graphics: preserve logical alpha when blending reordered targets (PPSA02721)

Co-authored-by: EmK530 <admin@emk530.net>

.github/workflows/build.yml

@@ -113,14 +113,15 @@ jobs:
         shell: cmd
         run: |
           cmake --build _Build/windows --target launcher ^
-            audio_out2_port_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests ^
+            audio_out2_port_tests pad_haptics_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests ^
+            shader_recompiler_compute_tests ^
             resource_tracking_tests scalar_provenance_tests ^
-            resource_materialization_tests shader_vertex_metadata_tests --parallel
+            resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests --parallel
 
       - name: Test
         shell: cmd
         run: |
-          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|virtual_memory_allocation)$"
+          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|pad_haptics|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$"
 
       - name: Install
         shell: cmd
@@ -201,16 +202,17 @@ jobs:
         shell: bash
         run: |
           cmake --build _Build/macos \
-            --target launcher audio_out2_port_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+            --target launcher audio_out2_port_tests pad_haptics_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+              memory_tracker_tests page_manager_tests shader_recompiler_compute_tests \
               resource_tracking_tests scalar_provenance_tests \
-              resource_materialization_tests shader_vertex_metadata_tests \
+              resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests \
             --parallel
 
       - name: Test
         shell: bash
         run: |
           ctest --test-dir _Build/macos --output-on-failure \
-            -R '^(audio_out2_port|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|pad_haptics|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash
@@ -384,16 +386,16 @@ jobs:
         run: |
           cmake --build _Build/linux \
             --target launcher page_manager_tests memory_tracker_tests \
-              audio_out2_port_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+              audio_out2_port_tests pad_haptics_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
               resource_tracking_tests scalar_provenance_tests \
-              resource_materialization_tests shader_vertex_metadata_tests \
+              resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests \
             --parallel
 
       - name: Test
         shell: bash
         run: |
           ctest --test-dir _Build/linux --output-on-failure \
-            -R '^(audio_out2_port|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|pad_haptics|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash

3rdparty/CMakeLists.txt

@@ -28,7 +28,54 @@ FetchContent_Declare(
 
 FetchContent_MakeAvailable(xbyak zydis)
 
+# ZArchive provides random-access reads for compressed game dumps. Build only
+# the reader/writer library and its zstd dependency; the command-line tool is
+# not part of KytyPS5's runtime distribution.
+set(ZSTD_BUILD_PROGRAMS OFF CACHE BOOL "" FORCE)
+set(ZSTD_BUILD_SHARED OFF CACHE BOOL "" FORCE)
+set(ZSTD_BUILD_STATIC ON CACHE BOOL "" FORCE)
+set(ZSTD_BUILD_TESTS OFF CACHE BOOL "" FORCE)
+set(ZSTD_LEGACY_SUPPORT OFF CACHE BOOL "" FORCE)
+set(ZSTD_MULTITHREAD_SUPPORT OFF CACHE BOOL "" FORCE)
+FetchContent_Declare(
+	zstd
+	GIT_REPOSITORY https://github.com/facebook/zstd.git
+	GIT_TAG f8745da6ff1ad1e7bab384bd1f9d742439278e99
+	GIT_SHALLOW FALSE
+	SOURCE_SUBDIR build/cmake
+)
+FetchContent_Declare(
+	zarchive_source
+	GIT_REPOSITORY https://github.com/Exzap/ZArchive.git
+	GIT_TAG d2c717730092c7bf8cbb033b12fd4001b7c4d932
+	GIT_SHALLOW FALSE
+	PATCH_COMMAND "${GIT_EXECUTABLE}" apply --ignore-whitespace
+		"${CMAKE_CURRENT_SOURCE_DIR}/patches/zarchive-reader.patch"
+	SOURCE_SUBDIR kyty-library-only
+)
+FetchContent_MakeAvailable(zstd zarchive_source)
+
+add_library(zarchive STATIC
+	"${zarchive_source_SOURCE_DIR}/src/zarchivereader.cpp"
+	"${zarchive_source_SOURCE_DIR}/src/zarchivewriter.cpp"
+	"${zarchive_source_SOURCE_DIR}/src/sha_256.c"
+)
+add_library(ZArchive::zarchive ALIAS zarchive)
+target_include_directories(zarchive SYSTEM PUBLIC "${zarchive_source_SOURCE_DIR}/include")
+target_include_directories(zarchive PRIVATE "${zstd_SOURCE_DIR}/lib")
+target_link_libraries(zarchive PRIVATE libzstd_static)
+target_compile_features(zarchive PRIVATE cxx_std_20)
+if(MSVC)
+	target_compile_options(zarchive PRIVATE /W0)
+else()
+	target_compile_options(zarchive PRIVATE -w)
+endif()
+
 set(SDL_SHARED OFF CACHE BOOL "" FORCE)
+if(APPLE)
+	# SDL's built-in iconv fallback does not support CP932.
+	set(SDL_SYSTEM_ICONV ON CACHE BOOL "Use system iconv for CP932 conversion" FORCE)
+endif()
 add_subdirectory(SDL3)
 
 add_library(Vulkan-Headers INTERFACE)

3rdparty/patches/zarchive-reader.patch

@@ -0,0 +1,100 @@
+--- a/include/zarchive/zarchivecommon.h
++++ b/include/zarchive/zarchivecommon.h
+@@ -256,7 +256,7 @@
+ 
+ 			bool IsWithinValidRange(uint64_t fileSize) const
+ 			{
+-				return (offset + size) <= fileSize;
++				return offset <= fileSize && size <= fileSize - offset;
+ 			}
+ 		};
+ 
+--- a/src/zarchivereader.cpp
++++ b/src/zarchivereader.cpp
+@@ -14,6 +14,7 @@
+ 
+ static bool _ifstream_readBytes(std::ifstream& file, uint64_t offset, void* buffer, uint32_t size)
+ {
++	file.clear();
+ 	file.seekg(offset, std::ios_base::beg);
+ 	file.read((char*)buffer, size);
+ 	return file.gcount() == size;
+@@ -80,9 +81,27 @@
+ 	// verify file tree
+ 	if (fileTree[0].IsFile())
+ 		return nullptr; // first entry must be root directory
+-	auto rootName = GetName(nameTable, fileTree[0].GetNameOffset());
+-	if (!rootName.empty())
++	if (fileTree[0].GetNameOffset() != 0x7FFFFFFF)
+ 		return nullptr; // root node must not have a name
++	const uint64_t dataCapacity = (uint64_t)offsetRecords.size() * _ZARCHIVE::ENTRIES_PER_OFFSETRECORD * _ZARCHIVE::COMPRESSED_BLOCK_SIZE;
++	for (size_t i = 0; i < fileTree.size(); ++i)
++	{
++		auto& entry = fileTree[i];
++		if (i != 0)
++		{
++			auto name = GetName(nameTable, entry.GetNameOffset());
++			if (name.empty() || name == "." || name == ".." || name.find_first_of("/\\") != std::string_view::npos || name.find('\0') != std::string_view::npos)
++				return nullptr;
++		}
++		if (entry.IsFile())
++		{
++			if (entry.GetFileOffset() > dataCapacity || entry.GetFileSize() > dataCapacity - entry.GetFileOffset())
++				return nullptr;
++		}
++		else if (entry.directoryRecord.nodeStartIndex > fileTree.size() ||
++			entry.directoryRecord.count > fileTree.size() - entry.directoryRecord.nodeStartIndex)
++			return nullptr;
++	}
+ 	// read meta data
+ 	// todo
+ 
+@@ -238,7 +257,7 @@
+ 		uint32_t stepSize = std::min<uint32_t>(remainingBytes, _ZARCHIVE::COMPRESSED_BLOCK_SIZE - blockOffset);
+ 		CacheBlock* block = GetCachedBlock(blockIdx);
+ 		if (!block)
+-			return 0;
++			return bytesToRead - remainingBytes;
+ 		std::memcpy(bufferU8, block->data + blockOffset, stepSize);
+ 		rawReadOffset += stepSize;
+ 		remainingBytes -= stepSize;
+@@ -325,14 +344,18 @@
+ 	// determine offset and size of compressed block
+ 	auto& record = m_offsetRecords[recordIndex];
+ 	uint64_t offset = record.baseOffset;
++	if (offset > m_compressedDataSize)
++		return false;
+ 	for (uint32_t i = 0; i < recordSubIndex; i++)
+ 	{
+-		offset += (uint64_t)record.size[i];
+-		offset++;
++		const uint64_t size = (uint64_t)record.size[i] + 1;
++		if (size > m_compressedDataSize - offset)
++			return false;
++		offset += size;
+ 	}
+ 	uint32_t compressedSize = (uint32_t)record.size[recordSubIndex] + 1;
+ 	// load file data
+-	if ((offset + compressedSize) > m_compressedDataSize)
++	if (compressedSize > m_compressedDataSize - offset)
+ 		return false;
+ 	offset += m_compressedDataOffset;
+ 	if (compressedSize == _ZARCHIVE::COMPRESSED_BLOCK_SIZE)
+@@ -350,7 +373,7 @@
+ // returns empty view on failure
+ std::string_view ZArchiveReader::GetName(const std::vector<uint8_t>& nameTable, uint32_t nameOffset)
+ {
+-	if (nameOffset == 0x7FFFFFFF || nameOffset > nameTable.size())
++	if (nameOffset == 0x7FFFFFFF || nameOffset >= nameTable.size())
+ 		return "";
+ 	// parse header
+ 	uint16_t nameLength = nameTable[nameOffset] & 0x7F;
+@@ -359,7 +382,7 @@
+ 		// extended 2-byte length
+ 		if (nameOffset + 1 >= nameTable.size())
+ 			return "";
+-		nameLength |= ((uint16_t)nameTable[nameOffset] << 7);
++		nameLength |= ((uint16_t)nameTable[nameOffset + 1] << 7);
+ 		nameOffset += 2;
+ 	}
+ 	else

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.16...4.3)
+cmake_minimum_required(VERSION 3.18...4.3)
 
 if(CMAKE_VERSION VERSION_GREATER_EQUAL 4.0)
 	set(CMAKE_POLICY_VERSION_MINIMUM 3.5 CACHE STRING "Minimum policy version for bundled dependencies")
@@ -407,6 +407,16 @@ add_executable(page_manager_tests EXCLUDE_FROM_ALL
 target_link_libraries(page_manager_tests common)
 target_include_directories(page_manager_tests PRIVATE ${inc_headers})
 
+add_executable(archive_file_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/ArchiveFileTests.cpp"
+)
+target_link_libraries(archive_file_tests common ZArchive::zarchive)
+target_include_directories(archive_file_tests PRIVATE ${inc_headers})
+
+add_executable(avplayer_file_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/AvPlayerFileTests.cpp")
+target_link_libraries(avplayer_file_tests common FFmpeg::ffmpeg ZArchive::zarchive)
+target_include_directories(avplayer_file_tests PRIVATE ${inc_headers})
+
 add_executable(bit_array_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/BitArrayTests.cpp"
 )
@@ -490,6 +500,17 @@ add_executable(save_data_memory_tests EXCLUDE_FROM_ALL
 target_link_libraries(save_data_memory_tests common fmt::fmt)
 target_include_directories(save_data_memory_tests PRIVATE ${inc_headers})
 
+add_executable(ces_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/CesTests.cpp")
+target_link_libraries(ces_tests common fmt::fmt SDL3::SDL3-static)
+target_include_directories(ces_tests PRIVATE ${inc_headers})
+
+add_executable(save_data_dialog_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/SaveDataDialogTests.cpp"
+	"${KYTY_SOURCE_DIR}/libs/dialog.cpp"
+)
+target_link_libraries(save_data_dialog_tests common fmt::fmt)
+target_include_directories(save_data_dialog_tests PRIVATE ${inc_headers})
+
 add_kyty_full_emulator_test(http_uri_parse_tests
 	"${KYTY_TESTS_DIR}/HttpUriParseTests.cpp")
 
@@ -503,6 +524,7 @@ target_include_directories(event_queue_lifetime_tests PRIVATE ${inc_headers})
 
 add_kyty_full_emulator_test(kernel_file_system_tests
 	"${KYTY_TESTS_DIR}/KernelFileSystemTests.cpp")
+target_link_libraries(kernel_file_system_tests ZArchive::zarchive)
 
 add_executable(sync_on_address_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/SyncOnAddressTests.cpp"
@@ -569,6 +591,8 @@ if(BUILD_TESTING)
 	add_test(NAME image_page_table COMMAND $<TARGET_FILE:image_page_table_tests>)
 	add_test(NAME memory_tracker COMMAND $<TARGET_FILE:memory_tracker_tests>)
 	add_test(NAME page_manager COMMAND $<TARGET_FILE:page_manager_tests>)
+	add_test(NAME archive_file COMMAND $<TARGET_FILE:archive_file_tests>)
+	add_test(NAME avplayer_file COMMAND $<TARGET_FILE:avplayer_file_tests>)
 	add_test(NAME bit_array COMMAND $<TARGET_FILE:bit_array_tests>)
 	add_test(NAME lru_cache COMMAND $<TARGET_FILE:lru_cache_tests>)
 	add_test(NAME shader_vertex_metadata COMMAND $<TARGET_FILE:shader_vertex_metadata_tests>)
@@ -581,6 +605,8 @@ if(BUILD_TESTING)
 	add_test(NAME ngs2_sampler COMMAND $<TARGET_FILE:ngs2_sampler_tests>)
 	add_test(NAME pad_haptics COMMAND $<TARGET_FILE:pad_haptics_tests>)
 	add_test(NAME save_data_memory COMMAND $<TARGET_FILE:save_data_memory_tests>)
+	add_test(NAME ces COMMAND $<TARGET_FILE:ces_tests>)
+	add_test(NAME save_data_dialog COMMAND $<TARGET_FILE:save_data_dialog_tests>)
 	add_test(NAME http_uri_parse COMMAND $<TARGET_FILE:http_uri_parse_tests>)
 	add_test(NAME shader_recompiler_compute COMMAND $<TARGET_FILE:shader_recompiler_compute_tests>)
 	add_test(NAME shader_recompiler_alignbyte
@@ -618,6 +644,8 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --compute-meta-clear-only)
 	add_test(NAME buffer_cache_dirty_gc
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-gc-only)
+	add_test(NAME color_render_target_1d
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --rt1d-only)
 
 	if(WIN32)
 		# These tests still depend on the Windows multisample-depth path.
@@ -636,6 +664,8 @@ if(BUILD_TESTING)
 		image_page_table_tests
 		memory_tracker_tests
 		page_manager_tests
+		archive_file_tests
+		avplayer_file_tests
 		bit_array_tests
 		lru_cache_tests
 		shader_vertex_metadata_tests
@@ -650,6 +680,8 @@ if(BUILD_TESTING)
 		ngs2_sampler_tests
 		pad_haptics_tests
 		save_data_memory_tests
+		ces_tests
+		save_data_dialog_tests
 		http_uri_parse_tests
 	)
 endif()

README.md

@@ -7,6 +7,10 @@
 [![Status](https://img.shields.io/badge/status-active%20development-orange.svg)](#current-status)
 [![License](https://img.shields.io/badge/license-GPL--2.0-blue.svg)](LICENSE)
 
+**[Weekly updates](https://github.com/KytyPS5/KytyPS5/discussions/862)** — game progress, recent fixes and ongoing development.
+
+**[Development on Discord](https://discord.gg/UNrkMqGaBg)** — KytyPS5 development.
+
 KytyPS5 is a free and open-source PlayStation 5 emulator written in C++ for Windows and Linux,
 with experimental macOS support. It is based on a heavily modified version of
 [Kyty](https://github.com/InoriRus/Kyty). The project is in active development, and behavior
@@ -68,8 +72,8 @@ or graphical glitches, so please include the version you tested when reporting a
       <img src="docs/screenshots/ps5-02.png" width="300" alt="Demon's Souls running in KytyPS5">
     </td>
     <td align="center">
-      <strong>Hellboy</strong><br>
-      <img src="docs/screenshots/ps5-06.png" width="300" alt="Hellboy running in KytyPS5">
+      <strong>UFC 5</strong><br>
+      <img src="docs/screenshots/ps5-06.png" width="300" alt="UFC 5 running in KytyPS5">
     </td>
   </tr>
 </table>
@@ -232,9 +236,10 @@ cmake --build _Build/linux --target launcher --parallel
 cmake --install _Build/linux --prefix _Build/linux/install
 ```
 
-The configure step downloads the FFmpeg prebuilts and the `xbyak`/`zydis` sources, so it needs
-network access; a fully sandboxed `nix build` would require vendoring those inputs. A Vulkan 1.3
-driver must be available at runtime (on NixOS, `hardware.graphics.enable = true`).
+The configure step downloads the FFmpeg prebuilts and the `xbyak`, `zydis`, `zstd`, and ZArchive
+sources, so it needs network access; a fully sandboxed `nix build` would require vendoring those
+inputs. A Vulkan 1.3 driver must be available at runtime (on NixOS,
+`hardware.graphics.enable = true`).
 
 ### Building on macOS
 
@@ -328,17 +333,21 @@ open _Build/macos/install/KytyPS5.app  # or double-click in Finder
 ```
 
 On first launch, add one or more game folders in the global settings. The launcher searches those
-folders recursively for game directories containing `eboot.bin`. Select a detected game and run it
-from the game list.
+folders recursively for game directories containing `eboot.bin` and ZArchive (`.zar`) game dumps
+whose archive root contains `eboot.bin`. Select a detected game and run it from the game list.
+ZArchive dumps are mounted read-only and streamed directly; they do not need to be extracted first.
 
-The emulator can also be started directly with a legally obtained game directory or ELF file:
+The emulator can also be started directly with a legally obtained game directory, ELF file, or
+ZArchive dump:
 
 ```powershell
 .\_Build\windows\install\kyty_emulator.exe --game "D:\Games\ExampleGame"
+.\_Build\windows\install\kyty_emulator.exe --game "D:\Games\ExampleGame.zar"
 ```
 
 ```bash
 ./_Build/linux/install/kyty_emulator --game "/games/ExampleGame"
+./_Build/linux/install/kyty_emulator --game "/games/ExampleGame.zar"
 ```
 
 On macOS, the adjacent flat or app-bundled `libMoltenVK.dylib` is found automatically; no

src/common/CMakeLists.txt

@@ -26,6 +26,7 @@ target_link_libraries(common
 		cpuinfo
 		kyty_git_version
 		spdlog::spdlog
+		ZArchive::zarchive
 )
 
 get_property(inc_headers TARGET common PROPERTY INCLUDE_DIRECTORIES)

src/common/archive.cpp

@@ -0,0 +1,397 @@
+#include "common/archive.h"
+
+#include "common/stringUtils.h"
+
+#include <algorithm>
+#include <condition_variable>
+#include <cstring>
+#include <mutex>
+#include <optional>
+#include <string_view>
+#include <thread>
+#include <unordered_map>
+#include <utility>
+#include <zarchive/zarchivereader.h>
+
+// POSIX hosts and MinGW's pthread-backed libstdc++ have cancellable condition waits.
+// MSVC/ClangCL and MinGW's native Win32 threading backend do not use pthread waits.
+#if !defined(_WIN32) || defined(_GLIBCXX_GCC_GTHR_POSIX_H)
+#include <pthread.h>
+#define KYTY_ARCHIVE_PTHREAD_CANCELLATION
+#endif
+
+namespace Common {
+
+// Backends expose immutable metadata and offset-based reads. Cursor management, path routing,
+// reader lifetime and the host-stack worker are shared by every archive format.
+class ArchiveReader {
+public:
+	struct Entry {
+		uint64_t id;
+		uint64_t size;
+		bool     is_file;
+	};
+
+	virtual ~ArchiveReader()                                                       = default;
+	virtual std::optional<Entry>        Find(std::string_view member)              = 0;
+	virtual std::vector<File::DirEntry> List(std::string_view member)              = 0;
+	virtual uint64_t Read(uint64_t id, uint64_t offset, uint32_t size, void* data) = 0;
+};
+
+namespace {
+
+class ZArchiveReaderBackend final: public ArchiveReader {
+public:
+	explicit ZArchiveReaderBackend(std::unique_ptr<ZArchiveReader> reader)
+	    : m_reader(std::move(reader)) {}
+
+	std::optional<Entry> Find(std::string_view member) override {
+		const auto node = m_reader->LookUp(member);
+		if (node == ZARCHIVE_INVALID_NODE) {
+			return {};
+		}
+		return Entry {node, m_reader->GetFileSize(node), m_reader->IsFile(node)};
+	}
+
+	std::vector<File::DirEntry> List(std::string_view member) override {
+		const auto node = m_reader->LookUp(member);
+		if (!m_reader->IsDirectory(node)) {
+			return {};
+		}
+		const auto count = m_reader->GetDirEntryCount(node);
+		if (count > (1u << 20u)) {
+			return {};
+		}
+		std::vector<File::DirEntry> result;
+		result.reserve(count);
+		for (uint32_t index = 0; index < count; ++index) {
+			ZArchiveReader::DirEntry entry {};
+			if (m_reader->GetDirEntry(node, index, entry)) {
+				result.push_back({std::string(entry.name), entry.isFile});
+			}
+		}
+		return result;
+	}
+
+	uint64_t Read(uint64_t id, uint64_t offset, uint32_t size, void* data) override {
+		// ZArchiveReader already serializes access to its stream and decompression cache.
+		return m_reader->ReadFromFile(static_cast<ZArchiveNodeHandle>(id), offset, size, data);
+	}
+
+private:
+	std::unique_ptr<ZArchiveReader> m_reader;
+};
+
+std::shared_ptr<ArchiveReader> OpenZArchive(const std::filesystem::path& path) {
+	auto reader = std::unique_ptr<ZArchiveReader>(ZArchiveReader::OpenFromFile(path));
+	return reader ? std::make_shared<ZArchiveReaderBackend>(std::move(reader)) : nullptr;
+}
+
+struct ArchiveFormat {
+	std::string_view extension;
+	std::shared_ptr<ArchiveReader> (*open)(const std::filesystem::path&);
+};
+
+constexpr ArchiveFormat Formats[] = {{".zar", OpenZArchive}};
+using NativeView                  = std::basic_string_view<std::filesystem::path::value_type>;
+
+const ArchiveFormat* FindFormat(NativeView path) {
+	for (const auto& format: Formats) {
+		if (path.size() < format.extension.size()) {
+			continue;
+		}
+		const auto extension = path.substr(path.size() - format.extension.size());
+		if (std::equal(
+		        extension.begin(), extension.end(), format.extension.begin(),
+		        [](auto a, char b) { return (a >= 'A' && a <= 'Z' ? a + ('a' - 'A') : a) == b; })) {
+			return &format;
+		}
+	}
+	return nullptr;
+}
+
+size_t FindMarker(NativeView path) {
+	// Native files use this check on every operation. The usual path has no marker and allocates
+	// nothing.
+	for (auto marker = path.find('!'); marker != NativeView::npos;
+	     marker      = path.find('!', marker + 1)) {
+		if ((marker + 1 == path.size() || path[marker + 1] == '/' || path[marker + 1] == '\\') &&
+		    FindFormat(path.substr(0, marker)) != nullptr) {
+			return marker;
+		}
+	}
+	return NativeView::npos;
+}
+
+struct ParsedPath {
+	std::filesystem::path archive;
+	std::string           member;
+};
+
+std::optional<ParsedPath> ParsePath(const std::filesystem::path& path) {
+	const auto& native = path.native();
+	const auto  marker = FindMarker(native);
+	if (marker == NativeView::npos) {
+		return {};
+	}
+	ParsedPath parsed {.archive = std::filesystem::path(native.substr(0, marker))};
+	auto       member = PathToGenericString(std::filesystem::path(native.substr(marker + 1)));
+	std::replace(member.begin(), member.end(), '\\', '/');
+	std::string_view remaining = member;
+	while (!remaining.empty()) {
+		const auto separator = remaining.find('/');
+		const auto component = remaining.substr(0, separator);
+		remaining.remove_prefix(separator == std::string_view::npos ? remaining.size() : separator + 1);
+		if (component.empty() || component == ".") {
+			continue;
+		}
+		if (component == "..") {
+			if (parsed.member.empty()) {
+				return {};
+			}
+			const auto separator = parsed.member.rfind('/');
+			parsed.member.resize(separator == std::string::npos ? 0 : separator);
+		} else {
+			if (!parsed.member.empty()) {
+				parsed.member += '/';
+			}
+			parsed.member += component;
+		}
+	}
+	return parsed;
+}
+
+std::mutex                                                              g_readers_mutex;
+std::unordered_map<std::filesystem::path, std::weak_ptr<ArchiveReader>> g_readers;
+
+// Guest stacks can be only 64 KiB. Decompress on an ordinary host stack, then copy into guest
+// memory on the caller so GPU-tracked writes use its fault handler. A single worker bounds host
+// threads and avoids duplicating the archive's own block cache. Requests live on waiting callers.
+class ArchiveIoThread {
+public:
+	static ArchiveIoThread& Instance() {
+		// Guest exit can run static destructors while other guest threads are still reading.
+		// Keep the worker alive until process termination, as with the guest threads themselves.
+		static auto* instance = new ArchiveIoThread;
+		return *instance;
+	}
+
+	uint64_t Read(ArchiveReader& reader, uint64_t id, uint64_t offset, uint32_t size, void* data) {
+		Request          request {reader, id, offset, size, data};
+		std::unique_lock lock(m_mutex);
+		if (m_last != nullptr) {
+			m_last->next = &request;
+		} else {
+			m_first = &request;
+		}
+		m_last = &request;
+		m_wake.notify_one();
+		request.ready.wait(lock, [&] { return request.done; });
+		return request.result;
+	}
+
+	KYTY_CLASS_NO_COPY(ArchiveIoThread);
+
+private:
+	struct Request {
+		ArchiveReader&          reader;
+		uint64_t                id;
+		uint64_t                offset;
+		uint32_t                size;
+		void*                   data;
+		Request*                next   = nullptr;
+		uint64_t                result = 0;
+		bool                    done   = false;
+		std::condition_variable ready {};
+	};
+
+	ArchiveIoThread() {
+		std::thread([this] { Loop(); }).detach();
+	}
+
+	void Loop() {
+		for (;;) {
+			Request* request;
+			{
+				std::unique_lock lock(m_mutex);
+				m_wake.wait(lock, [this] { return m_first != nullptr; });
+				request = m_first;
+				m_first = request->next;
+				if (m_first == nullptr) {
+					m_last = nullptr;
+				}
+			}
+			const auto result =
+			    request->reader.Read(request->id, request->offset, request->size, request->data);
+			{
+				std::lock_guard lock(m_mutex);
+				request->result = result;
+				request->done   = true;
+				// Notify before releasing the mutex: the waiting caller owns the condition
+				// variable.
+				request->ready.notify_one();
+			}
+		}
+	}
+
+	std::mutex              m_mutex;
+	std::condition_variable m_wake;
+	Request*                m_first = nullptr;
+	Request*                m_last  = nullptr;
+};
+
+} // namespace
+
+struct ArchiveFile::Private {
+	std::shared_ptr<ArchiveReader> reader;
+	ArchiveReader::Entry           entry;
+	uint64_t                       position = 0;
+};
+
+ArchiveFile::ArchiveFile(std::unique_ptr<Private> p): m_p(std::move(p)) {}
+ArchiveFile::~ArchiveFile() = default;
+
+uint64_t ArchiveFile::Size() const {
+	return m_p->entry.size;
+}
+uint64_t ArchiveFile::Tell() const {
+	return m_p->position;
+}
+bool ArchiveFile::Seek(uint64_t offset) {
+	m_p->position = offset;
+	return true;
+}
+
+void ArchiveFile::Read(void* data, uint32_t size, uint32_t* bytes_read) {
+#if defined(KYTY_ARCHIVE_PTHREAD_CANCELLATION)
+	// A pending cancellation must not destroy the waiting request or its TLS buffer while the
+	// worker still uses them. Restore the caller's state only after the copy is complete.
+	int cancellation_state = PTHREAD_CANCEL_ENABLE;
+	pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cancellation_state);
+#endif
+	constexpr uint32_t                chunk_size = 1u << 20u;
+	thread_local std::vector<uint8_t> chunk;
+	uint32_t                          read = 0;
+	if (data != nullptr && size != 0 && m_p->position < Size()) {
+		const auto requested =
+		    static_cast<uint32_t>(std::min<uint64_t>(size, Size() - m_p->position));
+		if (chunk.size() < std::min(chunk_size, requested)) {
+			chunk.resize(std::min(chunk_size, requested));
+		}
+		while (read < requested) {
+			const auto step = std::min(chunk_size, requested - read);
+			const auto got  = ArchiveIoThread::Instance().Read(*m_p->reader, m_p->entry.id,
+			                                                   m_p->position, step, chunk.data());
+			if (got == 0 || got > step) {
+				break;
+			}
+			std::memcpy(static_cast<uint8_t*>(data) + read, chunk.data(), got);
+			read += static_cast<uint32_t>(got);
+			m_p->position += got;
+			if (got < step) {
+				break;
+			}
+		}
+	}
+	if (bytes_read != nullptr) {
+		*bytes_read = read;
+	}
+#if defined(KYTY_ARCHIVE_PTHREAD_CANCELLATION)
+	pthread_setcancelstate(cancellation_state, nullptr);
+#endif
+}
+
+std::filesystem::path MakeArchivePath(const std::filesystem::path& archive,
+                                      const std::filesystem::path& member) {
+	auto root = archive;
+	root += "!";
+	return member.empty() ? root : root / member.relative_path();
+}
+
+bool IsSupportedArchive(const std::filesystem::path& path) {
+	return FindFormat(path.native()) != nullptr;
+}
+
+bool IsArchivePath(const std::filesystem::path& path) {
+	return FindMarker(path.native()) != NativeView::npos;
+}
+
+std::filesystem::path GetArchiveHostPath(const std::filesystem::path& path) {
+	const auto marker = FindMarker(path.native());
+	return marker == NativeView::npos ? std::filesystem::path {}
+	                                  : std::filesystem::path(path.native().substr(0, marker));
+}
+
+std::shared_ptr<ArchiveReader> OpenArchive(const std::filesystem::path& path) {
+	const auto  archive = IsArchivePath(path) ? GetArchiveHostPath(path) : path;
+	const auto* format  = FindFormat(archive.native());
+	if (format == nullptr) {
+		return {};
+	}
+	std::error_code error;
+	auto            key = std::filesystem::absolute(archive, error);
+	if (error) {
+		return {};
+	}
+	{
+		std::lock_guard lock(g_readers_mutex);
+		if (auto it = g_readers.find(key); it != g_readers.end()) {
+			if (auto reader = it->second.lock()) {
+				return reader;
+			}
+		}
+	}
+	// Do not hold the cache mutex while opening an unrelated archive from disk.
+	auto reader = format->open(archive);
+	if (reader == nullptr) {
+		return {};
+	}
+	std::lock_guard lock(g_readers_mutex);
+	std::erase_if(g_readers, [](const auto& entry) { return entry.second.expired(); });
+	auto& cached = g_readers[key];
+	if (auto existing = cached.lock()) {
+		return existing;
+	}
+	cached = reader;
+	return reader;
+}
+
+std::optional<File::Info> GetArchiveInfo(const std::filesystem::path& path) {
+	const auto parsed = ParsePath(path);
+	if (parsed) {
+		if (auto reader = OpenArchive(parsed->archive)) {
+			if (auto entry = reader->Find(parsed->member)) {
+				return File::Info {entry->is_file, entry->size};
+			}
+		}
+	}
+	return {};
+}
+
+std::vector<File::DirEntry> GetArchiveDirEntries(const std::filesystem::path& path) {
+	const auto parsed = ParsePath(path);
+	if (parsed) {
+		if (auto reader = OpenArchive(parsed->archive)) {
+			return reader->List(parsed->member);
+		}
+	}
+	return {};
+}
+
+std::unique_ptr<ArchiveFile> OpenArchiveFile(const std::filesystem::path& path) {
+	auto parsed = ParsePath(path);
+	if (!parsed) {
+		return {};
+	}
+	auto       reader = OpenArchive(parsed->archive);
+	const auto entry  = reader ? reader->Find(parsed->member) : std::nullopt;
+	if (!entry || !entry->is_file) {
+		return {};
+	}
+	auto p     = std::make_unique<ArchiveFile::Private>();
+	p->reader  = std::move(reader);
+	p->entry   = *entry;
+	return std::unique_ptr<ArchiveFile>(new ArchiveFile(std::move(p)));
+}
+
+} // namespace Common

src/common/archive.h

@@ -0,0 +1,50 @@
+#ifndef KYTY_COMMON_ARCHIVE_H_
+#define KYTY_COMMON_ARCHIVE_H_
+
+#include "common/file.h"
+
+#include <cstdint>
+#include <filesystem>
+#include <memory>
+#include <vector>
+
+namespace Common {
+
+class ArchiveReader;
+
+class ArchiveFile {
+public:
+	~ArchiveFile();
+
+	[[nodiscard]] uint64_t Size() const;
+	[[nodiscard]] uint64_t Tell() const;
+	bool                   Seek(uint64_t offset);
+	void                   Read(void* data, uint32_t size, uint32_t* bytes_read);
+
+	KYTY_CLASS_NO_COPY(ArchiveFile);
+
+private:
+	struct Private;
+	explicit ArchiveFile(std::unique_ptr<Private> p);
+	std::unique_ptr<Private> m_p;
+
+	friend std::unique_ptr<ArchiveFile> OpenArchiveFile(const std::filesystem::path& path);
+};
+
+// Virtual paths use "archive.ext!/member" and retain native host-path encoding.
+[[nodiscard]] std::filesystem::path       MakeArchivePath(const std::filesystem::path& archive,
+                                                          const std::filesystem::path& member = {});
+[[nodiscard]] bool                        IsSupportedArchive(const std::filesystem::path& path);
+[[nodiscard]] bool                        IsArchivePath(const std::filesystem::path& path);
+[[nodiscard]] std::optional<File::Info>   GetArchiveInfo(const std::filesystem::path& path);
+[[nodiscard]] std::vector<File::DirEntry> GetArchiveDirEntries(const std::filesystem::path& path);
+[[nodiscard]] std::filesystem::path       GetArchiveHostPath(const std::filesystem::path& path);
+[[nodiscard]] std::unique_ptr<ArchiveFile> OpenArchiveFile(const std::filesystem::path& path);
+
+// Retain this handle while a mount or metadata scan is active to reuse the index and block cache.
+// Accepts either a host archive filename or a virtual path inside the archive.
+[[nodiscard]] std::shared_ptr<ArchiveReader> OpenArchive(const std::filesystem::path& path);
+
+} // namespace Common
+
+#endif /* KYTY_COMMON_ARCHIVE_H_ */

src/common/file.cpp

@@ -1,5 +1,6 @@
 #include "common/file.h"
 
+#include "common/archive.h"
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/platform/sysFileIO.h"
@@ -61,7 +62,8 @@ SysFileTimeStruct DateTimeToFileTimeUtc(const DateTime& date_time) {
 } // namespace
 
 struct File::FilePrivate {
-	sys_file_t* f;
+	sys_file_t*                  f;
+	std::unique_ptr<ArchiveFile> archive;
 };
 
 File::File(): m_p(std::make_unique<FilePrivate>()) {
@@ -85,11 +87,15 @@ File::File(const std::filesystem::path& name, Mode mode): m_p(std::make_unique<F
 }
 
 bool File::IsInvalid() const {
-	return m_p->f == nullptr;
+	return m_p->f == nullptr && m_p->archive == nullptr;
 }
 
 bool File::Create(const std::filesystem::path& name) {
-	EXIT_IF(m_p->f != nullptr);
+	EXIT_IF(!IsInvalid());
+
+	if (IsArchivePath(name)) {
+		return false;
+	}
 
 	m_file_name = name;
 
@@ -104,9 +110,16 @@ bool File::Create(const std::filesystem::path& name) {
 }
 
 bool File::Open(const std::filesystem::path& name, Mode mode) {
-	EXIT_IF(m_p->f != nullptr);
+	EXIT_IF(!IsInvalid());
 
 	m_file_name = name;
+	if (IsArchivePath(name)) {
+		if (mode != Mode::Read) {
+			return false;
+		}
+		m_p->archive = OpenArchiveFile(name);
+		return m_p->archive != nullptr;
+	}
 
 	switch (mode) {
 		case Mode::Read: m_p->f = SysFileOpenR(name); break;
@@ -124,7 +137,7 @@ bool File::Open(const std::filesystem::path& name, Mode mode) {
 }
 
 bool File::OpenInMem(void* buf, uint32_t buf_size) {
-	EXIT_IF(m_p->f != nullptr);
+	EXIT_IF(!IsInvalid());
 
 	m_p->f = SysFileOpen(static_cast<uint8_t*>(buf), buf_size);
 
@@ -137,7 +150,7 @@ bool File::OpenInMem(void* buf, uint32_t buf_size) {
 }
 
 bool File::CreateInMem() {
-	EXIT_IF(m_p->f != nullptr);
+	EXIT_IF(!IsInvalid());
 
 	m_p->f = SysFileCreate();
 
@@ -150,6 +163,7 @@ bool File::CreateInMem() {
 }
 
 void File::Close() {
+	m_p->archive.reset();
 	if (m_p->f != nullptr) {
 		SysFileClose(m_p->f);
 		m_p->f = nullptr;
@@ -157,7 +171,11 @@ void File::Close() {
 }
 
 uint64_t File::Size() const {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return m_p->archive->Size();
+	}
 
 	if (m_p->f == nullptr) {
 		return 0;
@@ -167,41 +185,76 @@ uint64_t File::Size() const {
 }
 
 uint64_t File::Remaining() const {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
 
-	return Size() - Tell();
+	const auto size     = Size();
+	const auto position = Tell();
+	return position < size ? size - position : 0;
 }
 
 uint64_t File::Size(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		const auto info = GetArchiveInfo(name);
+		return info ? info->size : 0;
+	}
 	return SysFileSize(name);
 }
 
+std::optional<File::Info> File::GetInfo(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		return GetArchiveInfo(name);
+	}
+	Info info {};
+	return SysFileGetInfo(name, &info.is_file, &info.size) ? std::optional(info) : std::nullopt;
+}
+
 bool File::Seek(uint64_t offset) {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return m_p->archive->Seek(offset);
+	}
 
 	return SysFileSeek(*m_p->f, offset);
 }
 
 bool File::Truncate(uint64_t size) {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return false;
+	}
 
 	return SysFileTruncate(*m_p->f, size);
 }
 
 bool File::Unlink() {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return false;
+	}
 
 	return SysFileUnlink(*m_p->f, m_file_name);
 }
 
 uint64_t File::Tell() const {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return m_p->archive->Tell();
+	}
 
 	return SysFileTell(*m_p->f);
 }
 
 void File::Read(void* data, uint32_t size, uint32_t* bytes_read) {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		m_p->archive->Read(data, size, bytes_read);
+		return;
+	}
 
 	if (m_p->f != nullptr) {
 		SysFileRead(data, size, *m_p->f, bytes_read);
@@ -209,7 +262,14 @@ void File::Read(void* data, uint32_t size, uint32_t* bytes_read) {
 }
 
 void File::Write(const void* data, uint32_t size, uint32_t* bytes_written) {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		if (bytes_written != nullptr) {
+			*bytes_written = 0;
+		}
+		return;
+	}
 
 	SysFileWrite(data, size, *m_p->f, bytes_written);
 }
@@ -257,24 +317,41 @@ static std::filesystem::path WithoutTrailingSeparator(
 
 
 bool File::IsDirectoryExisting(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		const auto info = GetArchiveInfo(path);
+		return info && !info->is_file;
+	}
 	return SysFileIsDirectoryExisting(WithoutTrailingSeparator(path));
 }
 
 bool File::IsFileExisting(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		const auto info = GetArchiveInfo(name);
+		return info && info->is_file;
+	}
 	return SysFileIsFileExisting(name);
 }
 
 bool File::CreateDirectory(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return false;
+	}
     return SysFileCreateDirectory(
         WithoutTrailingSeparator(path));
 }
 
 bool File::DeleteDirectory(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return false;
+	}
     return SysFileDeleteDirectory(
         WithoutTrailingSeparator(path));
 }
 
 bool File::CreateDirectories(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return false;
+	}
     const auto normalized_path =
         WithoutTrailingSeparator(path);
 
@@ -297,6 +374,9 @@ bool File::CreateDirectories(const std::filesystem::path& path) {
 }
 
 bool File::DeleteDirectories(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return false;
+	}
     const auto normalized_path =
         WithoutTrailingSeparator(path);
 
@@ -322,11 +402,18 @@ bool File::DeleteDirectories(const std::filesystem::path& path) {
 bool File::DeleteFile(
     const std::filesystem::path& name) // @suppress("Member declaration not found")
 {
+	if (IsArchivePath(name)) {
+		return false;
+	}
 	return SysFileDeleteFile(name);
 }
 
 bool File::Flush() {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return true;
+	}
 
 	return SysFileFlush(*m_p->f);
 }
@@ -348,6 +435,9 @@ std::vector<std::byte> File::ReadWholeBuffer() {
 }
 
 DateTime File::GetLastAccessTimeUTC(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		return GetLastAccessTimeUTC(GetArchiveHostPath(name));
+	}
 	SysTimeStruct t {};
 	SysFileToSystemTimeUtc(SysFileGetLastAccessTimeUtc(name), t);
 
@@ -359,6 +449,9 @@ DateTime File::GetLastAccessTimeUTC(const std::filesystem::path& name) {
 }
 
 DateTime File::GetLastWriteTimeUTC(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		return GetLastWriteTimeUTC(GetArchiveHostPath(name));
+	}
 	SysTimeStruct t {};
 	SysFileToSystemTimeUtc(SysFileGetLastWriteTimeUtc(name), t);
 
@@ -373,6 +466,10 @@ void File::GetLastAccessAndWriteTimeUTC(const std::filesystem::path& name, DateT
                                         DateTime* write) {
 	EXIT_IF(access == nullptr);
 	EXIT_IF(write == nullptr);
+	if (IsArchivePath(name)) {
+		GetLastAccessAndWriteTimeUTC(GetArchiveHostPath(name), access, write);
+		return;
+	}
 
 	SysTimeStruct     at {};
 	SysTimeStruct     wt {};
@@ -399,7 +496,12 @@ void File::GetLastAccessAndWriteTimeUTC(DateTime* access, DateTime* write) {
 	EXIT_IF(access == nullptr);
 	EXIT_IF(write == nullptr);
 
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		GetLastAccessAndWriteTimeUTC(m_file_name, access, write);
+		return;
+	}
 
 	if (m_p->f == nullptr) {
 		*access = DateTime();
@@ -429,23 +531,35 @@ void File::GetLastAccessAndWriteTimeUTC(DateTime* access, DateTime* write) {
 }
 
 bool File::SetLastAccessTimeUTC(const std::filesystem::path& name, const DateTime& dt) {
+	if (IsArchivePath(name)) {
+		return false;
+	}
 	auto f = DateTimeToFileTimeUtc(dt);
 	return SysFileSetLastAccessTimeUtc(name, f);
 }
 
 bool File::SetLastWriteTimeUTC(const std::filesystem::path& name, const DateTime& dt) {
+	if (IsArchivePath(name)) {
+		return false;
+	}
 	auto f = DateTimeToFileTimeUtc(dt);
 	return SysFileSetLastWriteTimeUtc(name, f);
 }
 
 bool File::SetLastAccessAndWriteTimeUTC(const std::filesystem::path& name, const DateTime& access,
                                         const DateTime& write) {
+	if (IsArchivePath(name)) {
+		return false;
+	}
 	auto af = DateTimeToFileTimeUtc(access);
 	auto wf = DateTimeToFileTimeUtc(write);
 	return SysFileSetLastAccessAndWriteTimeUtc(name, af, wf);
 }
 
 std::vector<File::DirEntry> File::GetDirEntries(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return GetArchiveDirEntries(path);
+	}
 	std::vector<sys_dir_entry_t> files;
 
 	SysFileGetDents(path, files);
@@ -463,12 +577,18 @@ std::vector<File::DirEntry> File::GetDirEntries(const std::filesystem::path& pat
 bool File::CopyFile(const std::filesystem::path& src,
                     const std::filesystem::path& dst) // @suppress("Member declaration not found")
 {
+	if (IsArchivePath(src) || IsArchivePath(dst)) {
+		return false;
+	}
 	return SysFileCopyFile(src, dst);
 }
 
 bool File::RenameFile(const std::filesystem::path& src,
                       const std::filesystem::path& dst) // @suppress("Member declaration not found")
 {
+	if (IsArchivePath(src) || IsArchivePath(dst)) {
+		return false;
+	}
 	if (IsFileExisting(dst)) {
 		DeleteFile(dst); // @suppress("Invalid arguments")
 	}
@@ -477,6 +597,9 @@ bool File::RenameFile(const std::filesystem::path& src,
 }
 
 void File::RemoveReadonly(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		return;
+	}
 	SysFileRemoveReadonly(name);
 }
 

src/common/file.h

@@ -7,6 +7,7 @@
 #include <cstddef>
 #include <filesystem>
 #include <memory>
+#include <optional>
 #include <string>
 #include <vector>
 
@@ -34,6 +35,10 @@ public:
 		std::string name;
 		bool        is_file;
 	};
+	struct Info {
+		bool     is_file;
+		uint64_t size;
+	};
 
 	File();
 	explicit File(const std::filesystem::path& name);
@@ -72,6 +77,7 @@ public:
 	std::vector<std::byte> ReadWholeBuffer();
 
 	static uint64_t Size(const std::filesystem::path& name);
+	static std::optional<Info> GetInfo(const std::filesystem::path& name);
 
 	static bool IsDirectoryExisting(const std::filesystem::path& path);
 	static bool IsFileExisting(const std::filesystem::path& name);

src/common/platform/sysFileIO.h

@@ -45,6 +45,7 @@ bool        SysFileTruncate(sys_file_t& f, uint64_t size); // NOLINT(google-runt
 bool        SysFileUnlink(sys_file_t&                  f,
                           const std::filesystem::path& name); // NOLINT(google-runtime-references)
 uint64_t    SysFileSize(const std::filesystem::path& file_name);
+bool        SysFileGetInfo(const std::filesystem::path& name, bool* is_file, uint64_t* size);
 bool        SysFileIsError(sys_file_t& f); // NOLINT(google-runtime-references)
 bool        SysFileIsDirectoryExisting(const std::filesystem::path& path);
 bool        SysFileIsFileExisting(const std::filesystem::path& name);

src/common/platform/sysLinuxFileIO.cpp

@@ -54,7 +54,7 @@ struct sys_file_t {
 #endif
 
 static std::filesystem::path get_internal_name(const std::filesystem::path& name) {
-	return name.is_absolute() ? name : (std::filesystem::path(".") / name);
+	return name.empty() || name.is_absolute() ? name : (std::filesystem::path(".") / name);
 }
 
 // Pass access-pattern hints to the host.
@@ -297,10 +297,21 @@ uint64_t SysFileSize(sys_file_t& f) {
 	return 0;
 }
 
+bool SysFileGetInfo(const std::filesystem::path& name, bool* is_file, uint64_t* size) {
+	const auto path = get_internal_name(name);
+	struct stat info {};
+	if (stat(path.c_str(), &info) != 0) {
+		return false;
+	}
+	*is_file = !S_ISDIR(info.st_mode);
+	*size = *is_file ? static_cast<uint64_t>(info.st_size) : 0;
+	return true;
+}
+
 uint64_t SysFileSize(const std::filesystem::path& file_name) {
-	sys_file_t* f    = SysFileOpenR(file_name);
-	uint64_t    size = SysFileSize(*f);
-	SysFileClose(f);
+	bool is_file;
+	uint64_t size = 0;
+	SysFileGetInfo(file_name, &is_file, &size);
 	return size;
 }
 
@@ -353,29 +364,15 @@ bool SysFileIsError(sys_file_t& f) {
 }
 
 bool SysFileIsDirectoryExisting(const std::filesystem::path& path) {
-	auto real_name     = get_internal_name(path);
-	auto real_name_str = real_name.string();
-
-	struct stat s {};
-
-	if (0 != stat(real_name_str.c_str(), &s)) {
-		return false;
-	}
-
-	return S_ISDIR(s.st_mode); // NOLINT
+	bool is_file;
+	uint64_t size;
+	return SysFileGetInfo(path, &is_file, &size) && !is_file;
 }
 
 bool SysFileIsFileExisting(const std::filesystem::path& name) {
-	auto real_name     = get_internal_name(name);
-	auto real_name_str = real_name.string();
-
-	struct stat s {};
-
-	if (0 != stat(real_name_str.c_str(), &s)) {
-		return false;
-	}
-
-	return !S_ISDIR(s.st_mode); // NOLINT
+	bool is_file;
+	uint64_t size;
+	return SysFileGetInfo(name, &is_file, &size) && is_file;
 }
 
 bool SysFileCreateDirectory(const std::filesystem::path& path) {

src/common/platform/sysWindowsFileIO.cpp

@@ -266,19 +266,27 @@ uint64_t SysFileSize(sys_file_t& f) {
 	return 0;
 }
 
-uint64_t SysFileSize(const std::filesystem::path& file_name) {
-	LARGE_INTEGER             s;
-	WIN32_FILE_ATTRIBUTE_DATA a;
-
-	auto wide = file_name.wstring();
-	if (GetFileAttributesExW(wide.c_str(), GetFileExInfoStandard, &a) == 0) {
-		return 0;
+bool SysFileGetInfo(const std::filesystem::path& name, bool* is_file, uint64_t* size) {
+	const bool directory_path = !name.empty() && !name.has_filename() && name != name.root_path();
+	WIN32_FILE_ATTRIBUTE_DATA info {};
+	if (GetFileAttributesExW(directory_path ? name.parent_path().c_str() : name.c_str(),
+	                         GetFileExInfoStandard, &info) == 0) {
+		return false;
 	}
+	const bool file = (info.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY) == 0;
+	if (directory_path && file) {
+		return false;
+	}
+	*is_file = file;
+	*size = *is_file ? ((static_cast<uint64_t>(info.nFileSizeHigh) << 32u) | info.nFileSizeLow) : 0;
+	return true;
+}
 
-	s.HighPart = static_cast<LONG>(a.nFileSizeHigh);
-	s.LowPart  = a.nFileSizeLow;
-
-	return s.QuadPart;
+uint64_t SysFileSize(const std::filesystem::path& file_name) {
+	bool is_file;
+	uint64_t size = 0;
+	SysFileGetInfo(file_name, &is_file, &size);
+	return size;
 }
 
 bool SysFileTruncate(sys_file_t& f, uint64_t size) {
@@ -346,17 +354,15 @@ bool SysFileIsError(sys_file_t& f) {
 }
 
 bool SysFileIsDirectoryExisting(const std::filesystem::path& path) {
-	auto  wide = path.wstring();
-	DWORD a    = GetFileAttributesW(wide.c_str());
-	return a != INVALID_FILE_ATTRIBUTES &&
-	       ((a & static_cast<DWORD>(FILE_ATTRIBUTE_DIRECTORY)) != 0u);
+	bool is_file;
+	uint64_t size;
+	return SysFileGetInfo(path, &is_file, &size) && !is_file;
 }
 
 bool SysFileIsFileExisting(const std::filesystem::path& name) {
-	auto  wide = name.wstring();
-	DWORD a    = GetFileAttributesW(wide.c_str());
-	return a != INVALID_FILE_ATTRIBUTES &&
-	       ((a & static_cast<DWORD>(FILE_ATTRIBUTE_DIRECTORY)) == 0u);
+	bool is_file;
+	uint64_t size;
+	return SysFileGetInfo(name, &is_file, &size) && is_file;
 }
 
 bool SysFileCreateDirectory(const std::filesystem::path& path) {

src/common/profiler.cpp

@@ -2,52 +2,13 @@
 
 #include "common/emulatorConfig.h"
 
-#include <algorithm>
 #include <common/TracyProtocol.hpp>
 #include <common/TracyVersion.hpp>
 #include <cstdio>
 #include <tracy/Tracy.hpp>
-#include <vector>
-
-namespace {
-
-thread_local std::vector<Profiler::ScopedBlock*> g_block_stack;
-
-void RemoveBlock(Profiler::ScopedBlock* block) {
-	auto block_it = std::find(g_block_stack.rbegin(), g_block_stack.rend(), block);
-	if (block_it != g_block_stack.rend()) {
-		g_block_stack.erase(std::next(block_it).base());
-	}
-}
-
-} // namespace
 
 namespace Profiler {
 
-ScopedBlock::ScopedBlock(const tracy::SourceLocationData* source_location) {
-	if (tracy::ProfilerAvailable()) {
-		m_zone.emplace(source_location, TRACY_CALLSTACK, true);
-		g_block_stack.push_back(this);
-	}
-}
-
-ScopedBlock::~ScopedBlock() {
-	End();
-}
-
-void ScopedBlock::End() {
-	if (m_zone.has_value()) {
-		m_zone.reset();
-		RemoveBlock(this);
-	}
-}
-
-void EndBlock() {
-	if (!g_block_stack.empty()) {
-		g_block_stack.back()->End();
-	}
-}
-
 void SetThreadName(const char* name) {
 	if (tracy::ProfilerAvailable() && name != nullptr) {
 		tracy::SetThreadName(name);

src/common/profiler.h

@@ -4,7 +4,6 @@
 #include "common/common.h"
 
 #include <cstdint>
-#include <optional>
 #include <tracy/Tracy.hpp> // IWYU pragma: export
 
 namespace profiler::colors {
@@ -23,22 +22,6 @@ inline constexpr uint32_t DeepOrangeA200 = 0xff6e40;
 
 namespace Profiler {
 
-class ScopedBlock {
-public:
-	explicit ScopedBlock(const tracy::SourceLocationData* source_location);
-	ScopedBlock(const ScopedBlock&)            = delete;
-	ScopedBlock& operator=(const ScopedBlock&) = delete;
-	ScopedBlock(ScopedBlock&&)                 = delete;
-	ScopedBlock& operator=(ScopedBlock&&)      = delete;
-	~ScopedBlock();
-
-	void End();
-
-private:
-	std::optional<tracy::ScopedZone> m_zone;
-};
-
-void EndBlock();
 void SetThreadName(const char* name);
 
 void Initialize();
@@ -66,19 +49,11 @@ struct Lifecycle {
 	    kyty_profiler_source_location_, line) {name, TracyFunction, TracyFile,                     \
 	                                           static_cast<uint32_t>(line),                        \
 	                                           KYTY_PROFILER_COLOR_OR_DEFAULT(0, __VA_ARGS__)};    \
-	Profiler::ScopedBlock KYTY_PROFILER_CONCAT(kyty_profiler_block_, line)(                        \
-	    &KYTY_PROFILER_CONCAT(kyty_profiler_source_location_, line))
+	tracy::ScopedZone KYTY_PROFILER_CONCAT(kyty_profiler_block_, line)(                           \
+	    &KYTY_PROFILER_CONCAT(kyty_profiler_source_location_, line),                              \
+	    TRACY_CALLSTACK, tracy::ProfilerAvailable())
 
-#define KYTY_PROFILER_FUNCTION(...) KYTY_PROFILER_FUNCTION_IMPL(__LINE__ __VA_OPT__(, ) __VA_ARGS__)
-#define KYTY_PROFILER_FUNCTION_IMPL(line, ...)                                                     \
-	static constexpr tracy::SourceLocationData KYTY_PROFILER_CONCAT(                               \
-	    kyty_profiler_source_location_, line) {nullptr, TracyFunction, TracyFile,                  \
-	                                           static_cast<uint32_t>(line),                        \
-	                                           KYTY_PROFILER_COLOR_OR_DEFAULT(0, __VA_ARGS__)};    \
-	Profiler::ScopedBlock KYTY_PROFILER_CONCAT(kyty_profiler_block_, line)(                        \
-	    &KYTY_PROFILER_CONCAT(kyty_profiler_source_location_, line))
-
-#define KYTY_PROFILER_END_BLOCK Profiler::EndBlock()
+#define KYTY_PROFILER_FUNCTION(...) KYTY_PROFILER_BLOCK(nullptr __VA_OPT__(, ) __VA_ARGS__)
 
 #define KYTY_PROFILER_THREAD(name) Profiler::SetThreadName(name)
 

src/common/threads.cpp

@@ -2,7 +2,6 @@
 
 #include "common/assert.h"
 
-#include <algorithm>
 #include <atomic>
 #include <cerrno>
 #include <chrono>             // IWYU pragma: keep
@@ -26,8 +25,7 @@
 #ifdef KYTY_WIN_CS
 #include <windows.h> // IWYU pragma: keep
 // IWYU pragma: no_include <winbase.h>
-constexpr DWORD    KYTY_CS_SPIN_COUNT          = 4000;
-constexpr uint64_t KYTY_SLEEP_SPIN_LIMIT_100NS = 500; // 50 us
+constexpr DWORD KYTY_CS_SPIN_COUNT = 4000;
 
 #ifndef CREATE_WAITABLE_TIMER_HIGH_RESOLUTION
 #define CREATE_WAITABLE_TIMER_HIGH_RESOLUTION 0x00000002
@@ -38,30 +36,6 @@ static void SleepHighResolution100ns(uint64_t units_100ns) {
 		return;
 	}
 
-	// Keep spinning only where a kernel transition is
-	// likely to cost more than the requested delay; ordinary millisecond sleeps use the
-	// per-thread high-resolution waitable timer below.
-	if (units_100ns <= KYTY_SLEEP_SPIN_LIMIT_100NS) {
-		LARGE_INTEGER frequency {};
-		LARGE_INTEGER start {};
-		if (QueryPerformanceFrequency(&frequency) != 0 && QueryPerformanceCounter(&start) != 0 &&
-		    frequency.QuadPart > 0) {
-			const auto wait_ticks =
-			    static_cast<LONGLONG>((static_cast<long double>(units_100ns) *
-			                           static_cast<long double>(frequency.QuadPart)) /
-			                          10000000.0L);
-			const auto    deadline = start.QuadPart + std::max<LONGLONG>(wait_ticks, 1);
-			LARGE_INTEGER now {};
-			do {
-				if (QueryPerformanceCounter(&now) == 0) {
-					break;
-				}
-				YieldProcessor();
-			} while (now.QuadPart < deadline);
-			return;
-		}
-	}
-
 	thread_local HANDLE timer = CreateWaitableTimerExW(
 	    nullptr, nullptr, CREATE_WAITABLE_TIMER_HIGH_RESOLUTION, TIMER_ALL_ACCESS);
 	if (timer == nullptr) {
@@ -131,14 +105,13 @@ static SleepConditionVariableCS_func_t ResolveSleepConditionVariableCS() {
 #endif
 
 #ifdef KYTY_POSIX_HIGH_RES_SLEEP
-// Spin for very short waits; use an absolute deadline for longer waits.
+// An absolute deadline preserves the requested sleep across signal interruptions.
 static void SleepHighResolutionNanos(uint64_t nanos) {
 	if (nanos == 0) {
 		return;
 	}
 
 	constexpr uint64_t NANOS_PER_SEC = 1000000000;
-	constexpr uint64_t SPIN_LIMIT_NS = 50000; // below this a context switch dominates
 
 	timespec deadline {};
 	if (clock_gettime(CLOCK_MONOTONIC, &deadline) != 0) {
@@ -150,17 +123,6 @@ static void SleepHighResolutionNanos(uint64_t nanos) {
 	deadline.tv_sec += static_cast<time_t>(target_nsec / NANOS_PER_SEC);
 	deadline.tv_nsec = static_cast<long>(target_nsec % NANOS_PER_SEC);
 
-	if (nanos <= SPIN_LIMIT_NS) {
-		timespec now {};
-		do {
-			if (clock_gettime(CLOCK_MONOTONIC, &now) != 0) {
-				return;
-			}
-		} while (now.tv_sec < deadline.tv_sec ||
-		         (now.tv_sec == deadline.tv_sec && now.tv_nsec < deadline.tv_nsec));
-		return;
-	}
-
 	while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME, &deadline, nullptr) == EINTR) {
 	}
 }

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -2266,7 +2266,6 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 	const bool gl2_writeback = ((gcr_cntl & GcrGl2Writeback) != 0);
 
 	auto trigger_interrupt = [&]() {
-		bool queued = false;
 		switch (interrupt_selector) {
 			case 0x00:
 			case 0x03: break;
@@ -2274,13 +2273,10 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 			case 0x02:
 			case 0x04:
 				cp.TriggerEopEventAtEndOfPipe(interrupt_context_id);
-				queued = true;
+				cp.BufferFlush();
 				break;
 			default: EXIT("unknown release_mem interrupt selector\n");
 		}
-		if (queued) {
-			cp.BufferFlush();
-		}
 	};
 
 	if (data_sel == 0 || interrupt_selector == 4) {
@@ -2318,7 +2314,9 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 		cp.WriteAtEndOfPipe32(cache_policy, event_write_dest, eop_event_type, cache_action,
 		                      event_index, event_source, dst_gpu_addr, static_cast<uint32_t>(value),
 		                      interrupt_selector, interrupt_context_id);
-		cp.BufferFlush();
+		if (interrupt_selector == 0x01 || interrupt_selector == 0x02) {
+			cp.BufferFlush();
+		}
 
 		return 7;
 	}

src/graphics/host_gpu/regionDefinitions.h

@@ -3,6 +3,7 @@
 
 #include "common/bitArray.h"
 #include "common/common.h"
+#include "kernel/memory.h"
 
 #include <compare>
 
@@ -10,7 +11,8 @@ namespace Libs::Graphics {
 
 constexpr uint64_t TRACKER_PAGE_SIZE    = 4ull * 1024ull;
 constexpr uint64_t TRACKER_REGION_SIZE  = 4ull * 1024ull * 1024ull;
-constexpr uint64_t TRACKER_ADDRESS_SIZE = 1ull << 40u;
+constexpr uint64_t TRACKER_ADDRESS_SIZE = 1ull << 44u;
+constexpr uint64_t LOWER_ADDRESS_SIZE   = 1ull << 40u;
 constexpr size_t   TRACKER_REGION_PAGES = TRACKER_REGION_SIZE / TRACKER_PAGE_SIZE;
 
 struct GuestRange {
@@ -19,8 +21,11 @@ struct GuestRange {
 
 	[[nodiscard]] constexpr bool Empty() const noexcept { return address == 0 && size == 0; }
 	[[nodiscard]] constexpr bool Valid() const noexcept {
-		return address != 0 && size != 0 && address < TRACKER_ADDRESS_SIZE &&
-		       size <= TRACKER_ADDRESS_SIZE - address;
+		if (address == 0 || size == 0) return false;
+		if (address < LOWER_ADDRESS_SIZE) return size <= LOWER_ADDRESS_SIZE - address;
+		constexpr auto base = LibKernel::Memory::kExtendedMemoryBase;
+		constexpr auto length = LibKernel::Memory::kExtendedMemorySize;
+		return address >= base && address - base < length && size <= length - (address - base);
 	}
 	[[nodiscard]] constexpr bool     ValidOrEmpty() const noexcept { return Empty() || Valid(); }
 	[[nodiscard]] constexpr uint64_t End() const noexcept { return address + size; }

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

@@ -55,7 +55,8 @@ template <bool insert>
 void BufferCache::ChangeRegister(BufferId id) {
 	auto& buffer = m_slot_buffers[id];
 	PageTable::PageRange pages {};
-	EXIT_IF(!PageTable::TryGetPageRange(buffer.CpuAddress(), buffer.Size(), pages));
+	EXIT_IF(!(GuestRange {buffer.CpuAddress(), buffer.Size()}.Valid()) ||
+	        !PageTable::TryGetPageRange(buffer.CpuAddress(), buffer.Size(), pages));
 	for (size_t page = pages.first; page < pages.last_exclusive; ++page) {
 		if constexpr (insert) {
 			m_page_table[page] = id;
@@ -64,6 +65,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		}
 	}
 	const auto size_pages = pages.last_exclusive - pages.first;
+	const auto table_offset = PageIndex(buffer.CpuAddress()) * sizeof(vk::DeviceAddress);
 	if constexpr (insert) {
 		const auto [it, inserted] = m_buffers.emplace(buffer.CpuAddress(), id);
 		(void)it;
@@ -75,7 +77,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		for (uint64_t i = 0; i < size_pages; ++i) {
 			addresses.push_back(buffer.BufferDeviceAddress() + (i << CACHING_PAGEBITS));
 		}
-		WriteDataBuffer(m_bda_pagetable_buffer, pages.first * sizeof(vk::DeviceAddress),
+		WriteDataBuffer(m_bda_pagetable_buffer, table_offset,
 		                addresses.data(), addresses.size() * sizeof(vk::DeviceAddress));
 	} else {
 		const auto found = m_buffers.find(buffer.CpuAddress());
@@ -84,7 +86,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		EXIT_IF(buffer.Size() > m_total_used_memory);
 		m_total_used_memory -= buffer.Size();
 		m_lru_cache.Free(buffer.lru_id);
-		m_bda_pagetable_buffer.Fill(pages.first * sizeof(vk::DeviceAddress),
+		m_bda_pagetable_buffer.Fill(table_offset,
 		                            size_pages * sizeof(vk::DeviceAddress), 0);
 		buffer.is_deleted = true;
 	}
@@ -312,10 +314,14 @@ BufferCache::OverlapResult BufferCache::ResolveOverlaps(uint64_t vaddr, uint64_t
 			// Reserve space in the incoming stream's direction of growth.
 			// The old buffer extending left of the request predicts growth to the right, and vice versa.
 			if (expands_left) {
-				end += std::min(StreamLeapSize, PageTable::kAddressSpaceSize - end);
+				end += std::min(StreamLeapSize, (vaddr < LOWER_ADDRESS_SIZE ? LOWER_ADDRESS_SIZE
+				                                       : LibKernel::Memory::kExtendedMemoryBase +
+				                                             LibKernel::Memory::kExtendedMemorySize) - end);
 			}
 			if (expands_right) {
-				const auto minimum = CACHING_PAGESIZE * 2;
+				const auto minimum = vaddr < LOWER_ADDRESS_SIZE
+				                         ? CACHING_PAGESIZE * 2
+				                         : LibKernel::Memory::kExtendedMemoryBase;
 				if (begin > minimum) {
 					begin -= std::min(StreamLeapSize, begin - minimum);
 				}
@@ -448,7 +454,8 @@ std::pair<Buffer*, uint64_t> BufferCache::ObtainBuffer(uint64_t vaddr, uint64_t
 		const auto alignment = std::max<uint64_t>(
 		    m_graphics.physical_device_properties.limits.minUniformBufferOffsetAlignment, 1);
 		auto [mapped, offset] = m_stream_buffer.Map(size, alignment, false);
-		if (mapped != nullptr && Libs::LibKernel::Memory::TryReadBacking(vaddr, mapped, size)) {
+		if (mapped != nullptr) {
+			std::memcpy(mapped, reinterpret_cast<const void*>(vaddr), size);
 			m_stream_buffer.Commit();
 			return {&m_stream_buffer, offset};
 		}

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

@@ -29,10 +29,22 @@ class BufferCache {
 public:
 	static constexpr uint32_t CACHING_PAGEBITS  = 14;
 	static constexpr uint64_t CACHING_PAGESIZE  = uint64_t {1} << CACHING_PAGEBITS;
-	static constexpr uint64_t CACHING_NUMPAGES  = uint64_t {1} << (40 - CACHING_PAGEBITS);
+	static constexpr uint64_t CACHING_NUMPAGES  = (LOWER_ADDRESS_SIZE + LibKernel::Memory::kExtendedMemorySize) >> CACHING_PAGEBITS;
 	static constexpr uint64_t BDA_PAGETABLE_SIZE =
 	    CACHING_NUMPAGES * sizeof(vk::DeviceAddress);
 
+	static constexpr uint64_t PageIndex(uint64_t address) {
+		return (address < LOWER_ADDRESS_SIZE
+		            ? address
+		            : address - LibKernel::Memory::kExtendedMemoryBase + LOWER_ADDRESS_SIZE) >>
+		       CACHING_PAGEBITS;
+	}
+	static constexpr uint64_t GuestAddress(uint64_t offset) {
+		return offset < LOWER_ADDRESS_SIZE
+		           ? offset
+		           : offset - LOWER_ADDRESS_SIZE + LibKernel::Memory::kExtendedMemoryBase;
+	}
+
 	BufferCache(GraphicContext& graphics, CommandScheduler& scheduler, PageManager& page_manager,
 	            TextureCache& texture_cache);
 	~BufferCache();
@@ -88,7 +100,7 @@ private:
 		bool                has_stream_leap;
 	};
 
-	using PageTable = MultiLevelPageTable<BufferId, CACHING_PAGEBITS, 40, 16>;
+	using PageTable = MultiLevelPageTable<BufferId, CACHING_PAGEBITS, 44, 20>;
 	static_assert(CACHING_PAGESIZE == (uint64_t {1} << PageTable::kPageBits));
 	void WriteDataBuffer(Buffer& buffer, uint64_t address, const void* source, uint64_t size);
 	void TouchBuffer(const Buffer& buffer);

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

@@ -134,8 +134,9 @@ void FaultManager::ProcessFaultBuffer() {
 		const auto* faults = std::bit_cast<const uint64_t*>(mapped);
 		const auto  count  = static_cast<uint32_t>(faults[0]);
 		for (uint32_t index = 1; index <= count; ++index) {
-			fault_ranges.Add(faults[index], BufferCache::CACHING_PAGESIZE);
-			LOGF("Accessed non-GPU cached memory at 0x%016" PRIx64 "\n", faults[index]);
+			const auto address = BufferCache::GuestAddress(faults[index]);
+			fault_ranges.Add(address, BufferCache::CACHING_PAGESIZE);
+			LOGF("Accessed non-GPU cached memory at 0x%016" PRIx64 "\n", address);
 		}
 		fault_ranges.ForEach([this](uint64_t start, uint64_t end) {
 			EXIT_IF(end - start > std::numeric_limits<uint32_t>::max());

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

@@ -14,10 +14,10 @@ SamplerCache::~SamplerCache() {
 	}
 }
 
-vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r) {
+vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r, bool integer_border) {
 	Common::LockGuard lock(m_mutex);
 
-	const SamplerKey key {r.fields[0], r.fields[1], r.fields[2], r.fields[3]};
+	const SamplerKey key {r.fields[0], r.fields[1], r.fields[2], r.fields[3], integer_border};
 	if (auto iter = m_samplers.find(key); iter != m_samplers.end()) {
 		return iter->second;
 	}
@@ -93,23 +93,23 @@ vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r) {
 		return vk::SamplerAddressMode::eClampToBorder;
 	};
 
-	vk::BorderColor border = vk::BorderColor::eIntTransparentBlack;
+	vk::BorderColor border = integer_border ? vk::BorderColor::eIntTransparentBlack
+	                                       : vk::BorderColor::eFloatTransparentBlack;
 	switch (static_cast<Prospero::SamplerBorderColor>(r.BorderColorType())) {
-		case Prospero::SamplerBorderColor::kTransBlack:
-			border = vk::BorderColor::eIntTransparentBlack;
-			break;
+		case Prospero::SamplerBorderColor::kTransBlack: break;
 		case Prospero::SamplerBorderColor::kOpaqueBlack:
-			border = vk::BorderColor::eIntOpaqueBlack;
+			border = integer_border ? vk::BorderColor::eIntOpaqueBlack
+			                        : vk::BorderColor::eFloatOpaqueBlack;
 			break;
 		case Prospero::SamplerBorderColor::kOpaqueWhite:
-			border = vk::BorderColor::eIntOpaqueWhite;
+			border = integer_border ? vk::BorderColor::eIntOpaqueWhite
+			                        : vk::BorderColor::eFloatOpaqueWhite;
 			break;
 		case Prospero::SamplerBorderColor::kFromTable:
 			LOGF(
 			    "temporary: approximating table border color as transparent black, index = %" PRIu16
 			    "\n",
 			    r.BorderColorPtr());
-			border = vk::BorderColor::eIntTransparentBlack;
 			break;
 		default: EXIT("unknown border color: %d", static_cast<int>(r.BorderColorType()));
 	}

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

@@ -24,10 +24,10 @@ public:
 	~SamplerCache();
 	KYTY_CLASS_NO_COPY(SamplerCache);
 
-	vk::Sampler GetSampler(const ShaderSamplerResource& r);
+	vk::Sampler GetSampler(const ShaderSamplerResource& r, bool integer_border);
 
 private:
-	using SamplerKey = std::array<uint32_t, 4>;
+	using SamplerKey = std::array<uint32_t, 5>;
 
 	struct SamplerKeyHash {
 		std::size_t operator()(const SamplerKey& key) const {

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

@@ -222,14 +222,6 @@ TextureCache::BindingType TextureCache::UploadBinding(const Image& image) {
 	return image.usage.storage ? BindingType::Storage : BindingType::Texture;
 }
 
-bool TextureCache::SafeToDownload(const Image& image) {
-	if (!image.SafeToDownload()) {
-		return false;
-	}
-	const auto range = image.info.data;
-	return !m_buffer_cache.HasGpuDirtyBytes(range.address, range.size);
-}
-
 ImageId TextureCache::InsertImage(const ImageInfo& info) {
 	const auto id = m_slot_images.insert(m_graphics, m_scheduler, info);
 	if (!info.data.Empty()) {
@@ -1360,7 +1352,7 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
 		if (ensure_valid && owner->depth_id) {
 			owner = m_slot_images.try_get(owner->depth_id);
 		}
-		if (owner == nullptr || (ensure_valid && !SafeToDownload(*owner))) {
+		if (owner == nullptr || (ensure_valid && !owner->SafeToDownload())) {
 			continue;
 		}
 		matches.push_back(id);
@@ -1751,7 +1743,7 @@ bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uin
 	std::scoped_lock lock {m_texture_cache.m_lock};
 	auto& image = m_texture_cache.m_slot_images[selected];
 	// The GPU thread owns image retirement; CPU invalidation can dirty this image after lookup.
-	if (!m_texture_cache.SafeToDownload(image)) {
+	if (!image.SafeToDownload()) {
 		return false;
 	}
 	if (!buffer.IsInBounds(image.info.data.address, 1)) {
@@ -1814,7 +1806,7 @@ bool TextureCache::DownloadImageMemory(ImageId id) {
 		return false;
 	}
 	auto transfer = BuildDownload(image);
-	if (!transfer.valid || !SafeToDownload(image)) {
+	if (!transfer.valid || !image.SafeToDownload()) {
 		return false;
 	}
 	const auto range    = image.info.data;
@@ -1875,7 +1867,10 @@ bool TextureCache::IsRegionGpuModified(uint64_t address, uint64_t size) {
 	std::scoped_lock lock {m_lock};
 	for (const auto id: FindImagesInRegion(address, size, false)) {
 		const auto& image = m_slot_images[id];
-		if (!image.depth_id && image.IsGpuModified()) {
+		// PPSA17168: S_LOAD_DWORD reads shader data at an address overlapping an old
+		// render target whose memory the CPU has reused. The cached image still retains
+		// its earlier GPU-modified flag.
+		if (!image.depth_id && image.IsGpuModified() && !image.IsDefinitelyCpuDirty()) {
 			return true;
 		}
 	}
@@ -2001,7 +1996,7 @@ void TextureCache::RunGarbageCollector() {
 				continue;
 			}
 			if (owner->IsGpuModified()) {
-				const bool safe = SafeToDownload(*owner);
+				const bool safe = owner->SafeToDownload();
 				if (safe && owner->info.IsTiled()) {
 					continue;
 				}

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

@@ -91,9 +91,9 @@ private:
 	};
 
 	using ImageIds       = InlinePageOwnerList<ImageId, 16>;
-	using ImagePageTable = MultiLevelPageTable<ImageIds, 20, 40, 10>;
+	using ImagePageTable = MultiLevelPageTable<ImageIds, 20, 44, 14>;
 
-	// Callers have validated the nonempty 40-bit range with TryGetPageRange.
+	// Callers have validated the nonempty 44-bit range with TryGetPageRange.
 	template <typename Func>
 	static void ForEachPage(uint64_t address, size_t size, Func&& func) {
 		using FuncReturn = typename std::invoke_result<Func, uint64_t>::type;
@@ -128,7 +128,6 @@ private:
 	[[nodiscard]] static bool SameBacking(const ImageInfo& cached, const ImageInfo& requested,
 	                                      bool exact_format);
 	[[nodiscard]] static BindingType UploadBinding(const Image& image);
-	[[nodiscard]] bool               SafeToDownload(const Image& image);
 
 	// Caller holds m_lock; it also serializes the per-image query epoch.
 	[[nodiscard]] ImageIds      FindImagesInRegion(uint64_t address, uint64_t size,

src/graphics/host_gpu/renderer/colorRenderTarget.cpp

@@ -21,8 +21,6 @@
 
 namespace Libs::Graphics {
 
-static std::atomic<uint32_t> g_render_color_log_count = 0;
-
 static bool DccAlphaOnMsb(const HW::ColorInfo& info) {
 	switch (info.format) {
 		case Prospero::ChannelLayout::k10_10_10_2:
@@ -262,21 +260,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		EXIT("render-target backing range is invalid\n");
 	}
 
-	const vk::Extent2D view_extent = {std::max(width >> rt.view.current_mip_level, 1u),
-	                                  std::max(height >> rt.view.current_mip_level, 1u)};
-
-	auto decision_log_id = g_render_color_log_count.fetch_add(1);
-	if (decision_log_id < 128) {
-		LOGF("RenderColorTarget: slot=%" PRIu32 " addr=0x%010" PRIx64 " size=0x%016" PRIx64
-		     " extent=%ux%ux%u view_mip=%u view_extent=%ux%u levels=%u pitch=%u"
-		     " fmt=0x%08" PRIx32 " nfmt=0x%08" PRIx32 " order=0x%08" PRIx32 " samples=%u tile=%s\n",
-		     rt_slot, rt.base.addr, backing_size, width, height, depth, rt.view.current_mip_level,
-		     view_extent.width, view_extent.height, levels, pitch,
-		     static_cast<uint32_t>(rt.info.format), static_cast<uint32_t>(rt.info.channel_type),
-		     static_cast<uint32_t>(rt.info.channel_order), samples, tile ? "tiled" : "linear");
-	}
-
-	TextureCache::ImageDesc desc {};
+	auto& desc = r.desc;
 	desc.type              = TextureCache::BindingType::RenderTarget;
 	desc.info.data         = {rt.base.addr, backing_size};
 	desc.info.pixel_format = target_format.format;
@@ -348,7 +332,6 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 	desc.view_info.layer_count = view.layer_count;
 	desc.view_info.usage       = vk::ImageUsageFlagBits::eColorAttachment;
 	auto& texture_cache        = m_context.GetTextureCache();
-	r.desc                     = std::move(desc);
 	r.guest_mip_level          = rt.view.current_mip_level;
 	r.guest_array_layer        = view.base_layer;
 	r.image_id                 = texture_cache.FindImage(r.desc, exact_format);

src/graphics/host_gpu/renderer/image/image.cpp

@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/image/image.h"
 
 #include "common/assert.h"
+#include "common/logging/log.h"
 #include "common/profiler.h"
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
@@ -10,7 +11,9 @@
 
 #include <algorithm>
 #include <array>
+#include <atomic>
 #include <cstdint>
+#include <fmt/format.h>
 #include <xxhash.h>
 
 namespace Libs::Graphics {
@@ -48,12 +51,29 @@ namespace {
 }
 
 [[nodiscard]] vk::ImageUsageFlags ImageUsageFlags(GraphicContext& graphics, const ImageInfo& info) {
+	auto usage = vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst;
 	if (info.IsBlock()) {
-		return vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst |
-		       vk::ImageUsageFlagBits::eSampled;
+		usage |= vk::ImageUsageFlagBits::eSampled;
+		if (graphics.supports_block_texel_view) {
+			const auto storage = usage | vk::ImageUsageFlagBits::eStorage;
+			if (graphics.GetImageFormatProperties(info.pixel_format, HostImageType(info.type),
+			                                      vk::ImageTiling::eOptimal, storage,
+			                                      ImageCreateFlags(graphics, info),
+			                                      nullptr) == vk::Result::eSuccess) {
+				usage = storage;
+			} else {
+				static std::atomic_flag warned = ATOMIC_FLAG_INIT;
+				if (!warned.test_and_set(std::memory_order_relaxed)) {
+					Log::WriteToConsoleAndLog(fmt::format(
+					    "Warning: format {} does not support storage access; block-compressed "
+					    "textures written by the guest will not render.\n",
+					    vk::to_string(info.pixel_format)));
+				}
+			}
+		}
+		return usage;
 	}
 	const auto properties = graphics.GetFormatProperties(info.pixel_format);
-	auto       usage = vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst;
 	if (HasFormatFeature(properties, vk::FormatFeatureFlagBits::eSampledImage)) {
 		usage |= vk::ImageUsageFlagBits::eSampled;
 	}
@@ -628,10 +648,8 @@ void Validate(const ImageInfo& info) {
 			break;
 		case ImageMetadataKind::Dcc:
 		case ImageMetadataKind::Cmask:
-			if (info.metadata.range.address == 0 ||
-			    info.metadata.range.address >= TRACKER_ADDRESS_SIZE ||
-			    (info.metadata.range.size != 0 &&
-			     info.metadata.range.size > TRACKER_ADDRESS_SIZE - info.metadata.range.address) ||
+			if (!GuestRange {info.metadata.range.address,
+			                 std::max<uint64_t>(info.metadata.range.size, 1)}.Valid() ||
 			    info.metadata.compression == VideoOutCompression::Unsupported) {
 				EXIT("invalid color metadata\n");
 			}

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

@@ -0,0 +1,58 @@
+#include "graphics/host_gpu/renderer/pipeline/blendMapping.h"
+
+#include "graphics/guest_gpu/hardwareContext.h"
+
+#include <initializer_list>
+
+namespace Libs::Graphics {
+
+bool BlendFactorIsDualSource(uint8_t factor) {
+	return factor >= static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Color) &&
+	       factor <= static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
+}
+
+namespace {
+
+bool BlendFactorIsConstantColor(uint8_t factor) {
+	return factor == static_cast<uint8_t>(Prospero::BlendFactor::kConstantColor) ||
+	       factor == static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusConstantColor);
+}
+
+} // namespace
+
+BlendMappingSupport ClassifyBlendMapping(const HW::BlendControl&                blend,
+                                         const Prospero::ColorComponentMapping& mapping) {
+	// Color constants are not swizzled with the exports; scalar constant alpha is unaffected.
+	if (!mapping.IsIdentity() && (BlendFactorIsConstantColor(blend.color_srcblend) ||
+	                              BlendFactorIsConstantColor(blend.color_destblend))) {
+		return BlendMappingSupport::Unsupported;
+	}
+	if (mapping.Map(3) == 3u) {
+		return BlendMappingSupport::Direct;
+	}
+	// Moving alpha requires the same equation for all channels.
+	if (blend.separate_alpha_blend && (blend.alpha_srcblend != blend.color_srcblend ||
+	                                   blend.alpha_destblend != blend.color_destblend ||
+	                                   blend.alpha_comb_fcn != blend.color_comb_fcn)) {
+		return BlendMappingSupport::Unsupported;
+	}
+	auto support = BlendMappingSupport::Direct;
+	for (const auto factor: {blend.color_srcblend, blend.color_destblend}) {
+		if (BlendFactorIsDualSource(factor)) {
+			return BlendMappingSupport::Unsupported;
+		}
+		switch (static_cast<Prospero::BlendFactor>(factor)) {
+			case Prospero::BlendFactor::kSrcAlpha:
+			case Prospero::BlendFactor::kOneMinusSrcAlpha:
+				support = BlendMappingSupport::SourceAlpha;
+				break;
+			case Prospero::BlendFactor::kDstAlpha:
+			case Prospero::BlendFactor::kOneMinusDstAlpha:
+			case Prospero::BlendFactor::kSrcAlphaSaturate: return BlendMappingSupport::Unsupported;
+			default: break;
+		}
+	}
+	return support;
+}
+
+} // namespace Libs::Graphics

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

@@ -0,0 +1,27 @@
+#ifndef EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_
+#define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_
+
+#include <cstdint>
+
+namespace Libs::Graphics {
+
+namespace HW {
+struct BlendControl;
+}
+namespace Prospero {
+struct ColorComponentMapping;
+}
+
+enum class BlendMappingSupport {
+	Direct,
+	SourceAlpha, // Requires logical alpha in the second blend source.
+	Unsupported,
+};
+
+bool                BlendFactorIsDualSource(uint8_t factor);
+BlendMappingSupport ClassifyBlendMapping(const HW::BlendControl&                blend,
+                                         const Prospero::ColorComponentMapping& mapping);
+
+} // namespace Libs::Graphics
+
+#endif // EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_

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

@@ -91,19 +91,6 @@ vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture, uint32_t el
 	return {nullptr, view, texture.layout};
 }
 
-static const char* ShaderStageResourceName(ShaderType stage) {
-	switch (stage) {
-		case ShaderType::Vertex: return "Vertex";
-		case ShaderType::Mesh: return "Mesh";
-		case ShaderType::Local: return "Local";
-		case ShaderType::TessellationControl: return "Hull";
-		case ShaderType::TessellationEvaluation: return "Domain";
-		case ShaderType::Pixel: return "Pixel";
-		case ShaderType::Compute: return "Compute";
-		default: return "Unknown";
-	}
-}
-
 static Prospero::ImageType TextureType(const ShaderTextureResource& descriptor) {
 	const auto type = descriptor.Type();
 	return type == Prospero::ImageType::kCube ? Prospero::ImageType::kColor2DArray : type;
@@ -126,8 +113,7 @@ static bool IsMultisampledTexture(Prospero::ImageType type) {
 
 static vk::DescriptorBufferInfo
 NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource& source,
-                    const ShaderRecompiler::IR::BufferResource& resource, ShaderType stage,
-                    uint32_t slot, uint32_t& buffer_offset) {
+                    const ShaderRecompiler::IR::BufferResource& resource, uint32_t& buffer_offset) {
 	buffer_offset = 0;
 
 	const auto& [address, size, id] = source;
@@ -152,16 +138,6 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	if (resource.written) {
 		context.GetTextureCache().InvalidateMemoryFromGPU(address, size);
 	}
-	const char* access = "Read";
-	if (resource.written && resource.read) {
-		access = "ReadWrite";
-	} else if (resource.written) {
-		access = "Write";
-	}
-	SetVulkanObjectNameF(
-	    graphics.device, result.buffer,
-	    "Kyty.{}.StorageBuffer[slot={} guest=0x{:016x} size=0x{:x} access={} formatted={}]",
-	    ShaderStageResourceName(stage), slot, address, size, access, resource.formatted);
 	return result;
 }
 
@@ -196,8 +172,8 @@ bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor, bo
 	constexpr uint32_t htile_control = 0x00280000u;
 	const uint32_t expected_control  = htile_control | (descriptor.MsaaDepth() ? (1u << 10u) : 0u);
 	const auto     metadata_addr     = descriptor.MetaAddr() << 8u;
-	return metadata_control == expected_control && metadata_addr != 0 &&
-	       metadata_addr < TRACKER_ADDRESS_SIZE && (metadata_addr & 0x7fffu) == 0 &&
+	return metadata_control == expected_control && GuestRange {metadata_addr, 1}.Valid() &&
+	       (metadata_addr & 0x7fffu) == 0 &&
 	       descriptor.TileMode() == Prospero::TileMode::kDepth;
 }
 
@@ -738,14 +714,15 @@ static vk::Sampler NativeSampler(RenderContext&                       context,
                                  const ShaderRecompiler::IR::CompiledShaderInfo& program,
                                  uint32_t index,
                                  const ShaderRecompiler::IR::DescriptorValue& value) {
-	auto descriptor = DecodeNativeDescriptor<ShaderSamplerResource>(value);
-	if (!program.info.samplers[index].depth_compare) {
+	auto        descriptor = DecodeNativeDescriptor<ShaderSamplerResource>(value);
+	const auto& sampler = program.info.samplers[index];
+	if (!sampler.depth_compare) {
 		descriptor.fields[0] &= ~(0x7u << 12u);
 	}
-	if (program.info.samplers[index].force_point_filtering) {
+	if (sampler.force_point_filtering) {
 		descriptor.SetPointFiltering();
 	}
-	return context.GetSamplerCache().GetSampler(descriptor);
+	return context.GetSamplerCache().GetSampler(descriptor, sampler.integer_border);
 }
 
 static vk::DescriptorBufferInfo NativeUpload(RenderContext&            context,
@@ -873,7 +850,7 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 		uint32_t buffer_offset = 0;
 		prepared.buffers.push_back(NativeStorageBuffer(m_context, prepared.buffer_sources[i],
 		                                               program.info.buffers[resource],
-		                                               program.stage, resource, buffer_offset));
+		                                               buffer_offset));
 		pack_memory_offset(i, buffer_offset);
 	}
 	if (ShaderRecompiler::IR::FindBinding(

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

@@ -5,11 +5,13 @@
 #include "common/file.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
+#include "common/threads.h"
 #include "graphics/guest_gpu/hardwareContext.h"
 #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/imageView.h"
+#include "graphics/host_gpu/renderer/pipeline/blendMapping.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
@@ -29,7 +31,6 @@
 #include <span>
 #include <spirv-tools/libspirv.hpp>
 #include <string_view>
-#include <tuple>
 #include <utility>
 #include <vector>
 #include <xxhash.h>
@@ -116,33 +117,22 @@ void DumpShaderSpirv(const char* stage_name, uint64_t shader_hash,
 }
 
 void DumpShaderOriginal(const char* stage_name, uint64_t shader_hash,
-                        std::span<const uint32_t> code, const std::string& decoded_dump) {
+                        std::span<const uint32_t> code) {
 	if (!Config::GraphicsDebugDumpEnabled()) {
 		return;
 	}
 	EXIT_IF(code.empty());
 	static std::atomic_int id = 0;
-	const auto base = Config::GetShaderLogFolder() / "original" /
-	                  fmt::format("{:04d}_new_shader_{}_{:016x}", id++, stage_name, shader_hash);
-	Common::File::CreateDirectories(base.parent_path());
-	for (const auto& [suffix, data, size]: {
-	         std::tuple {".bin", static_cast<const void*>(code.data()), code.size_bytes()},
-	         std::tuple {".rdna2", static_cast<const void*>(decoded_dump.data()),
-	                     decoded_dump.size()},
-	     }) {
-		if (size == 0) {
-			continue;
-		}
-		auto path = base;
-		path += suffix;
-		Common::File file(path);
-		if (file.IsInvalid()) {
-			const auto path_text = Common::PathToString(path);
-			LOGF_COLOR(Log::Color::BrightRed, "Can't create file: %s\n", path_text.c_str());
-		} else {
-			file.Write(data, size);
-		}
+	const auto path = Config::GetShaderLogFolder() / "original" /
+	                  fmt::format("{:04d}_new_shader_{}_{:016x}.bin", id++, stage_name, shader_hash);
+	Common::File::CreateDirectories(path.parent_path());
+	Common::File file(path);
+	if (file.IsInvalid()) {
+		const auto path_text = Common::PathToString(path);
+		LOGF_COLOR(Log::Color::BrightRed, "Can't create file: %s\n", path_text.c_str());
+		return;
 	}
+	file.Write(code.data(), code.size_bytes());
 }
 
 bool ValidateShaderSpirv(const char* label, uint64_t shader_hash,
@@ -177,6 +167,32 @@ bool ValidateShaderSpirv(const char* label, uint64_t shader_hash,
 
 } // namespace
 
+std::size_t PipelineCache::GraphicsPipelineKeyHash::operator()(const GraphicsPipelineKey& key) const {
+	std::size_t hash = 0;
+	PipelineKeyHash::Mix(hash, key.rendering.color_count);
+	for (uint32_t i = 0; i < key.rendering.color_count; i++) {
+		PipelineKeyHash::Mix(hash, static_cast<uint32_t>(key.rendering.color_formats[i]));
+	}
+	PipelineKeyHash::Mix(hash, static_cast<uint32_t>(key.rendering.depth_format));
+	PipelineKeyHash::Mix(hash, static_cast<uint32_t>(key.rendering.stencil_format));
+	for (const auto id: key.vertex_shader_ids) {
+		PipelineKeyHash::Mix(hash, id);
+	}
+	PipelineKeyHash::Mix(hash, key.ps_shader_id);
+	PipelineKeyHash::Mix(hash, key.vertex_input.binding_count);
+	for (uint32_t i = 0; i < key.vertex_input.binding_count; i++) {
+		PipelineKeyHash::Mix(hash, key.vertex_input.bindings[i].stride);
+		PipelineKeyHash::Mix(hash, key.vertex_input.bindings[i].instance);
+	}
+	PipelineKeyHash::Mix(hash, key.vertex_input.attribute_count);
+	for (uint32_t i = 0; i < key.vertex_input.attribute_count; i++) {
+		PipelineKeyHash::Mix(hash, key.vertex_input.attributes[i].offset);
+		PipelineKeyHash::Mix(hash, key.vertex_input.attributes[i].binding);
+	}
+	PipelineKeyHash::Mix(hash, XXH3_64bits(&key.static_params, sizeof(key.static_params)));
+	return hash;
+}
+
 struct PipelineCache::ProgramCache {
 	struct ProgramKey {
 		ShaderType            stage           = ShaderType::Unknown;
@@ -218,32 +234,20 @@ struct PipelineCache::ProgramCache {
 			PipelineKeyHash::Mix(hash, key.code_size);
 			PipelineKeyHash::Mix(hash, key.static_state.size());
 			// Bucket same-shape static variants by source. ProgramKey equality performs the one
-			// exact state comparison needed on a stable hit without hashing up to 429 words first.
+			// exact state comparison needed on a stable hit without hashing the full state first.
 			return hash;
 		}
 	};
 
-	static constexpr std::size_t MaxStaticKeyWords = 13 + ShaderVertexInputInfo::RES_MAX * 13;
+	static constexpr std::size_t MaxStaticKeyWords = 32 + ShaderVertexInputInfo::RES_MAX * 6;
 
-	Permutation CompilePermutation(const ShaderParams&                          params,
+	Permutation CompilePermutation(const char*                                  stage_name,
 	                               const ShaderRecompiler::CompileOptions&      options,
 	                               ShaderRecompiler::TranslateResult            translated,
 	                               ShaderRecompiler::IR::ResourceSpecialization specialization,
 	                               uint32_t push_data_start_dword) {
-		const char* stage_name = nullptr;
-		switch (options.stage) {
-			case ShaderType::Vertex: stage_name = "vs"; break;
-			case ShaderType::Mesh: stage_name = "ms"; break;
-			case ShaderType::Local: stage_name = "ls"; break;
-			case ShaderType::TessellationControl: stage_name = "hs"; break;
-			case ShaderType::TessellationEvaluation: stage_name = "ds"; break;
-			case ShaderType::Pixel: stage_name = "ps"; break;
-			case ShaderType::Compute: stage_name = "cs"; break;
-			default: EXIT("invalid pipeline shader stage\n");
-		}
 		auto result = ShaderRecompiler::CompileProgram(std::move(translated), options,
 		                                               specialization, push_data_start_dword);
-		DumpShaderOriginal(stage_name, options.shader_hash, params.code, result.decoded_dump);
 		if (!ValidateShaderSpirv(options.dump_label, options.shader_hash, result.spirv)) {
 			DumpShaderSpirv(stage_name, options.shader_hash, result.spirv);
 			EXIT("%s failed hash=0x%016" PRIx64 ": SPIR-V validation failed\n", options.dump_label,
@@ -321,14 +325,15 @@ struct PipelineCache::ProgramCache {
 			stage_input.compute = &input_info;
 		}
 		const char* label = nullptr;
+		const char* stage_name = nullptr;
 		switch (stage) {
-			case ShaderType::Vertex: label = "ShaderRecompiler VS"; break;
-			case ShaderType::Mesh: label = "ShaderRecompiler MS"; break;
-			case ShaderType::Local: label = "ShaderRecompiler LS"; break;
-			case ShaderType::TessellationControl: label = "ShaderRecompiler HS"; break;
-			case ShaderType::TessellationEvaluation: label = "ShaderRecompiler DS"; break;
-			case ShaderType::Pixel: label = "ShaderRecompiler PS"; break;
-			case ShaderType::Compute: label = "ShaderRecompiler CS"; break;
+			case ShaderType::Vertex: label = "ShaderRecompiler VS"; stage_name = "vs"; break;
+			case ShaderType::Mesh: label = "ShaderRecompiler MS"; stage_name = "ms"; break;
+			case ShaderType::Local: label = "ShaderRecompiler LS"; stage_name = "ls"; break;
+			case ShaderType::TessellationControl: label = "ShaderRecompiler HS"; stage_name = "hs"; break;
+			case ShaderType::TessellationEvaluation: label = "ShaderRecompiler DS"; stage_name = "ds"; break;
+			case ShaderType::Pixel: label = "ShaderRecompiler PS"; stage_name = "ps"; break;
+			case ShaderType::Compute: label = "ShaderRecompiler CS"; stage_name = "cs"; break;
 			default: EXIT("invalid pipeline shader stage\n");
 		}
 		ShaderRecompiler::CompileOptions options;
@@ -351,6 +356,7 @@ struct PipelineCache::ProgramCache {
 		} else {
 			options.wave_size = input_info.wave_size;
 		}
+		DumpShaderOriginal(stage_name, options.shader_hash, params.code);
 		auto translated = ShaderRecompiler::TranslateProgram(params.code, options);
 		if (translated.skip_dispatch) {
 			entry = programs.try_emplace(lookup_key, ShaderRecompiler::IR::ResourcePlan {}).first;
@@ -365,7 +371,7 @@ struct PipelineCache::ProgramCache {
 			    entry->second.specialization));
 		}
 		entry->second.permutations.push_back(CompilePermutation(
-		    params, options, std::move(translated), entry->second.specialization, push_data_cursor));
+		    stage_name, options, std::move(translated), entry->second.specialization, push_data_cursor));
 		const auto& permutation = entry->second.permutations.back();
 		input_info.stage = {.program = &permutation.program, .resources = &entry->second.resources};
 		permutation.program.bindings.AdvancePushData(push_data_cursor);
@@ -514,7 +520,6 @@ void PipelineCache::InitializeDriverCache() {
 }
 
 void PipelineCache::Save() {
-	Common::LockGuard lock(m_mutex);
 	if (m_driver_cache == nullptr) {
 		return;
 	}
@@ -605,21 +610,30 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 	}
 	ShaderParams pixel_params;
 	if (pixel_active) {
-		pixel_params = PrepareProgram(pixel_regs, sh, target_export_mapping, pixel_info);
-		const auto& blend          = context.GetBlendControl(0);
-		const auto  is_dual_source = [](uint8_t factor) {
-			return factor >= static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Color) &&
-			       factor <= static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
-		};
+		pixel_params      = PrepareProgram(pixel_regs, sh, target_export_mapping, pixel_info);
+		const auto& blend = context.GetBlendControl(0);
 		pixel_info.dual_source_blending =
 		    blend.enable && !context.GetRenderTarget(0).info.blend_bypass &&
-		    (is_dual_source(blend.color_srcblend) || is_dual_source(blend.color_destblend) ||
-		     (blend.separate_alpha_blend &&
-		      (is_dual_source(blend.alpha_srcblend) || is_dual_source(blend.alpha_destblend))));
+		    (BlendFactorIsDualSource(blend.color_srcblend) ||
+		     BlendFactorIsDualSource(blend.color_destblend) ||
+		     (blend.separate_alpha_blend && (BlendFactorIsDualSource(blend.alpha_srcblend) ||
+		                                     BlendFactorIsDualSource(blend.alpha_destblend))));
 		if (pixel_info.dual_source_blending) {
-			// MRT1 supplies a second blend source for the same render target as MRT0.
+			// MRT1 supplies the second blend source for target 0.
 			pixel_info.target_output_mode[1]    = pixel_info.target_output_mode[0];
 			pixel_info.target_export_mapping[1] = pixel_info.target_export_mapping[0];
+		} else if (blend.enable && !context.GetRenderTarget(0).info.blend_bypass &&
+		           pixel_info.target_output_mode[0] != 0 && pixel_info.target_output_mode[0] != 7 &&
+		           std::all_of(std::begin(pixel_info.target_output_mode) + 1,
+		                       std::end(pixel_info.target_output_mode),
+		                       [](uint8_t mode) { return mode == 0; }) &&
+		           ClassifyBlendMapping(blend, pixel_info.target_export_mapping[0]) ==
+		               BlendMappingSupport::SourceAlpha) {
+			// Preserve logical alpha when the export mapping moves it.
+			pixel_info.alpha_blend_source_remap = true;
+			pixel_info.dual_source_blending     = true;
+			pixel_info.target_output_mode[1]    = pixel_info.target_output_mode[0];
+			pixel_info.target_export_mapping[1] = {};
 		}
 	}
 	if (context.GetClipControl().clip_disable) {
@@ -636,7 +650,6 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 		    static_cast<float>(std::min(limits.maxViewportDimensions[1], 16384u)) * 0.5f;
 		clip.enabled = true;
 	}
-	Common::LockGuard lock(m_mutex);
 	uint32_t          push_data_cursor =
 	    mesh_active ? ShaderRecompiler::IR::PushData::MeshDrawDwordCount : 0;
 	GraphicsPrograms  result;
@@ -654,7 +667,6 @@ ShaderProgram PipelineCache::GetComputeProgram(const HW::ComputeShaderInfo& regs
                                                ShaderComputeInputInfo&      input_info) {
 	input_info.host_subgroup_size = m_graphics.SupportsComputeWave64() ? 64u : 32u;
 	const auto        params      = PrepareProgram(regs, sh, input_info);
-	Common::LockGuard lock(m_mutex);
 	uint32_t          push_data_cursor = 0;
 	return m_program_cache->Get(params, input_info, push_data_cursor);
 }
@@ -679,7 +691,6 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 	EXIT_IF(ps_active && !pixel_program);
 	const auto color_count = static_cast<uint32_t>(colors.size());
 
-	Common::LockGuard lock(m_mutex);
 	auto&             ctx = command.GetRegisters();
 
 	const HW::ModeControl& mc = ctx.GetModeControl();
@@ -719,7 +730,24 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 		static_params.alpha_comb_fcn[slot]       = bc.alpha_comb_fcn;
 		static_params.alpha_destblend[slot]      = bc.alpha_destblend;
 		static_params.separate_alpha_blend[slot] = bc.separate_alpha_blend;
-		static_params.blend_enable[slot]         = bc.enable && !rt.info.blend_bypass;
+		const bool alpha_remap =
+		    slot == 0 && ps_input_info != nullptr && ps_input_info->alpha_blend_source_remap;
+		static_params.blend_enable[slot] = bc.enable && !rt.info.blend_bypass;
+		if (static_params.blend_enable[slot] && !alpha_remap &&
+		    ClassifyBlendMapping(bc, colors[i].export_mapping) != BlendMappingSupport::Direct) {
+			static_params.blend_enable[slot] = false;
+			static std::atomic_bool warned = false;
+			if (!warned.exchange(true, std::memory_order_relaxed)) {
+				Log::WriteToConsoleAndLog(fmt::format(
+				    "Warning: blending disabled for unsupported color mapping "
+				    "(slot={} mapping=0x{:02x} color={}/{} alpha={}/{} separate={}).\n",
+				    slot, colors[i].export_mapping.packed, bc.color_srcblend, bc.color_destblend,
+				    bc.alpha_srcblend, bc.alpha_destblend, bc.separate_alpha_blend ? 1 : 0));
+			}
+		}
+		if (alpha_remap) {
+			static_params.blend_alpha_source_remap = true;
+		}
 	}
 	const bool with_depth =
 	    depth.desc.view_info.format != vk::Format::eUndefined && static_cast<bool>(depth.image_id);
@@ -783,24 +811,20 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 		        vs_input_info.resources_num > ShaderVertexInputInfo::RES_MAX);
 		key.vertex_input.binding_count   = static_cast<uint8_t>(vs_input_info.buffers_num);
 		key.vertex_input.attribute_count = static_cast<uint8_t>(vs_input_info.resources_num);
-		uint32_t attributes_num          = 0;
 		for (int binding = 0; binding < vs_input_info.buffers_num; binding++) {
 			const auto& buffer = vs_input_info.buffers[binding];
-			EXIT_IF(buffer.attr_num < 0 || buffer.attr_num > ShaderVertexInputBuffer::ATTR_MAX);
-			attributes_num += static_cast<uint32_t>(buffer.attr_num);
-			EXIT_IF(attributes_num > static_cast<uint32_t>(vs_input_info.resources_num));
 			key.vertex_input.bindings[binding] = {.stride   = buffer.stride,
 			                                      .instance = buffer.fetch_index != 0};
-			for (int attribute = 0; attribute < buffer.attr_num; attribute++) {
-				const auto index = buffer.attr_indices[attribute];
-				EXIT_IF(index < 0 || index >= vs_input_info.resources_num);
-				key.vertex_input.attributes[index] = {
-				    .offset  = buffer.attr_offsets[attribute],
-				    .binding = static_cast<uint8_t>(binding),
-				};
-			}
 		}
-		EXIT_IF(attributes_num != static_cast<uint32_t>(vs_input_info.resources_num));
+		for (int attribute = 0; attribute < vs_input_info.resources_num; attribute++) {
+			const auto binding = vs_input_info.resources_dst[attribute].buffer_index;
+			EXIT_IF(binding < 0 || binding >= vs_input_info.buffers_num);
+			key.vertex_input.attributes[attribute] = {
+			    .offset = static_cast<uint32_t>(vs_input_info.resources[attribute].Base48() -
+			                                    vs_input_info.buffers[binding].addr),
+			    .binding = static_cast<uint8_t>(binding),
+			};
+		}
 	}
 
 	if (auto iter = m_graphics_pipelines.find(key); iter != m_graphics_pipelines.end()) {
@@ -839,8 +863,6 @@ PipelineCache::GetComputePipeline(const ShaderComputeInputInfo& input_info,
 
 	EXIT_IF(!compute_program);
 
-	Common::LockGuard lock(m_mutex);
-
 	if (auto iter = m_compute_pipelines.find(compute_program.id);
 	    iter != m_compute_pipelines.end()) {
 		return *iter->second;

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

@@ -4,7 +4,6 @@
 #include "common/abi.h"
 #include "common/assert.h"
 #include "common/common.h"
-#include "common/threads.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 #include "graphics/shader/shader.h"
@@ -56,6 +55,7 @@ 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                       blend_alpha_source_remap                           = false;
 
 	bool operator==(const PipelineStaticParameters& other) const noexcept;
 };
@@ -65,7 +65,7 @@ struct PipelineStaticParameters {
 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) == 125);
+static_assert(sizeof(PipelineStaticParameters) == 126);
 
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};
@@ -103,6 +103,7 @@ struct ShaderProgram {
 	explicit operator bool() const { return id != 0 && module != nullptr; }
 };
 
+// The owning renderer serializes access, including saves while the GPU is running.
 class PipelineCache {
 public:
 	explicit PipelineCache(GraphicContext& graphics);
@@ -166,45 +167,10 @@ private:
 			hash ^= value + static_cast<std::size_t>(0x9e3779b97f4a7c15ull) + (hash << 6u) +
 			        (hash >> 2u);
 		}
-
-		static void MixStaticParams(std::size_t& hash, const PipelineStaticParameters& params) {
-			const auto* bytes = reinterpret_cast<const uint8_t*>(&params);
-			for (std::size_t i = 0; i < sizeof(params); i++) {
-				Mix(hash, bytes[i]);
-			}
-		}
-
-		static void MixRendering(std::size_t& hash, const PipelineRenderingState& rendering) {
-			Mix(hash, rendering.color_count);
-			for (uint32_t i = 0; i < rendering.color_count; i++) {
-				Mix(hash, static_cast<uint32_t>(rendering.color_formats[i]));
-			}
-			Mix(hash, static_cast<uint32_t>(rendering.depth_format));
-			Mix(hash, static_cast<uint32_t>(rendering.stencil_format));
-		}
 	};
 
 	struct GraphicsPipelineKeyHash {
-		std::size_t operator()(const GraphicsPipelineKey& key) const {
-			std::size_t hash = 0;
-			PipelineKeyHash::MixRendering(hash, key.rendering);
-			for (const auto id: key.vertex_shader_ids) {
-				PipelineKeyHash::Mix(hash, id);
-			}
-			PipelineKeyHash::Mix(hash, key.ps_shader_id);
-			PipelineKeyHash::Mix(hash, key.vertex_input.binding_count);
-			for (uint32_t i = 0; i < key.vertex_input.binding_count; i++) {
-				PipelineKeyHash::Mix(hash, key.vertex_input.bindings[i].stride);
-				PipelineKeyHash::Mix(hash, key.vertex_input.bindings[i].instance);
-			}
-			PipelineKeyHash::Mix(hash, key.vertex_input.attribute_count);
-			for (uint32_t i = 0; i < key.vertex_input.attribute_count; i++) {
-				PipelineKeyHash::Mix(hash, key.vertex_input.attributes[i].offset);
-				PipelineKeyHash::Mix(hash, key.vertex_input.attributes[i].binding);
-			}
-			PipelineKeyHash::MixStaticParams(hash, key.static_params);
-			return hash;
-		}
+		std::size_t operator()(const GraphicsPipelineKey& key) const;
 	};
 
 	GraphicContext&               m_graphics;
@@ -214,7 +180,6 @@ private:
 	std::unordered_map<GraphicsPipelineKey, std::unique_ptr<Pipeline>, GraphicsPipelineKeyHash>
 	                                                        m_graphics_pipelines;
 	std::unordered_map<uint64_t, std::unique_ptr<Pipeline>> m_compute_pipelines;
-	Common::Mutex m_mutex;
 
 	void InitializeDriverCache();
 };

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

@@ -137,14 +137,17 @@ static void GetInputFormat(const ShaderBufferResource& res, vk::Format& format,
 	}
 }
 
-static vk::BlendFactor GetBlendFactor(uint32_t factor) {
+static vk::BlendFactor GetBlendFactor(uint32_t factor, bool remap_source_alpha) {
 	switch (static_cast<Prospero::BlendFactor>(factor)) {
 		case Prospero::BlendFactor::kZero: return vk::BlendFactor::eZero;
 		case Prospero::BlendFactor::kOne: return vk::BlendFactor::eOne;
 		case Prospero::BlendFactor::kSrcColor: return vk::BlendFactor::eSrcColor;
 		case Prospero::BlendFactor::kOneMinusSrcColor: return vk::BlendFactor::eOneMinusSrcColor;
-		case Prospero::BlendFactor::kSrcAlpha: return vk::BlendFactor::eSrcAlpha;
-		case Prospero::BlendFactor::kOneMinusSrcAlpha: return vk::BlendFactor::eOneMinusSrcAlpha;
+		case Prospero::BlendFactor::kSrcAlpha:
+			return remap_source_alpha ? vk::BlendFactor::eSrc1Color : vk::BlendFactor::eSrcAlpha;
+		case Prospero::BlendFactor::kOneMinusSrcAlpha:
+			return remap_source_alpha ? vk::BlendFactor::eOneMinusSrc1Color
+			                          : vk::BlendFactor::eOneMinusSrcAlpha;
 		case Prospero::BlendFactor::kDstAlpha: return vk::BlendFactor::eDstAlpha;
 		case Prospero::BlendFactor::kOneMinusDstAlpha: return vk::BlendFactor::eOneMinusDstAlpha;
 		case Prospero::BlendFactor::kDstColor: return vk::BlendFactor::eDstColor;
@@ -397,17 +400,20 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		color_blend_attachment[i].colorWriteMask =
 		    vk::ColorComponentFlags {static_params.color_mask[i]};
 		color_blend_attachment[i].blendEnable = static_params.blend_enable[i] ? VK_TRUE : VK_FALSE;
+		// Only target 0 can use the second blend source carrying logical alpha.
+		const bool remap_source_alpha         = i == 0 && static_params.blend_alpha_source_remap;
 		color_blend_attachment[i].srcColorBlendFactor =
-		    GetBlendFactor(static_params.color_srcblend[i]);
+		    GetBlendFactor(static_params.color_srcblend[i], remap_source_alpha);
 		color_blend_attachment[i].dstColorBlendFactor =
-		    GetBlendFactor(static_params.color_destblend[i]);
+		    GetBlendFactor(static_params.color_destblend[i], remap_source_alpha);
 		color_blend_attachment[i].colorBlendOp = GetBlendOp(static_params.color_comb_fcn[i]);
 		color_blend_attachment[i].srcAlphaBlendFactor =
-		    (static_params.separate_alpha_blend[i] ? GetBlendFactor(static_params.alpha_srcblend[i])
-		                                           : color_blend_attachment[i].srcColorBlendFactor);
+		    (static_params.separate_alpha_blend[i]
+		         ? GetBlendFactor(static_params.alpha_srcblend[i], remap_source_alpha)
+		         : color_blend_attachment[i].srcColorBlendFactor);
 		color_blend_attachment[i].dstAlphaBlendFactor =
 		    (static_params.separate_alpha_blend[i]
-		         ? GetBlendFactor(static_params.alpha_destblend[i])
+		         ? GetBlendFactor(static_params.alpha_destblend[i], remap_source_alpha)
 		         : color_blend_attachment[i].dstColorBlendFactor);
 		color_blend_attachment[i].alphaBlendOp =
 		    (static_params.separate_alpha_blend[i] ? GetBlendOp(static_params.alpha_comb_fcn[i])

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -249,8 +249,8 @@ static void LogDrawInputState(const CommandBuffer& buffer, const RenderColorInfo
 	for (int bi = 0; bi < vs_input_info.buffers_num; bi++) {
 		const auto& b = vs_input_info.buffers[bi];
 		LOGF("DrawInputState[%u]: vb[%d] addr=0x%010" PRIx64
-		     " stride=%u records=%u fetch_index=%u attr_num=%d\n",
-		     log_id, bi, b.addr, b.stride, b.num_records, b.fetch_index, b.attr_num);
+		     " stride=%u records=%u fetch_index=%u\n",
+		     log_id, bi, b.addr, b.stride, b.num_records, b.fetch_index);
 
 		const auto* bytes = reinterpret_cast<const uint8_t*>(b.addr);
 		if (bytes != nullptr && b.stride != 0) {
@@ -272,11 +272,13 @@ static void LogDrawInputState(const CommandBuffer& buffer, const RenderColorInfo
 				     raw[5], raw[6], raw[7], raw[8], flt[0], flt[1], flt[2], flt[3], flt[4], flt[5],
 				     flt[6], flt[7], flt[8]);
 
-				for (int ai = 0; ai < b.attr_num; ai++) {
-					const auto  res_index = b.attr_indices[ai];
-					const auto& r         = vs_input_info.resources[res_index];
-					const auto& rd        = vs_input_info.resources_dst[res_index];
-					const auto  offset    = b.attr_offsets[ai];
+				for (int ai = 0; ai < vs_input_info.resources_num; ai++) {
+					const auto& r  = vs_input_info.resources[ai];
+					const auto& rd = vs_input_info.resources_dst[ai];
+					if (rd.buffer_index != bi) {
+						continue;
+					}
+					const auto offset = static_cast<uint32_t>(r.Base48() - b.addr);
 					if (offset + 4u <= b.stride &&
 					    r.Format() == Prospero::BufferFormat::k8_8_8_8UNorm) {
 						uint32_t packed = 0;
@@ -296,14 +298,16 @@ static void LogDrawInputState(const CommandBuffer& buffer, const RenderColorInfo
 			}
 		}
 
-		for (int ai = 0; ai < b.attr_num; ai++) {
-			const auto  res_index = b.attr_indices[ai];
-			const auto& r         = vs_input_info.resources[res_index];
-			const auto& rd        = vs_input_info.resources_dst[res_index];
-			LOGF("DrawInputState[%u]: attr[%d] res=%d offset=%u dst=v%d regs=%d fetch_index=%u "
+		for (int ai = 0; ai < vs_input_info.resources_num; ai++) {
+			const auto& r  = vs_input_info.resources[ai];
+			const auto& rd = vs_input_info.resources_dst[ai];
+			if (rd.buffer_index != bi) {
+				continue;
+			}
+			LOGF("DrawInputState[%u]: attr[%d] offset=%u dst=v%d regs=%d fetch_index=%u "
 			     "sharp=%08" PRIx32 " %08" PRIx32 " %08" PRIx32 " %08" PRIx32 "\n",
-			     log_id, ai, res_index, b.attr_offsets[ai], rd.register_start, rd.registers_num,
-			     rd.fetch_index, r.fields[0], r.fields[1], r.fields[2], r.fields[3]);
+			     log_id, ai, static_cast<uint32_t>(r.Base48() - b.addr), rd.register_start,
+			     rd.registers_num, rd.fetch_index, r.fields[0], r.fields[1], r.fields[2], r.fields[3]);
 		}
 	}
 }
@@ -575,12 +579,17 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	return state;
 }
 
-static bool DrawHasActivePixelShader(const CommandBuffer& buffer) {
-	const auto& ctx              = buffer.GetRegisters();
-	const auto& sh_regs          = ctx.GetShaderRegisters();
-	const bool  has_color_output = (ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask) != 0;
-	return buffer.GetShaders().GetPs().ps_regs.data_addr != 0 &&
-	       (has_color_output || PixelShaderHasDepthOrCoverageSideEffects(sh_regs));
+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 = 0;
+	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
+		if (sh_regs.target_output_mode[slot] != 0 &&
+		    render_target_mask_slot(write_mask, slot) != 0) {
+			output_mask |= 1u << slot;
+		}
+	}
+	return output_mask;
 }
 
 enum class CbColorMode : uint8_t {
@@ -625,19 +634,22 @@ struct PreparedIndexBuffer {
 	vk::IndexType  type   = vk::IndexType::eUint16;
 };
 
-static uint64_t VertexBufferDescriptorSize(const ShaderVertexInputBuffer& buffer,
-                                           const ShaderVertexInputInfo& info) {
+static uint64_t VertexBufferDescriptorSize(int binding, const ShaderVertexInputInfo& info) {
+	const auto& buffer = info.buffers[binding];
 	if (buffer.stride != 0 || buffer.num_records == 0) {
 		return static_cast<uint64_t>(buffer.stride) * buffer.num_records;
 	}
 
 	uint64_t size = 0;
-	for (int i = 0; i < buffer.attr_num; i++) {
-		const auto& resource = info.resources[buffer.attr_indices[i]];
+	for (int i = 0; i < info.resources_num; i++) {
+		if (info.resources_dst[i].buffer_index != binding) {
+			continue;
+		}
+		const auto& resource = info.resources[i];
 		// RDNA2 OOB_SELECT=2 only checks NumRecords != 0. A constant attribute still
 		// fetches its entire format; NumRecords is not a byte count in this mode.
 		const uint64_t extent = resource.OutOfBounds() == 2
-		                            ? static_cast<uint64_t>(buffer.attr_offsets[i]) +
+		                            ? resource.Base48() - buffer.addr +
 		                                  ShaderRecompiler::Format::GetFormatInfo(resource.Format()).byte_size
 		                            : buffer.num_records;
 		size = std::max(size, extent);
@@ -674,7 +686,7 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 	uint32_t                                                      range_count = 0;
 	for (int i = 0; i < vs_input_info.buffers_num; i++) {
 		const auto& vertex = vs_input_info.buffers[i];
-		const auto  size   = VertexBufferDescriptorSize(vertex, vs_input_info);
+		const auto  size   = VertexBufferDescriptorSize(i, vs_input_info);
 		sizes[i]           = size;
 		if (size == 0) {
 			continue;
@@ -713,9 +725,6 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 		    Libs::LibKernel::Memory::ClampRangeSize(range.base_address, range.RequestedSize());
 		range.acquired_end = range.base_address + size;
 		range.binding      = cache.ObtainBuffer(range.base_address, size, false);
-		SetVulkanObjectNameF(
-		    buffer.GetContext().GetGraphics().device, range.binding.first->Handle(),
-		    "Kyty.VertexBufferRange[guest=0x{:016x} size=0x{:x}]", range.base_address, size);
 	}
 
 	// Rebuild slot bindings, offsetting non-empty slots into their acquired merged range.
@@ -747,10 +756,6 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 		prepared.buffers[i] = range->binding.first->Handle();
 		prepared.offsets[i] = range->binding.second + vertex.addr - range->base_address;
 		prepared.sizes[i]   = std::min(size, range->acquired_end - vertex.addr);
-		SetVulkanObjectNameF(
-		    buffer.GetContext().GetGraphics().device, prepared.buffers[i],
-		    "Kyty.VertexBuffer[slot={} guest=0x{:016x} size=0x{:x} stride={} records={}]", i,
-		    vertex.addr, size, vertex.stride, vertex.num_records);
 	}
 
 	return prepared;
@@ -847,7 +852,7 @@ static bool ResolvePrimitiveRestart(const CommandBuffer& buffer,
 }
 
 static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
-                           DrawRenderState& state) {
+                           uint32_t color_output_mask, DrawRenderState& state) {
 	auto& ctx    = buffer.GetRegisters();
 	auto& sh_ctx = buffer.GetShaders();
 
@@ -861,7 +866,7 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 	    target_export_mapping {};
 	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
 		const auto& rt = ctx.GetRenderTarget(slot);
-		if (rt.base.addr != 0 && render_target_mask_slot(ctx.GetRenderTargetMask(), slot) != 0) {
+		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)
@@ -880,8 +885,12 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCallInfo& draw,
                                             uint32_t            render_target_slice_offset,
 	                                        DrawRenderState& state) {
-	state.ps_active = DrawHasActivePixelShader(buffer);
-	RefreshShaders(buffer, draw, 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) {
@@ -890,6 +899,7 @@ bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCal
 			}
 		}
 	}
+	mrt_mask &= color_output_mask;
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "ResolveRenderColorTarget");
 	}
@@ -927,17 +937,11 @@ static PreparedIndexBuffer PrepareIndexBuffer(CommandBuffer&               buffe
 		auto& stream = buffer.GetContext().GetBufferCache().GetUtilityBuffer(MemoryUsage::Stream);
 		prepared.offset = stream.Copy(source.host_data, source.size, 16);
 		prepared.buffer = stream.Handle();
-		SetVulkanObjectNameF(buffer.GetContext().GetGraphics().device, prepared.buffer,
-		                     "Kyty.IndexBuffer[guest=transient size=0x{:x} type={}]", source.size,
-		                     static_cast<uint32_t>(source.type));
 	} else {
 		auto [buffer_ptr, offset] =
 		    buffer.GetContext().GetBufferCache().ObtainBuffer(source.address, source.size, false);
 		prepared.buffer = buffer_ptr->Handle();
 		prepared.offset = offset;
-		SetVulkanObjectNameF(buffer.GetContext().GetGraphics().device, prepared.buffer,
-		                     "Kyty.IndexBuffer[guest=0x{:016x} size=0x{:x} type={}]",
-		                     source.address, source.size, static_cast<uint32_t>(source.type));
 	}
 	return prepared;
 }

src/graphics/presentation/window/vulkanWindow.cpp

@@ -520,16 +520,20 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 #endif
 	depth_clip_control.depthClipControl = VK_TRUE;
 
-	auto features12  = WindowContext::RequiredVulkan12Features();
-	features12.pNext = &depth_clip_control;
-
+	const bool workgroup_layout_extension =
+	    HasExtension(device_extensions, VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME);
+	vk::PhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR supported_workgroup_layout {};
+	vk::PhysicalDeviceVulkan12Features supported_features12 {};
+	supported_features12.pNext = workgroup_layout_extension ? &supported_workgroup_layout : nullptr;
 	vk::PhysicalDeviceVulkan13Features supported_features13 {};
+	supported_features13.pNext = &supported_features12;
 
 	const auto robustness2_ext_enabled =
 	    HasExtension(device_extensions, VK_EXT_ROBUSTNESS_2_EXTENSION_NAME);
 
 	vk::PhysicalDeviceRobustness2FeaturesEXT supported_robustness2 {};
 	if (robustness2_ext_enabled) {
+		supported_robustness2.pNext = supported_features13.pNext;
 		supported_features13.pNext = &supported_robustness2;
 	}
 
@@ -557,6 +561,22 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		supported_features2.pNext = &provoking_vertex;
 	}
 	physical_device.getFeatures2(&supported_features2);
+
+	auto features12 = WindowContext::RequiredVulkan12Features();
+	features12.shaderSharedInt64Atomics = supported_features12.shaderSharedInt64Atomics;
+	vk::PhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR workgroup_layout {};
+	workgroup_layout.workgroupMemoryExplicitLayout =
+	    supported_workgroup_layout.workgroupMemoryExplicitLayout;
+	workgroup_layout.pNext = &depth_clip_control;
+	features12.pNext = workgroup_layout_extension ? static_cast<void*>(&workgroup_layout)
+	                                             : static_cast<void*>(&depth_clip_control);
+	if (!features12.shaderSharedInt64Atomics || !workgroup_layout.workgroupMemoryExplicitLayout) {
+		Log::WriteToConsoleAndLog(fmt::format(
+		    "WARNING: Native 64-bit LDS atomics are unavailable: shaderSharedInt64Atomics={}, "
+		    "workgroupMemoryExplicitLayout={}. Continuing with the selected Vulkan device.\n",
+		    features12.shaderSharedInt64Atomics != VK_FALSE,
+		    workgroup_layout.workgroupMemoryExplicitLayout != VK_FALSE));
+	}
 	graphics.mesh_shader_enabled = mesh_extension && supported_mesh.meshShader;
 
 	vk::PhysicalDeviceSubgroupSizeControlProperties subgroup_size_control {};
@@ -1049,6 +1069,7 @@ void WindowContext::CreateVulkan() {
 		for (const auto* extension: {VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
 		                             VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME,
 		                             VK_EXT_MESH_SHADER_EXTENSION_NAME,
+		                             VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME,
 		                             VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME}) {
 			if (HasExtension(available_extensions, extension)) {
 				device_extensions.push_back(extension);

src/graphics/presentation/window/window.cpp

@@ -817,6 +817,7 @@ void WindowRun() {
 	EXIT_IF(g_window == nullptr);
 
 	g_window->Run();
+	Common::LockGuard lock(g_window->render_context->GetMutex());
 	g_window->render_context->GetPipelineCache().Save();
 }
 

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -15,7 +15,6 @@
 #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 "graphics/shader/recompiler/ir/passes/SsaRewrite.h"
 
 #include <algorithm>
@@ -548,30 +547,40 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 
 	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " CFG BuildGraph\n", GetDumpLabel(options),
 	     StageName(options.stage), options.shader_hash);
-	auto cfg = CFG::BuildGraph(decoded);
+	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
 	     " loops=%" PRIu64 " back_edges=%" PRIu64 " elapsed_ms=%" PRIu64 "\n",
 	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
-	     static_cast<uint64_t>(cfg.blocks.size()), static_cast<uint64_t>(cfg.natural_loops.size()),
-	     static_cast<uint64_t>(cfg.back_edges.size()), phase_ms());
-	if (cfg.irreducible) {
-		LogDispatcherFallback(options, cfg, "build");
+	     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);
-		if (!CFG::Structurize(cfg)) {
-			LogDispatcherFallback(options, cfg, "structurize");
+		structured_cfg = CFG::Structurize(native_cfg);
+		if (structured_cfg.unsupported) {
+			native_cfg.unsupported = true;
+			native_cfg.failure_kind = structured_cfg.failure_kind;
+			native_cfg.failure_block = structured_cfg.failure_block;
+			native_cfg.unsupported_reason = structured_cfg.unsupported_reason;
+			LogDispatcherFallback(options, native_cfg, "structurize");
 		} else {
+			selected_cfg = &structured_cfg;
 			LOGF("%s structured CFG success: blocks=%" PRIu64 "\n", GetDumpLabel(options),
-			     static_cast<uint64_t>(cfg.blocks.size()));
+			     static_cast<uint64_t>(selected_cfg->blocks.size()));
 		}
 		LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " CFG Structurize blocks=%" PRIu64
 		     " loops=%" PRIu64 " elapsed_ms=%" PRIu64 "\n",
 		     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
-		     static_cast<uint64_t>(cfg.blocks.size()),
-		     static_cast<uint64_t>(cfg.natural_loops.size()), phase_ms());
+		     static_cast<uint64_t>(selected_cfg->blocks.size()),
+		     static_cast<uint64_t>(selected_cfg->natural_loops.size()), phase_ms());
 	}
 
+	const auto& cfg = *selected_cfg;
 	Frontend::EmbeddedFetchPlan embedded_fetch;
 	if ((options.stage == ShaderType::Vertex || options.stage == ShaderType::Local) &&
 	    options.input_info.vertex != nullptr && options.input_info.vertex->fetch_embedded) {
@@ -614,15 +623,21 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 		IR::EliminateDeadCode(ir.blocks);
 	}
 	LowerTessellationMemory(ir, options);
-	IR::BuildSrtPlan(ir);
-	IR::EliminateDeadCode(ir.blocks);
-	IR::TrackResources(ir);
+	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());
+		}
+	}
+	IR::TrackResources(ir, decoded, native_cfg);
 	IR::EliminateDeadCode(ir.blocks);
 	TranslateResult result;
 	result.program = std::move(ir);
 	if (options.dump_ir) {
 		result.decoded_dump = std::move(decoded_dump);
-		result.cfg_dump     = CFG::GraphToString(cfg);
+		result.cfg_dump     = std::move(cfg_dump);
 	}
 	return result;
 }
@@ -634,6 +649,35 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 	const auto emit_begin = std::chrono::steady_clock::now();
 	auto& ir = translated.program;
 	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.
+	for (auto& inst: ir.value_storage) {
+		inst.Invalidate();
+	}
+	for (auto* block: ir.blocks) {
+		for (auto& inst: *block) {
+			const auto op = inst.GetOpcode();
+			uint32_t first = 0;
+			if (op == IR::ValueOpcode::GetBufferResource) {
+				if (std::ranges::any_of(inst.Uses(), [&](const IR::Use& use) {
+					return ir.memory_info[use.user->Flags<IR::MemoryFlags>().index].kind ==
+					       IR::ResourceKind::IndirectBuffer;
+				})) {
+					continue;
+				}
+			} else if (op == IR::ValueOpcode::GetImageResource) {
+				const auto resource = inst.Flags<uint32_t>();
+				first = resource < ir.info.images.size() &&
+				                ir.info.images[resource].indirect_root == resource ? 1u : 0u;
+			} else if (op != IR::ValueOpcode::GetSamplerResource) {
+				continue;
+			}
+			for (size_t index = first; index < inst.NumArgs(); index++) {
+				inst.SetArg(index, IR::Value(0u));
+			}
+		}
+	}
+	ir.value_storage.clear();
 	IR::RemoveIdentities(ir.blocks);
 	IR::EliminateDeadCode(ir.blocks);
 

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

@@ -62,12 +62,7 @@ void ValidateNativeProgram(const IR::Program& program) {
 		Expect(Kind::BdaPagetable);
 		Expect(Kind::FaultBuffer);
 	}
-	const bool uses_flattened_runtime =
-	    !program.srt_reads.empty() ||
-	     std::ranges::any_of(program.info.images, [](const IR::ImageResource& image) {
-		     return image.indirect_search_iterations != 0u;
-	     });
-	if (uses_flattened_runtime) {
+	if (IR::UsesFlattenedSrt(program)) {
 		Expect(Kind::FlattenedSrt);
 	}
 	if (program.bindings.ShaderDataDwords() != 0 && !program.bindings.UsesPushData()) {
@@ -234,6 +229,12 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				if (kind != IR::ResourceKind::Lds && kind != IR::ResourceKind::Gds) {
 					Fail(program, "shared operation has invalid resource kind");
 				}
+				if (inst.GetOpcode() == IR::ValueOpcode::SharedAtomicOr64) {
+					if (kind != IR::ResourceKind::Lds || program.stage != ShaderType::Compute) {
+						Fail(program, "64-bit shared atomics require compute LDS");
+					}
+					requirements.shared_int64_atomics = true;
+				}
 				if (program.stage != ShaderType::Compute && program.stage != ShaderType::Mesh &&
 				    kind == IR::ResourceKind::Lds) {
 					requirements.function_lds = true;

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

@@ -307,6 +307,11 @@ uint32_t EmitConvertF32F64(EmitterState& state, uint32_t arg0) {
 	                  Unary(state, spv::OpBitcast, TypeF32(state), clamped), converted));
 }
 
+uint32_t EmitConvertF64F32(EmitterState& state, uint32_t arg0) {
+	return EmitNative<spv::OpFConvert, IR::Type::F64>(state,
+	                                                  EmitFlushF32DenormToSignedZero(state, arg0));
+}
+
 uint32_t EmitCompositeExtractU64(EmitterState& state, uint32_t arg0, IR::Value arg1) {
 	return EmitNative<spv::OpCompositeExtract, IR::Type::U32>(state, arg0, arg1.U32());
 }

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

@@ -203,9 +203,7 @@ uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {
 	const auto bits      = state.builder.AllocateId();
 	const auto abs_bits  = state.builder.AllocateId();
 	const auto sign_bits = state.builder.AllocateId();
-	const auto non_zero  = state.builder.AllocateId();
 	const auto subnormal = state.builder.AllocateId();
-	const auto flush     = state.builder.AllocateId();
 	const auto selected  = state.builder.AllocateId();
 	const auto ret       = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
@@ -213,12 +211,9 @@ uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {
 	                          ConstantU32(state, 0x7fffffffu));
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), sign_bits, bits,
 	                          ConstantU32(state, 0x80000000u));
-	state.builder.AddFunction(spv::OpINotEqual, TypeBool(state), non_zero, abs_bits,
-	                          ConstantU32(state, 0));
 	state.builder.AddFunction(spv::OpULessThan, TypeBool(state), subnormal, abs_bits,
 	                          ConstantU32(state, 0x00800000u));
-	state.builder.AddFunction(spv::OpLogicalAnd, TypeBool(state), flush, non_zero, subnormal);
-	state.builder.AddFunction(spv::OpSelect, TypeU32(state), selected, flush, sign_bits, bits);
+	state.builder.AddFunction(spv::OpSelect, TypeU32(state), selected, subnormal, sign_bits, bits);
 	state.builder.AddFunction(spv::OpBitcast, TypeF32(state), ret, selected);
 	return ret;
 }

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

@@ -452,6 +452,11 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (exp.kind == IR::ExportTargetKind::Null || exp.en == 0u) {
 		return;
 	}
+	// Skip dormant color exports after their valid mask; MRT1 is reserved for logical alpha.
+	if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
+	    exp.index != 0 && state.input_info.pixel->alpha_blend_source_remap) {
+		return;
+	}
 	EmitIfCondition(state, exec, [&]() {
 		const auto data = ctx.Arg(inst, 0);
 		if (exp.kind == IR::ExportTargetKind::Primitive) {
@@ -492,6 +497,18 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const bool uint_output = MrtOutputMode(state, exp) == 7u;
 		const auto vector_type = uint_output ? TypeU32Vector(state, 4) : TypeF32Vector(state, 4);
 		auto       value       = ExportVector(ctx, data, exp, uint_output);
+		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
+		    exp.index == 0 && !uint_output && state.input_info.pixel->alpha_blend_source_remap) {
+			// Broadcast logical alpha before swizzling the primary output.
+			const auto blend_output =
+			    OutputVariableForExport(state, {.kind = IR::ExportTargetKind::Mrt, .index = 1});
+			if (blend_output != 0) {
+				const auto alpha = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpVectorShuffle, vector_type, alpha, value, value,
+				                          3u, 3u, 3u, 3u);
+				state.builder.AddFunction(spv::OpStore, blend_output, alpha);
+			}
+		}
 		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
 		    exp.index < state.input_info.pixel->target_export_mapping.size()) {
 			const auto mapping = state.input_info.pixel->target_export_mapping[exp.index];
@@ -636,6 +653,29 @@ uint32_t EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	                                                ctx.Arg(inst, 1));
 }
 
+uint32_t EmitConditionRef(ValueEmitContext& ctx, const IR::Inst& inst) {
+	if (ctx.other_half == nullptr) return ctx.Arg(inst, 0);
+	// A native scalar branch makes one decision for both emulated wave halves.
+	if (ctx.half != 0) return ctx.other_half->Def(IR::Value(&inst));
+	const auto kind = inst.Flags<CFG::BranchCondition>();
+	if (kind == CFG::BranchCondition::ScalarInstruction) return ctx.Arg(inst, 0);
+	const auto ballot = ctx.Ballot(inst.Arg(0));
+	const auto low = ctx.state.builder.AllocateId();
+	const auto high = ctx.state.builder.AllocateId();
+	const auto combined = ctx.state.builder.AllocateId();
+	const auto result = ctx.state.builder.AllocateId();
+	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), low, ballot, 0);
+	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), high, ballot, 1);
+	const bool zero = kind == CFG::BranchCondition::ExecZero ||
+	                  kind == CFG::BranchCondition::VccZero || kind == CFG::BranchCondition::SccZero;
+	ctx.state.builder.AddFunction(zero ? spv::OpBitwiseAnd : spv::OpBitwiseOr,
+	                              TypeU32(ctx.state), combined, low, high);
+	ctx.state.builder.AddFunction(zero ? spv::OpIEqual : spv::OpINotEqual,
+	                              TypeBool(ctx.state), result, combined,
+	                              ConstantU32(ctx.state, zero ? ~0u : 0u));
+	return result;
+}
+
 uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate) {
 	return ctx.Ballot(predicate);
 }

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

@@ -2,6 +2,8 @@
 
 #include "graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h"
 
+#include <bit>
+
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 // Operand signatures distinguish SPIR-V IDs (uint32_t), raw literals/values
 // (IR::Value), and instructions whose metadata or operand loads must stay lazy.
@@ -29,7 +31,9 @@ uint32_t EmitConvertSigned32ToFloat(EmitterState& state, uint32_t arg0) {
 inline constexpr auto EmitConvertF32S32 = EmitConvertSigned32ToFloat<IR::Type::F32>;
 inline constexpr auto EmitConvertF64S32 = EmitConvertSigned32ToFloat<IR::Type::F64>;
 uint32_t              EmitConvertF32F64(EmitterState& state, uint32_t arg0);
+uint32_t              EmitConvertF64F32(EmitterState& state, uint32_t arg0);
 EMIT_NATIVE(ConvertF32U32, OpConvertUToF, F32, uint32_t)
+EMIT_NATIVE(ConvertF64U32, OpConvertUToF, F64, uint32_t)
 EMIT_NATIVE(CompositeConstructU64, OpCompositeConstruct, U64, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructU32x2, OpCompositeConstruct, U32x2, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructU32x3, OpCompositeConstruct, U32x3, uint32_t, uint32_t, uint32_t)
@@ -112,18 +116,34 @@ EMIT_NATIVE(LogicalOr, OpLogicalOr, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalAnd, OpLogicalAnd, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalXor, OpLogicalNotEqual, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalNot, OpLogicalNot, U1, uint32_t)
-EMIT_NATIVE(FPOrdEqual32, OpFOrdEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordEqual32, OpFUnordEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdNotEqual32, OpFOrdNotEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordNotEqual32, OpFUnordNotEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdLessThan32, OpFOrdLessThan, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordLessThan32, OpFUnordLessThan, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdGreaterThan32, OpFOrdGreaterThan, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordGreaterThan32, OpFUnordGreaterThan, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdLessThanEqual32, OpFOrdLessThanEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordLessThanEqual32, OpFUnordLessThanEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdGreaterThanEqual32, OpFOrdGreaterThanEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordGreaterThanEqual32, OpFUnordGreaterThanEqual, U1, uint32_t, uint32_t)
+template <spv::Op opcode>
+uint32_t EmitFloatCompare32(ValueEmitContext& ctx, const IR::Inst& inst) {
+	// SPIR-V's DenormFlushToZero mode does not require comparison operands to flush.
+	const bool flush = inst.Flags<IR::FPCompareFlags>().flush_input_denorms;
+	const auto operand = [&](size_t index) {
+		const auto value = inst.Arg(index);
+		if (flush && value.IsImmediate()) {
+			const auto bits = std::bit_cast<uint32_t>(value.F32Value());
+			return ConstantF32(ctx.state, (bits & 0x7fffffffu) < 0x00800000u
+			                                  ? bits & 0x80000000u : bits);
+		}
+		const auto id = ctx.Arg(inst, index);
+		return flush ? EmitFlushF32DenormToSignedZero(ctx.state, id) : id;
+	};
+	return EmitNative<opcode, IR::Type::U1>(ctx.state, operand(0), operand(1));
+}
+inline constexpr auto EmitFPOrdEqual32 = EmitFloatCompare32<spv::OpFOrdEqual>;
+inline constexpr auto EmitFPUnordEqual32 = EmitFloatCompare32<spv::OpFUnordEqual>;
+inline constexpr auto EmitFPOrdNotEqual32 = EmitFloatCompare32<spv::OpFOrdNotEqual>;
+inline constexpr auto EmitFPUnordNotEqual32 = EmitFloatCompare32<spv::OpFUnordNotEqual>;
+inline constexpr auto EmitFPOrdLessThan32 = EmitFloatCompare32<spv::OpFOrdLessThan>;
+inline constexpr auto EmitFPUnordLessThan32 = EmitFloatCompare32<spv::OpFUnordLessThan>;
+inline constexpr auto EmitFPOrdGreaterThan32 = EmitFloatCompare32<spv::OpFOrdGreaterThan>;
+inline constexpr auto EmitFPUnordGreaterThan32 = EmitFloatCompare32<spv::OpFUnordGreaterThan>;
+inline constexpr auto EmitFPOrdLessThanEqual32 = EmitFloatCompare32<spv::OpFOrdLessThanEqual>;
+inline constexpr auto EmitFPUnordLessThanEqual32 = EmitFloatCompare32<spv::OpFUnordLessThanEqual>;
+inline constexpr auto EmitFPOrdGreaterThanEqual32 = EmitFloatCompare32<spv::OpFOrdGreaterThanEqual>;
+inline constexpr auto EmitFPUnordGreaterThanEqual32 = EmitFloatCompare32<spv::OpFUnordGreaterThanEqual>;
 uint32_t              EmitFPIsNan32(EmitterState& state, uint32_t arg0);
 inline constexpr auto EmitFPCmpClass32 = EmitClassMaskF32;
 EMIT_NATIVE(FPAdd32, OpFAdd, F32, uint32_t, uint32_t)
@@ -182,6 +202,7 @@ uint32_t              EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& in
 uint32_t              EmitWqmU64(EmitterState& state, uint32_t value);
 uint32_t              EmitLaneId(EmitterState& state);
 uint32_t              EmitBallot(ValueEmitContext& ctx, IR::Value predicate);
+uint32_t              EmitConditionRef(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -201,6 +222,7 @@ void                  EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst
 void                  EmitStoreMemory(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitAtomic32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
+void                  EmitSharedAtomicOr64(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitSharedIncDec(ValueEmitContext& ctx, const IR::Inst& inst);

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

@@ -67,6 +67,7 @@ struct SpirvRequirements {
 	bool function_scratch             = false;
 	bool pixel_valid_mask             = false;
 	bool buffer_int64_atomics         = false;
+	bool shared_int64_atomics         = false;
 	bool coherent_buffers             = false;
 	bool float64                      = false;
 };
@@ -101,6 +102,7 @@ struct EmitterState {
 	uint32_t                                         shader_data_storage_variable = 0;
 	uint32_t                                         flattened_srt_variable  = 0;
 	uint32_t                                         lds_variable            = 0;
+	uint32_t                                         lds_u64_variable        = 0;
 	std::array<uint32_t, 2>                          scratch_variable {};
 	std::array<uint32_t, IR::ImageBindingCount>      image_variables {};
 	uint32_t                   sampler_variable                      = 0;
@@ -376,6 +378,7 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
 void EmitMemoryOffsets(EmitterState& state);
 
 uint32_t LdsDwordCount(const EmitterState& state);
+void EnsureLdsStorage(EmitterState& state);
 
 struct MemoryResourceAccess {
 	IR::ResourceKind      kind             = IR::ResourceKind::None;

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

@@ -146,7 +146,12 @@ uint32_t GuestAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Mem
 			ctx.Fail(inst, "has no address base pair");
 			return ConstantDeviceAddress(state, 0);
 		}
-		const auto base = DeviceAddressFromWords(state, ctx.Arg(*handle, 0), ctx.Arg(*handle, 1));
+		auto base_low = ctx.Arg(*handle, 0);
+		if (mem.kind == IR::ResourceKind::ScalarAddress) {
+			base_low = Binary(state, spv::OpBitwiseAnd, TypeU32(state), base_low,
+			                  ConstantU32(state, ~3u));
+		}
+		const auto base = DeviceAddressFromWords(state, base_low, ctx.Arg(*handle, 1));
 		address         = Binary(state, spv::OpIAdd, TypeScalarU64(state), base,
 		                         Unary(state, spv::OpUConvert, TypeScalarU64(state), low));
 	}
@@ -543,6 +548,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::SharedAtomicAnd32: return spv::OpAtomicAnd;
 		case IR::ValueOpcode::BufferAtomicOr32:
 		case IR::ValueOpcode::BufferAtomicOr64:
+		case IR::ValueOpcode::SharedAtomicOr64:
 		case IR::ValueOpcode::SharedAtomicOr32: return spv::OpAtomicOr;
 		case IR::ValueOpcode::BufferAtomicXor32:
 		case IR::ValueOpcode::SharedAtomicXor32: return spv::OpAtomicXor;
@@ -970,7 +976,14 @@ void DefineGetBdaPointer(EmitterState& state) {
 	state.builder.AddFunction(spv::OpFunctionParameter, type, address);
 	EmitLabel(state, entry_label);
 
-	const auto page64        = Binary(state, spv::OpShiftRightLogical, type, address,
+	const auto extended = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), address,
+	                             ConstantDeviceAddress(state, LibKernel::Memory::kExtendedMemoryBase));
+	const auto packed = Select(state, type, extended,
+	                           Binary(state, spv::OpISub, type, address,
+	                                  ConstantDeviceAddress(state,
+	                                      LibKernel::Memory::kExtendedMemoryBase - LOWER_ADDRESS_SIZE)),
+	                           address);
+	const auto page64        = Binary(state, spv::OpShiftRightLogical, type, packed,
 	                                  ConstantDeviceAddress(state, BufferCache::CACHING_PAGEBITS));
 	const auto page          = Unary(state, spv::OpUConvert, TypeU32(state), page64);
 	const auto entry_pointer = state.builder.AllocateId();
@@ -1052,6 +1065,32 @@ uint32_t EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst) {
 	    });
 }
 
+void EmitSharedAtomicOr64(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& state = ctx.state;
+	const auto& mem = ctx.Memory(inst);
+	EnsureLdsStorage(state);
+	EmitIfCondition(state, ctx.Arg(inst, 2), [&]() {
+		const auto address = Binary(state, spv::OpBitwiseAnd, TypeU32(state),
+		                            ByteAddress(ctx, inst, mem), ConstantU32(state, 0xfff8u));
+		const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), address,
+		                          ConstantU32(state, 3u));
+		const auto in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index,
+		                              ConstantU32(state, LdsDwordCount(state) / 2u));
+		EmitIfCondition(state, in_bounds, [&]() {
+			const auto pointer = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpAccessChain,
+			                          TypePointer(state, spv::StorageClassWorkgroup, TypeScalarU64(state)),
+			                          pointer, state.lds_u64_variable, ConstantU32(state, 0), index);
+			const auto value = Unary(state, spv::OpBitcast, TypeScalarU64(state), ctx.Arg(inst, 1));
+			state.builder.AddFunction(
+			    SpirvAtomicOpcode(inst.GetOpcode()), TypeScalarU64(state), state.builder.AllocateId(),
+			    pointer, ConstantU32(state, spv::ScopeWorkgroup),
+			    ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask |
+			                           spv::MemorySemanticsWorkgroupMemoryMask), value);
+		});
+	});
+}
+
 uint32_t EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto& mem       = ctx.Memory(inst);
 	const bool  max_value = inst.GetOpcode() == IR::ValueOpcode::BufferAtomicFMax32;
@@ -1099,16 +1138,27 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (ctx.half == 1) {
 		return ctx.other_half->Def(IR::Value(const_cast<IR::Inst*>(&inst)));
 	}
-	const auto m0 = ctx.Arg(inst, 0);
-	const auto base =
-	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), m0, ConstantU32(state, 16));
-	const auto size =
-	    Binary(state, spv::OpBitwiseAnd, TypeU32(state), m0, ConstantU32(state, 0xffffu));
-	const auto address = Binary(state, spv::OpIAdd, TypeU32(state), base,
-	                            ConstantU32(state, ctx.Memory(inst).offset));
+	const auto mem           = ctx.Memory(inst);
+	const auto offset        = mem.offset & 0xfffcu;
+	auto       address       = ConstantU32(state, offset);
+	uint32_t   region_bounds = 0;
+	if (mem.kind == IR::ResourceKind::Gds) {
+		const auto m0 = ctx.Arg(inst, 0);
+		const auto base =
+		    Binary(state, spv::OpShiftRightLogical, TypeU32(state), m0, ConstantU32(state, 16));
+		const auto size =
+		    Binary(state, spv::OpBitwiseAnd, TypeU32(state), m0, ConstantU32(state, 0xffffu));
+		address = Binary(state, spv::OpIAdd, TypeU32(state), base, address);
+		// The entire M0 region must fit the PS5's 48 KiB GDS partition.
+		region_bounds = AndCondition(
+		    state, Binary(state, spv::OpULessThan, TypeBool(state),
+		                  ConstantU32(state, offset + 3u), size),
+		    Binary(state, spv::OpULessThanEqual, TypeBool(state),
+		           Binary(state, spv::OpIAdd, TypeU32(state), base, size),
+		           ConstantU32(state, 0xc000u)));
+	}
 	const auto raw_index =
 	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), address, ConstantU32(state, 2));
-	const auto mem    = ctx.Memory(inst);
 	const auto access = PrepareMemoryResourceAccess(state, mem);
 	const auto index  = EmitMemoryElementIndex(state, access, raw_index);
 	const auto exec   = ctx.Arg(inst, 1);
@@ -1127,19 +1177,17 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
 	        : first;
 	const auto is_first       = Binary(state, spv::OpIEqual, TypeBool(state),
 	                                   EmitSubgroupLocalInvocationId(state), source_lane);
-	const auto storage_bounds = EmitMemoryElementInBounds(state, access, index);
-	const auto m0_bounds =
-	    mem.kind == IR::ResourceKind::Gds
-	        ? Binary(state, spv::OpINotEqual, TypeBool(state), size, ConstantU32(state, 0))
-	        : Binary(state, spv::OpULessThan, TypeBool(state),
-	                 ConstantU32(state, ctx.Memory(inst).offset + 3u), size);
+	auto bounds = EmitMemoryElementInBounds(state, access, index);
+	if (mem.kind == IR::ResourceKind::Gds) {
+		bounds = AndCondition(state, bounds, region_bounds);
+	}
 	const auto condition = AndCondition(
 	    state, is_first,
 	    AndCondition(state,
 	                 state.lane_count == 2 ? Binary(state, spv::OpINotEqual, TypeBool(state), count,
 	                                                ConstantU32(state, 0))
 	                                       : exec,
-	                 AndCondition(state, storage_bounds, m0_bounds)));
+	                 bounds));
 	const auto atomic = EmitValueOrZeroIfCondition(state, condition, [&]() {
 		const auto value = state.builder.AllocateId();
 		state.builder.AddFunction(append ? spv::OpAtomicIAdd : spv::OpAtomicISub, TypeU32(state),
@@ -1173,10 +1221,11 @@ void EmitReadConstBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (mem.planning_only) return;
 	auto& state        = ctx.state;
 	mem.kind           = IR::ResourceKind::ScalarBuffer;
-	const auto address = Binary(state, spv::OpIAdd, TypeU32(state), ctx.Arg(inst, 1),
-	                            ConstantU32(state, mem.offset));
-	const auto index =
-	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), address, ConstantU32(state, 2));
+	auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), ctx.Arg(inst, 1),
+	                    ConstantU32(state, 2));
+	if (mem.offset >= 4u) {
+		index = Binary(state, spv::OpIAdd, TypeU32(state), index, ConstantU32(state, mem.offset >> 2u));
+	}
 	const auto access    = PrepareMemoryResourceAccess(state, mem);
 	const auto element   = EmitMemoryElementIndex(state, access, index);
 	const auto condition = EmitMemoryElementInBounds(state, access, element);
@@ -1198,6 +1247,8 @@ void EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst) {
 		value = LoadWideBuffer(ctx, inst, buffer_components);
 	else if (shared_components > 1u)
 		value = LoadWideShared(ctx, inst, shared_components);
+	else if (mem.kind == IR::ResourceKind::ScalarAddress)
+		value = LoadBdaDword(ctx, GuestAddress(ctx, inst, mem));
 	else if (address_info.access == IR::AddressAccess::Read &&
 	         mem.kind != IR::ResourceKind::Scratch)
 		value = LoadBda(ctx, GuestAddress(ctx, inst, mem), ctx.Arg(inst, inst.NumArgs() - 1),

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

@@ -63,16 +63,34 @@ uint32_t LdsDwordCount(const EmitterState& state) {
 	return workgroup != nullptr ? workgroup->lds_size_dwords : 8192u;
 }
 
-static void EnsureLdsStorage(EmitterState& state) {
+void EnsureLdsStorage(EmitterState& state) {
 	if (state.lds_variable != 0) {
 		return;
 	}
 	if (ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
 		EXIT("function LDS was not prepared before SPIR-V function emission\n");
 	}
-	state.lds_variable = state.builder.DefineGlobalVariable(
-	    TypeU32ArrayPointer(state, spv::StorageClassWorkgroup, LdsDwordCount(state)),
-	    spv::StorageClassWorkgroup);
+	const auto define = [&](uint32_t type, uint32_t bytes) {
+		const auto array = state.builder.DecoratedType(
+		    spv::OpTypeArray, {{spv::OpDecorate, {spv::DecorationArrayStride, bytes}}}, type,
+		    ConstantU32(state, std::max(LdsDwordCount(state) * 4u / bytes, 1u)));
+		const auto block = state.builder.DecoratedType(
+		    spv::OpTypeStruct, {{spv::OpMemberDecorate, {0, spv::DecorationOffset, 0}},
+		                        {spv::OpDecorate, {spv::DecorationBlock}}}, array);
+		const auto variable = state.builder.DefineGlobalVariable(
+		    TypePointer(state, spv::StorageClassWorkgroup, block), spv::StorageClassWorkgroup);
+		state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationAliased);
+		return variable;
+	};
+	if (state.requirements.shared_int64_atomics) {
+		state.lds_variable = define(TypeU32(state), 4u);
+		state.lds_u64_variable = define(TypeScalarU64(state), 8u);
+		state.builder.AddName(state.lds_u64_variable, "lds_qwords");
+	} else {
+		state.lds_variable = state.builder.DefineGlobalVariable(
+		    TypeU32ArrayPointer(state, spv::StorageClassWorkgroup, LdsDwordCount(state)),
+		    spv::StorageClassWorkgroup);
+	}
 	state.builder.AddName(state.lds_variable, "lds_dwords");
 }
 
@@ -167,6 +185,11 @@ uint32_t EmitMemoryElementPointer(EmitterState& state, const MemoryResourceAcces
 		    : ShaderWorkgroupInput(state.program.stage, state.input_info) != nullptr
 		        ? spv::StorageClassWorkgroup
 		        : spv::StorageClassFunction;
+		if (access.kind == IR::ResourceKind::Lds && state.requirements.shared_int64_atomics) {
+			state.builder.AddFunction(spv::OpAccessChain, TypeU32ElementPointer(state, storage_class),
+			                          pointer, access.object_pointer, ConstantU32(state, 0), index);
+			return pointer;
+		}
 		state.builder.AddFunction(spv::OpAccessChain, TypeU32ElementPointer(state, storage_class),
 		                          pointer, access.object_pointer, index);
 		return pointer;

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

@@ -626,10 +626,15 @@ void DefineModule(EmitterState& state) {
 		state.builder.RequireCapability(spv::CapabilityPhysicalStorageBufferAddresses);
 		state.builder.RequireExtension("SPV_KHR_physical_storage_buffer");
 	}
-	if (state.requirements.buffer_int64_atomics) {
+	if (state.requirements.buffer_int64_atomics || state.requirements.shared_int64_atomics) {
 		state.builder.RequireCapability(spv::CapabilityInt64);
 		state.builder.RequireCapability(spv::CapabilityInt64Atomics);
 	}
+	if (state.requirements.shared_int64_atomics) {
+		state.builder.RequireVersion(0x00010400u);
+		state.builder.RequireExtension("SPV_KHR_workgroup_memory_explicit_layout");
+		state.builder.RequireCapability(spv::CapabilityWorkgroupMemoryExplicitLayoutKHR);
+	}
 	if (state.clip_distance_variable != 0) {
 		state.builder.RequireCapability(spv::CapabilityClipDistance);
 	}
@@ -684,8 +689,8 @@ void DefineModule(EmitterState& state) {
 	state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSignedZeroInfNanPreserve,
 	                               32u);
 	if (state.requirements.float64) {
-		EXIT_NOT_IMPLEMENTED(state.program.stage != ShaderType::Compute);
-		EXIT_NOT_IMPLEMENTED(state.input_info.compute->float_mode != 0xc0);
+		EXIT_NOT_IMPLEMENTED(state.program.stage == ShaderType::Compute &&
+		                     state.input_info.compute->float_mode != 0xc0);
 		// MODE=0xc0 uses round-to-nearest-even and preserves FP64 input/output denormals.
 		state.builder.RequireCapability(spv::CapabilityFloat64);
 		state.builder.RequireCapability(spv::CapabilityRoundingModeRTE);

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

@@ -97,30 +97,6 @@ void EmitReturn(ValueEmitContext& ctx) {
 	ctx.state.builder.AddFunction(spv::OpReturn);
 }
 
-uint32_t BranchCondition(ValueEmitContext& ctx, const IR::BlockInfo& info) {
-	// Scalar-instruction conditions already test the full wave's raw register values.
-	if (ctx.other_half == nullptr ||
-	    info.terminator.condition == CFG::BranchCondition::ScalarInstruction ||
-	    info.terminator.condition == CFG::BranchCondition::GotoVariable) {
-		return ctx.Def(info.condition);
-	}
-	const auto ballot = ctx.Ballot(info.condition);
-	const auto low    = ctx.state.builder.AllocateId();
-	const auto high   = ctx.state.builder.AllocateId();
-	const auto result = ctx.state.builder.AllocateId();
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), low, ballot, 0);
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), high, ballot, 1);
-	const auto kind     = info.terminator.condition;
-	const bool zero     = kind == CFG::BranchCondition::ExecZero ||
-	                      kind == CFG::BranchCondition::VccZero ||
-	                      kind == CFG::BranchCondition::SccZero;
-	const auto combined =
-	    EmitBinaryU32(ctx.state, zero ? spv::OpBitwiseAnd : spv::OpBitwiseOr, low, high);
-	ctx.state.builder.AddFunction(zero ? spv::OpIEqual : spv::OpINotEqual, TypeBool(ctx.state),
-	                              result, combined, ConstantU32(ctx.state, zero ? ~0u : 0u));
-	return result;
-}
-
 void EmitStructuredTerminator(ValueEmitContext& ctx, const IR::Block* block,
                               const IR::BlockInfo& info) {
 	const auto& program = ctx.state.program;
@@ -160,7 +136,7 @@ void EmitStructuredTerminator(ValueEmitContext& ctx, const IR::Block* block,
 				EmitReturn(ctx);
 				return;
 			}
-			const auto condition = BranchCondition(ctx, info);
+			const auto condition = ctx.Def(info.condition);
 			emit_merge();
 			ctx.state.builder.AddFunction(spv::OpBranchConditional, condition,
 			                              ctx.Label(true_block), ctx.Label(false_block));
@@ -193,7 +169,7 @@ uint32_t EmitDispatcherNextPc(ValueEmitContext& ctx, const DispatcherFunctionSta
 			EmitDispatcherTarget(ctx, dispatcher, block, term.false_block);
 			const auto selected = ctx.state.builder.AllocateId();
 			ctx.state.builder.AddFunction(
-			    spv::OpSelect, TypeU32(ctx.state), selected, BranchCondition(ctx, info),
+			    spv::OpSelect, TypeU32(ctx.state), selected, ctx.Def(info.condition),
 			    ConstantU32(ctx.state, term.true_block), ConstantU32(ctx.state, term.false_block));
 			return selected;
 		}

src/graphics/shader/recompiler/frontend/cfg/ShaderCFG.h

@@ -18,10 +18,17 @@ enum class BranchCondition {
 	ExecZero,
 	ExecNonZero,
 	ScalarInstruction,
-	GotoVariable,
+	Expression,
 	Unknown
 };
 
+struct ConditionExpression {
+	enum class Op { Constant, Variable, Native, Not, Or };
+	Op       op;
+	uint32_t lhs;
+	uint32_t rhs = UINT32_MAX;
+};
+
 enum class TerminatorKind { Branch, ConditionalBranch, IndirectBranch, Return, Unsupported };
 
 enum class FailureKind {
@@ -48,8 +55,7 @@ struct Terminator {
 	std::vector<uint32_t> indirect_targets;
 	std::vector<uint32_t> indirect_selector_values;
 	std::vector<uint32_t> indirect_selector_targets;
-	uint32_t              goto_variable = UINT32_MAX;
-	int32_t               goto_value    = -1;
+	uint32_t              expression    = UINT32_MAX;
 	bool                  loop_header   = false;
 };
 
@@ -59,10 +65,11 @@ struct BasicBlock {
 	uint32_t              end_pc     = 0;
 	uint32_t              inst_begin = 0;
 	uint32_t              inst_end   = 0;
+	struct Assignment { uint32_t variable; uint32_t expression; };
+	std::vector<Assignment> assignments;
 	std::vector<uint32_t> predecessors;
 	std::vector<uint32_t> successors;
 	std::vector<uint32_t> dominators;
-	std::vector<uint32_t> post_dominators;
 	Terminator            terminator;
 };
 
@@ -73,12 +80,9 @@ struct BackEdge {
 };
 
 struct NaturalLoop {
-	uint32_t              header         = UINT32_MAX;
-	uint32_t              latch          = UINT32_MAX;
-	uint32_t              merge          = UINT32_MAX;
-	uint32_t              continue_block = UINT32_MAX;
+	uint32_t              header = UINT32_MAX;
+	uint32_t              latch  = UINT32_MAX;
 	std::vector<uint32_t> body_blocks;
-	std::vector<uint32_t> exit_blocks;
 };
 
 struct StronglyConnectedComponent {
@@ -89,6 +93,7 @@ struct StronglyConnectedComponent {
 
 struct Graph {
 	std::vector<BasicBlock>                 blocks;
+	std::vector<ConditionExpression>        expressions;
 	std::vector<BackEdge>                   back_edges;
 	std::vector<NaturalLoop>                natural_loops;
 	std::vector<StronglyConnectedComponent> components;
@@ -105,14 +110,12 @@ struct Graph {
 	const BasicBlock* FindBlockByPc(uint32_t pc) const;
 	BasicBlock*       FindBlockByPc(uint32_t pc);
 	bool              Dominates(uint32_t dominator, uint32_t block) const;
-	bool              PostDominates(uint32_t post_dominator, uint32_t block) const;
-	uint32_t          FindNearestCommonPostDominator(uint32_t block_a, uint32_t block_b) const;
 };
 
 Graph       BuildGraph(const Decoder::Program& program);
-// Commits structured control flow on success; preserves the original graph with
-// failure diagnostics on failure. failure_block is an original block ID or UINT32_MAX.
-bool        Structurize(Graph& graph);
+// Returns structured control flow or failure diagnostics without changing the native graph.
+// On failure, failure_block is an original block ID or UINT32_MAX.
+Graph       Structurize(const Graph& graph);
 std::string BranchConditionToString(BranchCondition condition);
 std::string FailureKindToString(FailureKind kind);
 std::string GraphToString(const Graph& graph);

src/graphics/shader/recompiler/frontend/decode/MemoryOps.cpp

@@ -106,6 +106,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x39u, Opcode::DS_READ_I8, 1, 8, true},     {0x3au, Opcode::DS_READ_U8, 1, 8},
     {0x3bu, Opcode::DS_READ_I16, 1, 16, true},   {0x3cu, Opcode::DS_READ_U16, 1, 16},
     {0x3du, Opcode::DS_CONSUME, 1, 32},          {0x3eu, Opcode::DS_APPEND, 1, 32},
+    {0x4au, Opcode::DS_OR_B64, 2, 32},
     {0x4du, Opcode::DS_WRITE_B64, 2, 32},        {0x4eu, Opcode::DS_WRITE2_B64, 4, 32},
     {0x4fu, Opcode::DS_WRITE2ST64_B64, 4, 32},   {0x76u, Opcode::DS_READ_B64, 2, 32},
     {0x77u, Opcode::DS_READ2_B64, 4, 32},        {0x78u, Opcode::DS_READ2ST64_B64, 4, 32},
@@ -185,6 +186,7 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 		case Opcode::DS_AND_B32:
 		case Opcode::DS_AND_RTN_B32:
 		case Opcode::DS_OR_B32:
+		case Opcode::DS_OR_B64:
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.cpp

@@ -681,6 +681,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::DS_AND_B32:
 		case Opcode::DS_AND_RTN_B32:
 		case Opcode::DS_OR_B32:
+		case Opcode::DS_OR_B64:
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.h

@@ -156,6 +156,8 @@ enum class Opcode {
 	V_DOT2C_F32_F16,
 	V_CVT_F64_I32,
 	V_CVT_F32_F64,
+	V_CVT_F64_F32,
+	V_CVT_F64_U32,
 	V_CVT_F32_I32,
 	V_CVT_F32_U32,
 	V_CVT_U32_F32,
@@ -397,6 +399,7 @@ enum class Opcode {
 	V_CMPX_GT_F16,
 	V_CMPX_GE_F16,
 	V_CMPX_NGT_F16,
+	V_CMPX_NLE_F16,
 	V_CMPX_NEQ_F16,
 	V_CMPX_NLT_F16,
 	V_CMPX_LT_I32,
@@ -525,6 +528,7 @@ enum class Opcode {
 	DS_AND_B32,
 	DS_AND_RTN_B32,
 	DS_OR_B32,
+	DS_OR_B64,
 	DS_OR_RTN_B32,
 	DS_XOR_B32,
 	DS_XOR_RTN_B32,

src/graphics/shader/recompiler/frontend/decode/VectorAluOps.cpp

@@ -89,6 +89,8 @@ constexpr OpcodeMap VOP1_OPCODE_LIST[] = {
     {0x02u, Opcode::V_READFIRSTLANE_B32},
     {0x04u, Opcode::V_CVT_F64_I32},
     {0x0fu, Opcode::V_CVT_F32_F64},
+    {0x10u, Opcode::V_CVT_F64_F32},
+    {0x16u, Opcode::V_CVT_F64_U32},
     {0x2fu, Opcode::V_RCP_F64},
     {0x05u, Opcode::V_CVT_F32_I32},
     {0x06u, Opcode::V_CVT_F32_U32},
@@ -151,6 +153,8 @@ constexpr OpcodeMap VOP3_ENCODED_VOP1_OPCODE_LIST[] = {
     {0x02u, Opcode::V_READFIRSTLANE_B32},
     {0x04u, Opcode::V_CVT_F64_I32},
     {0x0fu, Opcode::V_CVT_F32_F64},
+    {0x10u, Opcode::V_CVT_F64_F32},
+    {0x16u, Opcode::V_CVT_F64_U32},
     {0x2fu, Opcode::V_RCP_F64},
     {0x05u, Opcode::V_CVT_F32_I32},
     {0x06u, Opcode::V_CVT_F32_U32},
@@ -258,7 +262,8 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0xd9u, Opcode::V_CMPX_LT_F16},        {0xdau, Opcode::V_CMPX_EQ_F16},
     {0xdbu, Opcode::V_CMPX_LE_F16},        {0xdcu, Opcode::V_CMPX_GT_F16},
     {0xdeu, Opcode::V_CMPX_GE_F16},        {0xfbu, Opcode::V_CMPX_NGT_F16},
-    {0xfdu, Opcode::V_CMPX_NEQ_F16},       {0xfeu, Opcode::V_CMPX_NLT_F16},
+    {0xfcu, Opcode::V_CMPX_NLE_F16},       {0xfdu, Opcode::V_CMPX_NEQ_F16},
+    {0xfeu, Opcode::V_CMPX_NLT_F16},
 };
 
 constexpr auto VOPC_OPS = Detail::MakeOpcodeTable<0x100>(VOPC_OPCODE_LIST);
@@ -461,6 +466,7 @@ bool IsVopcCompareExec(Opcode opcode);
 bool IsVop1FloatSourceOpcode(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::V_CVT_F32_F64:
+		case Opcode::V_CVT_F64_F32:
 		case Opcode::V_RCP_F64:
 		case Opcode::V_MOV_B32:
 		case Opcode::V_CVT_F32_F16:
@@ -721,6 +727,7 @@ void ApplyDppModifier(Operand& operand, uint32_t modifier, uint32_t encoding) {
 void DecodeVop1Dpp(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
                    uint32_t opcode, uint32_t vdst, Instruction& inst) {
 	if (inst.opcode == Opcode::V_CVT_F64_I32 || inst.opcode == Opcode::V_CVT_F32_F64 ||
+	    inst.opcode == Opcode::V_CVT_F64_F32 || inst.opcode == Opcode::V_CVT_F64_U32 ||
 	    inst.opcode == Opcode::V_RCP_F64) {
 		SetUnsupported(inst, Family::VOP1, opcode, "FP64 instructions do not support DPP");
 		return;
@@ -862,6 +869,7 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMPX_GT_F16:
 		case Opcode::V_CMPX_GE_F16:
 		case Opcode::V_CMPX_NGT_F16:
+		case Opcode::V_CMPX_NLE_F16:
 		case Opcode::V_CMPX_NEQ_F16:
 		case Opcode::V_CMPX_NLT_F16:
 		case Opcode::V_CMP_CLASS_F32:
@@ -1488,6 +1496,7 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_GT_F16:
 		case Opcode::V_CMPX_GE_F16:
 		case Opcode::V_CMPX_NGT_F16:
+		case Opcode::V_CMPX_NLE_F16:
 		case Opcode::V_CMPX_NEQ_F16:
 		case Opcode::V_CMPX_NLT_F16: return true;
 		default: return false;

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

@@ -44,7 +44,8 @@ void Translator::EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpc
 	    half ? IR::Value(ReadF16AsF32(inst.src0)) : ReadOperand(inst.src0, IR::Type::F32);
 	const auto rhs =
 	    half ? IR::Value(ReadF16AsF32(inst.src1)) : ReadOperand(inst.src1, IR::Type::F32);
-	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs})), false, cmpx);
+	const IR::FPCompareFlags flags{.flush_input_denorms = !half && flush_f32_inputs};
+	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs}, flags)), false, cmpx);
 }
 
 void Translator::EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered, bool cmpx) {

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

@@ -235,6 +235,7 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 				return index == 0u ? decoded.src0 : MakeImmediate(decoded.offset & 0xffffu);
 			case Decoder::Opcode::DS_CONSUME:
 			case Decoder::Opcode::DS_APPEND:
+				return decoded.gds ? MakeM0Operand() : MakeImmediate(0);
 			case Decoder::Opcode::DS_READ_ADDTID_B32: return MakeM0Operand();
 			case Decoder::Opcode::DS_WRITE_ADDTID_B32:
 				return index == 0u ? decoded.src1 : MakeM0Operand();
@@ -569,7 +570,10 @@ void Translator::DS_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opc
                            bool returns_value) {
 	const auto memory  = MemoryInfoFromDecoded(inst);
 	const auto address = ReadU32(MemorySourceAt(inst, 1));
-	const auto result  = ir.Emit(opcode, {address, ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
+	const auto data_src = MemorySourceAt(inst, 0);
+	const IR::Value value =
+	    memory.data_dwords == 2u ? IR::Value(ReadU64(data_src)) : IR::Value(ReadU32(data_src));
+	const auto result  = ir.Emit(opcode, {address, value, ir.GetExec()},
 	                             AddMemoryInfo(memory, inst.pc));
 	if (returns_value) {
 		WriteOperand(inst.dst, result);
@@ -845,8 +849,7 @@ void Translator::DS_MINMAX_F32(const Decoder::Instruction& inst, IR::ValueOpcode
 void Translator::DS_APPEND_CONSUME(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
 	const auto memory = MemoryInfoFromDecoded(inst);
 	WriteOperand(inst.dst, ir.Emit(opcode,
-	                               {ReadU32(MemorySourceAt(inst, 0)), ir.GetExec(), ir.GetExecLo(),
-	                                ir.GetExecHi()},
+	                               {ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
 	                               AddMemoryInfo(memory, inst.pc)));
 }
 
@@ -1003,6 +1006,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicAnd32, true);
 		case Decoder::Opcode::DS_OR_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr32, false);
+		case Decoder::Opcode::DS_OR_B64:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr64, false);
 		case Decoder::Opcode::DS_OR_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr32, true);
 		case Decoder::Opcode::DS_XOR_B32:

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

@@ -317,6 +317,7 @@ void Translator::WriteRawU32(const Decoder::Operand& operand, IR::U32 value) {
 		case Decoder::OperandKind::Sgpr: {
 			const auto reg = static_cast<IR::ScalarReg>(operand.reg);
 			ir.SetScalarReg(reg, value);
+			program.scalar_writes.push_back({current_pc, reg});
 			ir.SetScalarMaskTag(reg, IR::U1(IR::Value(false)));
 			if (IR::RegIndex(reg) > 0u) {
 				ir.SetScalarMaskTag(static_cast<IR::ScalarReg>(IR::RegIndex(reg) - 1u),
@@ -699,11 +700,13 @@ std::array<IR::U32, 2> Translator::WriteMask(const Decoder::Operand& operand, IR
 				                    IR::U1(IR::Value(false)));
 			}
 			ir.SetScalarReg(reg, mask[0]);
+			program.scalar_writes.push_back({current_pc, reg});
 			// A wave32 VALU mask destination must not overwrite the neighboring SGPR.
 			if ((write_64 || program.wave_size == 64u) &&
 			    IR::RegIndex(reg) + 1u < IR::NumScalarRegs) {
 				const auto high = static_cast<IR::ScalarReg>(IR::RegIndex(reg) + 1u);
 				ir.SetScalarReg(high, mask[1]);
+				program.scalar_writes.push_back({current_pc, high});
 				ir.SetThreadBitScalarReg(high, IR::U1(IR::Value(false)));
 				ir.SetScalarMaskTag(high, IR::U1(IR::Value(false)));
 			}
@@ -746,55 +749,61 @@ void Translator::WriteCompareResult(const Decoder::Operand& operand, IR::U1 valu
 	WriteMask(operand, ir.LogicalAnd(ir.GetExec(), value));
 }
 
-void Translator::AddBranchCondition(const CFG::BasicBlock& source, IR::BlockInfo& info) {
-	if (source.terminator.goto_value >= 0) {
-		if (source.terminator.goto_variable == UINT32_MAX) {
-			EXIT("block %u sets an invalid goto variable", source.id);
+void Translator::AddBranchCondition(const CFG::Graph& graph, const CFG::BasicBlock& source,
+                                    IR::BlockInfo& info) {
+	const auto native_condition = [&](CFG::BranchCondition kind) -> IR::U1 {
+		IR::U1 condition;
+		switch (kind) {
+			case CFG::BranchCondition::Always: condition = IR::U1(IR::Value(true)); break;
+			case CFG::BranchCondition::SccZero: condition = ir.LogicalNot(ir.GetScc()); break;
+			case CFG::BranchCondition::SccNonZero: condition = ir.GetScc(); break;
+			case CFG::BranchCondition::VccZero: condition = ir.LogicalNot(ir.GetVcc()); break;
+			case CFG::BranchCondition::VccNonZero: condition = ir.GetVcc(); break;
+			case CFG::BranchCondition::ExecZero: condition = ir.LogicalNot(ir.GetExec()); break;
+			case CFG::BranchCondition::ExecNonZero: condition = ir.GetExec(); break;
+			case CFG::BranchCondition::ScalarInstruction:
+				EXIT_IF(instruction_branch_condition.IsEmpty());
+				condition = instruction_branch_condition;
+				break;
+			default: EXIT("block %u has an invalid native branch condition", source.id);
 		}
-		ir.SetGotoVariable(source.terminator.goto_variable,
-		                   IR::U1(IR::Value(source.terminator.goto_value != 0)));
+		return IR::U1(ir.Emit(IR::ValueOpcode::ConditionRef, {condition}, kind));
+	};
+	const auto expression = [&](auto&& self, uint32_t index) -> IR::U1 {
+		const auto& value = graph.expressions.at(index);
+		switch (value.op) {
+			case CFG::ConditionExpression::Op::Constant: return IR::U1(IR::Value(value.lhs != 0));
+			case CFG::ConditionExpression::Op::Variable: return ir.GetGotoVariable(value.lhs);
+			case CFG::ConditionExpression::Op::Native:
+				return native_condition(static_cast<CFG::BranchCondition>(value.lhs));
+			case CFG::ConditionExpression::Op::Not: return ir.LogicalNot(self(self, value.lhs));
+			case CFG::ConditionExpression::Op::Or:
+				return ir.LogicalOr(self(self, value.lhs), self(self, value.rhs));
+		}
+		EXIT("invalid CFG condition expression");
+	};
+	for (const auto& assignment: source.assignments) {
+		ir.SetGotoVariable(assignment.variable, expression(expression, assignment.expression));
 	}
-	if (source.terminator.kind == CFG::TerminatorKind::IndirectBranch) {
-		if (source.terminator.indirect_selector_code != UINT32_MAX) {
-			info.indirect_target = ReadScalarCode(source.terminator.indirect_selector_code);
-		} else if (source.terminator.indirect_pc_sgpr != UINT32_MAX) {
-			info.indirect_target =
-			    ir.GetScalarReg(static_cast<IR::ScalarReg>(source.terminator.indirect_pc_sgpr));
+	const auto& term = source.terminator;
+	if (term.kind == CFG::TerminatorKind::IndirectBranch) {
+		if (term.indirect_selector_code != UINT32_MAX) {
+			info.indirect_target = ReadScalarCode(term.indirect_selector_code);
+		} else if (term.indirect_pc_sgpr != UINT32_MAX) {
+			info.indirect_target = ir.GetScalarReg(static_cast<IR::ScalarReg>(term.indirect_pc_sgpr));
 		} else {
 			EXIT("block %u has no indirect branch selector", source.id);
 		}
 		ir.Emit(IR::ValueOpcode::ReferenceU32, {info.indirect_target});
 		return;
 	}
-	if (source.terminator.kind != CFG::TerminatorKind::ConditionalBranch) {
-		return;
+	if (term.kind == CFG::TerminatorKind::ConditionalBranch) {
+		const auto condition = term.expression != UINT32_MAX
+		                           ? expression(expression, term.expression)
+		                           : native_condition(term.condition);
+		info.condition = condition;
+		ir.Emit(IR::ValueOpcode::Reference, {condition});
 	}
-	// EXEC and VCC are invocation-local Boolean masks. Branching on that Boolean lets inactive
-	// invocations leave the region without reconstructing a host-subgroup mask.
-	IR::U1 condition;
-	switch (source.terminator.condition) {
-		case CFG::BranchCondition::Always: condition = IR::U1(IR::Value(true)); break;
-		case CFG::BranchCondition::SccZero: condition = ir.LogicalNot(ir.GetScc()); break;
-		case CFG::BranchCondition::SccNonZero: condition = ir.GetScc(); break;
-		case CFG::BranchCondition::VccZero: condition = ir.LogicalNot(ir.GetVcc()); break;
-		case CFG::BranchCondition::VccNonZero: condition = ir.GetVcc(); break;
-		case CFG::BranchCondition::ExecZero: condition = ir.LogicalNot(ir.GetExec()); break;
-		case CFG::BranchCondition::ExecNonZero: condition = ir.GetExec(); break;
-		case CFG::BranchCondition::ScalarInstruction:
-			EXIT_IF(instruction_branch_condition.IsEmpty());
-			condition = instruction_branch_condition;
-			break;
-		case CFG::BranchCondition::GotoVariable:
-			if (source.terminator.goto_variable == UINT32_MAX) {
-				EXIT("block %u reads an invalid goto variable", source.id);
-			}
-			condition = ir.GetGotoVariable(source.terminator.goto_variable);
-			break;
-		case CFG::BranchCondition::Unknown:
-			EXIT("block %u has an unknown branch condition", source.id);
-	}
-	info.condition = condition;
-	ir.Emit(IR::ValueOpcode::Reference, {condition});
 }
 
 namespace {
@@ -867,7 +876,9 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
 	include_vector(inst.src2);
 	include_vector(inst.src3);
 	switch (inst.opcode) {
-		case Decoder::Opcode::V_CVT_F64_I32: include_vector(inst.dst, 2u); break;
+		case Decoder::Opcode::V_CVT_F64_I32:
+		case Decoder::Opcode::V_CVT_F64_F32:
+		case Decoder::Opcode::V_CVT_F64_U32: include_vector(inst.dst, 2u); break;
 		case Decoder::Opcode::V_FMA_F64: include_vector(inst.src2, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_MUL_F64: include_vector(inst.src1, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_RCP_F64: include_vector(inst.dst, 2u); [[fallthrough]];
@@ -1272,9 +1283,11 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			                      builtin(IR::StageInputKind::InstanceIndex));
 		}
 	}
+	const bool flush_f32_inputs = options.stage == ShaderType::Compute &&
+	                             (options.input_info.compute->float_mode & 0x10u) == 0;
 	for (const auto& cfg_block: cfg.blocks) {
 		const auto typed_index = block_indices.at(cfg_block.id);
-		Translator translator(result, result.blocks[typed_index], vector_limit);
+		Translator translator(result, result.blocks[typed_index], vector_limit, flush_f32_inputs);
 		for (uint32_t index = cfg_block.inst_begin; index < cfg_block.inst_end; index++) {
 			const auto& instruction = decoded.instructions[index];
 			if (IsCodeTableLoad(cfg, instruction.pc)) {
@@ -1297,7 +1310,7 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			}
 			translator.TranslateInstruction(instruction);
 		}
-		translator.AddBranchCondition(cfg_block, result.block_info[typed_index]);
+		translator.AddBranchCondition(cfg, cfg_block, result.block_info[typed_index]);
 	}
 	IR::ValidateProgram(result, false);
 	return result;

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

@@ -9,13 +9,15 @@ namespace Libs::Graphics::ShaderRecompiler::Frontend {
 
 class Translator {
 public:
-	Translator(IR::Program& program, IR::Block* block, uint32_t vector_limit)
-	    : program(program), ir(block), current_vector_limit(vector_limit) {}
+	Translator(IR::Program& program, IR::Block* block, uint32_t vector_limit,
+	           bool flush_f32_inputs)
+	    : program(program), ir(block), current_vector_limit(vector_limit),
+	      flush_f32_inputs(flush_f32_inputs) {}
 
 	void TranslateInstruction(const Decoder::Instruction& inst);
 	void TranslateEmbeddedFetch(const Decoder::Instruction& inst, uint32_t attribute,
 	                            uint32_t component_count, const ShaderBufferResource& resource);
-	void AddBranchCondition(const CFG::BasicBlock& source, IR::BlockInfo& info);
+	void AddBranchCondition(const CFG::Graph& graph, const CFG::BasicBlock& source, IR::BlockInfo& info);
 
 private:
 	const Decoder::Operand& SourceAt(const Decoder::Instruction& inst, uint32_t index);
@@ -261,6 +263,7 @@ private:
 	Decoder::Opcode current_opcode       = Decoder::Opcode::UNKNOWN;
 	uint32_t        current_pc           = 0;
 	uint32_t        current_vector_limit = 1;
+	bool            flush_f32_inputs;
 };
 
 } // namespace Libs::Graphics::ShaderRecompiler::Frontend

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

@@ -273,6 +273,9 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMPX_NGT_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, true, true);
 			return;
+		case O::V_CMPX_NLE_F16:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan32, true, true);
+			return;
 		case O::V_CMP_NEQ_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, true, false);
 			return;
@@ -299,6 +302,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CVT_F32_I32: V_CVT_F32_I32(inst); return;
 		case O::V_CVT_F64_I32: return FloatUnary(inst, IR::ValueOpcode::ConvertF64S32);
 		case O::V_CVT_F32_F64: return FloatUnary(inst, IR::ValueOpcode::ConvertF32F64);
+		case O::V_CVT_F64_F32: return FloatUnary(inst, IR::ValueOpcode::ConvertF64F32);
+		case O::V_CVT_F64_U32: return FloatUnary(inst, IR::ValueOpcode::ConvertF64U32);
 		case O::V_CVT_U32_F32: V_CVT_U32_F32(inst); return;
 		case O::V_CVT_I32_F32: V_CVT_I32_F32(inst); return;
 		case O::V_CVT_F16_F32: V_CVT_F16_F32(inst); return;

src/graphics/shader/recompiler/ir/Program.cpp

@@ -81,7 +81,8 @@ bool EquivalentValue(const ResourcePlan& program, Value left, Value right,
 		    program.memory_info[li] != program.memory_info[ri]) {
 			return false;
 		}
-	} else if (lhs->Flags<uint64_t>() != rhs->Flags<uint64_t>()) {
+	} else if (lhs->GetOpcode() != ValueOpcode::ReadConst &&
+	           lhs->Flags<uint64_t>() != rhs->Flags<uint64_t>()) {
 		return false;
 	}
 	for (size_t index = 0; index < lhs->NumArgs(); index++) {

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

@@ -147,6 +147,7 @@ struct SamplerResource {
 	uint32_t first_use_pc          = 0;
 	bool     force_point_filtering = false;
 	bool     depth_compare         = false;
+	bool     integer_border        = false;
 
 	bool operator==(const SamplerResource& other) const = default;
 };
@@ -548,7 +549,7 @@ struct ResourcePlan {
 	std::vector<SrtRead>                srt_reads;
 	std::vector<uint8_t>                clean_flat_slots;
 	bool                                requires_specialization_memory = false;
-	bool                                has_address_writes = false;
+	bool                                capture_specialization_reads = false;
 	bool                                srt_plan_complete          = false;
 	bool                                resource_tracking_complete = false;
 	ShaderInfo                          info;
@@ -582,10 +583,12 @@ struct Program: ResourcePlan {
 	CFG::FailureKind              cfg_failure_kind    = CFG::FailureKind::None;
 	std::string                   fallback_reason;
 	std::vector<BlockInfo>        block_info;
+	struct ScalarWrite { uint32_t pc; ScalarReg reg; };
+	std::vector<ScalarWrite>      scalar_writes;
 	// Typed memory and export instructions reference shader-local metadata by dense index.
 	// Decoder-only details (such as NSA register numbers) have already become IR operands.
 	std::vector<ExportInfo>       export_info;
-	std::vector<Value>            dynamic_reads;
+	bool                          has_address_writes = false;
 	bool                          shader_info_complete = false;
 	BindingLayout                 bindings;
 	bool                          binding_layout_complete = false;

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

@@ -173,6 +173,7 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::SharedAtomicUMax32:
 		case ValueOpcode::SharedAtomicAnd32:
 		case ValueOpcode::SharedAtomicOr32:
+		case ValueOpcode::SharedAtomicOr64:
 		case ValueOpcode::SharedAtomicXor32: return SharedAccess::Atomic;
 		case ValueOpcode::DataAppend: return SharedAccess::Append;
 		case ValueOpcode::DataConsume: return SharedAccess::Consume;
@@ -182,6 +183,7 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
 
 uint32_t SharedComponentCount(ValueOpcode opcode) {
 	switch (opcode) {
+		case ValueOpcode::SharedAtomicOr64:
 		case ValueOpcode::LoadSharedU32x2:
 		case ValueOpcode::WriteSharedU32x2: return 2u;
 		case ValueOpcode::LoadSharedU32x3:

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

@@ -52,6 +52,10 @@ struct PermlaneFlags {
 static_assert(sizeof(PermlaneFlags) <= sizeof(uint64_t));
 static_assert(std::is_trivially_copyable_v<PermlaneFlags>);
 
+struct FPCompareFlags {
+	bool flush_input_denorms = false;
+};
+
 struct MemoryFlags {
 	uint32_t index = 0;
 	uint32_t pc    = 0;

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

@@ -4,6 +4,7 @@ VALUE_OPCODE(Identity, Opaque, Opaque)
 VALUE_OPCODE(Void, Void)
 VALUE_OPCODE(Reference, Void, U1)
 VALUE_OPCODE(ReferenceU32, Void, U32)
+VALUE_OPCODE(ConditionRef, U1, U1)
 
 // Translation and register-state pseudo operations.
 VALUE_OPCODE(GetUserData, U32, ScalarReg)
@@ -65,7 +66,9 @@ VALUE_OPCODE(ConvertU32F32, U32, F32)
 VALUE_OPCODE(ConvertF32S32, F32, U32)
 VALUE_OPCODE(ConvertF64S32, F64, U32)
 VALUE_OPCODE(ConvertF32F64, F32, F64)
+VALUE_OPCODE(ConvertF64F32, F64, F32)
 VALUE_OPCODE(ConvertF32U32, F32, U32)
+VALUE_OPCODE(ConvertF64U32, F64, U32)
 VALUE_OPCODE(CompositeConstructU64, U64, U32, U32)
 VALUE_OPCODE(CompositeConstructU32x2, U32x2, U32, U32)
 VALUE_OPCODE(CompositeConstructU32x3, U32x3, U32, U32, U32)
@@ -273,9 +276,10 @@ VALUE_OPCODE(SharedAtomicSMax32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicUMax32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicAnd32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicOr32, U32, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicOr64, Void, U32, U64, U1)
 VALUE_OPCODE(SharedAtomicXor32, U32, U32, U32, U1)
-VALUE_OPCODE(DataAppend, U32, U32, U1, U32, U32)
-VALUE_OPCODE(DataConsume, U32, U32, U1, U32, U32)
+VALUE_OPCODE(DataAppend, U32, U32, U1)
+VALUE_OPCODE(DataConsume, U32, U32, U1)
 VALUE_OPCODE(SwizzleU32, U32, U32, U32, U1)
 
 VALUE_OPCODE(ImageQueryDimensions, U32x4, ImageResource, ImageAddress)

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

@@ -85,6 +85,16 @@ bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffe
 	return uses_gds;
 }
 
+bool UsesFlattenedSrt(const Program& program) {
+	return std::ranges::any_of(program.blocks, [](const Block* block) {
+		return std::ranges::any_of(*block, [](const Inst& inst) {
+			return inst.GetOpcode() == ValueOpcode::ReadConst;
+		});
+	}) || std::ranges::any_of(program.info.images, [](const ImageResource& image) {
+		return image.indirect_search_iterations != 0u;
+	});
+}
+
 void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 	if (!program.shader_info_complete || program.binding_layout_complete) {
 		EXIT("shader binding layout failed: %s", !program.shader_info_complete
@@ -145,12 +155,7 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 		AddBinding(next, DescriptorBindingKind::BdaPagetable);
 		AddBinding(next, DescriptorBindingKind::FaultBuffer);
 	}
-	const bool uses_flattened_runtime =
-	    !program.srt_reads.empty() ||
-	    std::ranges::any_of(program.info.images, [](const ImageResource& image) {
-		    return image.indirect_search_iterations != 0u;
-	    });
-	if (uses_flattened_runtime) {
+	if (UsesFlattenedSrt(program)) {
 		AddBinding(next, DescriptorBindingKind::FlattenedSrt);
 	}
 

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

@@ -9,6 +9,7 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword = 0);
 
 // Collect live direct buffer indices and report whether the shader accesses GDS.
 bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers);
+bool UsesFlattenedSrt(const Program& program);
 
 const DescriptorBinding* FindBinding(const BindingLayout& layout, DescriptorBindingKind kind);
 

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

@@ -630,6 +630,11 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 				}
 			}
 			return;
+		case ValueOpcode::ConditionRef: {
+			const auto value = Arg(inst, 0);
+			if (IsImmediate(value, Type::U1)) Replace(inst, value);
+			return;
+		}
 		case ValueOpcode::LogicalNot: {
 			const auto value = Arg(inst, 0);
 			if (IsImmediate(value, Type::U1)) {

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

@@ -107,14 +107,16 @@ Prospero::BufferFormat ImageConversionFormat(Prospero::BufferFormat format) {
 	                                                      : Prospero::BufferFormat::kInvalid;
 }
 
-bool RequiresPointSampler(const ImageResource& image) {
-	return image.numeric_class == Prospero::TextureNumericClass::Sint ||
-	       image.conversion_format != Prospero::BufferFormat::kInvalid;
-}
+enum class SamplerClass : uint8_t { Float, Integer, PointInteger };
 
-bool RequiresPointSampler(const ResourceSpecialization::Image& image) {
-	return image.numeric_class == Prospero::TextureNumericClass::Sint ||
-	       image.conversion_format != Prospero::BufferFormat::kInvalid;
+template <typename Image>
+SamplerClass ClassifySampler(const Image& image) {
+	if (image.numeric_class == Prospero::TextureNumericClass::Sint ||
+	    image.conversion_format != Prospero::BufferFormat::kInvalid) {
+		return SamplerClass::PointInteger;
+	}
+	return image.numeric_class == Prospero::TextureNumericClass::Uint ? SamplerClass::Integer
+	                                                               : SamplerClass::Float;
 }
 
 bool DescriptorIsCube(const DescriptorValue& descriptor) {
@@ -223,9 +225,9 @@ void MarkCleanFlatSlots(const ResourcePlan& program, const DescriptorSource* sou
 	}
 }
 
-bool ReadScalarTable(uint64_t base, uint64_t size, uint32_t dynamic_offset,
+bool ReadScalarTable(uint64_t base, uint64_t size, uint64_t dynamic_offset,
                      const SrtRuntime& runtime, std::span<uint32_t> words) {
-	const auto offset = static_cast<uint64_t>(dynamic_offset) & ~uint64_t {3};
+	const auto offset = dynamic_offset & ~uint64_t {3};
 	const auto count = std::min<uint64_t>(words.size(), offset < size ? (size - offset) / 4u : 0u);
 	std::ranges::fill(words.subspan(count), 0u);
 	if (count == 0u) {
@@ -304,26 +306,24 @@ bool MaterializeIndirectImage(const ResourcePlan& program,
 		    material.Stride() != indirect.selector_stride) {
 			return false;
 		}
-		// Enumerate every wrapped scalar-buffer offset that can pass the descriptor bounds.
-		const auto step = std::gcd<uint64_t>(indirect.selector_stride, uint64_t {1} << 32u);
-		const auto residue = static_cast<uint64_t>(indirect.selector_offset) % step;
-		const auto limit = std::min<uint64_t>(UINT32_MAX, material.GetSize() + 3u);
-		const auto probe_count = residue <= limit ? (limit - residue) / step + 1u : 0u;
+		// The first aligned offset includes the immediate added after shader U32 arithmetic.
+		const auto step = std::max<uint64_t>(4u,
+		    std::gcd<uint64_t>(indirect.selector_stride, uint64_t {1} << 32u));
+		const uint64_t first = indirect.selector_offset;
+		const auto size = material.GetSize();
+		const auto limit = std::min(first + (uint64_t {1} << 32u) - step, size >= 4u ? size - 4u : 0u);
+		const auto probe_count = size >= 4u && first <= limit ? (limit - first) / step + 1u : 0u;
 		if (probe_count > MaxIndirectImageProbes) {
 			return false;
 		}
 		keys.reserve(static_cast<size_t>(probe_count) + 1u);
 		keys.push_back(0u);
-		for (uint64_t offset = residue; offset <= limit && probe_count != 0u; offset += step) {
+		for (uint64_t probe = 0, offset = first; probe < probe_count; ++probe, offset += step) {
 			uint32_t key = 0;
-			if (!ReadScalarTable(material.Base48(), material.GetSize(),
-			                     static_cast<uint32_t>(offset), runtime, {&key, 1})) {
+			if (!ReadScalarTable(material.Base48(), size, offset, runtime, {&key, 1})) {
 				return false;
 			}
 			keys.push_back(key);
-			if (limit - offset < step) {
-				break;
-			}
 		}
 		std::ranges::sort(keys);
 		keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
@@ -380,8 +380,13 @@ bool MaterializeIndirectImage(const ResourcePlan& program,
 } // namespace
 
 struct SamplerPlan {
-	std::array<uint32_t, ShaderInfo::MaxSamplers> point_sampler {};
-	uint32_t                                      sampler_count = 0;
+	struct Binding {
+		uint32_t     source;
+		SamplerClass type;
+	};
+	std::array<std::array<uint32_t, 3>, ShaderInfo::MaxSamplers> mapping;
+	std::array<Binding, ShaderInfo::MaxSamplers>                bindings;
+	uint32_t                                                  sampler_count = 0;
 };
 
 struct ImageRemap {
@@ -595,11 +600,9 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 	if (!BuildSamplerPlan(program.info, specialization.images, sampler_plan)) {
 		return SpecializationFail("specialized sampler layout exceeds its resource limit");
 	}
-	for (uint32_t index = 0; index < program.info.samplers.size(); index++) {
-		const auto target = sampler_plan.point_sampler[index];
-		if (target != UINT32_MAX && target >= program.info.samplers.size()) {
-			snapshot.samplers.push_back(snapshot.samplers[index]);
-		}
+	for (uint32_t index = static_cast<uint32_t>(program.info.samplers.size());
+	     index < sampler_plan.sampler_count; index++) {
+		snapshot.samplers.push_back(snapshot.samplers[sampler_plan.bindings[index].source]);
 	}
 	ImageRemap(specialization).Apply(snapshot.images);
 	return true;
@@ -607,29 +610,29 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 
 template <typename Images>
 bool BuildSamplerPlan(const ShaderInfo& base, const Images& images, SamplerPlan& plan) {
-	if (base.samplers.size() > plan.point_sampler.size()) {
+	if (base.samplers.size() > plan.mapping.size()) {
 		return false;
 	}
 	std::array<uint8_t, ShaderInfo::MaxSamplers> usage {};
-	plan.point_sampler.fill(UINT32_MAX);
 	plan.sampler_count = static_cast<uint32_t>(base.samplers.size());
 	for (const auto& pair: base.sampled_pairs) {
 		if (pair.image >= images.size() || pair.sampler >= base.samplers.size()) {
 			return false;
 		}
-		usage[pair.sampler] |= RequiresPointSampler(images[pair.image]) ? 2u : 1u;
+		usage[pair.sampler] |= 1u << static_cast<uint32_t>(ClassifySampler(images[pair.image]));
 	}
 	for (uint32_t index = 0; index < base.samplers.size(); index++) {
-		if ((usage[index] & 2u) == 0u) {
-			continue;
-		}
-		if ((usage[index] & 1u) == 0u) {
-			plan.point_sampler[index] = index;
-		} else {
-			if (plan.sampler_count >= ShaderInfo::MaxSamplers) {
-				return false;
-			}
-			plan.point_sampler[index] = plan.sampler_count++;
+		auto& mapping = plan.mapping[index];
+		mapping.fill(UINT32_MAX);
+		const auto classes = usage[index] == 0u ? 1u : usage[index];
+		bool       first   = true;
+		for (uint32_t type = 0; type < mapping.size(); type++) {
+			if ((classes & (1u << type)) == 0u) continue;
+			const auto target = first ? index : plan.sampler_count++;
+			if (target >= ShaderInfo::MaxSamplers) return false;
+			mapping[type]         = target;
+			plan.bindings[target] = {index, static_cast<SamplerClass>(type)};
+			first                = false;
 		}
 	}
 	return true;
@@ -862,7 +865,6 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	plan.memory_info                = program.memory_info;
 	plan.srt_plan_complete          = program.srt_plan_complete;
 	plan.resource_tracking_complete = program.resource_tracking_complete;
-	plan.has_address_writes         = program.has_address_writes;
 
 	std::unordered_map<const Inst*, Inst*> cloned;
 	std::function<Value(Value)>            Clone = [&](Value value) -> Value {
@@ -921,17 +923,25 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		plan.uniform_fill.values[i] = Clone(plan.uniform_fill.values[i]);
 	}
 	plan.clean_flat_slots.resize(plan.srt_reads.size());
+	bool masked_image = false;
 	for (const auto& image: plan.info.images) {
 		const auto* source = Source(plan, image.source);
 		if (source == nullptr || !source->indirect_image.has_value()) {
 			continue;
 		}
 		plan.requires_specialization_memory = true;
+		masked_image |= !source->indirect_image->selector_mask.IsEmpty();
 		MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->material_source),
 		                   plan.clean_flat_slots, source->indirect_image->selector_mask);
 		MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->table_source),
 		                   plan.clean_flat_slots);
 	}
+	if (masked_image) {
+		plan.resource_tracking_complete &= !program.has_address_writes &&
+		    !std::ranges::any_of(plan.info.images, &ImageResource::written);
+		plan.capture_specialization_reads =
+		    std::ranges::any_of(plan.info.buffers, &BufferResource::written);
+	}
 	return plan;
 }
 
@@ -941,18 +951,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	    (program.requires_specialization_memory && runtime.read_specialization_memory == nullptr)) {
 		return false;
 	}
-	const bool masked_image = std::ranges::any_of(program.info.images, [&](const auto& image) {
-		const auto* source = Source(program, image.source);
-		return source != nullptr && source->indirect_image.has_value() &&
-		       !source->indirect_image->selector_mask.IsEmpty();
-	});
-	if (masked_image &&
-	    (program.has_address_writes ||
-	     std::ranges::any_of(program.info.images, &ImageResource::written))) {
-		return false;
-	}
-	const bool capture_reads = masked_image &&
-	    std::ranges::any_of(program.info.buffers, &BufferResource::written);
+	const bool capture_reads = program.capture_specialization_reads;
 	auto& reads = program.specialization_reads;
 	ReadCapture capture {runtime, reads};
 	SrtRuntime observed = runtime;
@@ -1105,25 +1104,18 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 	auto samplers      = program.info.samplers;
 	auto sampled_pairs = program.info.sampled_pairs;
 	samplers.reserve(sampler_plan.sampler_count);
-	for (uint32_t index = 0; index < program.info.samplers.size(); index++) {
-		const auto target = sampler_plan.point_sampler[index];
-		if (target == UINT32_MAX) {
-			continue;
-		}
-		if (target == index) {
-			samplers[index].force_point_filtering = true;
-		} else {
-			EXIT_IF(target != samplers.size());
-			auto sampler                  = samplers[index];
-			sampler.force_point_filtering = true;
-			samplers.push_back(std::move(sampler));
+	for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {
+		const auto& binding = sampler_plan.bindings[index];
+		if (index >= program.info.samplers.size()) {
+			samplers.push_back(program.info.samplers[binding.source]);
 		}
+		samplers[index].force_point_filtering = binding.type == SamplerClass::PointInteger;
+		samplers[index].integer_border        = binding.type != SamplerClass::Float;
 	}
 	for (auto& pair: sampled_pairs) {
-		if (RequiresPointSampler(images[pair.image])) {
-			EXIT_IF(sampler_plan.point_sampler[pair.sampler] == UINT32_MAX);
-			pair.sampler = sampler_plan.point_sampler[pair.sampler];
-		}
+		const auto type = static_cast<uint32_t>(ClassifySampler(images[pair.image]));
+		pair.sampler = sampler_plan.mapping[pair.sampler][type];
+		EXIT_IF(pair.sampler == UINT32_MAX);
 		samplers[pair.sampler].depth_compare |= images[pair.image].depth_compare;
 	}
 
@@ -1201,10 +1193,11 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 				inst.ReplaceUsesWith(Value(&*result));
 				continue;
 			}
-			if (image_opcode.needs_sampler && RequiresPointSampler(image) &&
+			if (image_opcode.needs_sampler &&
 			    memory.sampler < program.info.samplers.size()) {
-				EXIT_IF(sampler_plan.point_sampler[memory.sampler] == UINT32_MAX);
-				memory.sampler = sampler_plan.point_sampler[memory.sampler];
+				const auto type = static_cast<uint32_t>(ClassifySampler(image));
+				memory.sampler = sampler_plan.mapping[memory.sampler][type];
+				EXIT_IF(memory.sampler == UINT32_MAX);
 			}
 			EXIT_IF(image.indirect_root == memory.resource &&
 			        inst.GetOpcode() != ValueOpcode::ImageSampleRaw);

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

@@ -5,9 +5,8 @@
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-// Collects immutable resource topology from typed SSA handles, interns their resolved dwords in
-// descriptor_sources, then writes dense indices to handle flags and MemoryInfo.
-void TrackResources(Program& program);
+// Resolves native descriptor sources, plans their scalar reads, and assigns dense resource bindings.
+void TrackResources(Program& program, const Decoder::Program& decoded, const CFG::Graph& native_cfg);
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR
 

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

@@ -326,6 +326,8 @@ void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 }
 
 void CollectOutputs(const Program& program, ShaderStageInputInfo input_info, ShaderInfo& info) {
+	const bool alpha_remap = program.stage == ShaderType::Pixel && input_info.pixel != nullptr &&
+	                         input_info.pixel->alpha_blend_source_remap;
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
 			if (inst.GetOpcode() != ValueOpcode::SetAttribute) {
@@ -385,8 +387,14 @@ void CollectOutputs(const Program& program, ShaderStageInputInfo input_info, Sha
 					          export_info.index, fmt::format("out_param_{}", export_info.index));
 					break;
 				case ExportTargetKind::Mrt:
+					if (alpha_remap && export_info.index != 0) {
+						break;
+					}
 					AddOutput(info, StageOutputKind::Mrt, export_info.index, export_info.index,
 					          fmt::format("out_mrt_{}", export_info.index));
+					if (alpha_remap) {
+						AddOutput(info, StageOutputKind::Mrt, 1, 1, "out_mrt_1");
+					}
 					break;
 				default: break;
 			}

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

@@ -7,7 +7,6 @@
 #include <bit>
 #include <cmath>
 #include <cstring>
-#include <fmt/format.h>
 #include <unordered_set>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
@@ -23,21 +22,6 @@ namespace {
 
 constexpr uint64_t AddressMask = 0x0000ffffffffffffull;
 
-const char* StageName(ShaderType stage) {
-	switch (stage) {
-		case ShaderType::Vertex: return "vertex";
-		case ShaderType::Pixel: return "pixel";
-		case ShaderType::Fetch: return "fetch";
-		case ShaderType::Compute: return "compute";
-		default: return "unknown";
-	}
-}
-
-std::string Diagnostic(const ResourcePlan& program, uint32_t pc, const std::string& message) {
-	return fmt::format("shader SRT: hash=0x{:016x} stage={} pc=0x{:08x} {}", program.shader_hash,
-	                   StageName(program.stage), pc, message);
-}
-
 bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
 	if (base > AddressMask) {
 		return false;
@@ -89,6 +73,7 @@ bool IsRuntimeSelect(ValueOpcode op) {
 
 bool IsRuntimeUniformOp(ValueOpcode op) {
 	switch (op) {
+		case ValueOpcode::ConditionRef:
 		case ValueOpcode::BitCastU32F32:
 		case ValueOpcode::BitCastF32U32:
 		case ValueOpcode::ConvertU32F32:
@@ -312,151 +297,6 @@ private:
 	std::unordered_set<const Inst*> m_validated_dependencies;
 };
 
-class PlanBuilder {
-public:
-	explicit PlanBuilder(Program& program): m_program(program) {}
-
-	void Run() {
-		m_program.srt_reads.clear();
-		m_program.dynamic_reads.clear();
-		for (auto* block: m_program.blocks) {
-			for (auto& inst: *block) {
-				const auto op = inst.GetOpcode();
-				if (op == ValueOpcode::LoadAddressU32 || op == ValueOpcode::ReadConstBuffer) {
-					const auto flags = inst.Flags<MemoryFlags>();
-					if (flags.index < m_program.memory_info.size()) {
-						const auto kind       = m_program.memory_info[flags.index].kind;
-						const bool crosswired = (op == ValueOpcode::LoadAddressU32 &&
-						                         kind == ResourceKind::ScalarBuffer) ||
-						                        (op == ValueOpcode::ReadConstBuffer &&
-						                         kind == ResourceKind::ScalarAddress);
-						if (crosswired) {
-							Fail(flags.pc,
-							     fmt::format("{} has incompatible scalar memory metadata",
-							                 ValueOpcodeName(op)));
-						}
-					}
-				}
-				if (IsDescriptorHandle(inst.GetOpcode())) {
-					for (size_t index = 0; index < inst.NumArgs(); index++) {
-						Collect(inst.Arg(index), 0);
-					}
-				}
-			}
-		}
-		for (auto* block: m_program.blocks) {
-			for (auto& inst: *block) {
-				if (inst.GetOpcode() == ValueOpcode::LoadAddressU32 && IsRawRead(m_program, inst) &&
-				    inst.Arg(1).Resolve().IsImmediate() &&
-				    ValidateRuntimeValue(m_program, Value(&inst))) {
-					Collect(Value(&inst), inst.Flags<MemoryFlags>().pc);
-				}
-			}
-		}
-		PatchReads();
-	}
-
-private:
-	struct Patch {
-		Inst*    inst = nullptr;
-		uint32_t slot = 0;
-		bool     keep = false;
-	};
-
-	[[noreturn]] void Fail(uint32_t pc, const std::string& message) const {
-		const auto diagnostic = Diagnostic(m_program, pc, message);
-		EXIT("shader SRT planning failed: %s", diagnostic.c_str());
-		std::abort();
-	}
-
-	void Collect(Value value, uint32_t use_pc) {
-		value = value.Resolve();
-		if (value.IsImmediate()) {
-			return;
-		}
-		auto* inst = value.TryInstruction();
-		if (inst == nullptr) {
-			Fail(use_pc, "invalid typed planning value");
-		}
-		const auto cycle = std::ranges::find(m_visiting, inst);
-		if (cycle != m_visiting.end()) {
-			const auto contains_phi = std::any_of(cycle, m_visiting.end(), [](const Inst* value) {
-				return value->GetOpcode() == ValueOpcode::Phi;
-			});
-			if (contains_phi) {
-				return;
-			}
-			Fail(use_pc, fmt::format("cyclic typed planning value {} without a phi",
-			                         ValueOpcodeName(inst->GetOpcode())));
-		}
-		if (std::ranges::find(m_visited, inst) != m_visited.end()) {
-			return;
-		}
-		m_visiting.push_back(inst);
-		for (size_t index = 0; index < inst->NumArgs(); index++) {
-			Collect(inst->Arg(index), use_pc);
-		}
-		m_visiting.pop_back();
-		m_visited.push_back(inst);
-		if (!IsRawRead(m_program, *inst)) {
-			return;
-		}
-		const auto offset = inst->Arg(1).Resolve();
-		if (!offset.IsImmediate() || offset.GetType() != Type::U32) {
-			if (std::ranges::find(m_program.dynamic_reads, value) ==
-			    m_program.dynamic_reads.end()) {
-				m_program.dynamic_reads.push_back(value);
-			}
-			return;
-		}
-		for (uint32_t slot = 0; slot < m_program.srt_reads.size(); slot++) {
-			if (EquivalentValue(m_program, value, m_program.srt_reads[slot].value)) {
-				m_patches.push_back({inst, slot, false});
-				return;
-			}
-		}
-		const auto slot = static_cast<uint32_t>(m_program.srt_reads.size());
-		m_program.srt_reads.push_back({value, slot});
-		m_patches.push_back({inst, slot, true});
-	}
-
-	void PatchReads() {
-		for (const auto& patch: m_patches) {
-			auto* block = patch.inst->Parent();
-			auto& list  = block->Instructions();
-			auto  where =
-			    std::ranges::find_if(list, [&](const Inst& inst) { return &inst == patch.inst; });
-			const auto resource =
-			    Value(&*block->PrependNewInst(where, ValueOpcode::GetSrtResource));
-			const auto flat = Value(&*block->PrependNewInst(where, ValueOpcode::ReadConst,
-			                                                {resource, Value(patch.slot)}));
-			const auto uses = patch.inst->Uses();
-			for (const auto& use: uses) {
-				use.user->SetArg(use.operand, flat);
-			}
-			for (auto& info: m_program.block_info) {
-				if (info.condition.Resolve() == Value(patch.inst)) {
-					info.condition = flat;
-				}
-				if (info.indirect_target.Resolve() == Value(patch.inst)) {
-					info.indirect_target = flat;
-				}
-			}
-			if (patch.keep) {
-				const auto memory = patch.inst->Flags<MemoryFlags>().index;
-				if (memory < m_program.memory_info.size()) {
-					m_program.memory_info[memory].planning_only = true;
-				}
-				block->AppendNewInst(ValueOpcode::ReferenceU32, {Value(patch.inst)});
-			}
-		}
-	}
-
-	Program&           m_program;
-	std::vector<Inst*> m_visiting;
-	std::vector<Inst*> m_visited;
-	std::vector<Patch> m_patches;
-};
 
 } // namespace
 
@@ -616,16 +456,15 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 			return false;
 		}
 		const auto byte_offset =
-		    static_cast<uint64_t>(immediate) + static_cast<uint32_t>(offset);
-		const auto aligned = byte_offset & ~uint64_t {3};
+		    (static_cast<uint64_t>(immediate) & ~uint64_t {3}) + (static_cast<uint32_t>(offset) & ~3u);
 		const auto stride  = (static_cast<uint32_t>(high) >> 16u) & 0x3fffu;
 		const auto size = stride == 0u
 		                      ? static_cast<uint64_t>(static_cast<uint32_t>(records))
 		                      : static_cast<uint64_t>(stride) * static_cast<uint32_t>(records);
-		if (aligned > size || size - aligned < sizeof(uint32_t)) {
+		if (byte_offset > size || size - byte_offset < sizeof(uint32_t)) {
 			return false;
 		}
-		address = ((base & ~uint64_t {3}) + byte_offset) & ~uint64_t {3};
+		address = (base & ~uint64_t {3}) + byte_offset;
 	} else {
 		const auto relative = (immediate & ~int64_t {3}) +
 		                      static_cast<int64_t>(static_cast<uint32_t>(offset) & ~3u);
@@ -780,14 +619,18 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 			}
 			return false;
 		case ValueOpcode::FPOrdLessThanEqual32:
-			if (binary()) {
-				result = Float32(a) <= Float32(b);
-				return true;
-			}
-			return false;
 		case ValueOpcode::FPOrdGreaterThanEqual32:
 			if (binary()) {
-				result = Float32(a) >= Float32(b);
+				const auto operand = [&](uint64_t bits) {
+					if (inst.Flags<FPCompareFlags>().flush_input_denorms &&
+					    (bits & 0x7fffffffu) < 0x00800000u) {
+						bits &= 0x80000000u;
+					}
+					return Float32(bits);
+				};
+				result = inst.GetOpcode() == ValueOpcode::FPOrdLessThanEqual32
+				             ? operand(a) <= operand(b)
+				             : operand(a) >= operand(b);
 				return true;
 			}
 			return false;
@@ -970,6 +813,7 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::ConditionRef: return Arg(inst, 0, result);
 		case ValueOpcode::LogicalNot:
 			if (Arg(inst, 0, a)) {
 				result = a == 0u;
@@ -1061,13 +905,5 @@ bool ValidateRuntimeValue(const ResourcePlan& program, Value value, RuntimeValue
 	return RuntimeValidator(program, type).Run(value);
 }
 
-void BuildSrtPlan(Program& program) {
-	if (program.resource_tracking_complete) {
-		EXIT("shader SRT planning failed: cannot rebuild SRT after resource tracking");
-	}
-	program.srt_plan_complete = false;
-	PlanBuilder(program).Run();
-	program.srt_plan_complete = true;
-}
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -21,9 +21,6 @@ struct SrtRuntime {
 
 enum class RuntimeValueType { Any, Integer };
 
-// Collects reachable ReadConst values. Immediate offsets receive compact flat-buffer slots;
-// dynamic offsets remain explicit and are never assigned a fake slot.
-void BuildSrtPlan(Program& program);
 bool ValidateRuntimeValue(const ResourcePlan& program, Value value,
                           RuntimeValueType type = RuntimeValueType::Any);
 // Uses the strict reader for values that affect shader specialization.

src/graphics/shader/shader.cpp

@@ -336,14 +336,15 @@ static void ShaderDetectBuffers(ShaderVertexInputInfo& info) {
 	for (int ri = 0; ri < info.resources_num; ri++) {
 		const auto& r = info.resources[ri];
 
-		bool merged = false;
-		for (int bi = 0; bi < info.buffers_num; bi++) {
+		auto& destination = info.resources_dst[ri];
+		int bi = 0;
+		for (; bi < info.buffers_num; bi++) {
 			auto& b = info.buffers[bi];
 
 			uint64_t stride = b.stride;
 
 			if (stride == r.Stride() &&
-			    b.fetch_index == static_cast<uint32_t>(info.resources_dst[ri].fetch_index)) {
+			    b.fetch_index == destination.fetch_index) {
 				uint64_t rbase   = r.Base48();
 				uint64_t base    = std::min(rbase, b.addr);
 				uint64_t offset1 = rbase - base;
@@ -352,45 +353,33 @@ static void ShaderDetectBuffers(ShaderVertexInputInfo& info) {
 				if (offset1 < stride && offset2 < stride) {
 					EXIT_NOT_IMPLEMENTED(b.num_records != r.NumRecords());
 					b.addr = base;
-					EXIT_NOT_IMPLEMENTED(b.attr_num >= ShaderVertexInputBuffer::ATTR_MAX);
-					b.attr_indices[b.attr_num++] = ri;
-					merged                       = true;
 					break;
 				}
 			}
 		}
 
-		if (!merged) {
+		if (bi == info.buffers_num) {
 			EXIT_NOT_IMPLEMENTED(info.buffers_num >= ShaderVertexInputInfo::RES_MAX);
-			int bi                           = info.buffers_num++;
-			info.buffers[bi].addr            = r.Base48();
-			info.buffers[bi].stride          = r.Stride();
-			info.buffers[bi].num_records     = r.NumRecords();
-			info.buffers[bi].fetch_index     = info.resources_dst[ri].fetch_index;
-			info.buffers[bi].attr_num        = 1;
-			info.buffers[bi].attr_indices[0] = ri;
-		}
-	}
-
-	for (int bi = 0; bi < info.buffers_num; bi++) {
-		auto& b = info.buffers[bi];
-		for (int ri = 0; ri < b.attr_num; ri++) {
-			b.attr_offsets[ri] = info.resources[b.attr_indices[ri]].Base48() - b.addr;
+			info.buffers[bi] = {.addr = r.Base48(),
+			                    .stride = r.Stride(),
+			                    .num_records = r.NumRecords(),
+			                    .fetch_index = destination.fetch_index};
+			info.buffers_num++;
 		}
+		destination.buffer_index = bi;
 	}
 }
 
 static void ShaderApplyAttribSemantics(ShaderVertexInputInfo& info,
-                                       const ShaderSemantic*  input_semantics,
-                                       uint32_t num_input_semantics, const uint32_t* attrib,
-                                       const uint32_t* buffer) {
+                                       std::span<const ShaderSemantic> input_semantics,
+                                       const uint32_t* attrib, const uint32_t* buffer) {
 	KYTY_PROFILER_FUNCTION();
 
 	EXIT_IF(attrib == nullptr || buffer == nullptr);
 
 	const bool debug_dump = Config::GraphicsDebugDumpEnabled();
 
-	for (uint32_t i = 0; i < num_input_semantics; i++) {
+	for (uint32_t i = 0; i < input_semantics.size(); i++) {
 		const auto& in = input_semantics[i];
 
 		EXIT_NOT_IMPLEMENTED(in.static_vb_index == 1 || in.static_attribute == 1);
@@ -544,8 +533,7 @@ static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserS
 			     shader_addr);
 			return false;
 		}
-		ShaderApplyAttribSemantics(info, metadata.input_semantics.data(),
-		                           metadata.input_semantics_count, attrib, buffer);
+		ShaderApplyAttribSemantics(info, metadata.input_semantics, attrib, buffer);
 		ShaderDetectBuffers(info);
 	}
 	return true;
@@ -678,15 +666,8 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 		key.push_back(destination.registers_num);
 		key.push_back(destination.fetch_index);
 		key.push_back(static_cast<uint32_t>(destination.attr_id));
-		key.push_back(resource.Stride());
-		key.push_back(static_cast<uint32_t>(resource.SwizzleEnabled()));
-		key.push_back(resource.DstSelX());
-		key.push_back(resource.DstSelY());
-		key.push_back(resource.DstSelZ());
-		key.push_back(resource.DstSelW());
-		key.push_back(resource.RawFormat());
-		key.push_back(resource.OutOfBounds());
-		key.push_back(static_cast<uint32_t>(resource.AddTid()));
+		key.push_back(resource.fields[1] & 0xbfff0000u); // Stride and swizzle enable.
+		key.push_back(resource.fields[3] & 0x3087ffffu); // Channels, format, OOB, and add TID.
 	}
 }
 
@@ -712,6 +693,7 @@ void BuildStageStaticKey(const ShaderPixelInputInfo& info, std::vector<uint32_t>
 	key.push_back(static_cast<uint32_t>(info.ps_sample_mask_export_enable));
 	key.push_back(static_cast<uint32_t>(info.ps_early_z));
 	key.push_back(static_cast<uint32_t>(info.dual_source_blending));
+	key.push_back(static_cast<uint32_t>(info.alpha_blend_source_remap));
 	key.insert(key.end(), std::begin(info.target_output_mode), std::end(info.target_output_mode));
 	for (uint32_t base = 0; base < info.target_export_mapping.size(); base += 4u) {
 		uint32_t packed = 0;
@@ -907,8 +889,9 @@ void ShaderDbgDumpInputInfo(const ShaderVertexInputInfo& info) {
 
 		LOGF("\t\t register_start   = %d\n"
 		     "\t\t registers_num    = %d\n"
-		     "\t\t fetch_index      = %" PRIu32 "\n",
-		     rd.register_start, rd.registers_num, rd.fetch_index);
+		     "\t\t fetch_index      = %" PRIu32 "\n"
+		     "\t\t buffer_index     = %d\n",
+		     rd.register_start, rd.registers_num, rd.fetch_index, rd.buffer_index);
 		LOGF("\t\t fields           = %08" PRIx32 "%08" PRIx32 "%08" PRIx32 "%08" PRIx32 "\n",
 		     r.fields[3], r.fields[2], r.fields[1], r.fields[0]);
 		LOGF("\t\t Base()           = %" PRIx64 "\n"
@@ -934,14 +917,8 @@ void ShaderDbgDumpInputInfo(const ShaderVertexInputInfo& info) {
 		LOGF("\t\t addr        = %" PRIx64 "\n"
 		     "\t\t stride      = %" PRIu32 "\n"
 		     "\t\t num_records = %" PRIu32 "\n"
-		     "\t\t fetch_index = %" PRIu32 "\n"
-		     "\t\t attr_num    = %" PRId32 "\n",
-		     r.addr, r.stride, r.num_records, r.fetch_index, r.attr_num);
-		for (int j = 0; j < r.attr_num; j++) {
-			LOGF("\t\t attr_indices[%d]  = %d\n"
-			     "\t\t attr_offsets[%d]  = %u\n",
-			     j, r.attr_indices[j], j, r.attr_offsets[j]);
-		}
+		     "\t\t fetch_index = %" PRIu32 "\n",
+		     r.addr, r.stride, r.num_records, r.fetch_index);
 	}
 }
 

src/graphics/shader/shader.h

@@ -178,6 +178,8 @@ struct ShaderPixelInputInfo {
 	bool                                           ps_sample_mask_export_enable = false;
 	bool                                           ps_sample_shading            = false;
 	bool                                           dual_source_blending         = false;
+	// Export logical alpha through MRT1 for blending after channel swizzling.
+	bool                                           alpha_blend_source_remap     = false;
 	bool                                           ps_early_z                   = false;
 	bool                                           ps_execute_on_noop           = false;
 	ShaderStageRuntime                             stage;

src/graphics/shader/shaderBindings.h

@@ -174,15 +174,10 @@ struct ShaderSamplerResource {
 };
 
 struct ShaderVertexInputBuffer {
-	static constexpr int ATTR_MAX = 32;
-
-	uint64_t addr                   = 0;
-	uint32_t stride                 = 0;
-	uint32_t num_records            = 0;
-	uint32_t fetch_index            = 0;
-	int      attr_num               = 0;
-	int      attr_indices[ATTR_MAX] = {0};
-	uint32_t attr_offsets[ATTR_MAX] = {0};
+	uint64_t addr        = 0;
+	uint32_t stride      = 0;
+	uint32_t num_records = 0;
+	uint32_t fetch_index = 0;
 };
 
 struct ShaderVertexDestination {
@@ -190,6 +185,7 @@ struct ShaderVertexDestination {
 	int      registers_num  = 0;
 	int      attr_id        = -1;
 	uint32_t fetch_index    = 0;
+	int      buffer_index   = 0;
 };
 
 enum class ShaderStorageUsage {

src/graphics/shader/shaderVertexMetadata.cpp

@@ -68,13 +68,10 @@ bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sg
 		return ShaderError::Fail(error, "vertex semantic count is outside the supported domain");
 	}
 
-	const auto semantic_size =
-	    static_cast<uint64_t>(data.num_input_semantics) * sizeof(ShaderSemantic);
 	if (data.input_semantics == nullptr) {
 		return ShaderError::Fail(error, "missing vertex input semantics");
 	}
-	std::memcpy(next.input_semantics.data(), data.input_semantics, semantic_size);
-	next.input_semantics_count = data.num_input_semantics;
+	next.input_semantics = {data.input_semantics, data.num_input_semantics};
 	metadata                   = next;
 	return true;
 }

src/graphics/shader/shaderVertexMetadata.h

@@ -3,19 +3,18 @@
 
 #include "graphics/shader/shader.h"
 
-#include <array>
+#include <span>
 #include <string>
 
 namespace Libs::Graphics {
 
 struct ShaderVertexMetadata {
-	int                                                        vertex_buffer_reg = -1;
-	int                                                        vertex_attrib_reg = -1;
-	std::array<ShaderSemantic, ShaderVertexInputInfo::RES_MAX> input_semantics {};
-	uint32_t                                                   input_semantics_count = 0;
+	int                             vertex_buffer_reg = -1;
+	int                             vertex_attrib_reg = -1;
+	std::span<const ShaderSemantic> input_semantics;
 };
 
-// Copies the small AGC metadata subset used by the vertex path after validating every guest range.
+// 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);
 

src/kernel/fileSystem.cpp

@@ -1,5 +1,6 @@
 #include "kernel/fileSystem.h"
 
+#include "common/archive.h"
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/dateTime.h"
@@ -41,6 +42,7 @@ public:
 	struct MountPair {
 		std::filesystem::path dir;
 		std::string           point;
+		std::shared_ptr<Common::ArchiveReader> archive;
 	};
 
 	MountPoints() { EXIT_NOT_IMPLEMENTED(!Common::Thread::IsMainThread()); }
@@ -260,6 +262,7 @@ void MountPoints::Mount(const std::filesystem::path& folder, const std::string&
 	MountPair p;
 	p.dir   = folder;
 	p.point = point_str;
+	p.archive = Common::OpenArchive(folder);
 
 	m_mount_pairs.push_back(std::move(p));
 }
@@ -346,11 +349,13 @@ std::filesystem::path MountPoints::ResolvePath(const std::string& mounted_name)
 		}
 
 		const auto native_rel_path = Common::PathFromUtf8(rel_path);
+		if (Common::IsArchivePath(p.dir)) {
+			return p.dir / native_rel_path;
+		}
 
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		if (HasWindowsForbiddenFilenameCharacter(rel_path)) {
-			::printf("FileSystem: Windows-incompatible guest filename: %s\n",
-			         mounted_name.c_str());
+			::printf("FileSystem: Windows-incompatible guest filename: %s\n", mounted_name.c_str());
 		}
 		return p.dir / native_rel_path;
 #else
@@ -358,7 +363,8 @@ std::filesystem::path MountPoints::ResolvePath(const std::string& mounted_name)
 #endif
 	}
 
-	return mounted_name;
+	// Unmounted guest paths must not fall through to the host filesystem.
+	return {};
 }
 
 void Initialize() {
@@ -461,14 +467,24 @@ int KYTY_SYSV_ABI KernelOpen(const char* path, int flags, uint16_t mode) {
 	}
 
 	file->real_name = g_mount_points->ResolvePath(file->name);
+	if (file->real_name.empty()) {
+		g_files->DeleteDescriptor(descriptor);
+		return KERNEL_ERROR_ENOENT;
+	}
+	if (Common::IsArchivePath(file->real_name) &&
+	    (rw_mode != Common::File::Mode::Read || trunc || creat)) {
+		g_files->DeleteDescriptor(descriptor);
+		return KERNEL_ERROR_EROFS;
+	}
 
 	if (trunc && rw_mode == Common::File::Mode::Read) {
 		g_files->DeleteDescriptor(descriptor);
 		return KERNEL_ERROR_EACCES;
 	}
 
-	bool dir_exist  = Common::File::IsDirectoryExisting(file->real_name);
-	bool file_exist = Common::File::IsFileExisting(file->real_name);
+	const auto info = Common::File::GetInfo(file->real_name);
+	const bool dir_exist = info && !info->is_file;
+	const bool file_exist = info && info->is_file;
 
 	// A missing path opened without O_CREAT is ENOENT
 	if (!creat && !dir_exist && !file_exist) {
@@ -666,7 +682,7 @@ int64_t KYTY_SYSV_ABI KernelWrite(int d, const void* buf, size_t nbytes) {
 
 	auto* file = g_files->GetFile(d);
 
-	if (file == nullptr) {
+	if (file == nullptr || !file->opened || !file->writable) {
 		return KERNEL_ERROR_EBADF;
 	}
 
@@ -881,7 +897,7 @@ int64_t KYTY_SYSV_ABI KernelPwrite(int d, const void* buf, size_t nbytes, int64_
 
 	auto* file = g_files->GetFile(d);
 
-	if (file == nullptr) {
+	if (file == nullptr || !file->opened || !file->writable) {
 		return KERNEL_ERROR_EBADF;
 	}
 
@@ -1058,15 +1074,13 @@ int KYTY_SYSV_ABI KernelStat(const char* path, FileStat* sb) {
 
 	auto real_file_name = g_mount_points->ResolvePath(path);
 
-	bool is_dir  = Common::File::IsDirectoryExisting(real_file_name);
-	bool is_file = Common::File::IsFileExisting(real_file_name);
-
-	if (!is_dir && !is_file) {
+	const auto info = Common::File::GetInfo(real_file_name);
+	if (!info) {
 		LOGF("\t file not found\n");
 		return KERNEL_ERROR_ENOENT;
 	}
 
-	EXIT_NOT_IMPLEMENTED(is_dir && is_file);
+	const bool is_dir = !info->is_file;
 
 	FileStat stat {};
 	stat.st_mode = 0000777u | (is_dir ? 0040000u : 0100000u);
@@ -1080,7 +1094,7 @@ int KYTY_SYSV_ABI KernelStat(const char* path, FileStat* sb) {
 		stat.st_blksize = 512;
 		stat.st_blocks  = stat.st_size / 512;
 	} else {
-		stat.st_size    = static_cast<int64_t>(Common::File::Size(real_file_name));
+		stat.st_size    = static_cast<int64_t>(info->size);
 		stat.st_blksize = 512;
 		stat.st_blocks  = (stat.st_size + 511) / 512;
 
@@ -1217,6 +1231,9 @@ int KYTY_SYSV_ABI KernelUnlink(const char* path) {
 	}
 
 	auto real_file_name = g_mount_points->ResolvePath(path);
+	if (Common::IsArchivePath(real_file_name)) {
+		return KERNEL_ERROR_EROFS;
+	}
 
 	bool is_dir  = Common::File::IsDirectoryExisting(real_file_name);
 	bool is_file = Common::File::IsFileExisting(real_file_name);
@@ -1254,8 +1271,11 @@ int KYTY_SYSV_ABI KernelRename(const char* from, const char* to) {
 	auto to_path   = std::string(to);
 	auto real_from = g_mount_points->ResolvePath(from_path);
 	auto real_to   = g_mount_points->ResolvePath(to_path);
+	if (Common::IsArchivePath(real_from) || Common::IsArchivePath(real_to)) {
+		return KERNEL_ERROR_EROFS;
+	}
 
-	if (!Common::File::IsFileExisting(real_from)) {
+	if (real_to.empty() || !Common::File::IsFileExisting(real_from)) {
 		return KERNEL_ERROR_ENOENT;
 	}
 
@@ -1335,6 +1355,12 @@ int KYTY_SYSV_ABI KernelMkdir(const char* path, uint16_t mode) {
 	     path, mode);
 
 	auto real_name = g_mount_points->ResolvePath(std::string(path));
+	if (real_name.empty()) {
+		return KERNEL_ERROR_ENOENT;
+	}
+	if (Common::IsArchivePath(real_name)) {
+		return KERNEL_ERROR_EROFS;
+	}
 
 	if (Common::File::IsDirectoryExisting(real_name)) {
 		return KERNEL_ERROR_EEXIST;
@@ -1361,6 +1387,9 @@ int KYTY_SYSV_ABI KernelRmdir(const char* path) {
 	LOGF("\t path = %s\n", path);
 
 	auto real_name = g_mount_points->ResolvePath(std::string(path));
+	if (Common::IsArchivePath(real_name)) {
+		return KERNEL_ERROR_EROFS;
+	}
 
 	if (!Common::File::IsDirectoryExisting(real_name)) {
 		return KERNEL_ERROR_ENOENT;

src/kernel/fileSystem.h

@@ -54,6 +54,7 @@ struct Lifecycle {
 
 void                  Mount(const std::filesystem::path& folder, const std::string& point);
 void                  Umount(const std::string& folder_or_point);
+// Returns an empty path when the guest path has no mounted filesystem.
 std::filesystem::path GetRealFilename(const std::string& mounted_file_name);
 
 int KYTY_SYSV_ABI     KernelOpen(const char* path, int flags, uint16_t mode);

src/kernel/memory.cpp

@@ -78,8 +78,7 @@ static Graphics::RenderContext& GetGpuResources() {
 }
 
 static bool IsGpuAddressRange(uint64_t vaddr, uint64_t size) {
-	constexpr uint64_t GPU_ADDRESS_LIMIT = 1ull << 40u;
-	return vaddr != 0 && size != 0 && vaddr < GPU_ADDRESS_LIMIT && size < GPU_ADDRESS_LIMIT - vaddr;
+	return Graphics::GuestRange {vaddr, size}.Valid();
 }
 
 static void MapGpuRange(uint64_t vaddr, uint64_t size) {
@@ -660,6 +659,9 @@ static bool     g_test_fail_next_fixed_reserve_range_add     = false;
 
 class PhysicalMemory {
 public:
+	enum class AllocationKind { Direct, Pooled, Automatic };
+	using PhysicalRanges = std::vector<std::pair<uint64_t, uint64_t>>;
+
 	struct AllocatedBlock {
 		uint64_t            start_addr;
 		uint64_t            size;
@@ -671,7 +673,7 @@ public:
 		VirtualMemory::Mode mode;
 		GpuAccessMode       gpu_mode;
 		int                 memory_type;
-		bool                pool_expansion;
+		AllocationKind      kind;
 		char                name[KERNEL_MAXIMUM_NAME_LENGTH];
 	};
 
@@ -690,7 +692,10 @@ public:
 	}
 
 	bool Alloc(uint64_t search_start, uint64_t search_end, size_t len, size_t alignment,
-	           uint64_t* phys_addr_out, int memory_type, bool pool_expansion = false);
+	           uint64_t* phys_addr_out, int memory_type,
+	           AllocationKind kind = AllocationKind::Direct);
+	bool ReserveAutomatic(uint64_t size, PhysicalRanges* ranges);
+	void RestoreAutomatic(const PhysicalRanges& ranges);
 	bool Available(uint64_t search_start, uint64_t search_end, size_t alignment,
 	               uint64_t* phys_addr_out, uint64_t* size_out);
 	bool Release(uint64_t start, size_t len, uint64_t* vaddr, uint64_t* size,
@@ -718,12 +723,14 @@ public:
 	[[nodiscard]] const std::vector<AllocatedBlock>& GetMappings() const { return m_mappings; }
 
 private:
-	void ConsumeFreeRange(std::map<uint64_t, uint64_t>::iterator range, uint64_t start,
-	                      uint64_t size);
-	void AddFreeRange(uint64_t start, uint64_t size);
+	static void RemoveFreeRange(std::map<uint64_t, uint64_t>& ranges, uint64_t start,
+	                            uint64_t size);
+	static void AddFreeRange(std::map<uint64_t, uint64_t>& ranges, uint64_t start, uint64_t size);
+	void        ReclaimAutomatic(uint64_t start, uint64_t size);
 
 	std::map<uint64_t, AllocatedBlock> m_physical;
 	std::map<uint64_t, uint64_t>       m_free;
+	std::map<uint64_t, uint64_t>       m_automatic_free;
 	std::vector<AllocatedBlock>        m_mappings;
 	Common::Mutex                      m_mutex;
 };
@@ -1093,7 +1100,7 @@ void SetFlexibleMemorySize(uint64_t size) {
 }
 
 bool PhysicalMemory::Alloc(uint64_t search_start, uint64_t search_end, size_t len, size_t alignment,
-                           uint64_t* phys_addr_out, int memory_type, bool pool_expansion) {
+                           uint64_t* phys_addr_out, int memory_type, AllocationKind kind) {
 	if (phys_addr_out == nullptr) {
 		return false;
 	}
@@ -1128,10 +1135,13 @@ bool PhysicalMemory::Alloc(uint64_t search_start, uint64_t search_end, size_t le
 		b.prot           = 0;
 		b.mode           = VirtualMemory::Mode::NoAccess;
 		b.memory_type    = memory_type;
-		b.pool_expansion = pool_expansion;
+		b.kind           = kind;
 
-		ConsumeFreeRange(range, free_pos, len);
+		RemoveFreeRange(m_free, free_pos, len);
 		EXIT_IF(!m_physical.emplace(b.start_addr, b).second);
+		if (kind == AllocationKind::Automatic) {
+			AddFreeRange(m_automatic_free, free_pos, len);
+		}
 
 		*phys_addr_out = free_pos;
 		return true;
@@ -1180,35 +1190,99 @@ bool PhysicalMemory::Available(uint64_t search_start, uint64_t search_end, size_
 	return true;
 }
 
-void PhysicalMemory::ConsumeFreeRange(std::map<uint64_t, uint64_t>::iterator range, uint64_t start,
-                                      uint64_t size) {
-	const auto range_start = range->first;
-	const auto range_end   = range->first + range->second;
-	m_free.erase(range);
-	if (range_start < start) {
-		m_free.emplace(range_start, start - range_start);
+void PhysicalMemory::RemoveFreeRange(std::map<uint64_t, uint64_t>& ranges, uint64_t start,
+                                     uint64_t size) {
+	const auto end   = start + size;
+	auto       range = ranges.upper_bound(start);
+	if (range != ranges.begin()) {
+		--range;
 	}
-	if (start + size < range_end) {
-		m_free.emplace(start + size, range_end - start - size);
+	while (range != ranges.end() && range->first < end) {
+		const auto range_start = range->first;
+		const auto range_end   = range_start + range->second;
+		if (range_end <= start) {
+			++range;
+			continue;
+		}
+		range = ranges.erase(range);
+		if (range_start < start) {
+			ranges.emplace(range_start, start - range_start);
+		}
+		if (end < range_end) {
+			ranges.emplace(end, range_end - end);
+		}
 	}
 }
 
-void PhysicalMemory::AddFreeRange(uint64_t start, uint64_t size) {
+void PhysicalMemory::AddFreeRange(std::map<uint64_t, uint64_t>& ranges, uint64_t start,
+                                  uint64_t size) {
 	auto end  = start + size;
-	auto next = m_free.lower_bound(start);
-	if (next != m_free.begin()) {
+	auto next = ranges.lower_bound(start);
+	if (next != ranges.begin()) {
 		auto previous = std::prev(next);
 		if (previous->first + previous->second >= start) {
 			start = previous->first;
 			end   = std::max(end, previous->first + previous->second);
-			next  = m_free.erase(previous);
+			next  = ranges.erase(previous);
 		}
 	}
-	while (next != m_free.end() && next->first <= end) {
+	while (next != ranges.end() && next->first <= end) {
 		end  = std::max(end, next->first + next->second);
-		next = m_free.erase(next);
+		next = ranges.erase(next);
+	}
+	ranges.emplace(start, end - start);
+}
+
+bool PhysicalMemory::ReserveAutomatic(uint64_t size, PhysicalRanges* ranges) {
+	Common::LockGuard lock(m_mutex);
+	for (const auto& [start, available]: m_automatic_free) {
+		const auto length = std::min(size, available);
+		ranges->emplace_back(start, length);
+		size -= length;
+		if (size == 0) {
+			for (const auto& [offset, length]: *ranges) {
+				RemoveFreeRange(m_automatic_free, offset, length);
+			}
+			return true;
+		}
+	}
+	ranges->clear();
+	return false;
+}
+
+void PhysicalMemory::RestoreAutomatic(const PhysicalRanges& ranges) {
+	Common::LockGuard lock(m_mutex);
+	for (const auto& [start, size]: ranges) {
+		AddFreeRange(m_automatic_free, start, size);
+	}
+}
+
+void PhysicalMemory::ReclaimAutomatic(uint64_t start, uint64_t size) {
+	const auto end   = start + size;
+	auto       owner = m_physical.upper_bound(start);
+	if (owner != m_physical.begin()) {
+		--owner;
+	}
+	bool automatic = false;
+	for (; owner != m_physical.end() && owner->first < end; ++owner) {
+		const auto& block = owner->second;
+		if (block.kind != AllocationKind::Automatic || block.start_addr + block.size <= start) {
+			continue;
+		}
+		const auto first = std::max(start, block.start_addr);
+		const auto last  = std::min(end, block.start_addr + block.size);
+		AddFreeRange(m_automatic_free, first, last - first);
+		automatic = true;
+	}
+	if (automatic) {
+		for (const auto& mapping: m_mappings) {
+			const auto first = std::max(start, mapping.start_addr);
+			const auto last  = std::min(end, mapping.start_addr + mapping.size);
+			if (first < last) {
+				RemoveFreeRange(m_automatic_free, first, last - first);
+			}
+		}
 	}
-	m_free.emplace(start, end - start);
 }
 
 bool PhysicalMemory::Release(uint64_t start, size_t len, uint64_t* vaddr, uint64_t* size,
@@ -1225,18 +1299,22 @@ bool PhysicalMemory::Release(uint64_t start, size_t len, uint64_t* vaddr, uint64
 	}
 	auto  it = std::prev(next);
 	auto& b  = it->second;
-	if (b.pool_expansion || start < b.start_addr || start >= b.start_addr + b.size ||
-	    len > b.start_addr + b.size - start) {
+	if (b.kind == AllocationKind::Pooled || start < b.start_addr ||
+	    start >= b.start_addr + b.size || len > b.start_addr + b.size - start) {
 		return false;
 	}
 
+	if (b.kind == AllocationKind::Automatic) {
+		RemoveFreeRange(m_automatic_free, start, len);
+	}
+
 	if (start == b.start_addr && len == b.size) {
 		*vaddr    = b.map_vaddr;
 		*size     = b.map_size;
 		*gpu_mode = b.gpu_mode;
 
 		m_physical.erase(it);
-		AddFreeRange(start, len);
+		AddFreeRange(m_free, start, len);
 		return true;
 	}
 	if (start > b.start_addr && start + len < b.start_addr + b.size) {
@@ -1261,7 +1339,7 @@ bool PhysicalMemory::Release(uint64_t start, size_t len, uint64_t* vaddr, uint64
 		}
 
 		m_physical.emplace(right.start_addr, right);
-		AddFreeRange(start, len);
+		AddFreeRange(m_free, start, len);
 		return true;
 	}
 	if (start == b.start_addr && len < b.size) {
@@ -1278,7 +1356,7 @@ bool PhysicalMemory::Release(uint64_t start, size_t len, uint64_t* vaddr, uint64
 			remaining.map_size -= len;
 		}
 		m_physical.emplace(remaining.start_addr, remaining);
-		AddFreeRange(start, len);
+		AddFreeRange(m_free, start, len);
 		return true;
 	}
 	if (start > b.start_addr && start + len == b.start_addr + b.size) {
@@ -1290,7 +1368,7 @@ bool PhysicalMemory::Release(uint64_t start, size_t len, uint64_t* vaddr, uint64
 		if (b.map_vaddr != 0) {
 			b.map_size = b.size;
 		}
-		AddFreeRange(start, len);
+		AddFreeRange(m_free, start, len);
 		return true;
 	}
 
@@ -1311,6 +1389,7 @@ bool PhysicalMemory::Map(uint64_t vaddr, uint64_t phys_addr, size_t len, int pro
 		return false;
 	}
 	auto first = std::prev(next);
+	bool has_automatic = false;
 	while (current < phys_addr + len) {
 		auto block = m_physical.upper_bound(current);
 		if (block == m_physical.begin()) {
@@ -1318,10 +1397,11 @@ bool PhysicalMemory::Map(uint64_t vaddr, uint64_t phys_addr, size_t len, int pro
 		}
 		--block;
 		const auto block_end = block->second.start_addr + block->second.size;
-		if (block->second.pool_expansion || current < block->second.start_addr ||
+		if (block->second.kind == AllocationKind::Pooled || current < block->second.start_addr ||
 		    current >= block_end) {
 			return false;
 		}
+		has_automatic |= block->second.kind == AllocationKind::Automatic;
 		current = std::min<uint64_t>(phys_addr + len, block_end);
 		if (current < phys_addr + len) {
 			const auto following = std::next(block);
@@ -1342,6 +1422,9 @@ bool PhysicalMemory::Map(uint64_t vaddr, uint64_t phys_addr, size_t len, int pro
 	mapping.mode           = mode;
 	mapping.gpu_mode       = gpu_mode;
 	m_mappings.push_back(mapping);
+	if (has_automatic) {
+		RemoveFreeRange(m_automatic_free, phys_addr, len);
+	}
 
 	return true;
 }
@@ -1362,7 +1445,7 @@ bool PhysicalMemory::CanMapDirect(uint64_t phys_addr, size_t len) {
 		}
 		--block;
 		const auto block_end = block->second.start_addr + block->second.size;
-		if (block->second.pool_expansion || current < block->second.start_addr ||
+		if (block->second.kind == AllocationKind::Pooled || current < block->second.start_addr ||
 		    current >= block_end) {
 			return false;
 		}
@@ -1380,11 +1463,12 @@ bool PhysicalMemory::CanMapDirect(uint64_t phys_addr, size_t len) {
 bool PhysicalMemory::ReleasePoolExpansion(uint64_t phys_addr, size_t len) {
 	Common::LockGuard lock(m_mutex);
 	const auto        it = m_physical.find(phys_addr);
-	if (it == m_physical.end() || !it->second.pool_expansion || it->second.size != len) {
+	if (it == m_physical.end() || it->second.kind != AllocationKind::Pooled ||
+	    it->second.size != len) {
 		return false;
 	}
 	m_physical.erase(it);
-	AddFreeRange(phys_addr, len);
+	AddFreeRange(m_free, phys_addr, len);
 	return true;
 }
 
@@ -1407,7 +1491,7 @@ bool PhysicalMemory::GetAllocatedSpan(uint64_t phys_addr, size_t len,
 		}
 		--block;
 		const auto block_end = block->second.start_addr + block->second.size;
-		if (block->second.pool_expansion || current < block->second.start_addr ||
+		if (block->second.kind == AllocationKind::Pooled || current < block->second.start_addr ||
 		    current >= block_end) {
 			blocks->clear();
 			return false;
@@ -1463,49 +1547,39 @@ bool PhysicalMemory::Unmap(uint64_t vaddr, uint64_t size, GpuAccessMode* gpu_mod
 		}
 	};
 
-	size_t index = 0;
-	for (auto& b: m_mappings) {
-		if (b.map_vaddr == vaddr && b.map_size == size) {
-			*gpu_mode             = b.gpu_mode;
-			const auto host_vaddr = b.host_vaddr;
-			const auto host_size  = b.host_size;
-
+	for (size_t index = 0; index < m_mappings.size(); ++index) {
+		auto& b = m_mappings[index];
+		if (vaddr < b.map_vaddr || vaddr >= b.map_vaddr + b.map_size ||
+		    size > b.map_vaddr + b.map_size - vaddr) {
+			continue;
+		}
+		*gpu_mode             = b.gpu_mode;
+		const auto phys_addr  = b.start_addr + vaddr - b.map_vaddr;
+		const auto host_vaddr = b.host_vaddr;
+		const auto host_size  = b.host_size;
+		if (vaddr == b.map_vaddr && size == b.map_size) {
 			m_mappings.erase(m_mappings.begin() + static_cast<std::ptrdiff_t>(index));
 			set_host_release_if_last(host_vaddr, host_size);
-
-			return true;
-		}
-		if (vaddr > b.map_vaddr && vaddr + size < b.map_vaddr + b.map_size) {
-			*gpu_mode = b.gpu_mode;
-
+		} else if (vaddr > b.map_vaddr && vaddr + size < b.map_vaddr + b.map_size) {
 			AllocatedBlock right = b;
 			right.start_addr += (vaddr + size) - b.map_vaddr;
 			right.size      = b.map_vaddr + b.map_size - (vaddr + size);
 			right.map_size  = right.size;
 			right.map_vaddr = vaddr + size;
-
 			b.size     = vaddr - b.map_vaddr;
 			b.map_size = b.size;
 			m_mappings.push_back(right);
-			return true;
-		}
-		if (vaddr == b.map_vaddr && size < b.map_size) {
-			*gpu_mode = b.gpu_mode;
-
+		} else if (vaddr == b.map_vaddr) {
 			b.start_addr += size;
 			b.size -= size;
 			b.map_vaddr += size;
 			b.map_size -= size;
-			return true;
-		}
-		if (vaddr > b.map_vaddr && vaddr + size == b.map_vaddr + b.map_size) {
-			*gpu_mode = b.gpu_mode;
-
+		} else {
 			b.size     = vaddr - b.map_vaddr;
 			b.map_size = b.size;
-			return true;
 		}
-		index++;
+		ReclaimAutomatic(phys_addr, size);
+		return true;
 	}
 
 	return false;
@@ -2726,10 +2800,8 @@ int KYTY_SYSV_ABI KernelDirectMemoryQuery(int64_t offset, int flags, void* info,
 	return OK;
 }
 
-int KYTY_SYSV_ABI KernelAllocateDirectMemory(int64_t search_start, int64_t search_end, size_t len,
-                                             size_t alignment, int memory_type,
-                                             int64_t* phys_addr_out) {
-	PRINT_NAME();
+int AllocateDirectMemory(int64_t search_start, int64_t search_end, size_t len, size_t alignment,
+                         int memory_type, int64_t* phys_addr_out, bool automatic) {
 
 	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
 
@@ -2748,7 +2820,10 @@ int KYTY_SYSV_ABI KernelAllocateDirectMemory(int64_t search_start, int64_t searc
 	}
 
 	uint64_t addr = 0;
-	if (!g_physical_memory->Alloc(search_start, search_end, len, alignment, &addr, memory_type)) {
+	const auto kind = automatic ? PhysicalMemory::AllocationKind::Automatic
+	                            : PhysicalMemory::AllocationKind::Direct;
+	if (!g_physical_memory->Alloc(search_start, search_end, len, alignment, &addr, memory_type,
+	                              kind)) {
 		LOGF_COLOR(Log::Color::Red, "\t[Fail]\n");
 		return KERNEL_ERROR_EAGAIN;
 	}
@@ -2760,6 +2835,14 @@ int KYTY_SYSV_ABI KernelAllocateDirectMemory(int64_t search_start, int64_t searc
 	return OK;
 }
 
+int KYTY_SYSV_ABI KernelAllocateDirectMemory(int64_t search_start, int64_t search_end, size_t len,
+                                             size_t alignment, int memory_type,
+                                             int64_t* phys_addr_out) {
+	PRINT_NAME();
+	return AllocateDirectMemory(search_start, search_end, len, alignment, memory_type,
+	                            phys_addr_out);
+}
+
 int KYTY_SYSV_ABI KernelAllocateMainDirectMemory(size_t len, size_t alignment, int memory_type,
                                                  int64_t* phys_addr_out) {
 	PRINT_NAME();
@@ -3082,6 +3165,36 @@ int KYTY_SYSV_ABI KernelMapDirectMemory2(void** addr, size_t len, int type, int
 	return ret;
 }
 
+int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot) {
+	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
+	constexpr uint64_t                    page_size = 0x4000;
+	if (size == 0 || (vaddr & (page_size - 1)) != 0 || (size & (page_size - 1)) != 0 ||
+	    vaddr < kExtendedMemoryBase || vaddr >= kExtendedMemoryBase + kExtendedMemorySize ||
+	    size > kExtendedMemoryBase + kExtendedMemorySize - vaddr) {
+		return KERNEL_ERROR_EINVAL;
+	}
+
+	PhysicalMemory::PhysicalRanges ranges;
+	if (!g_physical_memory->ReserveAutomatic(size, &ranges)) {
+		return KERNEL_ERROR_EAGAIN;
+	}
+	uint64_t mapped = 0;
+	for (const auto& [offset, length]: ranges) {
+		void*     address = reinterpret_cast<void*>(vaddr + mapped);
+		const int result  = KernelMapDirectMemory2(&address, length, type, prot, 0x10,
+		                                           static_cast<int64_t>(offset), page_size);
+		if (result != OK) {
+			if (mapped != 0) {
+				EXIT_IF(KernelMunmap(vaddr, mapped) != OK);
+			}
+			g_physical_memory->RestoreAutomatic(ranges);
+			return result;
+		}
+		mapped += length;
+	}
+	return OK;
+}
+
 int KYTY_SYSV_ABI KernelMapNamedDirectMemory(void** addr, size_t len, int prot, int flags,
                                              int64_t direct_memory_start, size_t alignment,
                                              const char* name) {
@@ -3892,7 +4005,7 @@ int KYTY_SYSV_ABI KernelMemoryPoolExpand(int64_t search_start, int64_t search_en
 	uint64_t   phys_addr           = 0;
 	if (!g_physical_memory->Alloc(static_cast<uint64_t>(search_start),
 	                              static_cast<uint64_t>(search_end), len, effective_alignment,
-	                              &phys_addr, 0, true)) {
+	                              &phys_addr, 0, PhysicalMemory::AllocationKind::Pooled)) {
 		return KERNEL_ERROR_ENOMEM;
 	}
 

src/kernel/memory.h

@@ -24,6 +24,8 @@ struct Lifecycle {
 using callback_func_t = void (*)(uintptr_t addr, size_t size);
 
 constexpr uint32_t KERNEL_MAXIMUM_NAME_LENGTH = 32;
+constexpr uint64_t kExtendedMemoryBase       = 0x080000000000ull;
+constexpr uint64_t kExtendedMemorySize       = 512ull * 1024 * 1024 * 1024;
 
 struct VirtualQueryInfo {
 	uintptr_t start;
@@ -107,6 +109,9 @@ static_assert(sizeof(KernelMemoryPoolBlockStats) == 16,
 
 void                   RegisterCallbacks(callback_func_t alloc_func, callback_func_t free_func);
 void                   SetFlexibleMemorySize(uint64_t size);
+int AllocateDirectMemory(int64_t search_start, int64_t search_end, size_t size, size_t alignment,
+                         int memory_type, int64_t* phys_addr_out, bool automatic = false);
+int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot);
 bool                   TryWriteBacking(uint64_t vaddr, const void* data, uint64_t size);
 bool                   TryReadBacking(uint64_t vaddr, void* data, uint64_t size);
 bool                   TryReadGpuCleanBacking(uint64_t vaddr, void* data, uint64_t size);

src/kernel/memoryAddressSpace.inc

@@ -3,7 +3,6 @@
 
 static constexpr uint64_t HOST_SYSTEM_MANAGED_MIN  = 0x0000040000ull;
 static constexpr uint64_t HOST_SYSTEM_MANAGED_MAX  = 0x07ffffbfffull;
-static constexpr uint64_t HOST_SYSTEM_RESERVED_MIN = HOST_SYSTEM_MANAGED_MAX + 1u;
 #if defined(__APPLE__) && (defined(__x86_64__) || defined(_M_X64))
 static constexpr uint64_t HOST_SYSTEM_RESERVED_MAX = 0x0fbfffffffull;
 static constexpr uint64_t HOST_USER_MIN            = 0x7000000000ull;
@@ -1077,51 +1076,51 @@ private:
 #endif
 
 	void ReserveGuestRegions() {
+		const std::array<std::pair<uint64_t, uint64_t>, 3> regions {{
+		    {HOST_SYSTEM_MANAGED_MIN, HOST_SYSTEM_RESERVED_MAX + 1u},
+		    {HOST_USER_MIN, HOST_USER_MAX + 1u},
+		    {kExtendedMemoryBase, kExtendedMemoryBase + kExtendedMemorySize},
+		}};
+
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		constexpr uint64_t MinimumRegionSize = 0x1000000;
 		const auto         process           = GetCurrentProcess();
-		const auto         guest_end         = HOST_USER_MAX + 1u;
-		auto               next_addr         = HOST_SYSTEM_MANAGED_MIN;
-
-		while (next_addr < guest_end) {
-			MEMORY_BASIC_INFORMATION info {};
-			if (VirtualQuery(reinterpret_cast<const void*>(next_addr), &info, sizeof(info)) == 0) {
-				EXIT("failed to inspect guest address space at 0x%016" PRIx64 "\n", next_addr);
-			}
-
-			const auto region_base = reinterpret_cast<uint64_t>(info.BaseAddress);
-			const auto region_end =
-			    std::min<uint64_t>(guest_end, region_base + static_cast<uint64_t>(info.RegionSize));
-			const auto reserve_start =
-			    AlignUp(std::max<uint64_t>(region_base, HOST_SYSTEM_MANAGED_MIN), m_granularity);
-			const auto reserve_end = AlignDown(region_end, m_granularity);
-
-			if (info.State == MEM_FREE && reserve_end > reserve_start &&
-			    reserve_end - reserve_start > MinimumRegionSize) {
-				auto* ptr = VirtualAlloc2(
-				    process, reinterpret_cast<void*>(reserve_start), reserve_end - reserve_start,
-				    MEM_RESERVE | MEM_RESERVE_PLACEHOLDER, PAGE_NOACCESS, nullptr, 0);
-				if (reinterpret_cast<uint64_t>(ptr) != reserve_start) {
-					EXIT("failed to reserve guest address space at 0x%016" PRIx64
-					     ", size 0x%016" PRIx64 "\n",
-					     reserve_start, reserve_end - reserve_start);
+		for (const auto& [guest_start, guest_end]: regions) {
+			auto next_addr = guest_start;
+			while (next_addr < guest_end) {
+				MEMORY_BASIC_INFORMATION info {};
+				if (VirtualQuery(reinterpret_cast<const void*>(next_addr), &info, sizeof(info)) == 0) {
+					EXIT("failed to inspect guest address space at 0x%016" PRIx64 "\n", next_addr);
 				}
-				AddFreeUnlocked(reserve_start, reserve_end - reserve_start);
-				m_owned.emplace_back(reserve_start, reserve_end - reserve_start);
-			}
 
-			if (region_end <= next_addr) {
-				EXIT("guest address-space scan did not advance at 0x%016" PRIx64 "\n", next_addr);
+				const auto region_base = reinterpret_cast<uint64_t>(info.BaseAddress);
+				const auto region_end = std::min<uint64_t>(
+				    guest_end, region_base + static_cast<uint64_t>(info.RegionSize));
+				const auto reserve_start =
+				    AlignUp(std::max(region_base, guest_start), m_granularity);
+				const auto reserve_end = AlignDown(region_end, m_granularity);
+
+				if (info.State == MEM_FREE && reserve_end > reserve_start &&
+				    reserve_end - reserve_start > MinimumRegionSize) {
+					auto* ptr = VirtualAlloc2(
+					    process, reinterpret_cast<void*>(reserve_start), reserve_end - reserve_start,
+					    MEM_RESERVE | MEM_RESERVE_PLACEHOLDER, PAGE_NOACCESS, nullptr, 0);
+					if (reinterpret_cast<uint64_t>(ptr) != reserve_start) {
+						EXIT("failed to reserve guest address space at 0x%016" PRIx64
+						     ", size 0x%016" PRIx64 "\n",
+						     reserve_start, reserve_end - reserve_start);
+					}
+					AddFreeUnlocked(reserve_start, reserve_end - reserve_start);
+					m_owned.emplace_back(reserve_start, reserve_end - reserve_start);
+				}
+
+				if (region_end <= next_addr) {
+					EXIT("guest address-space scan did not advance at 0x%016" PRIx64 "\n", next_addr);
+				}
+				next_addr = region_end;
 			}
-			next_addr = region_end;
 		}
 #else
-		const std::array<std::pair<uint64_t, uint64_t>, 3> regions {{
-		    {HOST_SYSTEM_MANAGED_MIN, HOST_SYSTEM_MANAGED_MAX + 1u},
-		    {HOST_SYSTEM_RESERVED_MIN, HOST_SYSTEM_RESERVED_MAX + 1u},
-		    {HOST_USER_MIN, HOST_USER_MAX + 1u},
-		}};
-
 #if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
 		std::vector<std::pair<uint64_t, uint64_t>> occupied;
 		FILE*                                      maps = fopen("/proc/self/maps", "r");

src/kernel/pthread.cpp

@@ -512,15 +512,6 @@ static void SchedulerBackoffOnce() {
 #endif
 }
 
-static bool SleepMicroSchedulerBackoff(uint64_t microseconds) {
-	if (microseconds > 1) {
-		return false;
-	}
-
-	SchedulerBackoffOnce();
-	return true;
-}
-
 static void SleepMicroWithSignalPoll(uint64_t microseconds) {
 	if (microseconds == 0) {
 		KernelDispatchPendingSignalForCurrentThread();
@@ -529,9 +520,7 @@ static void SleepMicroWithSignalPoll(uint64_t microseconds) {
 
 	while (microseconds > 0) {
 		const auto step = std::min<uint64_t>(microseconds, SIGNAL_APC_POLL_MICROS);
-		if (!SleepMicroSchedulerBackoff(step)) {
-			Common::Thread::SleepMicro(step);
-		}
+		Common::Thread::SleepMicro(step);
 		microseconds -= step;
 		KernelDispatchPendingSignalForCurrentThread();
 	}
@@ -3842,28 +3831,16 @@ void KYTY_SYSV_ABI KernelSetThreadDtors(thread_dtors_func_t dtors) {
 }
 
 int KYTY_SYSV_ABI KernelUsleep(KernelUseconds microseconds) {
-	Common::Timer t;
-	t.Start();
 	SleepMicroWithSignalPoll(microseconds);
-	// double ts = t.GetTimeS();
-	// LOGF("\tactual: %g microseconds\n", ts * 1000000.0);
 	return OK;
 }
 
 unsigned int KYTY_SYSV_ABI KernelSleep(unsigned int seconds) {
-	PRINT_NAME();
-	LOGF("\tsleep: %u\n", seconds);
-	Common::Timer t;
-	t.Start();
 	SleepMicroWithSignalPoll(static_cast<uint64_t>(seconds) * 1000000ull);
-	double ts = t.GetTimeS();
-	LOGF("\tactual: %g seconds\n", ts);
 	return OK;
 }
 
 int KYTY_SYSV_ABI KernelNanosleep(const KernelTimespec* rqtp, KernelTimespec* rmtp) {
-	PRINT_NAME();
-
 	if (rqtp == nullptr) {
 		return KERNEL_ERROR_EFAULT;
 	}
@@ -3880,13 +3857,7 @@ int KYTY_SYSV_ABI KernelNanosleep(const KernelTimespec* rqtp, KernelTimespec* rm
 	uint64_t nanos =
 	    static_cast<uint64_t>(rqtp->tv_sec) * 1000000000ull + static_cast<uint64_t>(rqtp->tv_nsec);
 
-	LOGF("\tnanosleep: %" PRIu64 "\n", nanos);
-
-	Common::Timer t;
-	t.Start();
 	SleepNanoWithSignalPoll(nanos);
-	double ts = t.GetTimeS();
-	LOGF("\tactual: %g nanoseconds\n", ts * 1000000000.0);
 
 	if (rmtp != nullptr) {
 		rmtp->tv_sec  = 0;

src/launcher/CMakeLists.txt

@@ -20,7 +20,7 @@ add_executable(launcher WIN32
 )
 
 
-target_link_libraries(launcher Qt6::Concurrent Qt6::Network Qt6::Widgets Vulkan::Headers SDL3::SDL3-static kyty_git_version)
+target_link_libraries(launcher Qt6::Concurrent Qt6::Network Qt6::Widgets Vulkan::Headers SDL3::SDL3-static kyty_git_version common)
 if(WIN32)
 target_link_libraries(launcher setupapi)
 endif()

src/launcher/include/configurationItem.h

@@ -1,6 +1,7 @@
 #ifndef CONFIGURATION_ITEM_H
 #define CONFIGURATION_ITEM_H
 
+#include <QIcon>
 #include <QObject>
 #include <QTreeWidgetItem>
 
@@ -18,7 +19,7 @@ public:
 	explicit ConfigurationItem(std::unique_ptr<Configuration> info, QTreeWidget* parent);
 	~ConfigurationItem() override;
 
-	void Update();
+	void Update(bool reload_icon = false);
 	bool operator<(const QTreeWidgetItem& other) const override;
 
 	Configuration&                     GetInfo() { return *m_info; }
@@ -35,6 +36,7 @@ private:
 	void UpdateStatusIndicator();
 
 	std::unique_ptr<Configuration> m_info;
+	QIcon                          m_icon;
 	bool                           m_running          = false;
 	QComboBox*                     m_status_combo     = nullptr;
 	QLabel*                        m_status_indicator = nullptr;

src/launcher/include/gameContent.h

@@ -0,0 +1,28 @@
+#ifndef GAME_CONTENT_H
+#define GAME_CONTENT_H
+
+#include <QByteArray>
+#include <QString>
+#include <QStringList>
+
+#include <cstdint>
+#include <filesystem>
+
+namespace GameContent {
+
+inline constexpr uint64_t MaxMetadataSize      = uint64_t {1} << 20u;
+inline constexpr uint64_t MaxImageSize         = uint64_t {32} << 20u;
+inline constexpr uint64_t MaxTrophyPackageSize = uint64_t {128} << 20u;
+
+[[nodiscard]] std::filesystem::path ToPath(const QString& path);
+[[nodiscard]] QString               FromPath(const std::filesystem::path& path);
+[[nodiscard]] bool                  IsArchive(const QString& base);
+[[nodiscard]] std::filesystem::path Resolve(const QString& base, const QString& relative);
+[[nodiscard]] bool                  FileExists(const QString& base, const QString& relative);
+[[nodiscard]] QByteArray            ReadPath(const std::filesystem::path& path, uint64_t max_size);
+[[nodiscard]] QByteArray  ReadFile(const QString& base, const QString& relative, uint64_t max_size);
+[[nodiscard]] QStringList ListFiles(const QString& base, const QString& relative);
+
+} // namespace GameContent
+
+#endif // GAME_CONTENT_H

src/launcher/include/gameListTreeWidget.h

@@ -1,6 +1,8 @@
 #ifndef GAME_LIST_TREE_WIDGET_H
 #define GAME_LIST_TREE_WIDGET_H
 
+#include "gameContent.h"
+
 #include <QPaintEvent>
 #include <QPainter>
 #include <QPalette>
@@ -17,14 +19,15 @@ public:
 		viewport()->setAutoFillBackground(false);
 	}
 
-	void SetBackgroundImage(const QString& path) {
-		if (m_path == path) {
+	void SetBackgroundImage(const QString& key) {
+		if (m_path == key) {
 			return;
 		}
-		m_path   = path;
+		m_path   = key;
 		m_source = QPixmap();
-		if (!path.isEmpty()) {
-			m_source.load(path);
+		if (!key.isEmpty()) {
+			m_source.loadFromData(GameContent::ReadFile(key, QStringLiteral("sce_sys/pic0.png"),
+			                                            GameContent::MaxImageSize));
 		}
 		UpdateScaledBackground();
 		viewport()->update();

src/launcher/src/configurationItem.cpp

@@ -1,6 +1,7 @@
 #include "configurationItem.h"
 
 #include "configuration.h"
+#include "gameContent.h"
 
 #include <QApplication>
 #include <QComboBox>
@@ -14,6 +15,7 @@
 #include <QLabel>
 #include <QLineEdit>
 #include <QLocale>
+#include <QPixmap>
 #include <QSize>
 #include <QStringList>
 #include <QStyle>
@@ -154,7 +156,7 @@ ConfigurationItem::ConfigurationItem(std::unique_ptr<Configuration> info, QTreeW
 	m_comment_edit->setFrame(false);
 	parent->setItemWidget(this, CommentsColumn, m_comment_edit);
 
-	Update();
+	Update(true);
 	SetRunning(false);
 
 	setData(SizeColumn, Qt::UserRole, qint64(-1));
@@ -170,7 +172,14 @@ ConfigurationItem::ConfigurationItem(std::unique_ptr<Configuration> info, QTreeW
 		watcher->deleteLater();
 	});
 	watcher->setFuture(QtConcurrent::run([path = m_info->basedir]() -> qint64 {
-		if (path.isEmpty() || !QDir(path).exists()) {
+		if (path.isEmpty()) {
+			return -1;
+		}
+		const QFileInfo info(path);
+		if (GameContent::IsArchive(path)) {
+			return info.size();
+		}
+		if (!QDir(path).exists()) {
 			return -1;
 		}
 		qint64       bytes = 0;
@@ -188,7 +197,7 @@ ConfigurationItem::ConfigurationItem(std::unique_ptr<Configuration> info, QTreeW
 
 ConfigurationItem::~ConfigurationItem() = default;
 
-void ConfigurationItem::Update() {
+void ConfigurationItem::Update(bool reload_icon) {
 	const auto display_text = GetDisplayText(*m_info);
 	const auto path         = GetPathText(*m_info);
 
@@ -209,6 +218,12 @@ void ConfigurationItem::Update() {
 		m_comment_edit->setText(m_info->game_comment);
 	}
 
+	if (reload_icon) {
+		const auto icon_data = GameContent::ReadFile(
+		    m_info->basedir, QStringLiteral("sce_sys/icon0.png"), GameContent::MaxImageSize);
+		QPixmap icon;
+		m_icon = !icon_data.isEmpty() && icon.loadFromData(icon_data) ? QIcon(icon) : QIcon {};
+	}
 	UpdateIcon();
 	UpdateStatusIndicator();
 }
@@ -267,9 +282,8 @@ void ConfigurationItem::SetCompatibilityEditable(bool editable) {
 }
 
 void ConfigurationItem::UpdateIcon() {
-	const QString icon_file = QDir(m_info->basedir).filePath(QStringLiteral("sce_sys/icon0.png"));
-	if (QFileInfo::exists(icon_file)) {
-		setIcon(NameColumn, QIcon(icon_file));
+	if (!m_icon.isNull()) {
+		setIcon(NameColumn, m_icon);
 		return;
 	}
 

src/launcher/src/configurationListWidget.cpp

@@ -1,9 +1,11 @@
 #include "configurationListWidget.h"
 
+#include "common/archive.h"
 #include "compatibilityDatabase.h"
 #include "configuration.h"
 #include "configurationEditDialog.h"
 #include "configurationItem.h"
+#include "gameContent.h"
 #include "gameListTreeWidget.h"
 #include "inputMappingDialog.h"
 #include "mainDialog.h"
@@ -127,15 +129,6 @@ static QStringList SettingsStringList(const QVariant& value) {
 	return text.isEmpty() ? QStringList() : QStringList({text});
 }
 
-static QString GetPic0Path(const Configuration& info) {
-	if (info.basedir.isEmpty()) {
-		return {};
-	}
-
-	const auto path = QDir(info.basedir).filePath(QStringLiteral("sce_sys/pic0.png"));
-	return QFileInfo::exists(path) ? QFileInfo(path).absoluteFilePath() : QString();
-}
-
 static void ConfigureGameList(Ui::ConfigurationListWidget* ui) {
 	ui->cfgs_list->setAlternatingRowColors(false);
 	ui->cfgs_list->setAllColumnsShowFocus(true);
@@ -490,17 +483,16 @@ static QString GetFirmwareVersion(const QJsonObject& root) {
 	return version;
 }
 
-static GameMetadata GetGameMetadata(const QString& param_file, const QString& fallback) {
+static GameMetadata GetGameMetadata(const QByteArray& param_data, const QString& fallback) {
 	GameMetadata ret;
 	ret.title_name = fallback;
 
-	QFile param(param_file);
-	if (!param.open(QIODevice::ReadOnly)) {
+	if (param_data.isEmpty()) {
 		return ret;
 	}
 
 	QJsonParseError parse_error;
-	const auto      doc = QJsonDocument::fromJson(param.readAll(), &parse_error);
+	const auto      doc = QJsonDocument::fromJson(param_data, &parse_error);
 	if (parse_error.error != QJsonParseError::NoError || !doc.isObject()) {
 		return ret;
 	}
@@ -522,18 +514,18 @@ static GameMetadata GetGameMetadata(const QString& param_file, const QString& fa
 }
 
 static void SetGameFiles(Configuration& info, const QString& game_dir, const QString& game_path,
-                         const GameMetadata& metadata) {
-	QDir game(game_dir);
+                         const GameMetadata& metadata, bool archive) {
+	const QFileInfo game(game_dir);
 
 	info.game_path   = game_path;
-	info.basedir     = game.absolutePath();
+	info.basedir     = archive ? game.absoluteFilePath() : QDir(game_dir).absolutePath();
 	info.name        = metadata.title_name;
 	info.title_id    = metadata.title_id;
 	info.gameVersion = metadata.gameVersion;
 	info.firmwareVer = metadata.firmwareVer;
 
 	if (info.name.isEmpty()) {
-		info.name = game.dirName();
+		info.name = archive ? game.completeBaseName() : QDir(game_dir).dirName();
 	}
 }
 
@@ -609,6 +601,94 @@ void ConfigurationListWidget::ScanGameDirectory() {
 
 	const QString eboot_name = QStringLiteral("eboot.bin");
 	QSet<QString> found_games;
+	const auto add_game = [this, &found_games, &eboot_name,
+	                       &running_items](const QString& base, const QString& game_path,
+	                                       const QString& legacy_game_path, const QString& fallback,
+	                                       bool archive) {
+		const QString game_key = PathKey(game_path);
+		if (game_key.isEmpty() || found_games.contains(game_key)) {
+			return;
+		}
+		// Keep the index and decompression cache alive through validation, metadata and icon
+		// reads, after rejecting duplicate candidates from overlapping game folders.
+		const auto reader = archive ? Common::OpenArchive(GameContent::ToPath(base)) : nullptr;
+		if (archive && (reader == nullptr || !GameContent::FileExists(base, eboot_name))) {
+			return;
+		}
+		found_games.insert(game_key);
+
+		const auto metadata =
+		    GetGameMetadata(GameContent::ReadFile(base, QStringLiteral("sce_sys/param.json"),
+		                                          GameContent::MaxMetadataSize),
+		                    fallback);
+		auto info = std::make_unique<Configuration>();
+		info->custom_settings =
+		    FindCustomInfo(&m_custom_infos, game_path, legacy_game_path) != nullptr;
+
+		SetGameFiles(*info, base, game_path, metadata, archive);
+		const auto* compatibility = m_compatibility->Find(info->title_id);
+		if (compatibility != nullptr) {
+			info->game_status  = compatibility->status;
+			info->game_comment = compatibility->comment;
+		}
+
+		if (auto* item = running_items.value(game_key); item != nullptr) {
+			// Refresh metadata without losing the running row's identity.
+			const QSignalBlocker status_blocker(item->GetStatusCombo());
+			item->GetInfo().CopyGameInfoFrom(*info);
+			item->Update(true);
+			item->SetCompatibilityEditable(m_compatibility->IsLocal() &&
+			                               !item->GetInfo().title_id.trimmed().isEmpty());
+			return;
+		}
+
+		auto* item = new ConfigurationItem(std::move(info), m_ui->cfgs_list);
+		item->SetCompatibilityEditable(m_compatibility->IsLocal() &&
+		                               !item->GetInfo().title_id.trimmed().isEmpty());
+		connect(item->GetStatusCombo(), &QComboBox::currentIndexChanged, item,
+		        [this, item](int /*index*/) {
+			        if (!m_compatibility->IsLocal()) {
+				        return;
+			        }
+
+			        const auto& title_id = item->GetInfo().title_id;
+			        if (title_id.trimmed().isEmpty()) {
+				        return;
+			        }
+			        item->GetInfo().game_status = static_cast<Configuration::GameStatus>(
+			            item->GetStatusCombo()->currentData().toInt());
+			        item->Update();
+			        m_compatibility->SetStatus(title_id, item->GetInfo().game_status);
+			        m_ui->cfgs_list->setCurrentItem(item);
+			        SelectItem(item);
+		        });
+		connect(item->GetCommentEdit(), &QLineEdit::editingFinished, item, [this, item]() {
+			if (!m_compatibility->IsLocal()) {
+				return;
+			}
+
+			const auto& title_id = item->GetInfo().title_id;
+			if (title_id.trimmed().isEmpty()) {
+				return;
+			}
+			item->GetInfo().game_comment = item->GetCommentEdit()->text();
+			item->Update();
+			m_compatibility->SetComment(title_id, item->GetInfo().game_comment);
+			m_ui->cfgs_list->setCurrentItem(item);
+			SelectItem(item);
+		});
+	};
+	const auto scan_archives = [&add_game](const QDir& directory, const QDir& root) {
+		const auto entries = directory.entryInfoList(QDir::Files | QDir::NoSymLinks, QDir::Name);
+		for (const auto& archive: entries) {
+			const auto base = archive.absoluteFilePath();
+			if (!Common::IsSupportedArchive(GameContent::ToPath(base))) {
+				continue;
+			}
+			add_game(base, QDir::cleanPath(base), root.relativeFilePath(base),
+			         archive.completeBaseName(), true);
+		}
+	};
 
 	for (const auto& root_path: m_game_dirs) {
 		QDir root(root_path);
@@ -616,6 +696,7 @@ void ConfigurationListWidget::ScanGameDirectory() {
 			continue;
 		}
 
+		scan_archives(root, root);
 		QList<QDir> pending_dirs;
 		const auto  root_subdirs =
 		    root.entryInfoList(QDir::Dirs | QDir::NoDotAndDotDot | QDir::NoSymLinks);
@@ -625,74 +706,13 @@ void ConfigurationListWidget::ScanGameDirectory() {
 
 		while (!pending_dirs.isEmpty()) {
 			QDir game_dir = pending_dirs.takeFirst();
+			scan_archives(game_dir, root);
 
 			if (game_dir.exists(eboot_name)) {
 				const QString game_path        = NormalizeGameDirectory(game_dir.absolutePath());
 				const QString legacy_game_path = root.relativeFilePath(game_dir.absolutePath());
-				const QString game_key         = PathKey(game_path);
-				if (game_key.isEmpty() || found_games.contains(game_key)) {
-					continue;
-				}
-				found_games.insert(game_key);
-
-				auto metadata = GetGameMetadata(
-				    game_dir.filePath(QStringLiteral("sce_sys/param.json")), game_dir.dirName());
-
-				auto info = std::make_unique<Configuration>();
-				info->custom_settings =
-				    FindCustomInfo(&m_custom_infos, game_path, legacy_game_path) != nullptr;
-				SetGameFiles(*info, game_dir.absolutePath(), game_path, metadata);
-				const auto* compatibility = m_compatibility->Find(info->title_id);
-				if (compatibility != nullptr) {
-					info->game_status  = compatibility->status;
-					info->game_comment = compatibility->comment;
-				}
-
-				if (auto* item = running_items.value(game_key); item != nullptr) {
-					// Refresh metadata without losing the running row's identity.
-					const QSignalBlocker status_blocker(item->GetStatusCombo());
-					item->GetInfo().CopyGameInfoFrom(*info);
-					item->Update();
-					item->SetCompatibilityEditable(m_compatibility->IsLocal() &&
-					                               !item->GetInfo().title_id.trimmed().isEmpty());
-					continue;
-				}
-
-				auto* item = new ConfigurationItem(std::move(info), m_ui->cfgs_list);
-				item->SetCompatibilityEditable(m_compatibility->IsLocal() &&
-				                               !item->GetInfo().title_id.trimmed().isEmpty());
-				connect(item->GetStatusCombo(), &QComboBox::currentIndexChanged, item,
-				        [this, item](int /*index*/) {
-					        if (!m_compatibility->IsLocal()) {
-						        return;
-					        }
-
-					        const auto& title_id = item->GetInfo().title_id;
-					        if (title_id.trimmed().isEmpty()) {
-						        return;
-					        }
-					        item->GetInfo().game_status = static_cast<Configuration::GameStatus>(
-					            item->GetStatusCombo()->currentData().toInt());
-					        item->Update();
-					        m_compatibility->SetStatus(title_id, item->GetInfo().game_status);
-					        m_ui->cfgs_list->setCurrentItem(item);
-					        SelectItem(item);
-				        });
-				connect(item->GetCommentEdit(), &QLineEdit::editingFinished, item, [this, item]() {
-					if (!m_compatibility->IsLocal()) {
-						return;
-					}
-
-					const auto& title_id = item->GetInfo().title_id;
-					if (title_id.trimmed().isEmpty()) {
-						return;
-					}
-					item->GetInfo().game_comment = item->GetCommentEdit()->text();
-					item->Update();
-					m_compatibility->SetComment(title_id, item->GetInfo().game_comment);
-					m_ui->cfgs_list->setCurrentItem(item);
-					SelectItem(item);
-				});
+				add_game(game_dir.absolutePath(), game_path, legacy_game_path, game_dir.dirName(),
+				         false);
 				continue;
 			}
 
@@ -805,7 +825,8 @@ void ConfigurationListWidget::open_game_folder() {
 		return;
 	}
 
-	const QDir game_dir(item->GetInfo().basedir);
+	const auto base = item->GetInfo().basedir;
+	const QDir game_dir(GameContent::IsArchive(base) ? QFileInfo(base).absolutePath() : base);
 	if (!game_dir.exists()) {
 		QMessageBox::warning(this, tr("Open game folder"), tr("Game folder does not exist."));
 		return;
@@ -896,7 +917,7 @@ void ConfigurationListWidget::SelectItem(QTreeWidgetItem* witem) {
 	m_ui->edit_button->setEnabled(!item->IsRunning());
 
 	m_selected_item = item;
-	m_ui->cfgs_list->SetBackgroundImage(GetPic0Path(item->GetInfo()));
+	m_ui->cfgs_list->SetBackgroundImage(item->GetInfo().basedir);
 
 	emit Select();
 }
@@ -959,7 +980,8 @@ void ConfigurationListWidget::show_context_menu(const QPoint& pos) {
 
 	if (item != nullptr) {
 		action_run->setDisabled(item->IsRunning());
-		action_open_folder->setDisabled(!QDir(item->GetInfo().basedir).exists());
+		const auto& base = item->GetInfo().basedir;
+		action_open_folder->setDisabled(!QDir(base).exists() && !GameContent::IsArchive(base));
 		action_view_trophies->setDisabled(!has_trophy_data);
 		action_remove_save_data->setDisabled(item->IsRunning() || save_data_dirs.isEmpty());
 		action_edit->setDisabled(item->IsRunning());

src/launcher/src/gameContent.cpp

@@ -0,0 +1,80 @@
+#include "gameContent.h"
+
+#include "common/archive.h"
+#include "common/file.h"
+#include "common/stringUtils.h"
+
+#include <QFileInfo>
+
+#include <limits>
+
+namespace GameContent {
+
+std::filesystem::path ToPath(const QString& path) {
+#if defined(_WIN32)
+	return std::filesystem::path(path.toStdWString());
+#else
+	return std::filesystem::path(path.toStdString());
+#endif
+}
+
+QString FromPath(const std::filesystem::path& path) {
+#if defined(_WIN32)
+	return QString::fromStdWString(path.wstring());
+#else
+	return QString::fromStdString(Common::PathToString(path));
+#endif
+}
+
+bool IsArchive(const QString& base) {
+	const QFileInfo info(base);
+	return info.isFile() && Common::IsSupportedArchive(ToPath(base));
+}
+
+std::filesystem::path Resolve(const QString& base, const QString& relative) {
+	const auto root = ToPath(base);
+	return IsArchive(base) ? Common::MakeArchivePath(root, ToPath(relative))
+	                       : root / ToPath(relative);
+}
+
+bool FileExists(const QString& base, const QString& relative) {
+	return Common::File::IsFileExisting(Resolve(base, relative));
+}
+
+QByteArray ReadFile(const QString& base, const QString& relative, uint64_t max_size) {
+	return ReadPath(Resolve(base, relative), max_size);
+}
+
+QByteArray ReadPath(const std::filesystem::path& path, uint64_t max_size) {
+	Common::File file(path, Common::File::Mode::Read);
+	if (file.IsInvalid()) {
+		return {};
+	}
+
+	const auto size = file.Size();
+	if (size > max_size || size > static_cast<uint64_t>(std::numeric_limits<int>::max())) {
+		return {};
+	}
+
+	QByteArray data(static_cast<int>(size), Qt::Uninitialized);
+	uint32_t   bytes_read = 0;
+	file.Read(data.data(), static_cast<uint32_t>(size), &bytes_read);
+	return bytes_read == size ? data : QByteArray {};
+}
+
+QStringList ListFiles(const QString& base, const QString& relative) {
+	const auto  directory = Resolve(base, relative);
+	const bool  archive   = Common::IsArchivePath(directory);
+	QStringList result;
+	for (const auto& entry: Common::File::GetDirEntries(directory)) {
+		if (entry.is_file) {
+			const auto path = FromPath(directory / Common::PathFromUtf8(entry.name));
+			if (archive || !QFileInfo(path).isSymLink()) {
+				result.append(path);
+			}
+		}
+	}
+	return result;
+}
+
+} // namespace GameContent

src/launcher/src/mainDialog.cpp

@@ -3,6 +3,7 @@
 #include "configuration.h"
 #include "configurationItem.h"
 #include "configurationListWidget.h"
+#include "gameContent.h"
 #include "patchesDialog.h"
 #include "updateChecker.h"
 
@@ -260,7 +261,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	}
 
 	QString game = info.basedir;
-	if (!info.elf.isEmpty()) {
+	if (!info.elf.isEmpty() && !GameContent::IsArchive(info.basedir)) {
 		game = QDir(info.basedir).filePath(info.elf);
 	}
 	args << "--game" << game;
@@ -507,7 +508,7 @@ void MainDialogPrivate::Update() {
 	if (run_enabled) {
 		const auto& info = item->GetInfo();
 		auto        dir  = info.basedir;
-		run_enabled      = !dir.isEmpty() && QDir(dir).exists();
+		run_enabled      = !dir.isEmpty() && (QDir(dir).exists() || GameContent::IsArchive(dir));
 	}
 
 	m_ui->widget->SetRunEnabled(run_enabled);

src/launcher/src/trophyViewerDialog.cpp

@@ -1,13 +1,14 @@
 #include "trophyViewerDialog.h"
 
+#include "common/archive.h"
 #include "configuration.h"
+#include "gameContent.h"
 
 #include <QAbstractItemView>
 #include <QBrush>
 #include <QByteArray>
 #include <QDialogButtonBox>
 #include <QDir>
-#include <QFile>
 #include <QFileInfo>
 #include <QFont>
 #include <QHeaderView>
@@ -39,12 +40,14 @@
 
 namespace {
 
-constexpr quint32 UCP_MAGIC      = 0xb228c60a;
-constexpr quint32 UCP_VERSION    = 1;
-constexpr int     UCP_HEADER_LEN = 0x40;
-constexpr int     UCP_TOC_SKIP   = 0x20;
-constexpr int     UCP_ENTRY_LEN  = 0x40;
-constexpr int     UCP_NAME_LEN   = 0x20;
+constexpr quint32 UCP_MAGIC              = 0xb228c60a;
+constexpr quint32 UCP_VERSION            = 1;
+constexpr int     UCP_HEADER_LEN         = 0x40;
+constexpr int     UCP_TOC_SKIP           = 0x20;
+constexpr int     UCP_ENTRY_LEN          = 0x40;
+constexpr int     UCP_NAME_LEN           = 0x20;
+constexpr quint32 UCP_MAX_FILES          = 4096;
+constexpr quint64 UCP_MAX_EXTRACTED_SIZE = quint64 {64} << 20u;
 
 struct UcpEntry {
 	QString name;
@@ -98,14 +101,22 @@ static QString ReadFixedString(const QByteArray& data, qsizetype offset, qsizety
 	return QString::fromLatin1(data.constData() + offset, len);
 }
 
+static bool IsUsedUcpEntry(const QString& name) {
+	const auto lower = name.toCaseFolded();
+	return lower == QStringLiteral("tropconf.json") || lower == QStringLiteral("tropmeta.json") ||
+	       (lower.startsWith(QStringLiteral("tropmeta_")) &&
+	        lower.endsWith(QStringLiteral(".json"))) ||
+	       (lower.startsWith(QStringLiteral("trop")) && lower.endsWith(QStringLiteral(".png")));
+}
+
 static bool ReadUcp(const QString& file_name, QMap<QString, QByteArray>& files, QString& error) {
-	QFile file(file_name);
-	if (!file.open(QIODevice::ReadOnly)) {
+	const QByteArray data =
+	    GameContent::ReadPath(GameContent::ToPath(file_name), GameContent::MaxTrophyPackageSize);
+	if (data.isEmpty()) {
 		error = QObject::tr("Could not open %1").arg(QDir::toNativeSeparators(file_name));
 		return false;
 	}
 
-	const QByteArray data = file.readAll();
 	if (data.size() < UCP_HEADER_LEN) {
 		error = QObject::tr("%1 is too small to be a trophy package.")
 		            .arg(QFileInfo(file_name).fileName());
@@ -127,14 +138,18 @@ static bool ReadUcp(const QString& file_name, QMap<QString, QByteArray>& files,
 	}
 
 	const auto declared_size = qFromBigEndian<quint64>(data.constData() + 0x08);
-	if (declared_size > static_cast<quint64>(data.size())) {
+	if (declared_size < UCP_HEADER_LEN || declared_size > static_cast<quint64>(data.size())) {
 		error = QObject::tr("%1 is truncated.").arg(QFileInfo(file_name).fileName());
 		return false;
 	}
 
 	const auto file_count = qFromBigEndian<quint32>(data.constData() + 0x10);
+	if (file_count > UCP_MAX_FILES) {
+		error = QObject::tr("%1 contains too many files.").arg(QFileInfo(file_name).fileName());
+		return false;
+	}
 	const auto toc_offset = static_cast<quint64>(qFromBigEndian<quint32>(data.constData() + 0x14));
-	const auto data_size  = static_cast<quint64>(data.size());
+	const auto data_size  = declared_size;
 
 	const quint64 table_size = UCP_TOC_SKIP + static_cast<quint64>(file_count) * UCP_ENTRY_LEN;
 	if (toc_offset > data_size || table_size > data_size - toc_offset ||
@@ -144,6 +159,7 @@ static bool ReadUcp(const QString& file_name, QMap<QString, QByteArray>& files,
 		return false;
 	}
 
+	quint64 extracted_size = 0;
 	for (quint32 i = 0; i < file_count; i++) {
 		const auto entry_offset = static_cast<qsizetype>(toc_offset + UCP_TOC_SKIP +
 		                                                 static_cast<quint64>(i) * UCP_ENTRY_LEN);
@@ -161,9 +177,22 @@ static bool ReadUcp(const QString& file_name, QMap<QString, QByteArray>& files,
 			            .arg(QFileInfo(file_name).fileName(), entry.name);
 			return false;
 		}
+		if (!IsUsedUcpEntry(entry.name)) {
+			continue;
+		}
+
+		const auto entry_limit = entry.name.endsWith(QStringLiteral(".png"), Qt::CaseInsensitive)
+		                             ? GameContent::MaxImageSize
+		                             : GameContent::MaxMetadataSize;
+		if (entry.size > entry_limit || entry.size > UCP_MAX_EXTRACTED_SIZE - extracted_size) {
+			error = QObject::tr("%1 contains an oversized entry for %2.")
+			            .arg(QFileInfo(file_name).fileName(), entry.name);
+			return false;
+		}
 
 		files.insert(entry.name.toCaseFolded(), data.mid(static_cast<qsizetype>(entry.offset),
 		                                                 static_cast<qsizetype>(entry.size)));
+		extracted_size += entry.size;
 	}
 
 	return true;
@@ -392,30 +421,31 @@ static QStringList FindTrophyFiles(const Configuration* info) {
 		return {};
 	}
 
-	const QDir trophy_dir(QDir(info->basedir).filePath(QStringLiteral("sce_sys/trophy2")));
-	if (!trophy_dir.exists()) {
-		return {};
+	QStringList files;
+	for (const auto& file:
+	     GameContent::ListFiles(info->basedir, QStringLiteral("sce_sys/trophy2"))) {
+		const auto name = QFileInfo(file).fileName();
+		if (name.startsWith(QStringLiteral("trophy"), Qt::CaseInsensitive) &&
+		    name.endsWith(QStringLiteral(".ucp"), Qt::CaseInsensitive)) {
+			files.append(file);
+		}
 	}
-
-	const auto files =
-	    trophy_dir.entryInfoList({QStringLiteral("Trophy*.ucp"), QStringLiteral("trophy*.ucp")},
-	                             QDir::Files | QDir::NoSymLinks, QDir::Name | QDir::IgnoreCase);
+	files.sort(Qt::CaseInsensitive);
 
 	QStringList   trophy_files;
 	QSet<QString> seen;
 	for (const auto& file: files) {
-		auto key = file.canonicalFilePath();
+		auto key = QFileInfo(file).canonicalFilePath();
 		if (key.isEmpty()) {
-			key = file.absoluteFilePath();
+			key = file;
 		}
 		key = QDir::cleanPath(key).toCaseFolded();
 
 		if (!seen.contains(key)) {
 			seen.insert(key);
-			trophy_files.append(file.absoluteFilePath());
+			trophy_files.append(file);
 		}
 	}
-
 	return trophy_files;
 }
 
@@ -506,6 +536,7 @@ void TrophyViewerDialog::ShowForGame(const Configuration* info, QWidget* parent)
 }
 
 bool TrophyViewerDialog::LoadGame(const Configuration& info, QString& error) {
+	const auto reader = Common::OpenArchive(GameContent::ToPath(info.basedir));
 	const auto trophy_files = FindTrophyFiles(&info);
 	if (trophy_files.isEmpty()) {
 		error = tr("No trophy package found in sce_sys/trophy2.");

src/libs/ajm.cpp

@@ -161,10 +161,8 @@ int KYTY_SYSV_ABI AjmDecAt9ParseConfigData(const void*              config_data,
 		return AJM_ERROR_INVALID_PARAMETER;
 	}
 
-	uint8_t config[ATRAC9_CONFIG_DATA_SIZE] {};
-	std::memcpy(config, config_data, sizeof(config));
 	Atrac9CodecInfo codec_info {};
-	const int       init_result = Atrac9InitDecoder(handle, config);
+	const int       init_result = AjmAt9InitDecoder(handle, static_cast<const uint8_t*>(config_data));
 	const int       info_result =
 	    init_result == 0 ? Atrac9GetCodecInfo(handle, &codec_info) : init_result;
 	Atrac9ReleaseHandle(handle);

src/libs/ajm/atrac9_decoder.h

@@ -18,6 +18,19 @@ extern "C" {
 
 namespace Libs::Audio::Ajm {
 
+inline int AjmAt9InitDecoder(void* handle, const uint8_t* config_data) {
+	// Keep the guest configuration unchanged for metadata and decoder resets.
+	uint8_t config[ATRAC9_CONFIG_DATA_SIZE];
+	std::memcpy(config, config_data, sizeof(config));
+	const auto channel_config = (config[1] >> 1u) & 7u;
+	if (channel_config >= 6u) {
+		// LibAtrac9 lacks vibration layouts; their blocks match ordinary mono/dual-mono.
+		// Translate only the channel field, preserving the other configuration bits.
+		config[1] -= 6u << 1u;
+	}
+	return Atrac9InitDecoder(handle, config);
+}
+
 struct AjmDecAt9InitializeParameters {
 	uint8_t  config_data[ATRAC9_CONFIG_DATA_SIZE];
 	uint32_t reserved;
@@ -278,7 +291,7 @@ private:
 		std::memcpy(m_config_data, config_data, ATRAC9_CONFIG_DATA_SIZE);
 		m_has_config = true;
 
-		const int init_ret = Atrac9InitDecoder(m_handle, m_config_data);
+		const int init_ret = AjmAt9InitDecoder(m_handle, m_config_data);
 		if (init_ret != 0) {
 			m_is_initialized        = false;
 			result->result          = AJM_RESULT_CODEC_ERROR | AJM_RESULT_INVALID_DATA;

src/libs/audio.cpp

@@ -32,6 +32,7 @@ constexpr int AUDIO_OUT_PORT_TYPE_AUDIO3D   = 126;
 constexpr int AUDIO_OUT_PORT_TYPE_AUX       = 127;
 
 constexpr uint32_t AUDIO_OUT_PARAM_FORMAT_MASK = 0x000000ffu;
+constexpr uint64_t AUDIO_OUT_TARGET_LATENCY_US = 40000;
 
 constexpr int      AUDIO_IN_SILENT_STATE_DEVICE_NONE = 0x1;
 constexpr uint32_t AUDIO_IN_GRAIN_MAX_ASYNC          = 384;
@@ -368,11 +369,11 @@ bool Audio::QueueSdlAudio(PortOut* port, const void* data, bool blocking) {
 
 	uint32_t min_queued_size = 0;
 	if (blocking) {
-		constexpr uint64_t target_latency_us = 40000;
 		const auto buffer_us = port->freq != 0 ? (1000000ULL * port->samples_num) / port->freq : 0;
 		const auto buffers =
-		    buffer_us != 0 ? static_cast<uint32_t>((target_latency_us + buffer_us - 1) / buffer_us)
-		                   : 2u;
+		    buffer_us != 0
+		        ? static_cast<uint32_t>((AUDIO_OUT_TARGET_LATENCY_US + buffer_us - 1) / buffer_us)
+		        : 2u;
 		min_queued_size           = prepared_size * std::clamp(buffers, 2u, 16u);
 		const auto wait_start      = LibKernel::KernelGetProcessTime();
 		auto queued                = SDL_GetAudioStreamQueued(port->stream);
@@ -440,11 +441,14 @@ Audio::Id Audio::AudioOutOpen(int type, uint32_t samples_num, uint32_t freq, For
 				port.volume[i] = 32768;
 			}
 
-			if (type == AUDIO_OUT_PORT_TYPE_VIBRATION) {
-				port.haptics = Controller::DualSenseHaptics::Open(freq);
-			} else {
+			// Pad speaker ports keep the main output too; it plays them whenever no DualSense does.
+			if (type != AUDIO_OUT_PORT_TYPE_VIBRATION) {
 				OpenSdlDevice(&port);
 			}
+			if (type == AUDIO_OUT_PORT_TYPE_VIBRATION || type == AUDIO_OUT_PORT_TYPE_PADSPK) {
+				port.haptics =
+				    Controller::DualSenseHaptics::Open(freq, type == AUDIO_OUT_PORT_TYPE_PADSPK);
+			}
 
 			return Id::Create(id);
 		}
@@ -554,14 +558,22 @@ uint32_t Audio::AudioOutOutputs(OutputParam* params, uint32_t num, bool blocking
 	for (uint32_t i = 0; i < num; i++) {
 		auto& port = m_out_ports[params[i].handle.GetId()];
 
-		if (port.type == AUDIO_OUT_PORT_TYPE_VIBRATION) {
-			// Haptics never pace output; keep the stream alive against close and volume changes.
+		uint64_t controller_queued_us = 0;
+		if (port.haptics != nullptr) {
+			// Keep the stream alive against close and volume changes.
 			Common::LockGuard lock(m_mutex);
-			Controller::DualSenseHaptics::Queue(
+			controller_queued_us = Controller::DualSenseHaptics::Queue(
 			    port.haptics, Controller::GetActiveControllerId(), params[i].data, port.samples_num,
 			    static_cast<uint32_t>(port.channels_num), FormatIsFloat(port.format), port.volume);
-		} else {
+		}
+		if (controller_queued_us == 0) {
+			// No DualSense took it (e.g. it was unplugged); a pad speaker port plays on the main
+			// output instead, and a vibration port has none.
 			QueueSdlAudio(&port, params[i].data, blocking);
+		} else if (blocking && port.type == AUDIO_OUT_PORT_TYPE_PADSPK &&
+		           controller_queued_us > AUDIO_OUT_TARGET_LATENCY_US) {
+			// Vibration never paces output; the speaker is audible, so pace it like other ports.
+			Common::Thread::SleepMicro(controller_queued_us - AUDIO_OUT_TARGET_LATENCY_US);
 		}
 	}
 

src/libs/audio.h

@@ -70,6 +70,9 @@ int KYTY_SYSV_ABI AudioOut2PortGetState(AudioOut2PortHandle port, AudioOut2PortS
 int KYTY_SYSV_ABI AudioOut2GetSystemState(AudioOut2SystemState* state);
 int KYTY_SYSV_ABI AudioOut2UserCreate(uint32_t user_id, AudioOut2UserHandle* handle);
 int KYTY_SYSV_ABI AudioOut2UserDestroy(AudioOut2UserHandle handle);
+int KYTY_SYSV_ABI AudioOut2UserGetSupportedAttributes(AudioOut2UserHandle handle,
+                                                       uint32_t* context_attributes,
+                                                       uint32_t* port_attributes);
 size_t KYTY_SYSV_ABI AudioOut2GetSpeakerArrayMemorySize(uint32_t num_speakers, uint8_t is_3d,
                                                         uint8_t is_ambisonics);
 int KYTY_SYSV_ABI    AudioOut2SpeakerArrayCreate(AudioOut2SpeakerArrayHandle* handle,

src/libs/avPlayer.cpp

@@ -1,6 +1,6 @@
 #include "common/common.h"
+#include "common/file.h"
 #include "common/logging/log.h"
-#include "common/stringUtils.h"
 #include "kernel/fileSystem.h"
 #include "kernel/pthread.h"
 #include "libs/audio.h"
@@ -9,12 +9,14 @@
 #include <algorithm>
 #include <atomic>
 #include <cctype>
+#include <cerrno>
 #include <condition_variable>
 #include <cstddef>
 #include <cstdint>
 #include <cstdio>
 #include <cstring>
 #include <deque>
+#include <limits>
 #include <magic_enum.hpp>
 #include <memory>
 #include <mutex>
@@ -577,30 +579,40 @@ public:
 	explicit FileStreamer(AvPlayerFileReplacement f): file(f) {}
 	~FileStreamer() {
 		if (ctx != nullptr) {
+			av_freep(&ctx->buffer);
 			avio_context_free(&ctx);
 		}
-		if (opened && file.close != nullptr) {
+		if (opened) {
 			file.close(file.object_pointer);
 		}
 	}
 	bool Init(const std::string& path) {
-		if (file.open == nullptr || file.close == nullptr || file.read_offset == nullptr ||
-		    file.size == nullptr) {
+		if (file.open != nullptr) {
+			if (file.close == nullptr || file.read_offset == nullptr || file.size == nullptr ||
+			    file.open(file.object_pointer, path.c_str()) < 0) {
+				return false;
+			}
+			opened = true;
+			size   = file.size(file.object_pointer);
+		} else {
+			if (!local_file.Open(LibKernel::FileSystem::GetRealFilename(path),
+			                     Common::File::Mode::Read)) {
+				return false;
+			}
+			size = local_file.Size();
+		}
+		if (size == 0 || size > static_cast<uint64_t>(std::numeric_limits<int64_t>::max())) {
 			return false;
 		}
-		if (file.open(file.object_pointer, path.c_str()) < 0) {
-			return false;
-		}
-		opened = true;
-		size   = file.size(file.object_pointer);
-		if (size == 0) {
-			return false;
-		}
-		auto* buf = static_cast<uint8_t*>(av_malloc(4096));
+		constexpr int buffer_size = 64 * 1024;
+		auto*         buf         = static_cast<uint8_t*>(av_malloc(buffer_size));
 		if (buf == nullptr) {
 			return false;
 		}
-		ctx = avio_alloc_context(buf, 4096, 0, this, Read, nullptr, Seek);
+		ctx = avio_alloc_context(buf, buffer_size, 0, this, Read, nullptr, Seek);
+		if (ctx == nullptr) {
+			av_free(buf);
+		}
 		return ctx != nullptr;
 	}
 	AVIOContext* Context() const { return ctx; }
@@ -611,33 +623,50 @@ private:
 		if (s->pos >= s->size) {
 			return AVERROR_EOF;
 		}
-		len = static_cast<int>(std::min<uint64_t>(len, s->size - s->pos));
-		auto r =
-		    s->file.read_offset(s->file.object_pointer, buf, s->pos, static_cast<uint32_t>(len));
-		if (r <= 0) {
-			return r == 0 ? AVERROR_EOF : r;
+		len      = static_cast<int>(std::min<uint64_t>(len, s->size - s->pos));
+		int read = 0;
+		if (s->opened) {
+			read = s->file.read_offset(s->file.object_pointer, buf, s->pos,
+			                           static_cast<uint32_t>(len));
+		} else {
+			uint32_t bytes = 0;
+			s->local_file.Read(buf, static_cast<uint32_t>(len), &bytes);
+			read = static_cast<int>(bytes);
 		}
-		s->pos += static_cast<uint64_t>(r);
-		return r;
+		if (read <= 0 || read > len) {
+			return read < 0 ? read : AVERROR(EIO);
+		}
+		s->pos += static_cast<uint64_t>(read);
+		return read;
 	}
-	static int64_t Seek(void* opaque, int64_t off, int whence) {
+	static int64_t Seek(void* opaque, int64_t offset, int whence) {
 		auto* s = static_cast<FileStreamer*>(opaque);
 		if ((whence & AVSEEK_SIZE) != 0) {
 			return static_cast<int64_t>(s->size);
 		}
-		int64_t p = whence == SEEK_SET
-		                ? off
-		                : (whence == SEEK_CUR
-		                       ? static_cast<int64_t>(s->pos) + off
-		                       : (whence == SEEK_END ? static_cast<int64_t>(s->size) + off : -1));
-		if (p < 0) {
-			return -1;
+		whence &= ~AVSEEK_FORCE;
+		uint64_t base = 0;
+		switch (whence) {
+			case SEEK_SET: break;
+			case SEEK_CUR: base = s->pos; break;
+			case SEEK_END: base = s->size; break;
+			default: return AVERROR(EINVAL);
 		}
-		p      = std::min<int64_t>(p, static_cast<int64_t>(s->size));
-		s->pos = static_cast<uint64_t>(p);
-		return p;
+		const auto distance = offset < 0 ? uint64_t {0} - static_cast<uint64_t>(offset)
+		                                 : static_cast<uint64_t>(offset);
+		if ((offset < 0 && distance > base) ||
+		    (offset >= 0 && distance > std::numeric_limits<int64_t>::max() - base)) {
+			return AVERROR(EINVAL);
+		}
+		const auto position = offset < 0 ? base - distance : base + distance;
+		if (!s->opened && !s->local_file.Seek(position)) {
+			return AVERROR(EIO);
+		}
+		s->pos = position;
+		return static_cast<int64_t>(position);
 	}
 	AvPlayerFileReplacement file;
+	Common::File            local_file;
 	bool                    opened = false;
 	uint64_t                pos    = 0;
 	uint64_t                size   = 0;
@@ -675,25 +704,17 @@ public:
 			return static_cast<Source*>(opaque)->interrupt_io.load() ? 1 : 0;
 		};
 		raw->interrupt_callback.opaque = this;
-		if (file.open != nullptr) {
-			streamer = std::make_unique<FileStreamer>(file);
-			if (!streamer->Init(path)) {
-				avformat_free_context(raw);
-				return AVPLAYER_ERROR_OPERATION_FAILED;
-			}
-			raw->pb = streamer->Context();
-			if (auto rc = avformat_open_input(&raw, nullptr, nullptr, nullptr); rc < 0) {
-				LOGF("\t avformat_open_input callback failed: %s\n", fferr(rc).c_str());
-				return AVPLAYER_ERROR_OPERATION_FAILED;
-			}
-		} else {
-			auto real     = LibKernel::FileSystem::GetRealFilename(std::string(path.c_str()));
-			auto real_str = Common::PathToString(real);
-			if (auto rc = avformat_open_input(&raw, real_str.c_str(), nullptr, nullptr); rc < 0) {
-				LOGF("\t avformat_open_input failed: %s path=%s\n", fferr(rc).c_str(),
-				     real_str.c_str());
-				return AVPLAYER_ERROR_OPERATION_FAILED;
-			}
+		streamer                       = std::make_unique<FileStreamer>(file);
+		if (!streamer->Init(path)) {
+			avformat_free_context(raw);
+			return AVPLAYER_ERROR_OPERATION_FAILED;
+		}
+		raw->pb = streamer->Context();
+		if (const auto rc = avformat_open_input(&raw, file.open == nullptr ? path.c_str() : nullptr,
+		                                        nullptr, nullptr);
+		    rc < 0) {
+			LOGF("\t avformat_open_input failed: %s path=%s\n", fferr(rc).c_str(), path.c_str());
+			return AVPLAYER_ERROR_OPERATION_FAILED;
 		}
 		fmt = raw;
 		if (auto rc = avformat_find_stream_info(fmt, nullptr); rc < 0) {

src/libs/dialog.cpp

@@ -6,6 +6,7 @@
 #include "libs/errno.h"
 #include "libs/libs.h"
 
+#include <atomic>
 #include <cstdio>
 #include <cstring>
 #include <mutex>
@@ -306,6 +307,7 @@ LIB_NAME("SaveDataDialog", "SaveDataDialog");
 
 constexpr int SAVE_STATUS_NONE        = 0;
 constexpr int SAVE_STATUS_INITIALIZED = 1;
+constexpr int SAVE_STATUS_RUNNING     = 2;
 constexpr int SAVE_STATUS_FINISHED    = 3;
 constexpr int SAVE_RESULT_OK          = 0;
 constexpr int SAVE_BUTTON_ID_OK       = 1;
@@ -351,18 +353,20 @@ struct SaveDataDialogResult {
 	uint8_t  reserved[32];
 };
 
-static int   g_save_status    = SAVE_STATUS_NONE;
-static int   g_save_mode      = 0;
-static void* g_save_user_data = nullptr;
-static char  g_save_dir_name[sizeof(SaveDataDirName::data)] {};
+static std::atomic<int> g_save_status         = SAVE_STATUS_NONE;
+static bool             g_save_running_polled = false;
+static int              g_save_mode           = 0;
+static void*            g_save_user_data      = nullptr;
+static char             g_save_dir_name[sizeof(SaveDataDirName::data)] {};
 
 int KYTY_SYSV_ABI SaveDataDialogInitialize() {
 	PRINT_NAME();
 
-	g_save_status      = SAVE_STATUS_INITIALIZED;
-	g_save_mode        = 0;
-	g_save_user_data   = nullptr;
-	g_save_dir_name[0] = '\0';
+	g_save_status         = SAVE_STATUS_INITIALIZED;
+	g_save_running_polled = false;
+	g_save_mode           = 0;
+	g_save_user_data      = nullptr;
+	g_save_dir_name[0]    = '\0';
 
 	return OK;
 }
@@ -376,6 +380,14 @@ int KYTY_SYSV_ABI SaveDataDialogGetStatus() {
 int KYTY_SYSV_ABI SaveDataDialogUpdateStatus() {
 	PRINT_NAME();
 
+	// Some games require a RUNNING update before FINISHED.
+	if (g_save_status == SAVE_STATUS_RUNNING) {
+		if (g_save_running_polled) {
+			g_save_status = SAVE_STATUS_FINISHED;
+		}
+		g_save_running_polled = true;
+	}
+
 	return g_save_status;
 }
 
@@ -431,7 +443,8 @@ int KYTY_SYSV_ABI SaveDataDialogOpen(const void* param) {
 		}
 	}
 
-	g_save_status = SAVE_STATUS_FINISHED;
+	g_save_status         = SAVE_STATUS_RUNNING;
+	g_save_running_polled = false;
 
 	return OK;
 }
@@ -441,7 +454,8 @@ int KYTY_SYSV_ABI SaveDataDialogClose(const void* close_param) {
 
 	LOGF("\t close_param = 0x%016" PRIx64 "\n", reinterpret_cast<uint64_t>(close_param));
 
-	g_save_status = SAVE_STATUS_FINISHED;
+	g_save_status         = SAVE_STATUS_FINISHED;
+	g_save_running_polled = false;
 
 	return OK;
 }
@@ -455,10 +469,11 @@ int KYTY_SYSV_ABI SaveDataDialogIsReadyToDisplay() {
 int KYTY_SYSV_ABI SaveDataDialogTerminate() {
 	PRINT_NAME();
 
-	g_save_status      = SAVE_STATUS_NONE;
-	g_save_mode        = 0;
-	g_save_user_data   = nullptr;
-	g_save_dir_name[0] = '\0';
+	g_save_status         = SAVE_STATUS_NONE;
+	g_save_running_polled = false;
+	g_save_mode           = 0;
+	g_save_user_data      = nullptr;
+	g_save_dir_name[0]    = '\0';
 
 	return 0;
 }

src/libs/dualSenseHaptics.cpp

@@ -5,13 +5,17 @@
 
 #include <SDL3/SDL.h>
 #include <algorithm>
+#include <array>
 #include <cmath>
+#include <map>
 #include <vector>
 
 namespace Libs::Controller::DualSenseHaptics {
 
 struct Stream {
 	uint32_t           freq       = 0;
+	bool               speaker    = false;
+	int                controller = -1;
 	SDL_AudioStream*   sdl        = nullptr;
 	SDL_AudioDeviceID  device     = 0;
 	uint64_t           next_check = 0;
@@ -20,19 +24,26 @@ struct Stream {
 
 namespace {
 
-constexpr int FRAME_BYTES = 4 * sizeof(float);
-Common::Mutex g_mutex;
-int           g_controller       = -1;
-int           g_playback_streams = 0;
-uint8_t       g_large_motor = 0, g_small_motor = 0;
-uint64_t      g_rumble_until = 0, g_haptics_until = 0;
+constexpr int      FRAME_BYTES = 4 * sizeof(float);
+Common::Mutex      g_mutex;
+int                g_controller      = -1;
+int                g_haptics_streams = 0;
+uint8_t            g_large_motor = 0, g_small_motor = 0;
+uint64_t           g_rumble_until = 0, g_haptics_until = 0;
+int                g_speaker_controller = -1;
+std::map<int, int> g_speaker_streams;
 
-void ApplyRumble() { // Caller holds g_mutex; no calls into Audio or Controller.
+void ApplyRumble(bool force = false) { // Caller holds g_mutex; no calls into Audio or Controller.
 	SDL_LockJoysticks();
 	if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(g_controller));
 	    pad != nullptr) {
 		const auto now   = SDL_GetTicks();
 		const bool muted = g_haptics_until > now || g_rumble_until <= now;
+		if (force && !muted) {
+			// SDL caches motor strengths. Force a report after raw speaker routing changed the
+			// controller's motor mode without changing that cache.
+			(void)SDL_RumbleGamepad(pad, 0, 0, 0);
+		}
 		(void)SDL_RumbleGamepad(
 		    pad, muted ? 0 : g_large_motor * 0x101U, muted ? 0 : g_small_motor * 0x101U,
 		    g_rumble_until > now ? static_cast<uint32_t>(g_rumble_until - now) : 0);
@@ -42,14 +53,52 @@ void ApplyRumble() { // Caller holds g_mutex; no calls into Audio or Controller.
 
 void SelectController(int controller) {
 	if (g_controller != controller) {
-		g_haptics_until = 0;
+		g_haptics_until = g_rumble_until = 0;
 		ApplyRumble();
 		g_controller  = controller;
 		g_large_motor = g_small_motor = 0;
-		g_rumble_until                = 0;
 	}
 }
 
+bool RouteSpeaker(int controller, bool speaker) { // Caller holds g_mutex.
+	// Route the right channel to the speaker and mute the headphones with the values Linux uses
+	// without headphones (output report 0x02). Plugged-in headphones are not checked.
+	std::array<uint8_t, 38> report {};
+	report[0] = 0x80; // Audio control valid
+	if (speaker) {
+		report[0] |= 0x20; // Speaker volume valid
+		report[1]  = 0x80; // Audio control 2 valid
+		report[5]  = 0x64; // Speaker volume
+		report[7]  = 0x30; // Output path: right channel to the speaker, headphones muted
+		report[37] = 0x02; // Speaker preamp +6 dB
+	}
+	bool routed = false;
+	SDL_LockJoysticks();
+	if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(controller)); pad != nullptr) {
+		routed = SDL_SendGamepadEffect(pad, report.data(), static_cast<int>(report.size()));
+	}
+	SDL_UnlockJoysticks();
+	if (routed && controller == g_controller) {
+		ApplyRumble(true);
+	}
+	return routed;
+}
+
+bool SelectSpeaker(int controller) { // Caller holds g_mutex; -1 restores headphone routing.
+	if (g_speaker_controller == controller) {
+		return true;
+	}
+	if (g_speaker_controller != -1) {
+		RouteSpeaker(g_speaker_controller, false);
+	}
+	g_speaker_controller = -1;
+	if (controller != -1 && !RouteSpeaker(controller, true)) {
+		return false;
+	}
+	g_speaker_controller = controller;
+	return true;
+}
+
 void SDLCALL UpdateRumble(void*, SDL_AudioStream*, int, int) {
 	Common::LockGuard lock(g_mutex);
 	if (g_haptics_until != 0 && SDL_GetTicks() >= g_haptics_until) {
@@ -64,22 +113,55 @@ void CloseDevice(Stream* stream) {
 		SDL_DestroyAudioStream(stream->sdl);
 		stream->sdl = nullptr;
 		Common::LockGuard lock(g_mutex);
-		if (--g_playback_streams == 0 && g_haptics_until != 0) {
+		if (!stream->speaker && --g_haptics_streams == 0 && g_haptics_until != 0) {
 			g_haptics_until = 0;
 			ApplyRumble();
 		}
+		// Route back to the headphones once no speaker port plays on the controller.
+		if (stream->speaker && --g_speaker_streams.at(stream->controller) == 0) {
+			g_speaker_streams.erase(stream->controller);
+			if (g_speaker_controller == stream->controller) {
+				SelectSpeaker(-1);
+			}
+		}
 	}
 }
 
+bool CanUseDevice(int controller) {
+	// SDL exposes no portable association between a gamepad and its USB audio endpoint.
+	// Require one wired DualSense; guessing with multiple pads could play on the wrong one.
+	SDL_LockJoysticks();
+	auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(controller));
+	bool  usable =
+	    pad != nullptr && SDL_GetGamepadConnectionState(pad) == SDL_JOYSTICK_CONNECTION_WIRED;
+	int   count           = 0;
+	auto* pads            = SDL_GetGamepads(&count);
+	int   dualsense_count = 0;
+	bool  found           = false;
+	for (int i = 0; i < count; i++) {
+		if (SDL_GetGamepadTypeForID(pads[i]) == SDL_GAMEPAD_TYPE_PS5) {
+			dualsense_count++;
+			found |= pads[i] == static_cast<SDL_JoystickID>(controller);
+		}
+	}
+	SDL_free(pads);
+	SDL_UnlockJoysticks();
+	return usable && found && dualsense_count == 1;
+}
+
 SDL_AudioDeviceID FindDevice() {
 	int               count   = 0;
 	auto*             devices = SDL_GetAudioPlaybackDevices(&count);
 	SDL_AudioDeviceID device  = 0;
-	for (int i = 0; i < count && device == 0; i++) {
+	for (int i = 0; i < count; i++) {
 		const char*   name = SDL_GetAudioDeviceName(devices[i]);
 		SDL_AudioSpec spec {};
 		if (name != nullptr && SDL_strcasestr(name, "DualSense") != nullptr &&
 		    SDL_GetAudioDeviceFormat(devices[i], &spec, nullptr) && spec.channels == 4) {
+			if (device != 0) {
+				device = 0; // Multiple matching endpoints cannot be associated reliably either.
+				break;
+			}
 			device = devices[i];
 		}
 	}
@@ -87,13 +169,59 @@ SDL_AudioDeviceID FindDevice() {
 	return device;
 }
 
+void RefreshDevice(Stream* stream, int controller) {
+	if (stream->controller != controller) {
+		CloseDevice(stream);
+		stream->controller = controller;
+		stream->next_check = 0;
+	}
+	const auto now = SDL_GetTicks();
+	if (now < stream->next_check) {
+		return;
+	}
+	stream->next_check = now + 2000;
+	const auto device  = FindDevice();
+	if (stream->device == device && stream->sdl != nullptr) {
+		return;
+	}
+	CloseDevice(stream);
+	stream->device = device;
+	if (device == 0) {
+		return;
+	}
+	const SDL_AudioSpec desired {SDL_AUDIO_F32, 4, static_cast<int>(stream->freq)};
+	auto*               sdl = SDL_OpenAudioDeviceStream(device, &desired, UpdateRumble, nullptr);
+	SDL_AudioSpec       actual {};
+	bool ready = sdl != nullptr &&
+	             SDL_GetAudioDeviceFormat(SDL_GetAudioStreamDevice(sdl), &actual, nullptr) &&
+	             actual.channels == 4 && SDL_ResumeAudioStreamDevice(sdl);
+	if (ready) {
+		Common::LockGuard lock(g_mutex);
+		if (stream->speaker) {
+			ready = SelectSpeaker(controller);
+			if (ready) {
+				g_speaker_streams[controller]++;
+			}
+		} else {
+			g_haptics_streams++;
+		}
+	}
+	if (ready) {
+		stream->sdl = sdl;
+		LOGF("DualSenseHaptics: playing on '%s'\n", SDL_GetAudioDeviceName(device));
+	} else {
+		SDL_DestroyAudioStream(sdl);
+		LOGF("DualSenseHaptics: cannot open quad output: %s\n", SDL_GetError());
+	}
+}
+
 } // namespace
 
-Stream* Open(uint32_t freq) {
+Stream* Open(uint32_t freq, bool speaker) {
 	if (freq == 0 || !SDL_InitSubSystem(SDL_INIT_AUDIO)) {
 		return nullptr;
 	}
-	return new Stream {.freq = freq};
+	return new Stream {.freq = freq, .speaker = speaker};
 }
 
 void Close(Stream* stream) {
@@ -104,44 +232,30 @@ void Close(Stream* stream) {
 	}
 }
 
-void Queue(Stream* stream, int controller, const void* data, uint32_t frames, uint32_t channels,
-           bool is_float, const int* volume) {
+uint64_t Queue(Stream* stream, int controller, const void* data, uint32_t frames, uint32_t channels,
+               bool is_float, const int* volume) {
 	if (stream == nullptr || data == nullptr || frames == 0 || channels == 0 || volume == nullptr) {
-		return;
+		return 0;
 	}
-	if (SDL_GetGamepadTypeForID(static_cast<SDL_JoystickID>(controller)) != SDL_GAMEPAD_TYPE_PS5) {
+	if (!CanUseDevice(controller)) {
 		CloseDevice(stream);
-		return;
-	}
-	const auto now = SDL_GetTicks();
-	if (now >= stream->next_check) {
-		stream->next_check = now + 2000;
-		const auto device  = FindDevice();
-		if (stream->device != device || stream->sdl == nullptr) {
-			CloseDevice(stream);
-			stream->device = device;
-			if (device != 0) {
-				const SDL_AudioSpec desired {SDL_AUDIO_F32, 4, static_cast<int>(stream->freq)};
-				auto* sdl = SDL_OpenAudioDeviceStream(device, &desired, UpdateRumble, nullptr);
-				SDL_AudioSpec actual {};
-				if (sdl != nullptr &&
-				    SDL_GetAudioDeviceFormat(SDL_GetAudioStreamDevice(sdl), &actual, nullptr) &&
-				    actual.channels == 4 && SDL_ResumeAudioStreamDevice(sdl)) {
-					stream->sdl = sdl;
-					{
-						Common::LockGuard lock(g_mutex);
-						g_playback_streams++;
-					}
-					LOGF("DualSenseHaptics: playing on '%s'\n", SDL_GetAudioDeviceName(device));
-				} else {
-					SDL_DestroyAudioStream(sdl);
-					LOGF("DualSenseHaptics: cannot open quad output: %s\n", SDL_GetError());
-				}
-			}
-		}
+		stream->next_check = 0; // Look for the device as soon as a DualSense is active again.
+		return 0;
 	}
+	RefreshDevice(stream, controller);
 	if (stream->sdl == nullptr) {
-		return;
+		return 0;
+	}
+	if (stream->speaker) {
+		bool routed;
+		{
+			Common::LockGuard lock(g_mutex);
+			routed = SelectSpeaker(controller);
+		}
+		if (!routed) {
+			CloseDevice(stream);
+			return 0;
+		}
 	}
 	stream->buffer.assign(static_cast<size_t>(frames) * 4, 0.0f);
 	bool audible = false;
@@ -152,7 +266,8 @@ void Queue(Stream* stream, int controller, const void* data, uint32_t frames, ui
 			float      value  = is_float ? static_cast<const float*>(data)[index]
 			                             : static_cast<const int16_t*>(data)[index] / 32768.0f;
 			value *= volume[src_ch] / 32768.0f;
-			stream->buffer[static_cast<size_t>(frame) * 4 + ch + 2] = value;
+			// The speaker plays the right front channel; the actuators play the back ones.
+			stream->buffer[static_cast<size_t>(frame) * 4 + ch + (stream->speaker ? 0 : 2)] = value;
 			audible |= std::fabs(value) > 1.0f / 1024;
 		}
 	}
@@ -161,26 +276,29 @@ void Queue(Stream* stream, int controller, const void* data, uint32_t frames, ui
 		SDL_ClearAudioStream(stream->sdl);
 		queued = 0;
 	}
-	const int bytes = static_cast<int>(frames * FRAME_BYTES);
-	if (audible) {
+	const int  bytes = static_cast<int>(frames * FRAME_BYTES);
+	const auto queued_us =
+	    static_cast<uint64_t>(queued + bytes) * 1000000 / (stream->freq * FRAME_BYTES);
+	if (audible && !stream->speaker) {
 		Common::LockGuard lock(g_mutex);
 		if (SDL_GetGamepadTypeForID(static_cast<SDL_JoystickID>(controller)) !=
 		    SDL_GAMEPAD_TYPE_PS5) {
-			return;
+			return 0;
 		}
 		SelectController(controller);
 		const auto playback_now = SDL_GetTicks();
 		const bool starting     = g_haptics_until <= playback_now;
-		const auto duration     = (queued + bytes) * 1000ULL / (stream->freq * FRAME_BYTES);
 		g_haptics_until =
-		    std::max(g_haptics_until, playback_now + std::max<uint64_t>(duration, 250));
+		    std::max(g_haptics_until, playback_now + std::max<uint64_t>(queued_us / 1000, 250));
 		if (starting) {
 			ApplyRumble();
 		}
 	}
 	if (!SDL_PutAudioStreamData(stream->sdl, stream->buffer.data(), bytes)) {
 		CloseDevice(stream);
+		return 0;
 	}
+	return queued_us;
 }
 
 bool SetVibration(int controller, uint8_t large_motor, uint8_t small_motor) {
@@ -198,8 +316,8 @@ bool SetVibration(int controller, uint8_t large_motor, uint8_t small_motor) {
 
 void Shutdown() {
 	Common::LockGuard lock(g_mutex);
-	g_rumble_until = 0;
 	SelectController(-1);
+	SelectSpeaker(-1);
 }
 
 } // namespace Libs::Controller::DualSenseHaptics