4u)});
@@ -592,38 +581,28 @@ void TestWrittenDescriptorPredicateReads(bool memory_condition) {
   store_block.AppendNewInst(ValueOpcode::StoreBufferU32,
       {Value(&output), Value(0u), Value(0u), Value(0u), Value(1u), Value(true)})
       .SetFlags(MemoryFlags{.index = 1});
-  auto plan = ExtractResourcePlan(program);
+  auto plan = Extract(program);
   Check(plan.capture_specialization_reads == memory_condition && !plan.control_flow.empty(),
         "conditional writable descriptor used the wrong memory dependency policy");
-  struct Reads { uint32_t ordinary = 0; uint32_t strict = 0; bool clean = false; } reads;
+  struct Reads { uint32_t count = 0; bool readable = false; } reads;
   const std::array<uint32_t, 1> user_data{4u};
-  const SrtRuntime runtime{
-      .user_data = user_data,
-      .read_memory = +[](void *data, uint64_t, std::span<uint32_t> words) {
-        ++static_cast<Reads *>(data)->ordinary;
-        words[0] = 0x8000u;
-        return true;
-      },
-      .userdata = &reads,
-      .read_specialization_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
         auto &reads = *static_cast<Reads *>(data);
-        ++reads.strict;
-        if (!reads.clean || address != 0x1004u) return false;
+        ++reads.count;
+        if (!reads.readable || address != 0x1004u) return false;
         words[0] = 0x8000u;
         return true;
-      }};
+      }, &reads);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
-  for (const bool clean : {false, true}) {
-    reads = {.clean = clean};
-    // A failed predicate explores both edges; its dirty read also rejects the writable descriptor.
-    const bool expected = !memory_condition || clean;
-    Check(MaterializeResources(plan, runtime, snapshot, specialization) == expected &&
-              reads.ordinary == (memory_condition ? 0u : 1u) &&
-              reads.strict == (memory_condition ? 1u : 0u),
-          "writable descriptor changed reader policy or bypassed a dirty memory predicate");
-    if (!expected) continue;
-    Check(snapshot.buffers[0].dwords[0] == 0x8000u &&
+  for (const bool readable : {false, true}) {
+    reads = {.readable = readable};
+    // The predicate and the descriptor share one read. An unreadable one is zero, which
+    // makes a memory predicate skip the store block and leave its descriptor inactive.
+    Check(MaterializeResources(plan, {.user_data = user_data}, snapshot, specialization) &&
+              reads.count == 1,
+          "writable descriptor repeated or skipped its shared read");
+    Check(snapshot.buffers[0].dwords[0] == (readable ? 0x8000u : 0u) &&
               snapshot.specialization_reads == (memory_condition
                   ? std::vector<std::pair<uint64_t, uint64_t>>{{0x1004u, 4u}}
                   : std::vector<std::pair<uint64_t, uint64_t>>{}),
@@ -672,7 +651,7 @@ void TestSharedControlPredicates() {
       program.blocks[i]->terminator.true_block = program.blocks[i == 0 ? 1 : 3];
       program.blocks[i]->terminator.false_block = program.blocks[2];
     }
-    auto plan = ExtractResourcePlan(program);
+    auto plan = Extract(program);
     Check(plan.value_storage.size() <= 2u * depth + 12u &&
               plan.control_flow.size() == 4 &&
               plan.control_flow[0].condition == plan.control_flow[1].condition &&
@@ -680,16 +659,13 @@ void TestSharedControlPredicates() {
           "shared predicates expanded into trees or retained varying dependencies");
     for (const uint32_t flag : {0u, 1u}) {
       const std::array<uint32_t, 1> user_data{flag};
-      SrtWalker walker(plan, {.user_data = user_data});
       uint32_t result = 99;
       const bool conjunction = operation == ValueOpcode::LogicalAnd;
       const bool known = conjunction ? flag == 0 : flag != 0;
-      Check(walker.Evaluate(plan.control_flow[0].condition, result) == known &&
+      Check(EvaluateSrtValue(plan, {.user_data = user_data}, plan.control_flow[0].condition, result) == known &&
                 (!known || result == uint32_t{conjunction}),
             "shared predicate changed its uniform factor or cached runtime values");
-      std::vector<uint32_t> flat;
-      Check(walker.RefreshFlatBuffer(flat) &&
-                plan.visited_blocks == (known
+      Check(ReachedBlocks(plan, {.user_data = user_data}) == (known
                     ? (conjunction ? std::vector<uint8_t>{1, 1, 0, 1}
                                    : std::vector<uint8_t>{1, 0, 1, 0})
                     : std::vector<uint8_t>{1, 1, 1, 1}),
@@ -747,35 +723,28 @@ void TestNestedExecPredicatesReuseReads() {
   use.AppendNewInst(ValueOpcode::LoadBufferU32,
       {buffer, Value(0u), Value(0u), Value(0u), Value(true)})
       .SetFlags(MemoryFlags{.index = 1});
-  auto plan = ExtractResourcePlan(program);
-  struct Reads { uint32_t value = 0; bool clean = true; uint32_t strict = 0, ordinary = 0; } reads;
-  const SrtRuntime runtime{
-      .read_memory = +[](void *data, uint64_t, std::span<uint32_t>) {
-        ++static_cast<Reads *>(data)->ordinary;
-        return false;
-      },
-      .userdata = &reads,
-      .read_specialization_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+  auto plan = Extract(program);
+  struct Reads { uint32_t value = 0; bool readable = true; uint32_t count = 0; } reads;
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
         auto &reads = *static_cast<Reads *>(data);
-        ++reads.strict;
-        Check(reads.strict <= depth + 1,
-              "shared EXEC predicates repeated scalar reads beyond their distinct masks");
-        if (!reads.clean || address != 0x1000u || words.size() != 1) return false;
+        ++reads.count;
+        if (!reads.readable || address != 0x1000u || words.size() != 1) return false;
         words[0] = reads.value;
         return true;
-      }};
+      }, &reads);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
-  for (const auto state : {Reads{64u, true}, Reads{0u, true}, Reads{0u, false}, Reads{64u, true}}) {
+  for (const auto state : {Reads{64u, true}, Reads{0u, true}, Reads{64u, false}, Reads{64u, true}}) {
     reads = state;
-    Check(MaterializeResources(plan, runtime, snapshot, specialization),
+    Check(MaterializeResources(plan, {}, snapshot, specialization),
           "nested EXEC predicates rejected a valid resource snapshot");
-    const bool inactive = reads.clean && reads.value == 0;
+    // An unreadable count reads as zero.
+    const bool inactive = !reads.readable || reads.value == 0;
     Check(snapshot.buffers[0].dwords[0] == (inactive ? 0u : 0x8000u),
           "nested EXEC predicates lost an unknown branch or retained stale runtime facts");
-    Check(reads.strict > 0 && reads.ordinary == 0 &&
-              snapshot.specialization_reads.size() <= depth + 1,
-          "shared EXEC predicates repeated scalar reads beyond their distinct masks");
+    Check(reads.count == 1 && snapshot.specialization_reads ==
+              std::vector<std::pair<uint64_t, uint64_t>>{{0x1000u, 4u}},
+          "EXEC contexts repeated a mask-independent scalar read");
   }
 }
 
@@ -817,20 +786,12 @@ void TestFiniteImageRefreshReusesScalarReads() {
       .resource_class = ImageResourceClass::Sampled,
       .numeric_class = Libs::Graphics::Prospero::TextureNumericClass::Float,
       .dimension = Libs::Graphics::ShaderRecompiler::Decoder::ImageDimension::Dim2D});
-  auto plan = ExtractResourcePlan(program);
+  auto plan = Extract(program);
   struct Reads {
     std::array<uint32_t, 3> words{0x100u, 0x200u, 0x200u};
     std::array<uint32_t, 3> counts{};
-    uint32_t ordinary = 0;
   } reads;
-  const SrtRuntime runtime{
-      .read_memory = +[](void *data, uint64_t, std::span<uint32_t>) {
-        ++static_cast<Reads *>(data)->ordinary;
-        return false;
-      },
-      .userdata = &reads,
-      .read_specialization_memory = +[](void *data, uint64_t address,
-                                        std::span<uint32_t> words) {
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
         if (words.size() != 1 || address < 0x1000u || address >= 0x100cu ||
             (address & 3u) != 0) return false;
         auto &reads = *static_cast<Reads *>(data);
@@ -838,7 +799,8 @@ void TestFiniteImageRefreshReusesScalarReads() {
         ++reads.counts[index];
         words[0] = reads.words[index];
         return true;
-      }};
+      }, &reads);
+  const SrtRuntime runtime;
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   const auto capacities = [&] {
@@ -857,15 +819,15 @@ void TestFiniteImageRefreshReusesScalarReads() {
   };
   Check(MaterializeResources(plan, runtime, snapshot, specialization),
         "finite image materialization failed");
-  Check(reads.ordinary == 0 && reads.counts == std::array<uint32_t, 3>{1, 1, 1} &&
+  Check(reads.counts == std::array<uint32_t, 3>{1, 1, 1} &&
             snapshot.specialization_reads.size() == 3,
-        "finite image candidates repeated scalar reads or bypassed clean provenance");
+        "finite image candidates repeated or skipped scalar reads");
   check_mapping({0, 1, 1, 1});
   const auto warm_capacities = capacities();
   reads.words = {0x300u, 0x300u, 0x400u};
   Check(MaterializeResources(plan, runtime, snapshot, specialization),
         "finite image refresh failed after descriptor changes");
-  Check(reads.ordinary == 0 && reads.counts == std::array<uint32_t, 3>{2, 2, 2} &&
+  Check(reads.counts == std::array<uint32_t, 3>{2, 2, 2} &&
             snapshot.specialization_reads.size() == 3 &&
             snapshot.images[0].dwords[0] == 0x300u && snapshot.images[1].dwords[0] == 0x400u,
         "finite image refresh retained old scalar values or repeated reads");
@@ -967,37 +929,36 @@ void TestEmptyTraversalSkipsUnmappedDescriptor(bool independent_mask, bool count
       {source.dwords[0], source.dwords[1], source.dwords[2], source.dwords[3]});
   emit(use, ValueOpcode::LoadBufferU32, {buffer, Value(0u), Value(0u), Value(0u), Value(true)})
       .Instruction()->SetFlags(MemoryFlags{.index = buffer_index});
-  auto plan = ExtractResourcePlan(program);
-  struct Reads { uint32_t count = 0; bool clean = true; uint32_t descriptors = 0; } reads;
-  const SrtRuntime runtime{
-      .read_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+  auto plan = Extract(program);
+  struct Reads { uint32_t count = 0; bool readable = true; uint32_t descriptors = 0; } reads;
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
         constexpr std::array<uint32_t, 4> descriptor{0x8000u, 1u << 16u, 16u,
                                                    (3u << 28u) | 0x4dfacu};
+        auto &reads = *static_cast<Reads *>(data);
+        if (address == 0x1000u && reads.readable) {
+          words[0] = reads.count;
+          return true;
+        }
         if (address < 0x2000u || address >= 0x2010u || address % 4u || words.size() != 1)
           return false;
-        ++static_cast<Reads *>(data)->descriptors;
+        ++reads.descriptors;
         words[0] = descriptor[(address - 0x2000u) / 4u];
         return true;
-      },
-      .userdata = &reads,
-      .read_specialization_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
-        const auto &reads = *static_cast<Reads *>(data);
-        if (!reads.clean || address != 0x1000u || words.size() != 1) return false;
-        words[0] = reads.count;
-        return true;
-      }};
+      }, &reads);
+  const SrtRuntime runtime;
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
-  for (const auto state : {Reads{0u, true}, Reads{4u, true}, Reads{0u, false}, Reads{0u, true}}) {
+  for (const auto state : {Reads{0u, true}, Reads{4u, true}, Reads{4u, false}, Reads{0u, true}}) {
     reads = state;
     if (!counted_loop) {
       uint32_t condition = 99;
-      Check(!SrtWalker(plan, runtime).Evaluate(plan.control_flow[3].condition, condition),
+      Check(!EvaluateSrtValue(plan, runtime, plan.control_flow[3].condition, condition),
             "direct merge evaluation used missing or previous runtime reachability");
     }
     Check(MaterializeResources(plan, runtime, snapshot, specialization),
           "partial traversal proof rejected a valid resource snapshot");
-    const bool inactive = !independent_mask && reads.clean && reads.count == 0u;
+    // An unreadable count reads as zero.
+    const bool inactive = !independent_mask && (!reads.readable || reads.count == 0u);
     Check(snapshot.buffers[0].dwords[0] == (inactive ? 0u : 0x8000u),
           "empty traversal kept its descriptor or an uncertain traversal lost it");
     Check(specialization.buffers[0].zero_stride_oob == inactive,
@@ -1005,7 +966,7 @@ void TestEmptyTraversalSkipsUnmappedDescriptor(bool independent_mask, bool count
     Check(reads.descriptors == (inactive ? 0u : 4u),
           "inactive traversal read its descriptor or an uncertain traversal skipped it");
     if (inactive) {
-      Check(!snapshot.specialization_reads.empty(), "empty traversal lost its clean scalar read");
+      Check(!snapshot.specialization_reads.empty(), "empty traversal lost its captured scalar read");
       for (const auto &read : snapshot.specialization_reads)
         Check(read == std::pair<uint64_t, uint64_t>{0x1000u, 4u},
               "empty traversal proof captured an unrelated memory range");
@@ -1050,17 +1011,90 @@ void TestCyclicControlPredicatesRemainUnknown() {
   branch(entry, u, left, right);
   branch(left, a, right, exit);
   branch(right, b, left, exit);
-  auto plan = ExtractResourcePlan(program);
-  SrtWalker walker(plan, {});
+  auto plan = Extract(program);
   uint32_t value = 99;
-  Check(!walker.Evaluate(plan.control_flow[1].condition, value) &&
-            !walker.Evaluate(plan.control_flow[2].condition, value),
+  Check(!EvaluateSrtValue(plan, {}, plan.control_flow[1].condition, value) &&
+            !EvaluateSrtValue(plan, {}, plan.control_flow[2].condition, value),
         "cyclic varying control acquired an unsupported uniform value");
-  std::vector<uint32_t> flat;
-  Check(walker.RefreshFlatBuffer(flat) && plan.visited_blocks == std::vector<uint8_t>({1, 1, 1, 1}),
+  Check(ReachedBlocks(plan, {}) == std::vector<uint8_t>({1, 1, 1, 1}),
         "cyclic control lost a possible branch during reachability");
 }
 
+void TestBackEdgeReachesPhiPredecessor() {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  using Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind;
+  Program program;
+  program.stage = Libs::Graphics::ShaderType::Compute;
+  program.user_data_count = 3;
+  program.srt_plan_complete = program.resource_tracking_complete = true;
+  for (uint32_t i = 0; i < 4; ++i) AddValueBlock(program);
+  auto &entry = *program.blocks[0];
+  auto &header = *program.blocks[1];
+  auto &body = *program.blocks[2];
+  auto &exit = *program.blocks[3];
+  const auto branch = [](Block &block, Block &target, Value condition = {}, Block *other = nullptr) {
+    block.terminator.kind = other ? TerminatorKind::ConditionalBranch : TerminatorKind::Branch;
+    block.terminator.true_block = &target;
+    block.terminator.false_block = other;
+    block.condition = condition;
+    block.AddBranch(&target);
+    if (other) block.AddBranch(other);
+  };
+  // header: p = Phi(entry: 0, body: 1); table[1] == p enters body, which returns to header. The
+  // first pass evaluates p and loads table[1] before body is reached; the walk must revisit both,
+  // capture the load once and reach exit, which loads table[0].
+  program.memory_info.push_back({.kind = ResourceKind::ScalarAddress});
+  const auto table = Value(&entry.AppendNewInst(ValueOpcode::GetAddressResource,
+      {Value(&entry.AppendNewInst(ValueOpcode::GetUserData, {Value(ScalarReg(0))})),
+       Value(&entry.AppendNewInst(ValueOpcode::GetUserData, {Value(ScalarReg(1))}))}));
+  const auto load = [&](Block &block, uint32_t offset) -> Inst & {
+    auto &read = block.AppendNewInst(ValueOpcode::LoadAddressU32, {table, Value(offset), Value(0u), Value(true)});
+    read.SetFlags(MemoryFlags{.index = 0});
+    return read;
+  };
+  auto &p = header.AppendNewInst(ValueOpcode::Phi);
+  p.SetFlags(Type::U32);
+  p.AddPhiOperand(&entry, Value(0u));
+  p.AddPhiOperand(&body, Value(1u));
+  branch(entry, header);
+  branch(header, body, Value(&header.AppendNewInst(ValueOpcode::IEqual32, {Value(&load(header, 4u)), Value(&p)})),
+         &exit);
+  branch(body, header);
+  exit.terminator.kind = TerminatorKind::Return;
+  program.srt_reads.push_back({Value(&load(exit, 0u)), 0});
+  exit.AppendNewInst(ValueOpcode::ReadConst,
+                     {Value(&exit.AppendNewInst(ValueOpcode::GetSrtResource)), Value(0u)});
+  auto plan = Extract(program);
+  Check(plan.control_flow.size() == 4 && plan.control_phi_predecessors.size() == 1,
+        "back-edge plan lost its control flow or Phi");
+  plan.capture_specialization_reads = true;
+  const std::array<uint32_t, 2> user_data{0x1000u, 0u};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  {
+    // The table is unmapped: its loads read zero instead of faulting the walk.
+    FakeGuestMemory unmapped;
+    Check(MaterializeResources(plan, {.user_data = user_data}, snapshot, specialization) &&
+              snapshot.flattened_srt == std::vector<uint32_t>{0u},
+          "faulting load in a reachable block did not read zero");
+  }
+  for (uint32_t selected : {0u, 1u}) {
+    FakeGuestMemory fake(+[](void *userdata, uint64_t address, std::span<uint32_t> words) {
+          words[0] = address == 0x1000u ? 0x5678u : *static_cast<uint32_t *>(userdata);
+          return address == 0x1000u || address == 0x1004u;
+        }, &selected);
+    Check(MaterializeResources(plan, {.user_data = user_data}, snapshot, specialization) &&
+              snapshot.flattened_srt == std::vector<uint32_t>{0x5678u},
+          "back edge to an evaluated Phi left its exit unreachable");
+    Check(snapshot.specialization_reads ==
+              std::vector<std::pair<uint64_t, uint64_t>>{{0x1004u, 4u}, {0x1000u, 4u}},
+          "Phi restart kept the loads of the abandoned pass");
+    Check(ReachedBlocks(plan, {.user_data = user_data}) ==
+              (selected == 0u ? std::vector<uint8_t>{1, 1, 1, 1} : std::vector<uint8_t>{1, 1, 0, 1}),
+          "Phi restart followed a rejected edge or missed a reachable one");
+  }
+}
+
 void TestEmptyScalarBufferSpecialization() {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   struct Fixture {
@@ -1117,11 +1151,12 @@ void TestEmptyScalarBufferSpecialization() {
     return fixture;
   };
   auto fixture = make_fixture();
-  auto plan = ExtractResourcePlan(fixture.program);
+  auto plan = Extract(fixture.program);
   std::array<uint32_t, 4> descriptor{0x1000u, 8u << 16u, 0u, 3u << 28u};
   uint32_t reads = 0u;
-  const SrtRuntime runtime{.user_data = descriptor, .read_memory = RejectSpecializationRead,
-                           .userdata = &reads, .read_specialization_memory = RejectSpecializationRead};
+  const SrtRuntime runtime{.user_data = descriptor, .userdata = &reads,
+                           .read_specialization_memory = RejectSpecializationRead};
+  FakeGuestMemory fake(RejectSpecializationRead, &reads);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization, empty, populated;
   for (const uint32_t records : {0u, 1u, 0x20000000u, 0u}) {
@@ -1173,7 +1208,7 @@ void TestEmptyScalarBufferSpecialization() {
 void TestMixedSamplerVariantsShareRuntimeDescriptor() {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   auto program = MixedSamplerProgram();
-  auto plan = ExtractResourcePlan(program);
+  auto plan = Extract(program);
   std::array<uint32_t, 2> user_data{0x11111111u, 0x22222222u};
   const SrtRuntime runtime{.user_data = user_data};
   ResourceSnapshot snapshot;
@@ -1221,7 +1256,7 @@ void TestSignedStorageSpecialization() {
   program.descriptor_sources.push_back(source);
   program.info.images.push_back({.source = 0u, .resource_class = ImageResourceClass::Storage,
       .dimension = Libs::Graphics::ShaderRecompiler::Decoder::ImageDimension::Dim2D, .written = true});
-  auto plan = ExtractResourcePlan(program);
+  auto plan = Extract(program);
   std::array<uint32_t, 1> data{};
   const SrtRuntime runtime{.user_data = data};
   ResourceSnapshot snapshot;
@@ -1299,7 +1334,7 @@ void TestIndirectImageSamplerClasses(bool mixed) {
   program.descriptor_sources.push_back(root);
   program.info.samplers.resize(1);
   program.info.sampled_pairs.resize(1);
-  auto plan = ExtractResourcePlan(program);
+  auto plan = Extract(program);
   std::array<uint32_t, 1> user_data{0x11111111u};
   uint32_t reads = 0;
   const SrtRuntime runtime{.user_data = user_data, .userdata = &reads,
@@ -1336,6 +1371,48 @@ void TestIndirectImageSamplerClasses(bool mixed) {
   }
 }
 
+void TestShaderCacheRecords() {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  // Equivalent selector domains share one key group, which a load rebuilds.
+  Program selectors;
+  AddValueBlock(selectors);
+  for (const uint32_t stride : {4u, 8u, 4u}) {
+    DescriptorSource source;
+    source.indirect_descriptor = DescriptorSource::IndirectDescriptor{
+        .selector = DescriptorSource::IndirectDescriptor::SelectorRead{.stride = stride}};
+    selectors.descriptor_sources.push_back(source);
+  }
+  selectors.descriptor_sources.emplace_back();
+  const auto plan = Extract(selectors);
+  Check(plan.selector_groups == std::vector<uint32_t>{0u, 1u, 0u, UINT32_MAX} &&
+            plan.selector_keys.size() == 2u,
+        "loaded plan lost its selector key groups");
+
+  const std::array values{Value(), Value(ScalarReg(3)), Value(VectorReg(5)), Value(true),
+                          Value(uint8_t{7}), Value(uint16_t{9}), Value(11u), Value(uint64_t{1} << 40u),
+                          Value::F16(0x3c00u), Value::F32(-2.5f)};
+  Writer writer;
+  Io(writer, values);
+  Reader reader{writer.bytes};
+  std::array<Value, values.size()> loaded;
+  Io(reader, loaded);
+  Check(reader.ok && reader.bytes.empty() && loaded == values,
+        "immediate values changed through the shader cache");
+  // Damaged records: a missing instruction, a type that is no value type, an oversized count and
+  // an unknown opcode.
+  for (const uint8_t type : {1u, 3u}) {
+    const std::array<uint8_t, 12> bytes{type};
+    Reader damaged{bytes};
+    Value value;
+    Io(damaged, value);
+    Check(!damaged.ok, "damaged value was accepted");
+  }
+  ResourcePlan oversized, unknown;
+  Check(!LoadResourcePlan(std::array<uint8_t, 4>{0xffu, 0xffu, 0xffu, 0xffu}, oversized) &&
+            !LoadResourcePlan(std::array<uint8_t, 14>{1u, 0u, 0u, 0u, 0xffu, 0xffu}, unknown),
+        "damaged resource plan was accepted");
+}
+
 } // namespace
 
 namespace Common {
@@ -1353,10 +1430,10 @@ void DbgExit(int) { std::abort(); }
 } // namespace Common
 
 int main() {
-  TestMappedSrtUsesDirectReaderByDefault();
+  TestWalkerLoadsBypassSpecializationReader();
   TestNativeNestedSrtRefresh();
   TestIntegerRuntimeValueFollowsSrtReads();
-  TestSrtAliasesRetainReadPolicy();
+  TestSrtAliasesShareReads();
   TestUniformVectorDescriptorRead();
   TestExactReciprocalDescriptorArithmetic();
   TestUnbasedFlatCacheHitMaterializes();
@@ -1370,10 +1447,12 @@ int main() {
     TestEmptyTraversalSkipsUnmappedDescriptor(independent_mask);
   TestEmptyTraversalSkipsUnmappedDescriptor(false, true);
   TestCyclicControlPredicatesRemainUnknown();
+  TestBackEdgeReachesPhiPredecessor();
   TestEmptyScalarBufferSpecialization();
   TestMixedSamplerVariantsShareRuntimeDescriptor();
   TestSignedStorageSpecialization();
   for (const bool mixed : {false, true}) TestIndirectImageSamplerClasses(mixed);
+  TestShaderCacheRecords();
   std::puts("ResourceMaterializationTests: all cases passed");
   return 0;
 }

tests/ResourceTrackingTests.cpp

@@ -5,7 +5,7 @@
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceTracking.h"
 #include "graphics/shader/recompiler/ir/passes/ShaderInfoCollection.h"
-#include "graphics/shader/recompiler/ir/passes/SrtWalker.h"
+#include "SrtWalkerTestMemory.h"
 
 #include <array>
 #include <bit>
@@ -212,6 +212,12 @@ bool ReadLinearTestMemory(void *userdata, uint64_t address, std::span<uint32_t>
   return true;
 }
 
+// Serves walker loads from the memory a test runtime currently points at.
+bool ReadRuntimeMemory(void *userdata, uint64_t address, std::span<uint32_t> values) {
+  const auto &runtime = *static_cast<const SrtRuntime *>(userdata);
+  return runtime.read_specialization_memory(runtime.userdata, address, values);
+}
+
 std::unique_ptr<Fixture>
 MakeIndirectImageFixture(bool malformed, uint32_t table_stride = 32, uint32_t table_offset = 0,
                          uint32_t table_shader_offset = 0, bool nested = false,
@@ -424,9 +430,8 @@ void TestInvariantIndirectImageMaterialization(bool nested) {
     const auto handle = std::ranges::find_if(*wrap->block, [](const Inst &inst) {
       return inst.GetOpcode() == ValueOpcode::GetImageResource;
     });
-    SrtWalker walker(wrap->program, runtime);
     uint32_t offset = 0;
-    Check(walker.Evaluate(handle->Arg(0), offset) && offset == expected,
+    Check(EvaluateSrtValue(wrap->program, runtime, handle->Arg(0), offset) && offset == expected,
           "GPU descriptor offset projection lost U32 wrap or scalar immediate carry");
   }
   auto malformed = MakeIndirectImageFixture(true);
@@ -624,6 +629,7 @@ void TestBoundedImageViewEligibility() {
     plan.descriptor_sources.push_back(source);
   }
   plan.descriptor_sources[image_source].indirect_descriptor->sources = std::move(selected);
+  CompileSrtWalker(plan);
   Check(!MaterializeResources(plan, runtime, snapshot, specialization),
         "explicitly selected incompatible descriptor was silently normalized to null");
 }
@@ -873,9 +879,8 @@ void TestWaterfallImageTable() {
         std::array{UINT32_MAX, 0x5204u, UINT32_MAX}, std::array{0u, 0u, UINT32_MAX}}) {
     data[4] = index;
     data[3] = format;
-    SrtWalker walker(fixture->program, runtime);
     uint32_t offset = 0;
-    Check(walker.Evaluate(handle->Arg(0), offset) && offset == expected,
+    Check(EvaluateSrtValue(fixture->program, runtime, handle->Arg(0), offset) && offset == expected,
           "VMEM table projection lost record bounds, format disable or multiplication carry");
   }
   auto disabled = make(true, false);
@@ -1113,7 +1118,7 @@ void TestNestedPointerImageTable() {
   memory.fail_address = 0x1800u;
   *at(table) = 0u;
   Check(!MaterializeResources(plan, runtime, snapshot, specialization),
-        "unreadable null pointer row bypassed the strict pointee reader");
+        "unreadable null pointer row bypassed the specialization reader");
   *at(table) = objects;
 
   auto capped = make(4u);
@@ -1381,7 +1386,7 @@ void TestIndexedPointerImageTable() {
   *at(table) = 0u;
   *at(table + 4u) = 0u;
   Check(!MaterializeResources(plan, runtime, snapshot, specialization),
-        "explicit null pointer bypassed the strict child reader");
+        "explicit null pointer bypassed the specialization reader");
   auto reversed = make(Case::ReverseSigned);
   reversed->PlanAndTrack();
   for (const auto variant : {Case::MissingParent, Case::WrongParent, Case::MissingChild,
@@ -1398,8 +1403,9 @@ void TestIndexedPointerImageTable() {
   const auto counted_plan = ExtractResourcePlan(counted->program);
   constexpr uint32_t count_address = 0x1800u;
   std::array<uint32_t, 6> counted_data{table, 1u | (stride << 16u), 2u, 0x5204u, count_address, 1u};
-  const SrtRuntime counted_runtime{.user_data = counted_data, .read_memory = ReadLinearTestMemory, .userdata = &memory,
+  const SrtRuntime counted_runtime{.user_data = counted_data, .userdata = &memory,
                                   .read_specialization_memory = ReadLinearTestMemory};
+  FakeGuestMemory fake(ReadLinearTestMemory, &memory);
   *at(table) = objects;
   *at(table + 4u) = 0xbeef0001u;
   *at(count_address) = *at(count_address + 4u) = 2u;
@@ -1431,9 +1437,11 @@ void TestIndexedPointerImageTable() {
   Check(!MaterializeResources(counted_plan, counted_runtime, snapshot, specialization) && memory.watched_reads == 0u,
         "runtime parent count exceeding V# records was clamped or read candidates");
   *at(count_address) = 2u;
+  // An unreadable parent count reads as zero: an empty domain without candidate reads.
   memory.fail_address = high + count_address;
-  Check(!MaterializeResources(counted_plan, counted_runtime, snapshot, specialization) && memory.watched_reads == 0u,
-        "unreadable parent count was ignored before candidate reads");
+  Check(MaterializeResources(counted_plan, counted_runtime, snapshot, specialization) &&
+            snapshot.images.size() == 1u && memory.watched_reads == 0u,
+        "unreadable parent count did not read as an empty domain");
   memory.fail_address = UINT64_MAX;
   for (const auto variant : {Case::MissingCount, Case::DifferentCount, Case::UnprovenMask}) {
     auto unproved = make(variant);
@@ -1780,9 +1788,9 @@ void TestGuardedDirectImageTable() {
   };
   fill_table(memory, table);
   std::array<uint32_t, 2> user_data{0x1000u, 0u};
-  SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-                     .userdata = &memory,
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
                      .read_specialization_memory = ReadLinearTestMemory};
+  FakeGuestMemory fake(ReadRuntimeMemory, &runtime);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   auto packed = make_plan(Guard::Plain, 0u, true);
@@ -2352,10 +2360,9 @@ void TestBoundedComputeImageLoop() {
         (static_cast<uint32_t>(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
   }
   std::array<uint32_t, 3> user_data{0x1800u, 0u, 2u};
-  SrtRuntime runtime{.user_data = user_data,
-                     .read_memory = ReadLinearTestMemory,
-                     .userdata = &memory,
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
                      .read_specialization_memory = ReadLinearTestMemory};
+  FakeGuestMemory fake(ReadLinearTestMemory, &memory);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   for (const auto post : {PostTest::Signed, PostTest::Unsigned}) {
@@ -2668,8 +2675,8 @@ void TestDescendingImageRange() {
   std::array<uint32_t, 4> user_data {0x1000u, 0u, 0u, UINT32_MAX};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
-  const SrtRuntime runtime {.user_data = user_data, .read_memory = ReadLinearTestMemory,
-      .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  const SrtRuntime runtime {.user_data = user_data, .userdata = &memory,
+      .read_specialization_memory = ReadLinearTestMemory};
   for (const auto count : {0u, 2u, 0u}) {
     user_data[2] = count;
     Check(MaterializeResources(plan, runtime, snapshot, specialization) && snapshot.images.size() == 2u,
@@ -3176,7 +3183,7 @@ void TestUniformizedMaterialImageKeys() {
   make_plan(Variant::FirstLane);
   make_plan(Variant::FirstLaneDirect);
   plan = make_plan(Variant::FirstLaneAndNot);
-  Check(plan.requires_specialization_memory &&
+  Check(plan.capture_specialization_reads &&
             plan.descriptor_sources[plan.info.images[0].source]
                 .indirect_descriptor->selector_mask.Resolve().TryInstruction() != nullptr,
         "uniformized image mask did not survive extraction");
@@ -3207,10 +3214,8 @@ void TestUniformizedMaterialImageKeys() {
   user_data[9] = 16u << 16u;
   user_data[10] = 1u;
   memory.fail_address = base + 0xc00u + 3u * 0x90u;
-  const SrtRuntime runtime{
-      .user_data = user_data, .read_memory = ReadLinearTestMemory,
-      .userdata = &memory,
-      .read_specialization_memory = ReadLinearTestMemory};
+  const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+                           .read_specialization_memory = ReadLinearTestMemory};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
@@ -3388,7 +3393,7 @@ void TestIndexedMaterialImageKeys() {
     if (shared) {
       owner_key.ResolveInstruction()->ReplaceUsesWith(Value(12994u));
       uint32_t offset = 0u;
-      Check(SrtWalker(f.program, {}).Evaluate(shared_image.ResolveInstruction()->Arg(0), offset) &&
+      Check(EvaluateSrtValue(f.program, {}, shared_image.ResolveInstruction()->Arg(0), offset) &&
                 offset == 12994u * 32u,
             "inherited indexed selector changed GPU byte offsets into raw material keys");
     }
@@ -3811,10 +3816,9 @@ void TestUniformScalarBufferImage() {
               memory.words.begin() + (0x1010u + index * 48u) / 4u);
     memory.words[(0x1010u + index * 48u) / 4u] += index;
   }
-  SrtRuntime runtime{.user_data = user_data,
-                     .read_memory = ReadLinearTestMemory,
-                     .userdata = &memory,
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
                      .read_specialization_memory = ReadLinearTestMemory};
+  FakeGuestMemory fake(ReadLinearTestMemory, &memory);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   const auto materializes = [&](uint32_t address) {
@@ -3827,9 +3831,10 @@ void TestUniformScalarBufferImage() {
   memory.words[0x44u / 4u] = 0u;
   memory.words[0x1010u / 4u] = 0x30u;
   Check(materializes(0x30u), "changed scalar table memory reused the previous image descriptor");
+  // An unreadable key reads as zero and selects the first record instead of the second.
+  memory.words[0x44u / 4u] = 1u;
   memory.fail_address = 0x1044u;
-  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
-        "unavailable nested image table was accepted");
+  Check(materializes(0x30u), "unavailable nested image key did not read as zero");
 }
 
 void TestCompactScalarImageTable() {
@@ -4258,8 +4263,8 @@ void TestScalarBufferSelectorDomain() {
         (static_cast<uint32_t>(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
   }
   std::array<uint32_t, 6> user_data {0x1000u, 0u, 180u, 0u, 0x2000u, 0u};
-  const SrtRuntime runtime {.user_data = user_data, .read_memory = ReadLinearTestMemory,
-      .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  const SrtRuntime runtime {.user_data = user_data, .userdata = &memory,
+      .read_specialization_memory = ReadLinearTestMemory};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   memory.watched_address = 0x1000u + 144u;
@@ -4314,8 +4319,8 @@ void TestScalarBufferSelectorDomain() {
   user_data[1] = 2u << 16u;
   user_data[2] = UINT32_MAX;
   Check(MaterializeResources(plan,
-            {.user_data = user_data, .read_memory = read_sparse, .userdata = &sparse,
-             .read_specialization_memory = read_sparse}, snapshot, specialization) &&
+            {.user_data = user_data, .userdata = &sparse, .read_specialization_memory = read_sparse},
+            snapshot, specialization) &&
             sparse.offsets == std::vector<uint64_t>{144u, 0x80000090u} &&
             snapshot.images.size() == 2u,
         "large V# bounds extended the native U32 domain or invented an OOB zero key");
@@ -4644,7 +4649,8 @@ void TestCountedRowImageSelector() {
   at(MaterialBase + 256u) = 0x200u;
   at(MaterialBase + 260u) = static_cast<uint32_t>(Libs::Graphics::Prospero::BufferFormat::k8SScaled) << 20u;
   at(MaterialBase + 268u) = at(MaterialBase + 196u);
-  std::array<uint32_t, 12> user_data{KeyBase, 4u << 16u, 64u, 0x14004u,
+  // The walker needs every user-data DWORD the plan reads, including the zero s12-s14.
+  std::array<uint32_t, 15> user_data{KeyBase, 4u << 16u, 64u, 0x14004u,
       CountBase, 4u << 16u, 4u, 0x14004u, MaterialBase, 396u << 16u, 4u, 0x14004u};
   const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
       .read_specialization_memory = ReadLinearTestMemory};
@@ -5002,8 +5008,9 @@ void TestSharedAppendSelectorBounds() {
         (static_cast<uint32_t>(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
   }
   std::array<uint32_t, 9> user_data {0x1000u, 160u << 16u, 8u, 0u, 0x2000u, 0u, 0x1800u, 0u, 0u};
-  const SrtRuntime runtime {.user_data = user_data, .read_memory = ReadLinearTestMemory,
-      .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  SrtRuntime runtime {.user_data = user_data, .userdata = &memory,
+      .read_specialization_memory = ReadLinearTestMemory};
+  FakeGuestMemory fake(ReadRuntimeMemory, &runtime);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   memory.fail_address = memory.watched_address = 0x1090u;
@@ -5085,9 +5092,8 @@ void TestSharedAppendSelectorBounds() {
   user_data[2] = 65536u;
   user_data[4] = 0x50000000u;
   user_data[6] = 0x40000000u;
-  Check(MaterializeResources(plan,
-            {.user_data = user_data, .read_memory = read_boundary, .userdata = &boundary,
-             .read_specialization_memory = read_boundary}, snapshot, specialization) &&
+  runtime = {.user_data = user_data, .userdata = &boundary, .read_specialization_memory = read_boundary};
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
             boundary.keys_read == 65536u && snapshot.images.size() == 1u &&
             snapshot.images[0].dwords[0] == 0x101u,
         "shared append materialization rejected its exact nonwrapping probe limit");
@@ -5103,9 +5109,7 @@ void TestSharedAppendSelectorBounds() {
     for (const auto& test : overlaps) {
       boundary.counts = test.counts;
       boundary.keys_read = 0u;
-      Check(MaterializeResources(overlap_plan,
-                {.user_data = user_data, .read_memory = read_boundary, .userdata = &boundary,
-                 .read_specialization_memory = read_boundary}, snapshot, specialization) &&
+      Check(MaterializeResources(overlap_plan, runtime, snapshot, specialization) &&
                 boundary.keys_read == test.expected[layout] && snapshot.images.size() == 1u &&
                 snapshot.images[0].dwords[0] == (test.expected[layout] != 0u ? 0x101u : 0u),
             "shared append bounds lost overlap, included a touching list or retained stale keys");
@@ -5119,7 +5123,6 @@ void TestComputeBufferFill() {
     bool conditional = false;
     bool shifted = false;
     bool extra_store = false;
-    bool clean = false;
     bool branch = false;
   };
   const auto Run = [](Options options) {
@@ -5188,17 +5191,13 @@ void TestComputeBufferFill() {
       words[0] = memory.words[0];
       return true;
     };
-    Check(MaterializeResources(
-              plan,
-              {.user_data = userdata,
-               .read_memory = Read,
-               .userdata = &memory,
-               .read_specialization_memory = options.clean ? Read : nullptr},
-              snapshot, specialization),
-          "fill fixture did not materialize");
+    const auto materialize = [&](SrtMemoryReader reader) {
+      FakeGuestMemory fake(reader, &memory);
+      return MaterializeResources(plan, {.user_data = userdata}, snapshot, specialization);
+    };
+    Check(materialize(Read), "fill fixture did not materialize");
     const bool expected = !options.conditional && !options.shifted &&
-                          !options.extra_store && !options.branch &&
-                          (!options.scalar || options.clean);
+                          !options.extra_store && !options.branch;
     Check((snapshot.uniform_fill.words != 0) == expected,
           "fill proof accepted an unsafe store or missed the real GTA3 clear");
     if (expected) {
@@ -5207,21 +5206,19 @@ void TestComputeBufferFill() {
                 snapshot.uniform_fill.value ==
                     (options.scalar ? 0x40404040u : 0u),
             "fill proof lost address coverage or the actual stored scalar");
-      if (options.scalar && options.clean) {
-        Check(MaterializeResources(plan,
-                  {.user_data = userdata, .read_memory = Read, .userdata = &memory},
-                  snapshot, specialization) && snapshot.uniform_fill.words == 0,
-              "an unavailable clean value retained a previous uniform fill");
+      if (options.scalar) {
+        Check(materialize(nullptr) && snapshot.uniform_fill.words == 1 &&
+                  snapshot.uniform_fill.value == 0u,
+              "an unreadable fill value retained a previous uniform fill");
       }
     }
   };
   Run({});
-  Run({.scalar = true, .clean = true});
   Run({.scalar = true});
-  Run({.conditional = true, .clean = true});
-  Run({.shifted = true, .clean = true});
-  Run({.extra_store = true, .clean = true});
-  Run({.clean = true, .branch = true});
+  Run({.conditional = true});
+  Run({.shifted = true});
+  Run({.extra_store = true});
+  Run({.branch = true});
 }
 
 void TestDenseBufferTracking() {
@@ -5332,12 +5329,12 @@ void TestRuntimeUnsignedMinDescriptor() {
   SrtRuntime runtime{.user_data = user_data};
   DescriptorValue value;
   const auto source = fixture.program.info.buffers[0].source;
-  Check(SrtWalker(plan, runtime).EvaluateDescriptor(source, value) &&
+  Check(EvaluateSrtDescriptor(plan, runtime, source, value) &&
             value.dwords[3] == 0x100u,
         "runtime descriptor unsigned minimum did not clamp its first operand");
   user_data[0] = 0x80u;
   Check(
-      SrtWalker(plan, runtime).EvaluateDescriptor(source, value) &&
+      EvaluateSrtDescriptor(plan, runtime, source, value) &&
           value.dwords[3] == 0x80u,
       "runtime descriptor unsigned minimum did not preserve its first operand");
 }
@@ -5488,12 +5485,12 @@ void TestSampleAdjustSamplerScratch() {
   std::array<uint32_t, 4> user_data{4u, 1u, 2u, 0x80000abcu};
   SrtRuntime runtime{.user_data = user_data};
   DescriptorValue descriptor;
-  Check(SrtWalker(plan, runtime).EvaluateDescriptor(source, descriptor) &&
+  Check(EvaluateSrtDescriptor(plan, runtime, source, descriptor) &&
             descriptor.dwords[3] == 0x80000abcu,
         "SampleAdjust canonicalization lost sampler border fields");
   const auto shift_source = fixture.program.info.samplers[
       shift_sampler.Instruction()->Flags<uint32_t>()].source;
-  Check(SrtWalker(plan, runtime).EvaluateDescriptor(shift_source, descriptor) &&
+  Check(EvaluateSrtDescriptor(plan, runtime, shift_source, descriptor) &&
             std::ranges::equal(std::span(descriptor.dwords).first(4),
                                std::array{0x36u, 0xfff000u, 0x02500000u, 0u}),
         "SampleAdjust shift-only scratch was not reduced to its zero border payload");
@@ -5601,11 +5598,11 @@ void TestFmaskLoadSpecialization() {
         "FMASK load did not lower to a value vector");
   result = vector->Arg(0);
   uint32_t value = 0;
-  Check(SrtWalker(fixture.program, {.user_data = user_data}).Evaluate(result, value) &&
+  Check(EvaluateSrtValue(fixture.program, {.user_data = user_data}, result, value) &&
             value == 0x76543210u,
         "FMASK load did not return the native sample-to-fragment mapping");
   user_data[8] = 1;
-  Check(SrtWalker(fixture.program, {.user_data = user_data}).Evaluate(result, value) &&
+  Check(EvaluateSrtValue(fixture.program, {.user_data = user_data}, result, value) &&
             value == 0u,
         "inactive FMASK load did not preserve the execution mask");
   ShaderComputeInputInfo compute{};
@@ -5849,16 +5846,14 @@ void TestSrtFlatteningAndRuntimeMemoization() {
   std::array<uint32_t, 2> user_data{0x1000u, 0u};
   TestMemory memory;
   memory.words[1] = 0xdeadbeefu;
-  SrtRuntime runtime{.user_data = user_data,
-                     .read_memory = ReadTestMemory,
-                     .userdata = &memory};
+  SrtRuntime runtime{.user_data = user_data};
+  FakeGuestMemory fake(ReadTestMemory, &memory);
   auto plan = ExtractResourcePlan(fixture.program);
   DescriptorValue descriptor;
   std::vector<uint32_t> flat;
   const uint32_t request = fixture.program.info.buffers[0].source;
   const auto refresh = [&](const ResourcePlan& plan) {
-    SrtWalker walker(plan, runtime);
-    return walker.EvaluateDescriptor(request, descriptor) && walker.RefreshFlatBuffer(flat);
+    return WalkSrt(plan, runtime, flat, nullptr) && WalkedDescriptor(plan, request, descriptor);
   };
   Check(refresh(plan), "typed runtime source evaluation failed");
   Check(descriptor.dwords[0] == 0xdeadbeefu &&
@@ -5873,10 +5868,11 @@ void TestSrtFlatteningAndRuntimeMemoization() {
 
   memory.reads = 0;
   memory.fail_after = 0;
-  Check(!refresh(plan), "unavailable scalar memory was accepted");
+  Check(refresh(plan) && descriptor.dwords[0] == 0u && flat == std::vector<uint32_t>{0u},
+        "unavailable scalar memory did not read as zero");
   memory.fail_after = UINT32_MAX;
   Check(refresh(plan) && descriptor.dwords[0] == 0x12345678u && memory.reads == 1,
-        "failed runtime evaluation left a value marked as visiting");
+        "a zero read left a stale value in the next walk");
 
   auto moved = std::move(plan);
   memory.reads = 0;
@@ -5924,18 +5920,17 @@ void TestScalarDataReadControlFlow() {
         {address, Value(0u), Value(0u), Value(true)}, fixture.AddMemory(scalar, 4), use);
     fixture.PlanAndTrack();
     auto plan = ExtractResourcePlan(fixture.program);
-    Check(plan.srt_reads.size() == 1 && !fixture.program.info.uses_dma &&
-              bool(plan.srt_code) == (comparison == ValueOpcode::Void),
-          "scalar data lost flattening or conditional reads entered the unconditional JIT");
+    Check(plan.srt_reads.size() == 1 && !fixture.program.info.uses_dma && plan.srt_code,
+          "scalar data lost flattening or its walker");
     uint32_t reads = 0;
     std::array<uint32_t, 2> user_data {};
-    SrtRuntime runtime {.user_data = user_data,
-        .read_memory = +[](void* data, uint64_t address, std::span<uint32_t> words) {
+    SrtRuntime runtime {.user_data = user_data};
+    FakeGuestMemory fake(+[](void* data, uint64_t address, std::span<uint32_t> words) {
           ++*static_cast<uint32_t*>(data);
           if (address == 0) return false;
           words[0] = 42;
           return true;
-        }, .userdata = &reads};
+        }, &reads);
     ResourceSnapshot snapshot;
     ResourceSpecialization specialization;
     for (const uint64_t pointer : {uint64_t{0}, uint64_t{0x1000}, uint64_t{1} << 32u}) {
@@ -5977,11 +5972,10 @@ void TestDynamicSrtReadRemainsExplicit() {
   std::array<uint32_t, 3> user_data{0x1000u, 0u, 4u};
   TestMemory memory;
   memory.words[1] = 0xabcdef01u;
-  SrtRuntime runtime{.user_data = user_data,
-                     .read_memory = ReadTestMemory,
-                     .userdata = &memory};
+  SrtRuntime runtime{.user_data = user_data};
+  FakeGuestMemory fake(ReadTestMemory, &memory);
   DescriptorValue value;
-  Check(SrtWalker(plan, runtime).EvaluateDescriptor(fixture.program.info.buffers[0].source, value) &&
+  Check(EvaluateSrtDescriptor(plan, runtime, fixture.program.info.buffers[0].source, value) &&
             value.dwords[0] == 0xabcdef01u && memory.reads == 1,
         "dynamic typed scalar descriptor source was not evaluated");
 
@@ -6091,8 +6085,9 @@ void TestBoundedFormattedBufferStore() {
     memory.words[word + 3u] = static_cast<uint32_t>(Prospero::BufferFormat::k32UInt) << 12u | 4u;
   }
   std::array<uint32_t, 2> user_data{static_cast<uint32_t>(memory.base), 0u};
-  SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-                     .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+                     .read_specialization_memory = ReadLinearTestMemory};
+  FakeGuestMemory fake(ReadLinearTestMemory, &memory);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   const auto materialize = [&] { return MaterializeResources(plan, runtime, snapshot, specialization); };
@@ -6316,8 +6311,9 @@ void TestBoundedBufferRecords(uint32_t d32_components) {
   }
   std::array<uint32_t, 6> user_data{uint32_t(memory.base) + table_offset, 8u << 16u,
                                    4u * 192u / 8u, 0x5204u, uint32_t(memory.base), 0u};
-  SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-                     .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+                     .read_specialization_memory = ReadLinearTestMemory};
+  FakeGuestMemory fake(ReadLinearTestMemory, &memory);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   const auto materialize = [&] { return MaterializeResources(plan, runtime, snapshot, specialization); };
@@ -6498,8 +6494,8 @@ void TestGuardedWaterfallBufferRecords() {
   std::copy(poison.begin(), poison.end(), memory.words.begin() + (table_offset + 480u) / 4u);
   std::array<uint32_t, 8> user_data{uint32_t(memory.base) + table_offset, 488u << 16u,
       2u, 0x5204u, 0x9000u, 4u << 16u, 32u, 0x5204u};
-  SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-      .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+      .read_specialization_memory = ReadLinearTestMemory};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   for (const auto guard : {Guard::Exact, Guard::Reversed, Guard::Greater, Guard::GreaterEqual,
@@ -6617,8 +6613,8 @@ void TestGpuSelectedBufferRecords() {
     // SMEM aligns the base and dynamic offset independently; its V# format is ignored.
     std::array<uint32_t, 8> user_data{uint32_t(memory.base) + table_offset + 3u, 0u, 0u, 0u,
         uint32_t(memory.fail_address), 12u << 16u, 129600u, 0x5204u};
-    SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-        .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+    SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+        .read_specialization_memory = ReadLinearTestMemory};
     ResourceSnapshot snapshot;
     ResourceSpecialization specialization;
     const auto materialize = [&] { return MaterializeResources(plan, runtime, snapshot, specialization); };
@@ -6709,8 +6705,8 @@ void TestWorkgroupSelectedBufferRecords() {
   std::array<uint32_t, 8> user_data{uint32_t(memory.base) + table, 192u << 16u, 2u, 0x5204u,
       uint32_t(memory.base) + selector, 12u << 16u, MaxIndirectDescriptorProbes, 0x5204u};
   memory.watched_address = memory.base;
-  SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-      .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory,
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+      .read_specialization_memory = ReadLinearTestMemory,
       .indirect_args_address = memory.base};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
@@ -6938,8 +6934,9 @@ void TestGpuSelectedFormattedRecordStore() {
   memory.fail_address = memory.base + 0x2000u;
   std::array<uint32_t, 8> user_data{uint32_t(memory.base) + table_offset, 592u << 16u, 6u, 0x5204u,
       uint32_t(memory.fail_address), 4u << 16u, 64u, 0x5204u};
-  SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-      .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+      .read_specialization_memory = ReadLinearTestMemory};
+  FakeGuestMemory fake(ReadRuntimeMemory, &runtime);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   const auto materialize = [&] { return MaterializeResources(plan, runtime, snapshot, specialization); };
@@ -7127,9 +7124,8 @@ void TestGpuSelectedFormattedRecordStore() {
   std::array<uint32_t, 8> prefix_data{uint32_t(prefix_memory.base) + table_offset - 48u, 0u, 0u, 0u,
       uint32_t(prefix_memory.base) + selector_offset, 8u << 16u, 12u, 0x5204u};
   std::array counts{2u, 1u, 1u};
-  SrtRuntime prefix_runtime{.user_data = prefix_data, .read_memory = read_prefix,
-      .userdata = &prefix_memory, .read_specialization_memory = read_prefix,
-      .workgroup_counts = counts};
+  runtime = {.user_data = prefix_data, .userdata = &prefix_memory,
+      .read_specialization_memory = read_prefix, .workgroup_counts = counts};
   prefix_memory.watched_address = prefix_memory.base + selector_offset + 4u;
   for (const auto prefix : {Prefix::Forward, Prefix::Reversed}) {
     const auto selected = make(16u, 16u, false, true, prefix);
@@ -7143,7 +7139,7 @@ void TestGpuSelectedFormattedRecordStore() {
     const auto refresh = [&] {
       prefix_memory.watched_reads = 0u;
       prefix_memory.read_attempts = 0u;
-      return MaterializeResources(selected_plan, prefix_runtime, snapshot, specialization);
+      return MaterializeResources(selected_plan, runtime, snapshot, specialization);
     };
     for (const uint32_t first : {0u, 2u, 4u, 0u}) {
       prefix_memory.words[first_offset / 4u] = first;
@@ -7155,9 +7151,12 @@ void TestGpuSelectedFormattedRecordStore() {
         Check(std::ranges::find(snapshot.specialization_reads, read) != snapshot.specialization_reads.end(),
               "runtime first or selected prefix escaped specialization read capture");
     }
+    // An unreadable runtime first reads as zero.
+    prefix_memory.words[first_offset / 4u] = 4u;
     prefix_memory.fail_address = prefix_memory.base + first_offset;
-    Check(!refresh() && prefix_memory.read_attempts == 1u,
-          "unreadable runtime first reached the selector payload");
+    Check(refresh() && snapshot.buffers.size() == selected_plan.info.buffers.size() + 1u &&
+              snapshot.buffers[wide_root].dwords[0] == 0x100000u,
+          "unreadable runtime first did not read as zero");
     prefix_memory.fail_address = UINT64_MAX;
     prefix_memory.words[first_offset / 4u] = 6u;
     Check(!refresh(), "an actually selected invalid numeric descriptor was silently omitted");
@@ -7257,8 +7256,8 @@ void TestFiniteBufferSelection() {
   }
   memory.fail_address = memory.base + 32u + 5u * 16u;
   const std::array<uint32_t, 2> user_data{static_cast<uint32_t>(memory.base), 0u};
-  SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-                     .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+                     .read_specialization_memory = ReadLinearTestMemory};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   Check(MaterializeResources(plan, runtime, snapshot, specialization) && snapshot.buffers.size() == 5u,
@@ -7339,8 +7338,8 @@ void TestPackedScalarBufferSelection() {
   // No descriptor exists after key 18; broad enumeration would read this address.
   memory.fail_address = memory.base + 0x170u;
   const std::array<uint32_t, 2> user_data{static_cast<uint32_t>(memory.base), 0u};
-  const SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-                          .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+  const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+                          .read_specialization_memory = ReadLinearTestMemory};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   const auto materialize = [&] { return MaterializeResources(plan, runtime, snapshot, specialization); };
@@ -7497,38 +7496,27 @@ void TestConditionalSamplerPhi() {
         for (uint32_t word = 448 / 4; word < (480 + 16) / 4; ++word) {
           memory.words[word] = 0x400u + word;
         }
-        const SrtRuntime runtime{.read_memory = ReadLinearTestMemory,
-                                 .userdata = &memory,
-                                 .read_specialization_memory =
-                                     ReadLinearTestMemory};
+        FakeGuestMemory fake(ReadLinearTestMemory, &memory);
         // Sampler selection compares a signed value against zero.
         for (const auto flag : {-1, 0, 1, INT32_MIN, INT32_MAX}) {
           memory.words[(0x1000 + 196) / 4] = std::bit_cast<uint32_t>(flag);
           const uint32_t first = (flag < 0) != reverse ? 480 / 4 : 448 / 4;
           memory.fail_address = 0x1000 + (first == 448 / 4 ? 480u : 448u);
           DescriptorValue selected;
-          SrtWalker clean(plan, CleanRuntime(runtime));
-          Check(SrtWalker(plan, runtime, &clean).EvaluateDescriptor(source, selected),
+          Check(EvaluateSrtDescriptor(plan, {}, source, selected),
                 "conditional sampler did not survive detached plan lifetime");
           for (uint32_t word = 0; word < 4; ++word) {
             Check(selected.dwords[word] == memory.words[first + word],
                   "conditional sampler chose the wrong incoming descriptor");
           }
         }
-        DescriptorValue selected;
-        auto no_clean_reader = runtime;
-        no_clean_reader.read_specialization_memory = nullptr;
-        {
-          SrtWalker clean(plan, CleanRuntime(no_clean_reader));
-          Check(!SrtWalker(plan, no_clean_reader, &clean).EvaluateDescriptor(source, selected),
-                "conditional sampler used unchecked memory for its predicate");
-        }
+        // An unreadable negative predicate reads as zero and selects the nonnegative edge.
+        memory.words[(0x1000 + 196) / 4] = UINT32_MAX;
         memory.fail_address = 0x2000 + 196;
-        {
-          SrtWalker clean(plan, CleanRuntime(runtime));
-          Check(!SrtWalker(plan, runtime, &clean).EvaluateDescriptor(source, selected),
-                "conditional sampler ignored unavailable coherent predicate memory");
-        }
+        DescriptorValue selected;
+        Check(EvaluateSrtDescriptor(plan, {}, source, selected) &&
+                  selected.dwords[0] == memory.words[reverse ? 480 / 4 : 448 / 4],
+              "conditional sampler did not read unavailable predicate memory as zero");
       }
     }
     CheckFatal([&] { ConditionalSamplerPlan(diamond, false, false, true); },
@@ -7592,10 +7580,9 @@ void TestFiniteImagePhiCycle() {
         "finite image key did not preserve the loop-carried selection");
   std::array<uint32_t, 16> data;
   for (uint32_t i = 0; i < data.size(); ++i) data[i] = 100u + i;
-  SrtWalker walker(plan, SrtRuntime{.user_data = data});
   for (uint32_t candidate = 0; candidate < 2; ++candidate) {
     DescriptorValue descriptor;
-    Check(walker.EvaluateDescriptor(finite->sources[candidate], descriptor),
+    Check(EvaluateSrtDescriptor(plan, {.user_data = data}, finite->sources[candidate], descriptor),
           "finite image candidate is not host-readable");
     for (uint32_t word = 0; word < 8; ++word)
       Check(descriptor.dwords[word] == data[candidate * 8 + word],
@@ -7666,8 +7653,8 @@ void TestPackedDescriptorWaterfall() {
       memory.words[word + 1u] = 0x00fff000u;
       memory.words[word + 2u] = (row & 1u) ? 0x0a500000u : 0u;
     }
-    const SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-        .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+    const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+        .read_specialization_memory = ReadLinearTestMemory};
     ResourceSnapshot snapshot;
     ResourceSpecialization specialization;
     Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
@@ -7795,8 +7782,9 @@ void TestFiniteImageWaterfall() {
           (uint32_t(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
     }
     std::array<uint32_t, 4> user_data{uint32_t(memory.base) + 256u, 8u << 16u, 60u, 0x5204u};
-    SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
-        .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+    SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+        .read_specialization_memory = ReadLinearTestMemory};
+    FakeGuestMemory fake(ReadLinearTestMemory, &memory);
     ResourceSnapshot snapshot;
     ResourceSpecialization specialization, populated;
     for (const uint32_t count : {0u, 60u, 0u, 60u}) {
@@ -8048,7 +8036,7 @@ void TestLoopCycleEnteredThroughRuntimeValue() {
   const std::array<uint32_t, 1> user_data{0x4000u};
   DescriptorValue descriptor;
   Check(fixture.program.info.buffers.size() == 1u &&
-            SrtWalker(plan, {.user_data = user_data}).EvaluateDescriptor(
+            EvaluateSrtDescriptor(plan, {.user_data = user_data},
                 fixture.program.info.buffers[0].source, descriptor) &&
             descriptor.dwords[0] == user_data[0],
         "runtime-rooted invariant loop lost its buffer source");
@@ -8079,7 +8067,7 @@ void TestInvariantLoopPhi() {
   std::array<uint32_t, 1> user_data{0x12345678u};
   SrtRuntime runtime{.user_data = user_data};
   DescriptorValue descriptor;
-  Check(SrtWalker(plan, runtime).EvaluateDescriptor(fixture.program.info.buffers[0].source, descriptor) &&
+  Check(EvaluateSrtDescriptor(plan, runtime, fixture.program.info.buffers[0].source, descriptor) &&
             descriptor.dwords[0] == user_data[0],
         "loop-invariant descriptor phi was not evaluated through typed SSA");
 }
@@ -8196,8 +8184,7 @@ void TestBoundedRelativeRegisterWrites() {
     const std::array<uint32_t, 3> user_data{0x1000u, 0u, 72u};
     SrtRuntime runtime{.user_data = user_data};
     DescriptorValue descriptor;
-    Check(SrtWalker(plan, runtime).EvaluateDescriptor(
-              fixture.program.info.buffers[0].source, descriptor) &&
+    Check(EvaluateSrtDescriptor(plan, runtime, fixture.program.info.buffers[0].source, descriptor) &&
               descriptor.dwords[2] == 72u,
           "impossible relative writes left a false descriptor dependency");
   }
@@ -8412,6 +8399,7 @@ void TestConditionalBufferMaterialization() {
   TestMemory memory;
   SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
                      .read_specialization_memory = ReadTestMemory};
+  FakeGuestMemory fake(ReadTestMemory, &memory);
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
@@ -8422,12 +8410,11 @@ void TestConditionalBufferMaterialization() {
   Check(snapshot.user_data == std::vector<uint32_t>(user_data.begin(), user_data.end()),
         "resource reachability changed native shader user data");
 
+  // The walker needs every user-data DWORD it may read, even for an untaken branch.
   runtime.user_data = std::span(user_data).first(4);
-  Check(MaterializeResources(plan, runtime, snapshot, specialization),
-        "untaken branch evaluated its unavailable descriptor");
-  memory.words[0] = 1;
   Check(!MaterializeResources(plan, runtime, snapshot, specialization),
-        "taken branch accepted an unavailable descriptor");
+        "truncated user data was accepted");
+  memory.words[0] = 1;
 
   runtime.user_data = user_data;
   const auto CheckActive = [&] {
@@ -8438,9 +8425,13 @@ void TestConditionalBufferMaterialization() {
           "potentially executed buffer descriptor was discarded");
   };
   CheckActive();
-  memory.words[0] = 0;
+  // An unreadable control word reads as zero and skips the branch.
   memory.fail_after = memory.reads;
-  CheckActive();
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            snapshot.buffers.size() == 2 && snapshot.buffers[1].dwords == std::array<uint32_t, 8>{},
+        "unreadable scalar control word did not read as zero");
+  // Walker loads do not depend on the specialization reader.
+  memory.fail_after = UINT32_MAX;
   runtime.read_specialization_memory = nullptr;
   CheckActive();
 }
@@ -8515,8 +8506,9 @@ void TestGuardedScalarDescriptorReads() {
         return ReadLinearTestMemory(&memory.data, address, words);
       };
       const std::array<uint32_t, 5> user_data{0x1000u, 0u, 64u, 0u, 0x1040u};
-      const SrtRuntime runtime{.user_data = user_data, .read_memory = Read,
-                               .userdata = &memory, .read_specialization_memory = Read};
+      const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+                               .read_specialization_memory = Read};
+      FakeGuestMemory fake(Read, &memory);
       memory.data.words[0xa8 / 4] = 0x2000u;
       memory.data.words[0xb0 / 4] = 4u;
       ResourceSnapshot snapshot;
@@ -8526,8 +8518,9 @@ void TestGuardedScalarDescriptorReads() {
                   std::ranges::all_of(snapshot.flattened_srt, [](auto word) { return word == 0; }),
               "disabled feature speculatively dereferenced its null BVH table");
         memory.data.words[7] = 1;
-        Check(!MaterializeResources(plan, runtime, snapshot, specialization) && memory.null_reads == 1,
-              "active feature accepted an unreadable BVH table");
+        Check(MaterializeResources(plan, runtime, snapshot, specialization) && memory.null_reads == 1 &&
+                  std::ranges::all_of(snapshot.flattened_srt, [](auto word) { return word == 0; }),
+              "active feature did not read its unreadable null BVH table as zero");
       }
       memory.data.words[0x40 / 4] = 0x1080u;
       memory.data.watched_address = 0x10a8u;
@@ -8557,19 +8550,24 @@ void TestConservativeBufferReachability() {
   TestMemory memory;
   const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
                            .read_specialization_memory = ReadTestMemory};
+  FakeGuestMemory fake(ReadTestMemory, &memory);
   for (const auto use : {ConditionalBufferUse::Shared, ConditionalBufferUse::Loop,
                          ConditionalBufferUse::Writable}) {
     auto plan = ConditionalBufferPlan(use);
     ResourceSnapshot snapshot;
     ResourceSpecialization specialization;
-    std::array<uint32_t, 4> expected{};
-    if (use != ConditionalBufferUse::Writable)
-      std::copy_n(user_data.begin() + 4, expected.size(), expected.begin());
-    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
-              snapshot.buffers.size() == 2 &&
-              std::equal(expected.begin(), expected.end(),
-                         snapshot.buffers[1].dwords.begin()),
-          "shared or loop-dependent resource was pruned, or inactive scalar branch retained");
+    // With a zero control word, the loop body's only other predecessor is itself.
+    for (const uint32_t flag : {0u, 1u}) {
+      memory.words[0] = flag;
+      std::array<uint32_t, 4> expected{};
+      if (use == ConditionalBufferUse::Shared || flag != 0u)
+        std::copy_n(user_data.begin() + 4, expected.size(), expected.begin());
+      Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+                snapshot.buffers.size() == 2 &&
+                std::equal(expected.begin(), expected.end(),
+                           snapshot.buffers[1].dwords.begin()),
+            "shared or entered loop resource was pruned, or an untaken branch or loop retained");
+    }
   }
 }
 

tests/ScalarProvenanceTests.cpp

@@ -4,7 +4,7 @@
 #include "graphics/shader/recompiler/ir/passes/ReadLaneElimination.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceTracking.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
-#include "graphics/shader/recompiler/ir/passes/SrtWalker.h"
+#include "SrtWalkerTestMemory.h"
 
 #include <cstdlib>
 #include <iostream>
@@ -95,6 +95,16 @@ bool ReadMemory(void *userdata, uint64_t address, std::span<uint32_t> values) {
   return true;
 }
 
+// Evaluates the DWORDs of a descriptor source that no resource consumes.
+bool EvaluateSource(ResourcePlan &plan, const SrtRuntime &runtime, uint32_t source,
+                    DescriptorValue &result) {
+  const auto &words = plan.descriptor_sources[source];
+  result = {.dword_count = words.dword_count};
+  for (uint32_t word = 0; word < words.dword_count; ++word)
+    if (!EvaluateSrtValue(plan, runtime, words.dwords[word], result.dwords[word])) return false;
+  return true;
+}
+
 Value Address(Fixture &fixture, Value low, Value high, uint32_t block = 0) {
   return fixture.Emit(ValueOpcode::GetAddressResource, {low, high}, 0, block);
 }
@@ -132,9 +142,9 @@ void TestImmediateFlatteningAndGvn() {
         "flattened raw read was not kept as a planning-only root");
 
   TestMemory memory_image{{{0x1020u, 0xfeedbeefu}}};
-  SrtRuntime runtime{.read_memory = ReadMemory, .userdata = &memory_image};
+  FakeGuestMemory fake(ReadMemory, &memory_image);
   std::vector<uint32_t> flat;
-  Check(SrtWalker(ExtractResourcePlan(fixture.program), runtime).RefreshFlatBuffer(flat), "flattened SRT walk failed");
+  Check(WalkSrt(ExtractResourcePlan(fixture.program), {}, flat, nullptr), "flattened SRT walk failed");
   Check(flat == std::vector<uint32_t>{0xfeedbeefu} && memory_image.reads == 1,
         "flattened SRT did not evaluate its canonical read once");
 }
@@ -148,9 +158,9 @@ void TestRawScalarComponentAlignment() {
   fixture.Plan();
 
   TestMemory memory_image{{{0x1000u, 0x12345678u}}};
-  SrtRuntime runtime{.read_memory = ReadMemory, .userdata = &memory_image};
+  FakeGuestMemory fake(ReadMemory, &memory_image);
   std::vector<uint32_t> flat;
-  Check(SrtWalker(ExtractResourcePlan(fixture.program), runtime).RefreshFlatBuffer(flat), "raw scalar SRT walk failed");
+  Check(WalkSrt(ExtractResourcePlan(fixture.program), {}, flat, nullptr), "raw scalar SRT walk failed");
   Check(
       flat == std::vector<uint32_t>{0x12345678u} && memory_image.reads == 1,
       "raw scalar base, immediate, and offset were not aligned independently");
@@ -203,9 +213,9 @@ void TestNestedSrtWalk() {
   fixture.Plan();
 
   TestMemory memory_image{{{0x1000u, 0x2000u}, {0x2000u, 0xabcdef01u}}};
-  SrtRuntime runtime{.read_memory = ReadMemory, .userdata = &memory_image};
+  FakeGuestMemory fake(ReadMemory, &memory_image);
   std::vector<uint32_t> flat;
-  Check(SrtWalker(ExtractResourcePlan(fixture.program), runtime).RefreshFlatBuffer(flat), "nested SRT walk failed");
+  Check(WalkSrt(ExtractResourcePlan(fixture.program), {}, flat, nullptr), "nested SRT walk failed");
   Check(flat == std::vector<uint32_t>({0x2000u, 0xabcdef01u}),
         "nested typed SRT reads were not evaluated in dependency order");
 }
@@ -227,8 +237,9 @@ void TestShaderBaseAndUserData() {
   const std::array user_data{0x20u};
   SrtRuntime runtime{.user_data = user_data,
                      .shader_base = 0x12345678abcdef00ull};
+  auto plan = ExtractResourcePlan(fixture.program);
   DescriptorValue result;
-  Check(SrtWalker(ExtractResourcePlan(fixture.program), runtime).EvaluateDescriptor(0, result),
+  Check(EvaluateSource(plan, runtime, 0, result),
         "shader-relative descriptor evaluation failed");
   Check(result.dword_count == 3 && result.dwords[0] == 0xabcdef00u &&
             result.dwords[1] == 0x12345678u && result.dwords[2] == 0x24u,
@@ -249,8 +260,9 @@ void TestCarryAndBitFields() {
   fixture.program.descriptor_sources.push_back(
       {.dwords = {low, high, inserted, sign}, .dword_count = 4});
 
+  auto plan = ExtractResourcePlan(fixture.program);
   DescriptorValue result;
-  Check(SrtWalker(ExtractResourcePlan(fixture.program), {}).EvaluateDescriptor(0, result),
+  Check(EvaluateSource(plan, {}, 0, result),
         "carry and bit-field descriptor evaluation failed");
   Check(result.dwords[0] == 1u && result.dwords[1] == 1u &&
             result.dwords[2] == 0x89abcdefu && result.dwords[3] == 0xffffffffu,
@@ -273,12 +285,12 @@ void TestInvariantAndDivergentPhi() {
   fixture.program.descriptor_sources.push_back(
       {.dwords = {Value(&divergent)}, .dword_count = 1});
 
-  const auto plan = ExtractResourcePlan(fixture.program);
+  auto plan = ExtractResourcePlan(fixture.program);
   DescriptorValue result;
-  Check(SrtWalker(plan, {}).EvaluateDescriptor(0, result) &&
+  Check(EvaluateSource(plan, {}, 0, result) &&
             result.dwords[0] == 7u,
         "loop-invariant typed phi was rejected");
-  Check(!SrtWalker(plan, {}).EvaluateDescriptor(1, result),
+  Check(!EvaluateSource(plan, {}, 1, result),
         "divergent phi was accepted");
 }
 
@@ -294,7 +306,7 @@ void TestOperandMutationLifecycle() {
   source->ReplaceUsesWith(Value(0xaau));
   uint32_t result = 0;
   Check(source->UseCount() == 0 &&
-            SrtWalker(fixture.program, {}).Evaluate(inserted, result) && result == 0x3au,
+            EvaluateSrtValue(fixture.program, {}, inserted, result) && result == 0x3au,
         "inline operand replacement changed the evaluated value");
 
   for (const auto opcode : {ValueOpcode::GetImageResource, ValueOpcode::MakeImageAddress}) {
@@ -333,7 +345,7 @@ void TestOperandMutationLifecycle() {
   const auto sum = fixture.Emit(ValueOpcode::IAdd32, {Value(&phi), Value(5u)}, 0, 5);
   phi.ReplaceUsesWith(Value(6u));
   Check(source->UseCount() == 0 && phi.NumPhiBlocks() == 0 && phi.NumArgs() == 1 &&
-            SrtWalker(fixture.program, {}).Evaluate(sum, result) && result == 11,
+            EvaluateSrtValue(fixture.program, {}, sum, result) && result == 11,
         "Phi replacement retained incoming uses or changed the consumer value");
 }
 
@@ -374,8 +386,9 @@ void TestRuntime64BitDescriptorOps() {
   fixture.program.descriptor_sources.push_back(
       {.dwords = {low, high}, .dword_count = 2});
 
+  auto plan = ExtractResourcePlan(fixture.program);
   DescriptorValue result;
-  Check(SrtWalker(ExtractResourcePlan(fixture.program), {}).EvaluateDescriptor(0, result) &&
+  Check(EvaluateSource(plan, {}, 0, result) &&
             result.dwords[0] == 0xabcdu && result.dwords[1] == 0x1235u,
         "64-bit typed descriptor arithmetic evaluation is incorrect");
 }
@@ -448,13 +461,13 @@ void TestUniformFirstLaneSamplerLod() {
   std::array<uint32_t, 7> user_data{};
   user_data[6] = 3u;
   SrtRuntime runtime{.user_data = user_data};
-  const auto plan = ExtractResourcePlan(fixture.program);
+  auto plan = ExtractResourcePlan(fixture.program);
   DescriptorValue result;
-  Check(SrtWalker(plan, runtime).EvaluateDescriptor(0, result) &&
+  Check(EvaluateSource(plan, runtime, 0, result) &&
             result.dwords[0] == 0x00300300u,
         "uniform sampler LOD evaluated incorrectly");
   user_data[6] = 20u;
-  Check(SrtWalker(plan, runtime).EvaluateDescriptor(0, result) &&
+  Check(EvaluateSource(plan, runtime, 0, result) &&
             result.dwords[0] == 0x00ffffffu,
         "uniform sampler LOD clamp evaluated incorrectly");
 }
@@ -489,13 +502,12 @@ void TestFloatComparisonDescriptorInputs() {
     fixture.program.descriptor_sources.push_back(
         {.dwords = {low, high}, .dword_count = 2});
 
-    const auto plan = ExtractResourcePlan(fixture.program);
+    auto plan = ExtractResourcePlan(fixture.program);
     for (size_t index = 0; index < inputs.size(); index++) {
       const auto &input = inputs[index];
       const std::array user_data{input.bits};
       DescriptorValue result;
-      Check(SrtWalker(plan, {.user_data = user_data})
-                    .EvaluateDescriptor(0, result) &&
+      Check(EvaluateSource(plan, {.user_data = user_data}, 0, result) &&
                 result.dwords[0] == (flush && index < 4 ? 1 : input.less_equal) &&
                 result.dwords[1] == (flush && index < 4 ? 1 : input.greater_equal),
             "descriptor comparison disagrees with native FP32 input mode");
@@ -551,7 +563,6 @@ void TestConstantBufferBounds() {
     uint32_t immediate;
     uint32_t stride;
     uint32_t records;
-    bool valid;
     uint32_t expected;
     uint32_t reads;
   };
@@ -565,17 +576,17 @@ void TestConstantBufferBounds() {
     ++probe.attempts;
     return ReadMemory(&probe.memory, address, values);
   };
-  // PS5 scalar-buffer loads return zero per out-of-range DWORD; an unreadable
-  // in-range address remains a failed memory read.
-  for (const auto &test : {Case{12u, 0u, 0u, 16u, true, 0xa5a5a5a5u, 1u},
-                           Case{3u, 1u, 0u, 16u, true, 0x12345678u, 1u},
-                           Case{0xfffffffcu, 4u, 0u, 16u, true, 0u, 0u},
-                           Case{16u, 0u, 0u, 16u, true, 0u, 0u},
-                           Case{12u, 0u, 0u, 14u, true, 0u, 0u},
-                           Case{0u, 0u, 0u, 0u, true, 0u, 0u},
-                           Case{8u, 0u, 0u, 16u, false, 0u, 1u},
-                           Case{12u, 0u, 48u, 32768u, true, 0xa5a5a5a5u, 1u},
-                           Case{0x6339c0u, 16u, 48u, 32768u, true, 0u, 0u}}) {
+  // PS5 scalar-buffer loads return zero per out-of-range DWORD without an access;
+  // an unreadable in-range address is accessed and reads as zero too.
+  for (const auto &test : {Case{12u, 0u, 0u, 16u, 0xa5a5a5a5u, 1u},
+                           Case{3u, 1u, 0u, 16u, 0x12345678u, 1u},
+                           Case{0xfffffffcu, 4u, 0u, 16u, 0u, 0u},
+                           Case{16u, 0u, 0u, 16u, 0u, 0u},
+                           Case{12u, 0u, 0u, 14u, 0u, 0u},
+                           Case{0u, 0u, 0u, 0u, 0u, 0u},
+                           Case{8u, 0u, 0u, 16u, 0u, 1u},
+                           Case{12u, 0u, 48u, 32768u, 0xa5a5a5a5u, 1u},
+                           Case{0x6339c0u, 16u, 48u, 32768u, 0u, 0u}}) {
     Fixture fixture;
     const auto memory = fixture.AddMemory(ResourceKind::ScalarBuffer, test.immediate);
     const auto buffer =
@@ -588,12 +599,11 @@ void TestConstantBufferBounds() {
     fixture.Plan();
 
     MemoryProbe memory_image;
-    SrtRuntime runtime{.read_memory = read_memory, .userdata = &memory_image};
+    FakeGuestMemory fake(read_memory, &memory_image);
     std::vector<uint32_t> flat;
-    Check(SrtWalker(ExtractResourcePlan(fixture.program), runtime).RefreshFlatBuffer(flat) == test.valid,
+    Check(WalkSrt(ExtractResourcePlan(fixture.program), {}, flat, nullptr),
           "constant-buffer walk misaligned or wrapped its offset components");
-    Check((!test.valid || flat == std::vector<uint32_t>{test.expected}) &&
-              memory_image.attempts == test.reads,
+    Check(flat == std::vector<uint32_t>{test.expected} && memory_image.attempts == test.reads,
           "constant-buffer walk read the wrong word or accessed an out-of-bounds address");
   }
 }
@@ -684,7 +694,8 @@ void TestCompleteUMinReduction() {
     const auto cross = fixture.Emit(ValueOpcode::Permlane16U32,
         {test == Case::CrossSource ? source : prefix,
          Value(test == Case::CrossSelectorLow ? 0xfffffff0u : UINT32_MAX),
-         Value(test == Case::CrossSelectorHigh ? 0xfffffff0u : UINT32_MAX), cross_exec});
+         Value(test == Case::CrossSelectorHigh ? 0xfffffff0u : UINT32_MAX), cross_exec,
+         Value(0u)});
     cross.Instruction()->SetFlags(PermlaneFlags{
         .x16 = test != Case::CrossRow, .fetch_inactive = test == Case::CrossFetch,
         .bound_control = test != Case::CrossBound});
@@ -831,8 +842,9 @@ void TestUndefinedRuntimeValueFails() {
 
   fixture.program.descriptor_sources.push_back(
       {.dwords = {undef}, .dword_count = 1});
+  auto plan = ExtractResourcePlan(fixture.program);
   DescriptorValue result;
-  Check(!SrtWalker(ExtractResourcePlan(fixture.program), {}).EvaluateDescriptor(0, result),
+  Check(!EvaluateSource(plan, {}, 0, result),
         "undefined typed descriptor source was accepted");
 }
 

tests/ShaderVertexMetadataTests.cpp

@@ -40,9 +40,7 @@ void CheckRejected(const ShaderMappedData& data, const char* text) {
 	output.vertex_buffer_reg = 37;
 	output.vertex_attrib_reg = 41;
 	output.input_semantics   = prior_semantics;
-	std::string error;
-	Check(!ShaderReadVertexMetadata(data, 64, output, &error), text);
-	Check(!error.empty(), "metadata rejection omitted its diagnostic");
+	Check(ShaderReadVertexMetadata(data, 64, output) != nullptr, text);
 	Check(output.vertex_buffer_reg == 37 && output.vertex_attrib_reg == 41 &&
 	          output.input_semantics.data() == prior_semantics.data() &&
 	          output.input_semantics.size() == prior_semantics.size(),
@@ -52,8 +50,7 @@ void CheckRejected(const ShaderMappedData& data, const char* text) {
 void TestValidAndInvalidMetadata() {
 	Fixture              fixture;
 	ShaderVertexMetadata output;
-	std::string          error;
-	Check(ShaderReadVertexMetadata(fixture.mapped, 64, output, &error),
+	Check(ShaderReadVertexMetadata(fixture.mapped, 64, output) == nullptr,
 	      "valid AGC vertex metadata was rejected");
 	Check(output.vertex_buffer_reg == 2 && output.vertex_attrib_reg == 4 &&
 	          output.input_semantics.size() == 1,
@@ -64,7 +61,7 @@ void TestValidAndInvalidMetadata() {
 	    AGC_ILLEGAL_DIRECT_OFFSET;
 	buffer_only.mapped.input_semantics     = nullptr;
 	buffer_only.mapped.num_input_semantics = 0;
-	Check(ShaderReadVertexMetadata(buffer_only.mapped, 64, output, &error),
+	Check(ShaderReadVertexMetadata(buffer_only.mapped, 64, output) == nullptr,
 	      "vertex-buffer-only AGC metadata was rejected");
 	Check(output.vertex_buffer_reg == 2 && output.vertex_attrib_reg == -1 &&
 	          output.input_semantics.empty(),

tests/SlotVectorTests.cpp

@@ -0,0 +1,103 @@
+#include "common/slotVector.h"
+
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <utility>
+#include <vector>
+
+namespace {
+
+void Check(bool value, const char *message) {
+  if (!value) {
+    std::fprintf(stderr, "SlotVectorTests: failed: %s\n", message);
+    std::abort();
+  }
+}
+
+// Non-movable like the cached Image/Buffer; counts live objects.
+struct Probe {
+  Probe(int *live, uint32_t value) : live(live), value(value) { ++*live; }
+  ~Probe() { --*live; }
+  KYTY_CLASS_NO_COPY(Probe);
+
+  int *live;
+  uint32_t value;
+};
+
+static_assert(sizeof(Common::SlotId) == sizeof(uint32_t));
+
+void TestInsertEraseReuse() {
+  int live = 0;
+  Common::SlotVector<Probe> slots;
+  Check(!slots.is_allocated(Common::SlotId {}) && slots.try_get(Common::SlotId {}) == nullptr &&
+            !slots.is_allocated(Common::SlotId {0}),
+        "an empty vector reported an allocated slot");
+
+  const auto a = slots.insert(&live, 0u);
+  const auto b = slots.insert(&live, 1u);
+  const auto c = slots.insert(&live, 2u);
+  Check(a.index == 0 && b.index == 1 && c.index == 2 && live == 3,
+        "fresh slots were not handed out in order");
+  Check(slots.is_allocated(b) && slots.try_get(b) == &slots[b] && slots[b].value == 1,
+        "an inserted slot did not expose its object");
+  Check(!slots.is_allocated(Common::SlotId {3}), "an unused slot reported an object");
+
+  slots.erase(b);
+  slots.erase(a);
+  Check(live == 1 && !slots.is_allocated(a) && !slots.is_allocated(b) &&
+            slots.try_get(b) == nullptr && slots.is_allocated(c),
+        "erase did not destroy and free exactly its slot");
+
+  // Freed indices are reused most recently freed first, with the same plain id.
+  const auto reused_a = slots.insert(&live, 0u);
+  const auto reused_b = slots.insert(&live, 1u);
+  const auto d = slots.insert(&live, 3u);
+  Check(reused_a == a && reused_b == b && d.index == 3 && live == 4,
+        "freed slots were not reused in LIFO order");
+  Check(slots[reused_a].value == 0 && slots[reused_b].value == 1,
+        "a reused slot did not expose its new object");
+}
+
+void TestForEachAndStability() {
+  int live = 0;
+  {
+    Common::SlotVector<Probe> slots;
+    const auto first = slots.insert(&live, 0u);
+    Probe *const first_address = &slots[first];
+
+    // Cross several chunks; earlier objects must keep their address.
+    std::vector<Common::SlotId> ids {first};
+    for (uint32_t i = 1; i < 1000; ++i) {
+      ids.push_back(slots.insert(&live, i));
+    }
+    Check(&slots[first] == first_address && live == 1000,
+          "growing the vector moved a stored object");
+
+    std::vector<uint32_t> expected;
+    for (uint32_t i = 0; i < 1000; ++i) {
+      if (i % 3 == 0) {
+        slots.erase(ids[i]);
+      } else {
+        expected.push_back(i);
+      }
+    }
+    std::vector<uint32_t> visited;
+    std::as_const(slots).ForEach([&](Common::SlotId id, const Probe &probe) {
+      Check(probe.value == id.index, "ForEach passed an id that does not match its object");
+      visited.push_back(id.index);
+    });
+    Check(visited == expected && live == static_cast<int>(expected.size()),
+          "ForEach did not visit exactly the live slots in index order");
+  }
+  Check(live == 0, "destroying the vector did not destroy every live object exactly once");
+}
+
+} // namespace
+
+int main() {
+  TestInsertEraseReuse();
+  TestForEachAndStability();
+  std::puts("SlotVectorTests: all cases passed");
+  return 0;
+}

tests/SrtWalkerTestMemory.h

@@ -0,0 +1,113 @@
+#ifndef KYTY_TESTS_SRTWALKERTESTMEMORY_H_
+#define KYTY_TESTS_SRTWALKERTESTMEMORY_H_
+
+#include "graphics/shader/recompiler/ir/passes/SrtWalker.h"
+
+#include <xbyak/xbyak.h>
+
+#include <csignal>
+#include <cstdlib>
+#include <utility>
+#include <vector>
+
+#if defined(__linux__)
+#include <ucontext.h>
+#elif defined(_WIN32)
+#include <windows.h>
+#endif
+
+// Serves SRT walker loads from fake guest addresses: a walker load that faults completes with
+// the test reader's DWORD, or with zero like the emulator when the reader fails or is null.
+class FakeGuestMemory {
+public:
+  explicit FakeGuestMemory(Libs::Graphics::ShaderRecompiler::IR::SrtMemoryReader reader = nullptr,
+                           void *userdata = nullptr) {
+    Current() = {reader, userdata};
+#if defined(__linux__)
+    struct sigaction action {};
+    action.sa_sigaction = Handle;
+    action.sa_flags = SA_SIGINFO;
+    sigaction(SIGSEGV, &action, &m_previous);
+#elif defined(_WIN32)
+    m_handler = AddVectoredExceptionHandler(1, Handle);
+#endif
+  }
+  ~FakeGuestMemory() {
+#if defined(__linux__)
+    sigaction(SIGSEGV, &m_previous, nullptr);
+#elif defined(_WIN32)
+    RemoveVectoredExceptionHandler(m_handler);
+#endif
+    Current() = {};
+  }
+  FakeGuestMemory(const FakeGuestMemory &) = delete;
+  FakeGuestMemory &operator=(const FakeGuestMemory &) = delete;
+
+private:
+  using Reader = std::pair<Libs::Graphics::ShaderRecompiler::IR::SrtMemoryReader, void *>;
+  static Reader &Current() {
+    static thread_local Reader reader;
+    return reader;
+  }
+  static uint32_t Read(uint64_t address) {
+    uint32_t word = 0;
+    const auto [reader, userdata] = Current();
+    return reader != nullptr && reader(userdata, address, {&word, 1}) ? word : 0u;
+  }
+#if defined(__linux__)
+  static void Handle(int, siginfo_t *, void *context) {
+    auto &registers = static_cast<ucontext_t *>(context)->uc_mcontext.gregs;
+    const auto length =
+        Libs::Graphics::ShaderRecompiler::IR::SrtWalkerLoadLength(registers[REG_RIP]);
+    if (length == 0) std::abort();
+    registers[REG_RAX] = Read(registers[REG_R10]);
+    registers[REG_RIP] += length;
+  }
+  struct sigaction m_previous {};
+#elif defined(_WIN32)
+  static LONG CALLBACK Handle(EXCEPTION_POINTERS *info) {
+    auto &context = *info->ContextRecord;
+    const auto length = Libs::Graphics::ShaderRecompiler::IR::SrtWalkerLoadLength(context.Rip);
+    if (info->ExceptionRecord->ExceptionCode != EXCEPTION_ACCESS_VIOLATION || length == 0)
+      return EXCEPTION_CONTINUE_SEARCH;
+    context.Rax = Read(context.R10);
+    context.Rip += length;
+    return EXCEPTION_CONTINUE_EXECUTION;
+  }
+  void *m_handler = nullptr;
+#endif
+};
+
+// Evaluates one value through a walker compiled for it, keeping the plan's resource walker.
+inline bool EvaluateSrtValue(Libs::Graphics::ShaderRecompiler::IR::ResourcePlan &plan,
+                             const Libs::Graphics::ShaderRecompiler::IR::SrtRuntime &runtime,
+                             Libs::Graphics::ShaderRecompiler::IR::Value value, uint32_t &result) {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  const auto *inst = value.Resolve().TryInstruction();
+  const auto output = inst != nullptr ? inst->WalkOutput() : UINT32_MAX;
+  auto code = std::move(plan.srt_code);
+  auto scratch = std::move(plan.srt_scratch);
+  const auto user_data = plan.srt_user_data;
+  const auto captures = plan.srt_captures;
+  CompileSrtWalker(plan, {&value, 1});
+  std::vector<uint32_t> flat;
+  const bool known = WalkSrt(plan, runtime, flat, nullptr) && WalkedValue(plan, value, result);
+  plan.srt_code = std::move(code);
+  plan.srt_scratch = std::move(scratch);
+  plan.srt_user_data = user_data;
+  plan.srt_captures = captures;
+  if (inst != nullptr) inst->SetWalkOutput(output);
+  return known;
+}
+
+// Walks the plan's resources and reads back one descriptor.
+inline bool EvaluateSrtDescriptor(const Libs::Graphics::ShaderRecompiler::IR::ResourcePlan &plan,
+                                  const Libs::Graphics::ShaderRecompiler::IR::SrtRuntime &runtime,
+                                  uint32_t source,
+                                  Libs::Graphics::ShaderRecompiler::IR::DescriptorValue &result) {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  std::vector<uint32_t> flat;
+  return WalkSrt(plan, runtime, flat, nullptr) && WalkedDescriptor(plan, source, result);
+}
+
+#endif // KYTY_TESTS_SRTWALKERTESTMEMORY_H_

tests/VirtualMemoryAllocationTests.cpp

@@ -720,6 +720,44 @@ void TestBackingReadExcludesWritesAndUnmap() {
 	std::printf("[host]    %-48s ok\n", test);
 }
 
+void TestTrackerProtectSharesBackingReads() {
+	using namespace Libs::LibKernel::Memory;
+	const char* test = "TrackerProtectSharesBackingReads";
+	const auto  base = MapNamedFlexible(test, SceKernelPageSize, SceKernelProtCpuRw,
+	                                    "tracker_protect_shares_reads");
+	std::counting_semaphore<2> entered {0};
+	std::counting_semaphore<2> resume {0};
+	g_backing_reads_entered  = &entered;
+	g_continue_backing_reads = &resume;
+	TestSetBackingReadCallback(ParkBackingRead);
+	uint64_t    value = 0;
+	std::thread reader([&] { CopySparseMemory(base, &value, sizeof(value)); });
+	const bool  reader_parked = entered.try_acquire_for(std::chrono::seconds(5));
+	std::counting_semaphore<2> done {0};
+	bool                       tracker_ok   = false;
+	int                        guest_result = OK;
+	std::thread                protector([&] {
+		tracker_ok = ProtectGuestHostMemory(base, SceKernelPageSize,
+		                                    Common::VirtualMemory::Mode::ReadWrite);
+		done.release();
+		guest_result =
+		    KernelMprotect(reinterpret_cast<void*>(base), SceKernelPageSize, SceKernelProtCpuRw);
+		done.release();
+	});
+	const bool shared   = done.try_acquire_for(std::chrono::seconds(5));
+	const bool excluded = !done.try_acquire_for(std::chrono::milliseconds(100));
+	resume.release();
+	reader.join();
+	protector.join();
+	TestSetBackingReadCallback(nullptr);
+	CheckOk(test, KernelMunmap(base, SceKernelPageSize), "KernelMunmap");
+	Check(test, reader_parked && shared && tracker_ok,
+	      "tracker protection waited for a backing read");
+	Check(test, excluded && guest_result == OK,
+	      "guest mprotect completed while a backing read held its lock");
+	std::printf("[host]    %-48s ok\n", test);
+}
+
 void TestGuestAddressSpaceHasNoFixedFallback() {
 	const char* test            = "GuestAddressSpaceHasNoFixedFallback";
 	const auto  unowned_address = reinterpret_cast<void*>(0x10000);
@@ -4894,6 +4932,7 @@ int main(int argc, char** argv) {
 	if (argc == 2 && std::strcmp(argv[1], "--backing-transfers-only") == 0) {
 		RunTest(TestConcurrentBackingReads);
 		RunTest(TestBackingReadExcludesWritesAndUnmap);
+		RunTest(TestTrackerProtectSharesBackingReads);
 		RunTest(TestSparseBackingReadPreservesResidency);
 		RunTest(TestCopySparseMemoryMappings);
 		RunTest(TestSparseReadDuringDirectCommit);
@@ -4953,6 +4992,7 @@ int main(int argc, char** argv) {
 	RunTest(TestSparseReadDuringDirectCommit);
 	RunTest(TestConcurrentBackingReads);
 	RunTest(TestBackingReadExcludesWritesAndUnmap);
+	RunTest(TestTrackerProtectSharesBackingReads);
 	RunTest(TestGuestAddressSpaceHasNoFixedFallback);
 	RunTest(TestGuestFreeRangeSearchDoesNotUnderflow);
 	RunTest(TestFlexibleMemoryCapacityIsBootFixed);

tests/shaderCfgTests.cpp

@@ -8,6 +8,7 @@
 #include "graphics/host_gpu/renderer/image/imageView.h"
 #include "graphics/host_gpu/renderer/image/textureCommon.h"
 #include "graphics/host_gpu/renderer/pipeline/blendMapping.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
 #include "graphics/shader/recompiler/backend/spirv/SpirvEmitter.h"
@@ -24,13 +25,13 @@
 #include "graphics/shader/recompiler/ir/passes/ResourceTracking.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
 #include "graphics/shader/recompiler/ir/passes/ShaderInfoCollection.h"
-#include "graphics/shader/recompiler/ir/passes/SrtWalker.h"
 #include "graphics/shader/recompiler/ir/passes/SsaRewrite.h"
 #include "graphics/shader/shader.h"
 #include "graphics/shader/shaderCompiler.h"
 #include "libs/agc.h"
 #include "spirv-tools/libspirv.hpp"
 #include "xxhash.h"
+#include "SrtWalkerTestMemory.h"
 
 #include <algorithm>
 #include <array>
@@ -135,8 +136,7 @@ struct TestCompileResult {
 TestCompileResult RecompileForTest(
     std::span<const uint32_t> code,
     const ShaderRecompiler::CompileOptions &options,
-    ShaderRecompiler::IR::SrtMemoryReader read_memory = nullptr,
-    void *read_memory_data = nullptr, uint32_t push_data_start_dword = 0) {
+    uint32_t push_data_start_dword = 0) {
   auto translated = ShaderRecompiler::TranslateProgram(code, options);
   auto plan = ShaderRecompiler::IR::ExtractResourcePlan(translated.program);
   ShaderRecompiler::IR::ResourceSnapshot resources;
@@ -144,8 +144,6 @@ TestCompileResult RecompileForTest(
   const ShaderRecompiler::IR::SrtRuntime runtime{
       .user_data = options.user_data,
       .shader_base = reinterpret_cast<uint64_t>(code.data()),
-      .read_memory = read_memory,
-      .userdata = read_memory_data,
       .read_specialization_memory = ReadHostTestMemory,
   };
   Check(ShaderRecompiler::IR::MaterializeResources(
@@ -166,10 +164,10 @@ void CompilePixelRuntime(const ShaderParams &params,
 
   options.input_info.pixel = &input_info;
   auto result = RecompileForTest(params.code, options);
-  static std::deque<ShaderRecompiler::IR::CompiledShaderInfo> programs;
+  static std::deque<ShaderModule> modules;
   static std::deque<ShaderRecompiler::IR::ResourceSnapshot> resources;
-  programs.push_back(std::move(result.program).TakeCompiledInfo());
-  input_info.stage.program = &programs.back();
+  modules.push_back({.info = std::move(result.program).TakeCompiledInfo()});
+  input_info.stage.module = &modules.back();
   resources.push_back(std::move(result.resources));
   input_info.stage.resources = &resources.back();
 }
@@ -1160,7 +1158,7 @@ void TestNativeShaderResourceDependencies() {
   set_buffer(0, 0x1000, 16, 3);
   set_buffer(1, 0x1800, 4, 7);
   set_buffer(2, 0x2000, 0, 64);
-  const ShaderRecompiler::IR::CompiledShaderInfo compiled{
+  const ShaderModule compiled{.info = {
       .stage = program.stage,
       .shader_hash = program.shader_hash,
       .wave_size = program.wave_size,
@@ -1169,8 +1167,8 @@ void TestNativeShaderResourceDependencies() {
       .scratch_dwords = program.scratch_dwords,
       .info = program.info,
       .bindings = program.bindings,
-  };
-  ShaderStageRuntime runtime{.program = &compiled, .resources = &resources};
+  }};
+  ShaderStageRuntime runtime{.module = &compiled, .resources = &resources};
   Check(HasShaderBufferWrites(runtime),
         "graphics/compute write predicate lost nonempty written buffers");
   set_buffer(0, 0, 16, 3);
@@ -1588,11 +1586,6 @@ void SetImageTestFormat(std::array<uint32_t, 64> *data, uint32_t srsrc,
   (*data)[format_dword] = static_cast<uint32_t>(format) << 20u;
 }
 
-bool ReadZeroTestMemory(void *, uint64_t, std::span<uint32_t> values) {
-  std::ranges::fill(values, 0u);
-  return true;
-}
-
 constexpr uint32_t EncodeVop2(uint32_t opcode, uint32_t dst, uint32_t src0,
                               uint32_t src1) {
   return ((opcode & 0x3fu) << 25u) | ((dst & 0xffu) << 17u) |
@@ -1914,6 +1907,13 @@ void TestVertexBufferGrouping() {
               input.resources[2].Base48() == input.buffers[0].addr,
           "a later lower-address attribute or changed table left stale vertex offsets");
   }
+  // The vertex table metadata is parsed when the shader is mapped.
+  user_data.direct_resource_count = 0;
+  (void)PrepareProgram(regs, HW::Context{}, HW::UserConfig{}, input);
+  Check(input.fetch_attrib_reg == 2, "a draw reread the vertex metadata of a mapped shader");
+  ShaderMapUserData(regs.es_regs.data_addr, mapped);
+  (void)PrepareProgram(regs, HW::Context{}, HW::UserConfig{}, input);
+  Check(!input.fetch_embedded, "remapping a shader kept its previous vertex metadata");
 }
 
 void TestNggVertexEntryState() {
@@ -5408,7 +5408,8 @@ void TestNewShaderRecompilerMemoryFamilyTranslation() {
   options.dump_ir = true;
   options.user_data = user_data;
 
-  auto result = RecompileForTest(shader, options, ReadZeroTestMemory);
+  FakeGuestMemory zero_memory;
+  auto result = RecompileForTest(shader, options);
   Check((result.decoded_dump.find("s_load_dword") != std::string::npos),
         "new decoder did not decode SMEM dword load");
   Check((result.decoded_dump.find("s_buffer_load_dword") != std::string::npos),
@@ -9223,7 +9224,7 @@ void TestSharedReturnPreservesDescriptorDominance() {
   user_data[29] = static_cast<uint32_t>(address >> 32u);
   auto options = MakeCompileOptions(ShaderType::Pixel);
   options.user_data = user_data;
-  auto result = RecompileForTest(shader, options, ReadHostTestMemory);
+  auto result = RecompileForTest(shader, options);
   Check(!result.program.dispatcher_fallback &&
             result.resources.buffers.size() == 1u &&
             result.resources.buffers[0].dwords[0] == 0x2000u,
@@ -9283,7 +9284,7 @@ void TestNestedSelectionPreservesDescriptorSources() {
     auto options = MakeCompileOptions(ShaderType::Compute);
     options.user_data_base = 8;
     options.user_data = user_data;
-    const auto result = RecompileForTest(shader, options, ReadHostTestMemory);
+    const auto result = RecompileForTest(shader, options);
     Check(!result.program.dispatcher_fallback && result.resources.buffers.size() == 4u,
           "nested descriptor selection lost its structured resource bindings");
     for (size_t i = 0; i < result.program.info.buffers.size(); ++i) {
@@ -9359,7 +9360,7 @@ void TestSharedExitPreservesNativeDescriptorSources() {
     user_data[2] = arm;
     auto options = MakeCompileOptions(ShaderType::Pixel);
     options.user_data = user_data;
-    const auto result = RecompileForTest(shader, options, ReadHostTestMemory);
+    const auto result = RecompileForTest(shader, options);
     Check(!result.program.dispatcher_fallback && result.resources.buffers.size() == 2u,
           "shared-exit descriptor selection lost its structured resource bindings");
     for (size_t i = 0; i < result.program.info.buffers.size(); ++i) {
@@ -9405,7 +9406,7 @@ void TestNativeScalarReadDescriptorPlanning() {
     user_data[8] = mode;
     auto options = MakeCompileOptions(ShaderType::Pixel);
     options.user_data = user_data;
-    const auto result = RecompileForTest(shader, options, ReadHostTestMemory);
+    const auto result = RecompileForTest(shader, options);
     const auto expected = mode == 2u ? table : std::array<uint32_t, 4>{};
     Check(!result.program.dispatcher_fallback && result.resources.buffers.size() == 1u &&
               std::ranges::equal(result.resources.flattened_srt, expected),
@@ -9507,7 +9508,7 @@ void TestNativeGuardedSamplerSource(bool inline_constants) {
     auto options = MakeCompileOptions(ShaderType::Pixel);
     options.user_data_base = 8;
     options.user_data = user_data;
-    const auto result = RecompileForTest(code, options, ReadHostTestMemory);
+    const auto result = RecompileForTest(code, options);
     const auto expected = mode == 2u ? sampler : std::array<uint32_t, 4>{};
     Check(!result.program.dispatcher_fallback && result.resources.samplers.size() == 1u &&
               result.resources.samplers[0].dword_count == expected.size() &&
@@ -9615,8 +9616,7 @@ void TestBoundedMaterialBufferStores(bool scalar_key) {
   ShaderRecompiler::IR::ResourceSnapshot snapshot;
   ShaderRecompiler::IR::ResourceSpecialization specialization;
   const ShaderRecompiler::IR::SrtRuntime runtime{
-      .user_data = user_data, .read_memory = ReadHostTestMemory,
-      .read_specialization_memory = ReadHostTestMemory};
+      .user_data = user_data, .read_specialization_memory = ReadHostTestMemory};
   const auto materialize = [&] {
     return ShaderRecompiler::IR::MaterializeResources(plan, runtime, snapshot, specialization);
   };
@@ -9701,7 +9701,7 @@ void TestNativeDescriptorProvenanceKeepsGpuSelection() {
   options.input_info.compute = &input_info;
   options.user_data = user_data;
   for (const auto &shader : {selected, partial, loop}) {
-    const auto result = RecompileForTest(shader, options, ReadHostTestMemory);
+    const auto result = RecompileForTest(shader, options);
     Check(result.program.info.uses_dma && result.program.info.buffers.size() == 1u &&
               result.program.info.buffers[0].written,
           "GPU-selected scalar descriptor was materialized as a host binding");
@@ -10818,11 +10818,10 @@ void TestGraphicsComputeGetreg() {
   ShaderMappedData mapped{};
   mapped.code_size_bytes = sizeof(shader);
   ShaderMapUserData(regs.cs_regs.data_addr, mapped);
-  HW::ShaderRegisters sh{};
   ShaderComputeInputInfo graphics{}, async{};
   async.async_compute = true;
-  (void)PrepareProgram(regs, sh, graphics);
-  (void)PrepareProgram(regs, sh, async);
+  (void)PrepareProgram(regs, graphics);
+  (void)PrepareProgram(regs, async);
   Check(!graphics.async_compute && async.async_compute &&
             MakeStageStaticKey(graphics) != MakeStageStaticKey(async) &&
             MakeStageStaticKey(graphics).size() == MakeStageStaticKey(async).size(),
@@ -10864,7 +10863,7 @@ void TestMemoryFedScalarLeafCall() {
   struct Memory {
     std::array<uint32_t, 128> words{};
     uint64_t end = base + 0x10c;
-    bool clean = true;
+    bool readable = true;
   } memory;
   std::copy_n(std::array{0xe0040060u, 0x0010000fu, 8u, 0x5204u}.begin(),
               4, memory.words.begin());
@@ -10875,13 +10874,14 @@ void TestMemoryFedScalarLeafCall() {
   memory.words[66] = EncodeSop1(0x21, 125, 14);
   const auto read = [](void *data, uint64_t address, std::span<uint32_t> out) {
     const auto& m = *static_cast<Memory*>(data);
-    if (!m.clean || address < base || address > m.end || out.size_bytes() > m.end - address)
+    if (!m.readable || address < base || address > m.end || out.size_bytes() > m.end - address)
       return false;
     std::memcpy(out.data(), reinterpret_cast<const uint8_t*>(m.words.data()) + address - base,
                 out.size_bytes());
     return true;
   };
-  IR::SrtRuntime runtime{.read_memory = read, .userdata = &memory, .read_specialization_memory = read};
+  FakeGuestMemory fake(read, &memory);
+  IR::SrtRuntime runtime{.userdata = &memory, .read_specialization_memory = read};
   const auto options = MakeCompileOptions(ShaderType::Compute);
   const uint32_t gpu_callable[] = {EncodeVop1(0x02, 4, 256),
       EncodeSmem0(0x09, 14, 0), 4u << 25u, EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
@@ -10965,14 +10965,14 @@ void TestMemoryFedScalarLeafCall() {
   Check(source.revision == 3 && source.reads.back().second == 8,
         "shorter exact-boundary leaf was not decoded after a failed old-span read");
 #if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
-  memory.clean = false;
+  memory.readable = false;
   ExpectFatal([&] { (void)RefreshShaderSource(source, runtime); },
               "unreadable call dependencies reused stale linked code");
-  memory.clean = true;
-  auto ordinary = runtime;
-  ordinary.read_specialization_memory = nullptr;
-  ExpectFatal([&] { (void)RefreshShaderSource(source, ordinary); },
-              "call target query accepted an ordinary non-strict memory reader");
+  memory.readable = true;
+  auto no_reader = runtime;
+  no_reader.read_specialization_memory = nullptr;
+  ExpectFatal([&] { (void)RefreshShaderSource(source, no_reader); },
+              "leaf code was linked without a specialization memory reader");
   std::vector<uint32_t> escaping(std::begin(shader), std::end(shader) - 1);
   escaping.insert(escaping.end(), {EncodeVop1(0x01, 1, 14),
       EncodeMubuf0(0x1c, 0, false), EncodeMubuf1(1, 0, 0), EncodeSopp(0x01)});
@@ -10988,9 +10988,10 @@ void TestMemoryFedScalarLeafCall() {
       EncodeSMovB32(15, 129), EncodeSop2(0x0f, 14, 128, 128),
       EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
   auto pair_source = PrepareShaderSource(high_writer, options);
-  IR::SrtWalker pair_walker(pair_source.call_targets, runtime);
+  std::vector<uint32_t> flat;
   uint32_t high = 1;
-  Check(pair_walker.Evaluate(pair_source.calls.front().target[1], high) && high == 0,
+  Check(IR::WalkSrt(pair_source.call_targets, runtime, flat, nullptr) &&
+            IR::WalkedValue(pair_source.call_targets, pair_source.calls.front().target[1], high) && high == 0,
         "scalar pair producer used a stale preceding high-half definition");
   const uint32_t joined[] = {
       EncodeSopp(0x04, 2), EncodeSMovB32(14, 129), EncodeSopp(0x02, 1),
@@ -11136,7 +11137,7 @@ void TestMeshExportStorage() {
     options.back_code = split ? std::span{back} : std::span<const uint32_t>{};
     const auto result = RecompileForTest(
         split ? std::span{front} : std::span<const uint32_t>{monolithic},
-        options, nullptr, nullptr, push_data_start);
+        options, push_data_start);
     CheckSpirvBinaryValidates(result.spirv);
     const auto &layout = result.program.bindings;
     Check(layout.UsesPushData() == (push_data_start == PushData::MeshDrawDwordCount) &&
@@ -11284,13 +11285,11 @@ void TestMergedShaderUserDataSnapshot() {
   const auto plan = IR::ExtractResourcePlan(program);
   for (const auto *params : {&first, &second}) {
     const auto user_data = std::span(params->user_data).first(params->user_data_count);
-    const IR::SrtRuntime runtime{.user_data = user_data,
-                                 .read_memory = ReadHostTestMemory};
-    IR::SrtWalker walker(plan, runtime);
+    const IR::SrtRuntime runtime{.user_data = user_data};
     IR::DescriptorValue front_descriptor, back_descriptor;
     const auto &table = params == &first ? first_table : second_table;
-    Check(walker.EvaluateDescriptor(program.info.buffers[0].source, front_descriptor) &&
-              walker.EvaluateDescriptor(program.info.buffers[1].source, back_descriptor) &&
+    Check(EvaluateSrtDescriptor(plan, runtime, program.info.buffers[0].source, front_descriptor) &&
+              EvaluateSrtDescriptor(plan, runtime, program.info.buffers[1].source, back_descriptor) &&
               std::equal(user_data.begin() + 8, user_data.end(),
                          front_descriptor.dwords.begin()) &&
               std::equal(table.begin(), table.end(), back_descriptor.dwords.begin()),
@@ -13824,7 +13823,7 @@ void TestAutomaticObjectPrimitiveId() {
             PushData::GraphicsDrawDwords(true, false) == PushData::MeshDrawDwordCount,
         "graphics draw data reserved an incorrect prefix");
   for (const uint32_t start : {prefix, PushData::DwordCount}) {
-    const auto result = RecompileForTest(shader, options, nullptr, nullptr, start);
+    const auto result = RecompileForTest(shader, options, start);
     Check(result.program.bindings.UsesPushData() == (start == prefix),
           "object ID changed shader-data push/spill selection");
     Check(std::ranges::any_of(result.program.info.outputs, [](const StageOutput& output) {
@@ -13841,7 +13840,7 @@ void TestAutomaticObjectPrimitiveId() {
   }
 #if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
   pixel.target_shader_mask = 0xf0000000u;
-  ExpectFatal([&] { (void)RecompileForTest(shader, options, nullptr, nullptr, prefix); },
+  ExpectFatal([&] { (void)RecompileForTest(shader, options, prefix); },
               "automatic RT7 silently overlapped an explicit MRT7 export");
 #endif
 }
@@ -14489,33 +14488,6 @@ void CheckFlattenedReadSlots(const ShaderRecompiler::IR::Program &program,
   }
 }
 
-bool ReadSrtHostDword(void *, uint64_t address, std::span<uint32_t> values) {
-  if (address == 0 || values.empty()) {
-    return false;
-  }
-  std::memcpy(values.data(), reinterpret_cast<const void *>(address), values.size_bytes());
-  return true;
-}
-
-struct SrtHostRange {
-  const uint32_t *data;
-  size_t count;
-};
-
-bool ReadSrtHostRangeDword(void *userdata, uint64_t address, std::span<uint32_t> values) {
-  const auto *range = static_cast<const SrtHostRange *>(userdata);
-  if (range == nullptr || range->data == nullptr || range->count == 0 || values.empty()) {
-    return false;
-  }
-  const auto base = reinterpret_cast<uint64_t>(range->data);
-  const auto size = range->count * sizeof(uint32_t);
-  if (address < base || values.size_bytes() > size || address - base > size - values.size_bytes()) {
-    return false;
-  }
-  std::memcpy(values.data(), reinterpret_cast<const void *>(address), values.size_bytes());
-  return true;
-}
-
 void TestTypedDescriptorRealWideMoveTranslation() {
   const uint32_t shader[] = {
       EncodeSop1(0x04, 0, 20), // s_mov_b64 s[0:1], s[20:21]
@@ -14581,7 +14553,7 @@ void TestComputeImageFill() {
   options.input_info.compute = &compute;
   options.user_data = userdata;
   enum class Mutation { None, Offset, WrongAxis, VaryingValue, Predicate, ExtraStore, A16 };
-  const auto Run = [&](Mutation mutation, bool clean = true) {
+  const auto Run = [&](Mutation mutation) {
     auto translated = ShaderRecompiler::TranslateProgram(shader, options);
     auto &program = translated.program;
     for (auto *block : program.blocks) {
@@ -14618,13 +14590,11 @@ void TestComputeImageFill() {
     auto plan = ExtractResourcePlan(program);
     ResourceSnapshot snapshot;
     ResourceSpecialization specialization;
-    Check(MaterializeResources(plan,
-              {.user_data = userdata, .read_memory = ReadHostTestMemory,
-               .read_specialization_memory = clean ? ReadHostTestMemory : nullptr},
-              snapshot, specialization),
+    // The fill value is a walker load, so it needs no specialization memory reader.
+    Check(MaterializeResources(plan, {.user_data = userdata}, snapshot, specialization),
           "captured stencil clear did not materialize");
     const auto &fill = snapshot.uniform_fill;
-    const bool expected = mutation == Mutation::None && clean;
+    const bool expected = mutation == Mutation::None;
     Check((fill.kind == UniformFillKind::Image) == expected,
           "image fill proof accepted an unsafe store or missed the captured clear");
     if (expected) {
@@ -14634,7 +14604,6 @@ void TestComputeImageFill() {
     }
   };
   Run(Mutation::None);
-  Run(Mutation::None, false);
   for (const auto mutation : {Mutation::Offset, Mutation::WrongAxis, Mutation::VaryingValue,
                               Mutation::Predicate, Mutation::ExtraStore, Mutation::A16})
     Run(mutation);
@@ -14835,27 +14804,22 @@ void TestGpuProducedWritableDescriptor() {
         "native GPU-built writable descriptor acquired an unnecessary alias proof");
   struct Reads {
     std::array<uint32_t, 3> header{0x26c0u, 0u, 151u};
-    uint32_t ordinary = 0;
-    uint32_t strict = 0;
+    uint32_t count = 0;
   } reads;
-  const SrtRuntime runtime{
-      .user_data = user_data,
-      .read_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
-        if (words.size() != 1) return false;
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
         auto &reads = *static_cast<Reads *>(data);
-        const uint32_t index = address == 0x300315028ull ? 0u
-                             : address == 0x30031502cull ? 1u
-                             : address == 0x300315074ull ? 2u : 3u;
-        if (index == 3u) return false;
-        ++reads.ordinary;
+        uint32_t index = 0;
+        switch (address) {
+        case 0x300315028ull: index = 0u; break;
+        case 0x30031502cull: index = 1u; break;
+        case 0x300315074ull: index = 2u; break;
+        default: return false;
+        }
+        ++reads.count;
         words[0] = reads.header[index];
         return true;
-      },
-      .userdata = &reads,
-      .read_specialization_memory = +[](void *data, uint64_t, std::span<uint32_t>) {
-        ++static_cast<Reads *>(data)->strict;
-        return false; // Current GPU bytes are available only through the ordinary reader.
-      }};
+      }, &reads);
+  const SrtRuntime runtime{.user_data = user_data};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   const auto check = [&](uint32_t low, uint32_t high, uint32_t count, uint32_t reads_expected) {
@@ -14866,8 +14830,7 @@ void TestGpuProducedWritableDescriptor() {
               snapshot.buffers[0].dwords[1] == high &&
               snapshot.buffers[0].dwords[2] == count &&
               snapshot.buffers[0].dwords[3] == 0x16204u &&
-              snapshot.specialization_reads.empty() && reads.strict == 0 &&
-              reads.ordinary == reads_expected,
+              snapshot.specialization_reads.empty() && reads.count == reads_expected,
           "GPU-produced descriptor lost current scalar data or repeated its reads");
   };
   check(0x003176c0u, 0x00800003u, 151u, 3u);
@@ -14929,25 +14892,29 @@ void TestUniformSelectedWritableDescriptor() {
   auto plan = ExtractResourcePlan(translated.program);
   Check(plan.info.buffers.size() == 1 && plan.info.buffers[0].written,
         "uniform address selection lost its writable buffer");
-  const auto read = +[](void *data, uint64_t address, std::span<uint32_t> words) {
-    const bool alternate = *static_cast<uint32_t *>(data) == 0;
-    if (words.size() != 1) return false;
+  // An unavailable read would complete as zero, so the reader counts them.
+  struct Memory { const uint32_t *flag; uint32_t unavailable = 0; } memory{&user_data[2]};
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
+    auto &memory = *static_cast<Memory *>(data);
+    const bool alternate = *memory.flag == 0;
+    bool available = false;
     switch (address) {
-    case 0x1040: words[0] = 8u; return alternate;
-    case 0x1044: words[0] = 1u; return alternate;
-    case 0x1080: words[0] = 3u; return !alternate;
-    case 0x109c: words[0] = 2u; return !alternate;
-    case 0x1084: words[0] = 10u; return true;
-    default: return false;
+    case 0x1040: words[0] = 8u; available = alternate; break;
+    case 0x1044: words[0] = 1u; available = alternate; break;
+    case 0x1080: words[0] = 3u; available = !alternate; break;
+    case 0x109c: words[0] = 2u; available = !alternate; break;
+    case 0x1084: words[0] = 10u; available = true; break;
+    default: break;
     }
-  };
-  const SrtRuntime runtime{.user_data = user_data, .read_memory = read,
-                           .userdata = &user_data[2], .read_specialization_memory = read};
+    memory.unavailable += !available;
+    return available;
+  }, &memory);
+  const SrtRuntime runtime{.user_data = user_data};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   for (const uint32_t flag : {1u, 0u, 1u}) {
     user_data[2] = flag;
-    Check(MaterializeResources(plan, runtime, snapshot, specialization),
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) && memory.unavailable == 0,
           "uniform selection evaluated an unavailable untaken scalar source");
     const auto &descriptor = snapshot.buffers[0].dwords;
     Check(descriptor[0] == (flag ? 0xe0u : 4u) &&
@@ -14977,18 +14944,16 @@ void TestUniformSelectedDescriptorLoadAddress() {
   Check(plan.srt_reads.size() == 4u && plan.info.buffers.size() == 1u &&
             plan.info.buffers[0].written && !plan.capture_specialization_reads,
         "conditional scalar address duplicated or detached descriptor reads");
-  const auto read = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+  FakeGuestMemory fake(+[](void *data, uint64_t address, std::span<uint32_t> words) {
     const auto flag = *static_cast<uint32_t *>(data);
     const uint64_t base = flag ? 0x2000u : 0x1000u;
     const std::array<uint32_t, 4> descriptor{0x3000u + flag * 0x1000u,
                                             4u << 16u, 16u, 0x16204u};
-    if (words.size() != 1u || address < base || address >= base + 16u ||
-        (address & 3u) != 0u) return false;
+    if (address < base || address >= base + 16u || (address & 3u) != 0u) return false;
     words[0] = descriptor[(address - base) / 4u];
     return true;
-  };
-  const SrtRuntime runtime{.user_data = user_data, .read_memory = read,
-                           .userdata = &user_data[0]};
+  }, &user_data[0]);
+  const SrtRuntime runtime{.user_data = user_data};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
   for (const uint32_t flag : {0u, 1u, 0u}) {
@@ -15058,7 +15023,7 @@ void TestBoundedScalarMaterialImageKeys() {
     uint32_t stride = 160;
     uint32_t records = 409;
     uint64_t unavailable = 0;
-    bool clean = true;
+    bool readable = true;
     bool mixed = false;
     bool same_key = false;
     uint32_t key_reads = 0;
@@ -15066,7 +15031,7 @@ void TestBoundedScalarMaterialImageKeys() {
   } memory;
   const auto read = +[](void *data, uint64_t address, std::span<uint32_t> words) {
     auto &m = *static_cast<Memory *>(data);
-    if (!m.clean || (m.unavailable >= address &&
+    if (!m.readable || (m.unavailable >= address &&
         m.unavailable - address < words.size_bytes())) return false;
     if (words.size() == 1 && address >= 0x10a0u && address < 0x10c0u) {
       const uint32_t descriptors[] = {0x2000u, m.stride << 16u, m.records, 0x5204u,
@@ -15101,6 +15066,7 @@ void TestBoundedScalarMaterialImageKeys() {
     }
     return false;
   };
+  FakeGuestMemory fake(read, &memory);
   const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
                            .read_specialization_memory = read};
   ResourceSnapshot snapshot;
@@ -15138,9 +15104,11 @@ void TestBoundedScalarMaterialImageKeys() {
   Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
             memory.key_reads == 0 && memory.images == 0 && snapshot.images.size() == 2,
         "empty scalar material loop retained image probes");
-  memory = {.clean = false};
-  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
-        "dirty scalar material proof was accepted");
+  // Unreadable descriptor and count reads are zero, which leaves the loop empty.
+  memory = {.readable = false};
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            memory.key_reads == 0 && memory.images == 0 && snapshot.images.size() == 2,
+        "unreadable scalar material count did not read as zero");
   for (const uint64_t unavailable : {0x2094u, 0x211cu}) {
     memory = {.unavailable = unavailable};
     Check(!MaterializeResources(plan, runtime, snapshot, specialization),
@@ -15218,12 +15186,14 @@ void TestTypedDescriptorRealCarryAndScalarLoads() {
   std::array<uint32_t, 64> carry_user_data{};
   const uint64_t shader_base = 0x00000012fffffffbull;
   const uint64_t expected_pc = shader_base + 5u;
-  ShaderRecompiler::IR::SrtRuntime carry_runtime{carry_user_data, shader_base,
-                                                 nullptr, nullptr};
-  ShaderRecompiler::IR::DescriptorValue carry_value;
-  Check(ShaderRecompiler::IR::SrtWalker(ShaderRecompiler::IR::ExtractResourcePlan(carry_ir), carry_runtime).EvaluateDescriptor(carry_source_index, carry_value) &&
-            carry_value.dwords[0] == static_cast<uint32_t>(expected_pc) &&
-            carry_value.dwords[1] == static_cast<uint32_t>(expected_pc >> 32u),
+  const ShaderRecompiler::IR::SrtRuntime carry_runtime{carry_user_data, shader_base};
+  auto carry_plan = ShaderRecompiler::IR::ExtractResourcePlan(carry_ir);
+  const auto &carry_words = carry_plan.descriptor_sources[carry_source_index].dwords;
+  uint32_t carry_low = 0, carry_high = 0;
+  Check(EvaluateSrtValue(carry_plan, carry_runtime, carry_words[0], carry_low) &&
+            EvaluateSrtValue(carry_plan, carry_runtime, carry_words[1], carry_high) &&
+            carry_low == static_cast<uint32_t>(expected_pc) &&
+            carry_high == static_cast<uint32_t>(expected_pc >> 32u),
         "S_GETPC shader-base or add/addc carry evaluation was incorrect");
 
   const uint32_t load_shader[] = {
@@ -15285,19 +15255,21 @@ void TestTypedDescriptorRealCarryAndScalarLoads() {
   BuildTypedPlan(inline_sampler_shader,
                  static_cast<uint32_t>(std::size(inline_sampler_shader)),
                  inline_sampler_ir);
+  // The walk also reads the sampled image's user-data descriptor.
+  auto user_data = ImageTestUserData();
   ShaderRecompiler::IR::DescriptorValue sampler;
-  ShaderRecompiler::IR::SrtRuntime runtime;
+  const ShaderRecompiler::IR::SrtRuntime runtime{user_data};
   Check(inline_sampler_ir.info.samplers.size() == 1 &&
             TypedDescriptorSource(inline_sampler_ir,
                                   inline_sampler_ir.info.samplers[0].source) !=
                 nullptr &&
-            ShaderRecompiler::IR::SrtWalker(ShaderRecompiler::IR::ExtractResourcePlan(inline_sampler_ir), runtime).EvaluateDescriptor(inline_sampler_ir.info.samplers[0].source, sampler) &&
+            EvaluateSrtDescriptor(ShaderRecompiler::IR::ExtractResourcePlan(inline_sampler_ir), runtime,
+                                  inline_sampler_ir.info.samplers[0].source, sampler) &&
             sampler.dwords[0] == 0 && sampler.dwords[1] == 0x00fff000u &&
             sampler.dwords[2] == 0x09500000u && sampler.dwords[3] == 0,
         "real inline sampler construction was unresolved or evaluated "
         "incorrectly");
 
-  auto user_data = ImageTestUserData();
   auto options = MakeCompileOptions(ShaderType::Pixel);
   options.user_data = user_data;
   auto result = RecompileForTest(inline_sampler_shader, options);
@@ -15325,9 +15297,8 @@ void TestSrtWalkerRealSmemTranslation() {
   user_data[8] = static_cast<uint32_t>(address);
   user_data[9] = static_cast<uint32_t>(address >> 32u);
   std::vector<uint32_t> flat;
-  const ShaderRecompiler::IR::SrtRuntime runtime{user_data, 0, ReadSrtHostDword,
-                                                 nullptr};
-  const auto walked = ShaderRecompiler::IR::SrtWalker(ShaderRecompiler::IR::ExtractResourcePlan(ir), runtime).RefreshFlatBuffer(flat);
+  const ShaderRecompiler::IR::SrtRuntime runtime{user_data};
+  const auto walked = ShaderRecompiler::IR::WalkSrt(ShaderRecompiler::IR::ExtractResourcePlan(ir), runtime, flat, nullptr);
   Check(walked, "SRT walk failed");
   Check(flat.size() == table.size() &&
             std::equal(flat.begin(), flat.end(), table.begin()),
@@ -15353,10 +15324,9 @@ void TestSrtWalkerVccBaseTranslation() {
   const auto address = reinterpret_cast<uint64_t>(table.data());
   user_data[27] = static_cast<uint32_t>(address);
   user_data[28] = static_cast<uint32_t>(address >> 32u);
-  const ShaderRecompiler::IR::SrtRuntime runtime{user_data, 0, ReadSrtHostDword,
-                                                 nullptr};
+  const ShaderRecompiler::IR::SrtRuntime runtime{user_data};
   std::vector<uint32_t> flat;
-  Check(ShaderRecompiler::IR::SrtWalker(ShaderRecompiler::IR::ExtractResourcePlan(ir), runtime).RefreshFlatBuffer(flat), "SRT walk failed");
+  Check(ShaderRecompiler::IR::WalkSrt(ShaderRecompiler::IR::ExtractResourcePlan(ir), runtime, flat, nullptr), "SRT walk failed");
   Check(flat.size() == table.size() &&
             std::equal(flat.begin(), flat.end(), table.begin()),
         "typed SSA lost an SMEM base copied through VCC");
@@ -15382,17 +15352,16 @@ void TestSrtWalkerRealSBufferTranslation() {
   user_data[9] = static_cast<uint32_t>(address >> 32u);
   user_data[10] = sizeof(table);
   std::vector<uint32_t> flat;
-  const ShaderRecompiler::IR::SrtRuntime runtime{user_data, 0, ReadSrtHostDword,
-                                                 nullptr};
+  const ShaderRecompiler::IR::SrtRuntime runtime{user_data};
   const auto plan = ShaderRecompiler::IR::ExtractResourcePlan(ir);
-  const auto walked = ShaderRecompiler::IR::SrtWalker(plan, runtime).RefreshFlatBuffer(flat);
+  const auto walked = ShaderRecompiler::IR::WalkSrt(plan, runtime, flat, nullptr);
   Check(walked, "SRT walk failed");
   Check(flat.size() == 4 &&
             std::equal(flat.begin(), flat.end(), table.begin()),
         "real S_BUFFER_LOAD walk did not align offset components independently");
 
   user_data[10] = 3 * sizeof(uint32_t);
-  const auto bounds_walked = ShaderRecompiler::IR::SrtWalker(plan, runtime).RefreshFlatBuffer(flat);
+  const auto bounds_walked = ShaderRecompiler::IR::WalkSrt(plan, runtime, flat, nullptr);
   Check(bounds_walked && flat == std::vector<uint32_t>{table[0], table[1], table[2], 0u},
         "real S_BUFFER_LOAD walk did not zero only the out-of-bounds DWORD");
   CheckFlattenedReadSlots(
@@ -15408,7 +15377,7 @@ void TestSrtWalkerRealSBufferTranslation() {
                  static_cast<uint32_t>(std::size(negative_shader)),
                  negative_ir);
   user_data[10] = sizeof(table);
-  Check(!ShaderRecompiler::IR::SrtWalker(ShaderRecompiler::IR::ExtractResourcePlan(negative_ir), runtime).RefreshFlatBuffer(flat),
+  Check(!ShaderRecompiler::IR::WalkSrt(ShaderRecompiler::IR::ExtractResourcePlan(negative_ir), runtime, flat, nullptr),
         "real S_BUFFER_LOAD walk accepted a negative immediate");
 }
 
@@ -15443,11 +15412,9 @@ void TestScalarMemorySourcesCapturedBeforeWrites() {
     if (opcode == 0x0a) {
       user_data[base + 2u] = sizeof(table);
     }
-    SrtHostRange range{table.data(), table.size()};
-    const ShaderRecompiler::IR::SrtRuntime runtime{
-        user_data, 0, ReadSrtHostRangeDword, &range};
+    const ShaderRecompiler::IR::SrtRuntime runtime{user_data};
     std::vector<uint32_t> flat;
-    Check(ShaderRecompiler::IR::SrtWalker(ShaderRecompiler::IR::ExtractResourcePlan(ir), runtime).RefreshFlatBuffer(flat), "SRT walk failed");
+    Check(ShaderRecompiler::IR::WalkSrt(ShaderRecompiler::IR::ExtractResourcePlan(ir), runtime, flat, nullptr), "SRT walk failed");
     Check(flat.size() == table.size() &&
               std::equal(flat.begin(), flat.end(), table.begin()),
           "overlapping scalar-memory load did not capture its sources before "
@@ -15822,13 +15789,12 @@ void TestComputeLdsAllocationIdentity() {
   ShaderMappedData mapped{};
   mapped.code_size_bytes = sizeof(shader);
   ShaderMapUserData(regs.cs_regs.data_addr, mapped);
-  HW::ShaderRegisters sh{};
 
   const auto compile = [&](uint16_t encoded_lds_size,
                            uint32_t expected_dwords) {
     regs.cs_regs.lds_size = encoded_lds_size;
     ShaderComputeInputInfo input_info{};
-    const auto params = PrepareProgram(regs, sh, input_info);
+    const auto params = PrepareProgram(regs, input_info);
     Check(input_info.lds_size_dwords == expected_dwords,
           "COMPUTE_PGM_RSRC2 LDS allocation units were not decoded");
     auto options = MakeCompileOptions(ShaderType::Compute);
@@ -15894,7 +15860,7 @@ void TestComputeLdsAllocationIdentity() {
   ShaderMapUserData(regs.cs_regs.data_addr, mapped);
   regs.cs_regs.lds_size = decode_lds_field(lds_896_rsrc2);
   ShaderComputeInputInfo scratch_info{};
-  const auto scratch_params = PrepareProgram(regs, sh, scratch_info);
+  const auto scratch_params = PrepareProgram(regs, scratch_info);
   Check(scratch_info.scratch_size_dwords == 7,
         "AGC per-thread scratch size was not propagated");
   auto scratch_options = MakeCompileOptions(ShaderType::Compute);
@@ -15960,8 +15926,8 @@ void TestPixelProgramCacheBindingIdentity() {
         first_params.hash == XXH3_64bits(shader.data(), shader.size_bytes()) &&
             first_params.hash == second_params.hash &&
             first_key == MakeStageStaticKey(second_info) &&
-            first_info.stage.program != nullptr &&
-            second_info.stage.program != nullptr,
+            first_info.stage.module != nullptr &&
+            second_info.stage.module != nullptr,
         "pixel program matching did not preserve an identical source/static identity");
     return std::pair {first_params.hash, first_key};
   };
@@ -16007,7 +15973,7 @@ void TestGraphicsPushConstantPlacement() {
 
   constexpr uint32_t PushDataStart = 7;
   auto placed =
-      RecompileForTest(shader, options, nullptr, nullptr, PushDataStart);
+      RecompileForTest(shader, options, PushDataStart);
   Check(placed.program.bindings.UsesPushData() &&
             placed.program.bindings.push_data_start_dword == PushDataStart &&
             placed.program.bindings.descriptor_counts[static_cast<size_t>(BindingKind::ShaderData)] == 0,
@@ -16016,7 +15982,7 @@ void TestGraphicsPushConstantPlacement() {
         "SPIR-V push-data block did not use the canonical zero base");
   CheckSpirvBinaryValidates(placed.spirv);
 
-  auto exact = RecompileForTest(shader, options, nullptr, nullptr, 28);
+  auto exact = RecompileForTest(shader, options, 28);
   Check(exact.program.bindings.UsesPushData() &&
             exact.program.bindings.push_data_start_dword == 28 &&
             exact.program.bindings.ShaderDataDwords() == 4,
@@ -16026,8 +15992,8 @@ void TestGraphicsPushConstantPlacement() {
   Check(cursor == 32, "exact-fit shader data advanced the cursor incorrectly");
   CheckSpirvBinaryValidates(exact.spirv);
 
-  auto spill = RecompileForTest(shader, options, nullptr, nullptr, 29);
-  auto full = RecompileForTest(shader, options, nullptr, nullptr, 32);
+  auto spill = RecompileForTest(shader, options, 29);
+  auto full = RecompileForTest(shader, options, 32);
   Check(!spill.program.bindings.UsesPushData() &&
             spill.program.bindings.descriptor_counts[static_cast<size_t>(BindingKind::ShaderData)] == 1 &&
             spill.program.bindings == full.program.bindings && spill.spirv == full.spirv,