gpu: preserve finite buffer descriptors across sparse PRT holes

Avoid readbacks for compute metadata fills

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

src/common/CMakeLists.txt

@@ -43,6 +43,10 @@ target_link_libraries(common
 		ZArchive::zarchive
 		nlohmann_json::nlohmann_json
 )
+if(APPLE)
+	# hostStats.cpp reads GPU utilization from the IOAccelerator registry entries.
+	target_link_libraries(common PRIVATE "-framework CoreFoundation" "-framework IOKit")
+endif()
 
 get_property(inc_headers TARGET common PROPERTY INCLUDE_DIRECTORIES)
 set(check_headers ${CMAKE_CURRENT_SOURCE_DIR})

src/common/debugCounters.h

@@ -0,0 +1,117 @@
+#ifndef KYTY_COMMON_DEBUG_COUNTERS_H_
+#define KYTY_COMMON_DEBUG_COUNTERS_H_
+
+#include <atomic>
+#include <chrono>
+#include <cstdint>
+
+namespace Common::DebugCounters {
+
+static_assert(std::atomic<uint64_t>::is_always_lock_free && std::atomic<bool>::is_always_lock_free);
+
+// Collect events only while the panel is visible. Relaxed, lock-free atomics keep updates
+// cheap and safe in signal handlers. Rates come from comparing two snapshots.
+enum class Counter {
+	// GPU work.
+	Draws,
+	DrawInstances,
+	IndirectDraws,
+	Dispatches,
+	QueueSubmits,
+	GpuWaits, // CPU blocked on a GPU timeline semaphore.
+	GpuWaitNs,
+	GpuFaults, // Guest writes to GPU-tracked memory.
+	ShadersCompiled,
+	ShaderCompileNs,
+	PipelinesCreated,
+	PipelineCreateNs,
+	// Caches.
+	TextureUploadBytes,
+	TextureDownloadBytes,
+	TextureEvictions,
+	BufferUploadBytes,
+	BufferDownloadBytes,
+	BufferEvictions,
+	// Presentation.
+	FlipsRequested,
+	FlipsRejected, // Guest flip queue was full.
+	LateVblanks,   // The present thread fell behind the vblank rate.
+	SwapchainRecreations,
+	DroppedPresents,
+	// Emulator.
+	EmulatedInstructions, // Host CPU lacked an instruction the guest used.
+	UnresolvedImportCalls,
+	GuestFileReadBytes,
+	GuestFileWriteBytes,
+	AudioUnderruns,
+	Count
+};
+
+enum class Peak { ShaderCompileNs, PipelineCreateNs, Count };
+
+namespace Detail {
+inline std::atomic<bool>     g_enabled {false};
+inline std::atomic<uint64_t> g_counters[static_cast<int>(Counter::Count)] {};
+inline std::atomic<uint64_t> g_peaks[static_cast<int>(Peak::Count)] {};
+} // namespace Detail
+
+[[nodiscard]] inline bool Enabled() noexcept {
+	return Detail::g_enabled.load(std::memory_order_relaxed);
+}
+
+inline void SetEnabled(bool enabled) noexcept {
+	Detail::g_enabled.store(enabled, std::memory_order_relaxed);
+}
+
+inline void Add(Counter counter, uint64_t value = 1) noexcept {
+	if (!Enabled()) {
+		return;
+	}
+	Detail::g_counters[static_cast<int>(counter)].fetch_add(value, std::memory_order_relaxed);
+}
+
+class ScopedTimer final {
+public:
+	ScopedTimer(Counter count, Counter duration, Peak peak = Peak::Count)
+	    : m_count(count), m_duration(duration), m_peak(peak),
+	      m_start(Enabled() ? std::chrono::steady_clock::now()
+	                        : std::chrono::steady_clock::time_point {}) {}
+	~ScopedTimer() {
+		if (m_start == std::chrono::steady_clock::time_point {} || !Enabled()) {
+			return;
+		}
+		const auto elapsed =
+		    static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::nanoseconds>(
+		                              std::chrono::steady_clock::now() - m_start)
+		                              .count());
+		Add(m_count);
+		Add(m_duration, elapsed);
+		if (m_peak != Peak::Count) {
+			auto& peak     = Detail::g_peaks[static_cast<int>(m_peak)];
+			auto  previous = peak.load(std::memory_order_relaxed);
+			while (previous < elapsed &&
+			       !peak.compare_exchange_weak(previous, elapsed, std::memory_order_relaxed)) {
+			}
+		}
+	}
+	ScopedTimer(const ScopedTimer&)            = delete;
+	ScopedTimer& operator=(const ScopedTimer&) = delete;
+
+private:
+	Counter                               m_count;
+	Counter                               m_duration;
+	Peak                                  m_peak;
+	std::chrono::steady_clock::time_point m_start;
+};
+
+[[nodiscard]] inline uint64_t Get(Counter counter) noexcept {
+	return Detail::g_counters[static_cast<int>(counter)].load(std::memory_order_relaxed);
+}
+
+[[nodiscard]] inline uint64_t Get(Peak peak) noexcept {
+	return Detail::g_peaks[static_cast<int>(peak)].load(std::memory_order_relaxed);
+}
+
+} // namespace Common::DebugCounters
+
+#endif /* KYTY_COMMON_DEBUG_COUNTERS_H_ */

src/common/emulatorConfig.cpp

@@ -1,6 +1,7 @@
 #include "common/emulatorConfig.h"
 
 #include "common/assert.h"
+#include "common/debugCounters.h"
 
 #include <algorithm>
 #include <memory>
@@ -16,6 +17,7 @@ void Initialize() {
 }
 
 void Shutdown() {
+	Common::DebugCounters::SetEnabled(false);
 	g_config.reset();
 }
 
@@ -27,6 +29,7 @@ void Load(const ConfigOptions& cfg) {
 	EXIT_IF(cfg.keyboard_player > static_cast<uint32_t>(MAX_PLAYERS));
 
 	*g_config = cfg;
+	Common::DebugCounters::SetEnabled(cfg.perf_overlay_enabled);
 }
 
 uint32_t GetScreenWidth() {

src/common/emulatorConfig.h

@@ -89,6 +89,7 @@ struct ConfigOptions {
 	bool                   tessellation_enabled        = false;
 	bool                   trophy_enabled              = true;
 	uint32_t               trophy_duration_seconds     = 6;
+	bool                   perf_overlay_enabled            = false;
 	bool                   playgo_hack_enabled         = false;
 	bool                   skip_notice_screen          = false;
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS

src/common/hostStats.cpp

@@ -0,0 +1,580 @@
+#include "common/hostStats.h"
+
+#include "common/config.h"
+
+#include <algorithm>
+#include <chrono>
+#include <thread>
+#include <unordered_map>
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h> // IWYU pragma: keep
+// clang-format off
+#include <psapi.h>
+// clang-format on
+#elif defined(__APPLE__)
+#include <CoreFoundation/CoreFoundation.h>
+#include <IOKit/IOKitLib.h>
+#include <cstring>
+#include <libproc.h>
+#include <mach/mach.h>
+#include <notify.h>
+#include <sys/resource.h>
+#include <sys/sysctl.h>
+#include <unistd.h>
+#else
+#include <filesystem>
+#include <fstream>
+#include <sstream>
+#include <sys/resource.h>
+#include <unistd.h>
+#endif
+
+namespace Common {
+
+namespace {
+
+// Cumulative counters; rates come from the difference between two samples.
+struct Counters {
+	struct Thread {
+		uint64_t    cpu_ns;
+		std::string name;
+	};
+	std::optional<uint64_t>              process_cpu_ns;
+	std::optional<uint64_t>              system_busy;
+	std::optional<uint64_t>              system_total;
+	std::optional<uint64_t>              context_switches;
+	std::optional<uint64_t>              page_faults;
+	std::optional<uint64_t>              page_ins;
+	std::optional<uint64_t>              disk_read;
+	std::optional<uint64_t>              disk_write;
+	std::unordered_map<uint64_t, Thread> threads;
+	uint64_t                             wall_ns = 0;
+};
+
+uint64_t WallNanoseconds() {
+	return static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::nanoseconds>(
+	                                 std::chrono::steady_clock::now().time_since_epoch())
+	                                 .count());
+}
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+
+uint64_t FileTimeTo100Ns(const FILETIME& time) {
+	return (static_cast<uint64_t>(time.dwHighDateTime) << 32u) | time.dwLowDateTime;
+}
+
+struct PlatformState {};
+
+void SamplePlatform(PlatformState& /*state*/, HostStats& stats, Counters& now) {
+	const auto process = GetCurrentProcess();
+	FILETIME   creation {};
+	FILETIME   exit {};
+	FILETIME   kernel {};
+	FILETIME   user {};
+	if (GetProcessTimes(process, &creation, &exit, &kernel, &user) != 0) {
+		now.process_cpu_ns = (FileTimeTo100Ns(kernel) + FileTimeTo100Ns(user)) * 100u;
+	}
+	FILETIME idle {};
+	if (GetSystemTimes(&idle, &kernel, &user) != 0) {
+		// Kernel time includes idle time.
+		now.system_total = FileTimeTo100Ns(kernel) + FileTimeTo100Ns(user);
+		now.system_busy  = *now.system_total - FileTimeTo100Ns(idle);
+	}
+
+	PROCESS_MEMORY_COUNTERS memory {};
+	if (K32GetProcessMemoryInfo(process, &memory, sizeof(memory)) != 0) {
+		stats.process_memory   = memory.WorkingSetSize;
+		stats.process_resident = memory.WorkingSetSize;
+		now.page_faults        = memory.PageFaultCount;
+	}
+	IO_COUNTERS io {};
+	if (GetProcessIoCounters(process, &io) != 0) {
+		now.disk_read  = io.ReadTransferCount;
+		now.disk_write = io.WriteTransferCount;
+	}
+
+	MEMORYSTATUSEX system {};
+	system.dwLength = sizeof(system);
+	if (GlobalMemoryStatusEx(&system) != 0) {
+		stats.system_memory_total = system.ullTotalPhys;
+		stats.system_memory_used  = system.ullTotalPhys - system.ullAvailPhys;
+		stats.system_free         = system.ullAvailPhys;
+		stats.memory_pressure     = system.dwMemoryLoad >= 90   ? MemoryPressure::Critical
+		                            : system.dwMemoryLoad >= 80 ? MemoryPressure::Warning
+		                                                        : MemoryPressure::Normal;
+	}
+}
+
+#else
+
+void SampleRusage(Counters& now) {
+	rusage usage {};
+	if (getrusage(RUSAGE_SELF, &usage) != 0) {
+		return;
+	}
+	const auto to_ns = [](const timeval& time) {
+		return static_cast<uint64_t>(time.tv_sec) * 1000000000u +
+		       static_cast<uint64_t>(time.tv_usec) * 1000u;
+	};
+	now.process_cpu_ns   = to_ns(usage.ru_utime) + to_ns(usage.ru_stime);
+	now.context_switches = static_cast<uint64_t>(usage.ru_nvcsw + usage.ru_nivcsw);
+	now.page_faults      = static_cast<uint64_t>(usage.ru_minflt + usage.ru_majflt);
+	now.page_ins         = static_cast<uint64_t>(usage.ru_majflt);
+}
+
+#if defined(__APPLE__)
+
+template <typename T>
+std::optional<T> SysctlValue(const char* name) {
+	T      value {};
+	size_t size = sizeof(value);
+	if (sysctlbyname(name, &value, &size, nullptr, 0) != 0 || size != sizeof(value)) {
+		return std::nullopt;
+	}
+	return value;
+}
+
+struct PlatformState {
+	mach_port_t             host          = mach_host_self();
+	int                     thermal_token = -1;
+	bool                    thermal_ready = false;
+	std::string             gpu_model;
+	std::optional<uint32_t> gpu_cores;
+
+	~PlatformState() {
+		if (thermal_ready) {
+			notify_cancel(thermal_token);
+		}
+		mach_port_deallocate(mach_task_self(), host);
+	}
+};
+
+std::optional<int64_t> DictionaryNumber(CFDictionaryRef dictionary, CFStringRef key) {
+	const auto* value  = CFDictionaryGetValue(dictionary, key);
+	int64_t     number = 0;
+	if (value == nullptr || CFGetTypeID(value) != CFNumberGetTypeID() ||
+	    !CFNumberGetValue(static_cast<CFNumberRef>(value), kCFNumberSInt64Type, &number) ||
+	    number < 0) {
+		return std::nullopt;
+	}
+	return number;
+}
+
+std::string RegistryString(io_object_t service, CFStringRef key) {
+	std::string result;
+	auto*       value = IORegistryEntryCreateCFProperty(service, key, kCFAllocatorDefault, 0);
+	if (value == nullptr) {
+		return result;
+	}
+	if (CFGetTypeID(value) == CFStringGetTypeID()) {
+		char buffer[128] {};
+		if (CFStringGetCString(static_cast<CFStringRef>(value), buffer, sizeof(buffer),
+		                       kCFStringEncodingUTF8)) {
+			result = buffer;
+		}
+	} else if (CFGetTypeID(value) == CFDataGetTypeID()) {
+		// Intel Macs store the model as NUL-terminated data.
+		const auto* data   = static_cast<CFDataRef>(value);
+		const auto* bytes  = reinterpret_cast<const char*>(CFDataGetBytePtr(data));
+		const auto  length = static_cast<size_t>(CFDataGetLength(data));
+		result.assign(bytes, strnlen(bytes, length));
+	}
+	CFRelease(value);
+	return result;
+}
+
+// IOAccelerator exposes whole-device statistics; there is no per-process figure.
+void SampleGpu(PlatformState& state, HostStats& stats) {
+	io_iterator_t iterator = 0;
+	if (IOServiceGetMatchingServices(MACH_PORT_NULL, IOServiceMatching("IOAccelerator"),
+	                                 &iterator) != KERN_SUCCESS) {
+		return;
+	}
+	const auto take_max = [](std::optional<double>& target, std::optional<int64_t> value) {
+		if (value) {
+			target = std::max(target.value_or(0.0), std::min(100.0, static_cast<double>(*value)));
+		}
+	};
+	const auto add = [](std::optional<uint64_t>& target, std::optional<int64_t> value) {
+		if (value) {
+			target = target.value_or(0) + static_cast<uint64_t>(*value);
+		}
+	};
+	for (io_object_t service = IOIteratorNext(iterator); service != 0;
+	     service             = IOIteratorNext(iterator)) {
+		if (state.gpu_model.empty()) {
+			state.gpu_model = RegistryString(service, CFSTR("model"));
+			auto*   cores   = IORegistryEntryCreateCFProperty(service, CFSTR("gpu-core-count"),
+			                                                  kCFAllocatorDefault, 0);
+			int64_t count   = 0;
+			if (cores != nullptr) {
+				if (CFGetTypeID(cores) == CFNumberGetTypeID() &&
+				    CFNumberGetValue(static_cast<CFNumberRef>(cores), kCFNumberSInt64Type,
+				                     &count) &&
+				    count > 0) {
+					state.gpu_cores = static_cast<uint32_t>(count);
+				}
+				CFRelease(cores);
+			}
+		}
+		auto* statistics = IORegistryEntryCreateCFProperty(service, CFSTR("PerformanceStatistics"),
+		                                                   kCFAllocatorDefault, 0);
+		if (statistics != nullptr) {
+			if (CFGetTypeID(statistics) == CFDictionaryGetTypeID()) {
+				const auto* values = static_cast<CFDictionaryRef>(statistics);
+				take_max(stats.gpu_percent,
+				         DictionaryNumber(values, CFSTR("Device Utilization %")));
+				take_max(stats.gpu_renderer_percent,
+				         DictionaryNumber(values, CFSTR("Renderer Utilization %")));
+				take_max(stats.gpu_tiler_percent,
+				         DictionaryNumber(values, CFSTR("Tiler Utilization %")));
+				add(stats.gpu_memory_in_use,
+				    DictionaryNumber(values, CFSTR("In use system memory")));
+				add(stats.gpu_memory_allocated,
+				    DictionaryNumber(values, CFSTR("Alloc system memory")));
+				add(stats.gpu_recoveries, DictionaryNumber(values, CFSTR("recoveryCount")));
+			}
+			CFRelease(statistics);
+		}
+		IOObjectRelease(service);
+	}
+	IOObjectRelease(iterator);
+	stats.gpu_model = state.gpu_model;
+	stats.gpu_cores = state.gpu_cores;
+}
+
+void SampleThreads(HostStats& stats, Counters& now) {
+	thread_act_array_t     threads = nullptr;
+	mach_msg_type_number_t count   = 0;
+	const auto             task    = mach_task_self();
+	if (task_threads(task, &threads, &count) != KERN_SUCCESS) {
+		return;
+	}
+	stats.thread_count = count;
+	for (mach_msg_type_number_t i = 0; i < count; i++) {
+		thread_identifier_info_data_t identifier {};
+		mach_msg_type_number_t        size = THREAD_IDENTIFIER_INFO_COUNT;
+		thread_basic_info_data_t      basic {};
+		mach_msg_type_number_t        basic_size = THREAD_BASIC_INFO_COUNT;
+		if (thread_info(threads[i], THREAD_IDENTIFIER_INFO,
+		                reinterpret_cast<thread_info_t>(&identifier), &size) == KERN_SUCCESS &&
+		    thread_info(threads[i], THREAD_BASIC_INFO, reinterpret_cast<thread_info_t>(&basic),
+		                &basic_size) == KERN_SUCCESS) {
+			const auto to_ns = [](const time_value_t& time) {
+				return static_cast<uint64_t>(time.seconds) * 1000000000u +
+				       static_cast<uint64_t>(time.microseconds) * 1000u;
+			};
+			auto& thread  = now.threads[identifier.thread_id];
+			thread.cpu_ns = to_ns(basic.user_time) + to_ns(basic.system_time);
+			thread_extended_info_data_t extended {};
+			mach_msg_type_number_t      extended_size = THREAD_EXTENDED_INFO_COUNT;
+			if (thread_info(threads[i], THREAD_EXTENDED_INFO,
+			                reinterpret_cast<thread_info_t>(&extended),
+			                &extended_size) == KERN_SUCCESS) {
+				thread.name = std::string(extended.pth_name,
+				                          strnlen(extended.pth_name, sizeof(extended.pth_name)));
+			}
+		}
+		mach_port_deallocate(task, threads[i]);
+	}
+	vm_deallocate(task, reinterpret_cast<vm_address_t>(threads), count * sizeof(thread_act_t));
+}
+
+void SamplePlatform(PlatformState& state, HostStats& stats, Counters& now) {
+	SampleRusage(now);
+
+	host_cpu_load_info_data_t load {};
+	mach_msg_type_number_t    load_count = HOST_CPU_LOAD_INFO_COUNT;
+	if (host_statistics(state.host, HOST_CPU_LOAD_INFO, reinterpret_cast<host_info_t>(&load),
+	                    &load_count) == KERN_SUCCESS) {
+		const uint64_t busy = static_cast<uint64_t>(load.cpu_ticks[CPU_STATE_USER]) +
+		                      load.cpu_ticks[CPU_STATE_SYSTEM] + load.cpu_ticks[CPU_STATE_NICE];
+		now.system_busy     = busy;
+		now.system_total    = busy + load.cpu_ticks[CPU_STATE_IDLE];
+	}
+
+	// Physical footprint is what Activity Monitor shows in its Memory column.
+	task_vm_info_data_t    vm_info {};
+	mach_msg_type_number_t vm_info_count = TASK_VM_INFO_COUNT;
+	if (task_info(mach_task_self(), TASK_VM_INFO, reinterpret_cast<task_info_t>(&vm_info),
+	              &vm_info_count) == KERN_SUCCESS) {
+		stats.process_memory      = vm_info.phys_footprint;
+		stats.process_resident    = vm_info.resident_size;
+		stats.process_compressed  = vm_info.compressed;
+		stats.process_file_backed = vm_info.external;
+		stats.process_virtual     = vm_info.virtual_size;
+	}
+	rusage_info_v2 usage {};
+	if (proc_pid_rusage(getpid(), RUSAGE_INFO_V2, reinterpret_cast<rusage_info_t*>(&usage)) == 0) {
+		now.disk_read  = usage.ri_diskio_bytesread;
+		now.disk_write = usage.ri_diskio_byteswritten;
+	}
+
+	stats.system_memory_total   = SysctlValue<uint64_t>("hw.memsize");
+	vm_size_t              page = 0;
+	vm_statistics64_data_t vm {};
+	mach_msg_type_number_t count = HOST_VM_INFO64_COUNT;
+	if (host_page_size(state.host, &page) == KERN_SUCCESS &&
+	    host_statistics64(state.host, HOST_VM_INFO64, reinterpret_cast<host_info64_t>(&vm),
+	                      &count) == KERN_SUCCESS) {
+		// Activity Monitor's "Memory Used": app memory + wired + compressed.
+		const uint64_t app       = vm.internal_page_count > vm.purgeable_count
+		                               ? vm.internal_page_count - vm.purgeable_count
+		                               : 0;
+		stats.system_memory_used = (app + vm.wire_count + vm.compressor_page_count) * page;
+		stats.system_wired       = static_cast<uint64_t>(vm.wire_count) * page;
+		stats.system_compressed  = static_cast<uint64_t>(vm.compressor_page_count) * page;
+		stats.system_cached =
+		    (static_cast<uint64_t>(vm.external_page_count) + vm.purgeable_count) * page;
+		stats.system_free = static_cast<uint64_t>(vm.free_count) * page;
+	}
+
+	if (const auto swap = SysctlValue<xsw_usage>("vm.swapusage")) {
+		stats.swap_used  = swap->xsu_used;
+		stats.swap_total = swap->xsu_total;
+	}
+
+	// 1 = normal, 2 = warning, 4 = critical, as reported to dispatch memory-pressure sources.
+	if (const auto level = SysctlValue<int>("kern.memorystatus_vm_pressure_level")) {
+		stats.memory_pressure = *level >= 4   ? MemoryPressure::Critical
+		                        : *level >= 2 ? MemoryPressure::Warning
+		                                      : MemoryPressure::Normal;
+	}
+
+	// 0 nominal, 1 moderate, 2 heavy, 3 trapping, 4 sleeping.
+	if (!state.thermal_ready) {
+		state.thermal_ready = notify_register_check("com.apple.system.thermalpressurelevel",
+		                                            &state.thermal_token) == NOTIFY_STATUS_OK;
+	}
+	uint64_t thermal = 0;
+	if (state.thermal_ready &&
+	    notify_get_state(state.thermal_token, &thermal) == NOTIFY_STATUS_OK) {
+		stats.thermal_pressure = thermal >= 3   ? ThermalPressure::Critical
+		                         : thermal == 2 ? ThermalPressure::Heavy
+		                         : thermal == 1 ? ThermalPressure::Moderate
+		                                        : ThermalPressure::Nominal;
+	}
+
+	SampleThreads(stats, now);
+	SampleGpu(state, stats);
+}
+
+#else
+
+using ProcValues = std::unordered_map<std::string, uint64_t>;
+
+// Reads "<key>: <number> [kB]" lines; kB values are returned in bytes.
+ProcValues ReadProcValues(const char* path) {
+	ProcValues    values;
+	std::ifstream file(path);
+	std::string   line;
+	while (std::getline(file, line)) {
+		const auto colon = line.find(':');
+		if (colon == std::string::npos) {
+			continue;
+		}
+		std::istringstream stream(line.substr(colon + 1));
+		uint64_t           number = 0;
+		std::string        unit;
+		if (stream >> number) {
+			stream >> unit;
+			values[line.substr(0, colon)] = unit == "kB" ? number * 1024u : number;
+		}
+	}
+	return values;
+}
+
+std::optional<uint64_t> Find(const ProcValues& values, const char* key) {
+	const auto it = values.find(key);
+	return it != values.end() ? std::optional(it->second) : std::nullopt;
+}
+
+struct PlatformState {};
+
+void SampleThreads(HostStats& stats, Counters& now) {
+	static const double ns_per_tick = 1e9 / static_cast<double>(std::max(sysconf(_SC_CLK_TCK), 1L));
+	std::error_code     error;
+	uint32_t            count = 0;
+	for (std::filesystem::directory_iterator it("/proc/self/task", error), end; !error && it != end;
+	     it.increment(error)) {
+		count++;
+		std::ifstream stat_file(it->path() / "stat");
+		std::string   stat;
+		uint64_t      id = 0;
+		if (!(stat_file >> id)) {
+			continue; // The thread may have exited during enumeration.
+		}
+		std::getline(stat_file, stat);
+		const auto close = stat.rfind(')');
+		if (close == std::string::npos) {
+			continue;
+		}
+		// Fields after the command name start at field 3 (state); utime and stime are 14 and 15.
+		std::istringstream fields(stat.substr(close + 2));
+		std::string        field;
+		for (int index = 3; index < 14 && fields >> field; index++) {
+		}
+		uint64_t user   = 0;
+		uint64_t system = 0;
+		if (!(fields >> user >> system)) {
+			continue;
+		}
+		auto& thread  = now.threads[id];
+		thread.cpu_ns = static_cast<uint64_t>(static_cast<double>(user + system) * ns_per_tick);
+		std::ifstream name_file(it->path() / "comm");
+		std::getline(name_file, thread.name);
+	}
+	if (!error) {
+		stats.thread_count = count;
+	}
+}
+
+void SamplePlatform(PlatformState& /*state*/, HostStats& stats, Counters& now) {
+	SampleRusage(now);
+
+	std::ifstream cpu_file("/proc/stat");
+	std::string   label;
+	if (cpu_file >> label && label == "cpu") {
+		uint64_t total = 0;
+		uint64_t idle  = 0;
+		uint64_t value = 0;
+		for (int index = 0; index < 8 && cpu_file >> value; index++) {
+			total += value;
+			if (index == 3 || index == 4) { // idle, iowait
+				idle += value;
+			}
+		}
+		now.system_total = total;
+		now.system_busy  = total - idle;
+	}
+
+	const auto status         = ReadProcValues("/proc/self/status");
+	stats.process_memory      = Find(status, "VmRSS");
+	stats.process_resident    = Find(status, "VmRSS");
+	stats.process_file_backed = Find(status, "RssFile");
+	stats.process_compressed  = Find(status, "VmSwap");
+	stats.process_virtual     = Find(status, "VmSize");
+
+	const auto io  = ReadProcValues("/proc/self/io");
+	now.disk_read  = Find(io, "read_bytes");
+	now.disk_write = Find(io, "write_bytes");
+
+	const auto memory    = ReadProcValues("/proc/meminfo");
+	const auto total     = Find(memory, "MemTotal");
+	const auto available = Find(memory, "MemAvailable");
+	if (total && available && *available <= *total) {
+		stats.system_memory_total = total;
+		stats.system_memory_used  = *total - *available;
+	}
+	stats.system_free = Find(memory, "MemFree");
+	if (const auto cached = Find(memory, "Cached")) {
+		stats.system_cached = *cached + Find(memory, "Buffers").value_or(0);
+	}
+	const auto swap_total = Find(memory, "SwapTotal");
+	const auto swap_free  = Find(memory, "SwapFree");
+	if (swap_total && swap_free && *swap_free <= *swap_total) {
+		stats.swap_total = swap_total;
+		stats.swap_used  = *swap_total - *swap_free;
+	}
+
+	// Pressure stall information: share of the last 10 s some task waited on memory.
+	std::ifstream pressure("/proc/pressure/memory");
+	std::string   kind;
+	std::string   average;
+	double        stalled = 0.0;
+	if (pressure >> kind >> average && kind == "some" && average.rfind("avg10=", 0) == 0 &&
+	    std::istringstream(average.substr(6)) >> stalled) {
+		stats.memory_pressure = stalled >= 20.0  ? MemoryPressure::Critical
+		                        : stalled >= 5.0 ? MemoryPressure::Warning
+		                                         : MemoryPressure::Normal;
+	} else if (total && available && *total != 0) {
+		const double free_share = static_cast<double>(*available) / static_cast<double>(*total);
+		stats.memory_pressure   = free_share < 0.10   ? MemoryPressure::Critical
+		                          : free_share < 0.25 ? MemoryPressure::Warning
+		                                              : MemoryPressure::Normal;
+	}
+
+	SampleThreads(stats, now);
+}
+
+#endif
+#endif
+
+constexpr size_t BUSIEST_THREADS = 6;
+
+} // namespace
+
+struct HostStatsSampler::Impl {
+	PlatformState state;
+	Counters      last;
+	uint32_t      logical_cores = std::max(std::thread::hardware_concurrency(), 1u);
+	bool          has_last      = false;
+};
+
+HostStatsSampler::HostStatsSampler(): m_impl(std::make_unique<Impl>()) {}
+
+HostStatsSampler::~HostStatsSampler() = default;
+
+HostStats HostStatsSampler::Sample() {
+	HostStats stats;
+	stats.logical_cores = m_impl->logical_cores;
+
+	Counters now;
+	now.wall_ns = WallNanoseconds();
+	SamplePlatform(m_impl->state, stats, now);
+
+	const auto& last    = m_impl->last;
+	const auto  elapsed = static_cast<double>(now.wall_ns - last.wall_ns) / 1e9;
+	if (m_impl->has_last && elapsed > 0.0) {
+		const auto rate = [&](std::optional<uint64_t> current,
+		                      std::optional<uint64_t> previous) -> std::optional<double> {
+			if (!current || !previous || *current < *previous) {
+				return std::nullopt;
+			}
+			return static_cast<double>(*current - *previous) / elapsed;
+		};
+		if (const auto cpu = rate(now.process_cpu_ns, last.process_cpu_ns)) {
+			stats.process_cpu_percent =
+			    std::clamp(*cpu / 1e9 * 100.0 / stats.logical_cores, 0.0, 100.0);
+		}
+		if (now.system_total && last.system_total && now.system_busy && last.system_busy &&
+		    *now.system_total > *last.system_total && *now.system_busy >= *last.system_busy) {
+			stats.system_cpu_percent =
+			    std::clamp(static_cast<double>(*now.system_busy - *last.system_busy) /
+			                   static_cast<double>(*now.system_total - *last.system_total) * 100.0,
+			               0.0, 100.0);
+		}
+		stats.context_switches_per_s = rate(now.context_switches, last.context_switches);
+		stats.page_faults_per_s      = rate(now.page_faults, last.page_faults);
+		stats.page_ins_per_s         = rate(now.page_ins, last.page_ins);
+		stats.disk_read_per_s        = rate(now.disk_read, last.disk_read);
+		stats.disk_write_per_s       = rate(now.disk_write, last.disk_write);
+
+		for (const auto& [id, thread]: now.threads) {
+			// A thread missing from the last sample started during this interval.
+			const auto     previous = last.threads.find(id);
+			const uint64_t before   = previous != last.threads.end() ? previous->second.cpu_ns : 0;
+			if (thread.cpu_ns < before) {
+				continue;
+			}
+			const double percent =
+			    static_cast<double>(thread.cpu_ns - before) / 1e9 / elapsed * 100.0;
+			stats.busiest_threads.push_back(
+			    {thread.name.empty() ? "thread " + std::to_string(id) : thread.name, percent});
+		}
+		std::sort(stats.busiest_threads.begin(), stats.busiest_threads.end(),
+		          [](const auto& a, const auto& b) { return a.cpu_percent > b.cpu_percent; });
+		if (stats.busiest_threads.size() > BUSIEST_THREADS) {
+			stats.busiest_threads.resize(BUSIEST_THREADS);
+		}
+	}
+	m_impl->last     = std::move(now);
+	m_impl->has_last = true;
+	return stats;
+}
+
+} // namespace Common

src/common/hostStats.h

@@ -0,0 +1,86 @@
+#ifndef KYTY_COMMON_HOST_STATS_H_
+#define KYTY_COMMON_HOST_STATS_H_
+
+#include <cstdint>
+#include <memory>
+#include <optional>
+#include <string>
+#include <vector>
+
+namespace Common {
+
+enum class MemoryPressure { Unknown, Normal, Warning, Critical };
+enum class ThermalPressure { Unknown, Nominal, Moderate, Heavy, Critical };
+
+struct ThreadUsage {
+	std::string name;
+	double      cpu_percent = 0.0; // Of one core.
+};
+
+// Host-wide and process-wide resource usage. Missing values are std::nullopt; rates are per
+// second since the previous sample.
+struct HostStats {
+	uint32_t logical_cores = 0;
+
+	// Processing.
+	std::optional<double>    process_cpu_percent; // Share of all logical cores, 0-100.
+	std::optional<double>    system_cpu_percent;  // Whole machine, 0-100.
+	std::optional<uint32_t>  thread_count;
+	std::vector<ThreadUsage> busiest_threads; // Busiest first.
+	std::optional<double>    context_switches_per_s;
+	std::optional<double>    page_faults_per_s;
+	std::optional<double>    page_ins_per_s; // Faults that read from disk.
+	ThermalPressure          thermal_pressure = ThermalPressure::Unknown;
+
+	// Emulator process memory.
+	std::optional<uint64_t> process_memory; // Physical footprint / resident set.
+	std::optional<uint64_t> process_resident;
+	std::optional<uint64_t> process_compressed; // Compressed or swapped out.
+	std::optional<uint64_t> process_file_backed;
+	std::optional<uint64_t> process_virtual; // Reserved address space.
+
+	// System memory.
+	std::optional<uint64_t> system_memory_used;
+	std::optional<uint64_t> system_memory_total;
+	std::optional<uint64_t> system_wired;
+	std::optional<uint64_t> system_compressed;
+	std::optional<uint64_t> system_cached; // File cache, reclaimable.
+	std::optional<uint64_t> system_free;
+	std::optional<uint64_t> swap_used;
+	std::optional<uint64_t> swap_total;
+	MemoryPressure          memory_pressure = MemoryPressure::Unknown;
+
+	// Disk I/O of the emulator process.
+	std::optional<double> disk_read_per_s;
+	std::optional<double> disk_write_per_s;
+
+	// GPU, whole device (macOS IOAccelerator).
+	std::string             gpu_model;
+	std::optional<uint32_t> gpu_cores;
+	std::optional<double>   gpu_percent;
+	std::optional<double>   gpu_renderer_percent;
+	std::optional<double>   gpu_tiler_percent;
+	std::optional<uint64_t> gpu_memory_in_use; // System memory the GPU driver has in use.
+	std::optional<uint64_t> gpu_memory_allocated;
+	std::optional<uint64_t> gpu_recoveries; // GPU resets since boot.
+};
+
+// May block for a few milliseconds (thread enumeration, OS registry reads), so call it off the
+// render thread. The first sample has no rates.
+class HostStatsSampler final {
+public:
+	HostStatsSampler();
+	~HostStatsSampler();
+	HostStatsSampler(const HostStatsSampler&)            = delete;
+	HostStatsSampler& operator=(const HostStatsSampler&) = delete;
+
+	[[nodiscard]] HostStats Sample();
+
+private:
+	struct Impl;
+	std::unique_ptr<Impl> m_impl;
+};
+
+} // namespace Common
+
+#endif /* KYTY_COMMON_HOST_STATS_H_ */

src/common/platform/sysLinuxFileIO.cpp

@@ -367,7 +367,7 @@ bool SysFileDeleteFile(const std::filesystem::path& name) {
 
 bool SysFileFlush(sys_file_t& f) {
 	if (f.type == SYS_FILE_FILE && f.f != nullptr) {
-		return (fflush(f.f) == 0);
+		return fflush(f.f) == 0 && fsync(fileno(f.f)) == 0;
 	}
 
 	return false;

src/emulator.cpp

@@ -89,42 +89,6 @@ static void MountSandboxDirs() {
 	MountOrCreateDir("_TempData/" + title_id, "/temp");
 }
 
-static bool ClearDirectoryContents(const std::filesystem::path& dir) {
-	bool ok = true;
-
-	for (const auto& entry: Common::File::GetDirEntries(dir)) {
-		if (entry.name == "." || entry.name == "..") {
-			continue;
-		}
-
-		auto path = dir / entry.name;
-
-		if (entry.is_file) {
-			Common::File::RemoveReadonly(path);
-			ok = Common::File::DeleteFile(path) && ok;
-		} else {
-			ok = ClearDirectoryContents(path) && ok;
-			ok = Common::File::DeleteDirectory(path) && ok;
-		}
-	}
-
-	return ok;
-}
-
-static void ClearDebugTextureFolder() {
-	const std::string debug_texture_folder = "_Textures";
-
-	if (!Common::File::IsDirectoryExisting(debug_texture_folder)) {
-		Common::File::CreateDirectories(debug_texture_folder);
-		return;
-	}
-
-	if (!ClearDirectoryContents(debug_texture_folder)) {
-		LOGF_COLOR(Log::Color::BrightYellow, "TextureDump: failed to completely clear %s\n",
-		           debug_texture_folder.c_str());
-	}
-}
-
 static void Init(const Config::ConfigOptions& cfg, const std::filesystem::path& param_json,
                  Common::Subsystems& subsystems) {
 	EXIT_IF(!Common::Thread::IsMainThread());
@@ -198,8 +162,6 @@ void Run(const RunOptions& options) {
 	Common::Subsystems subsystems(true);
 	Init(options.config, param_json, subsystems);
 
-	ClearDebugTextureFolder();
-
 	PrintSystemInfo();
 	std::string title_id;
 	Loader::SystemContentParamSfoGetString("TITLE_ID", &title_id);

src/graphics/guest_gpu/graphicsRun.cpp

@@ -1,6 +1,7 @@
 #include "graphics/guest_gpu/graphicsRun.h"
 
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -342,12 +343,11 @@ void CommandProcessor::WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw
 		return;
 	}
 
-	if (write_one_address) {
-		for (uint32_t i = 0; i < dw_num; i++) {
-			dst[0] = src[i];
-		}
-	} else {
-		memcpy(dst, src, static_cast<size_t>(dw_num) * sizeof(uint32_t));
+	// Completion labels can be reused immediately; memcpy may write a destination twice.
+	auto* output = static_cast<volatile uint32_t*>(dst);
+	for (uint32_t i = 0; i < dw_num; i++) {
+		*output = src[i];
+		output += !write_one_address;
 	}
 }
 
@@ -906,6 +906,7 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 	if (draw_count == 0) {
 		return;
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::IndirectDraws, draw_count);
 
 	const auto args_size = indexed ? sizeof(DrawIndexedIndirectArgs) : sizeof(DrawIndirectArgs);
 	EXIT_NOT_IMPLEMENTED(stride_in_bytes < args_size);

src/graphics/host_gpu/graphicContext.h

@@ -8,6 +8,7 @@
 
 #include <map>
 #include <mutex>
+#include <span>
 #include <tuple>
 #include <vector>
 #include <vk_mem_alloc.h>
@@ -99,6 +100,14 @@ struct GraphicContext {
 	[[nodiscard]] bool CanReportMemoryUsage() const noexcept { return memory_budget_ext_enabled; }
 	[[nodiscard]] uint64_t GetDeviceMemoryUsage() const;
 	[[nodiscard]] uint64_t GetTotalMemoryBudget() const;
+
+	struct HeapUsage {
+		uint64_t usage        = 0; // Whole process, as reported by VK_EXT_memory_budget.
+		uint64_t budget       = 0;
+		uint32_t allocations  = 0;
+		bool     device_local = false;
+	};
+	[[nodiscard]] uint32_t GetHeapUsage(std::span<HeapUsage, VK_MAX_MEMORY_HEAPS> heaps) const;
 	[[nodiscard]] bool     CreateImage(const vk::ImageCreateInfo& info, VulkanImage& image);
 	void                   DeleteImage(VulkanImage& image);
 

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

@@ -2,6 +2,7 @@
 
 #include "common/alignment.h"
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
 #include "graphics/guest_gpu/graphicsRun.h"
@@ -177,6 +178,7 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 	if (copies.empty()) {
 		return false;
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::BufferDownloadBytes, total_size);
 
 	auto [mapped, offset] = m_download_buffer.Map(total_size, 64);
 	std::unique_ptr<Buffer> temporary;
@@ -467,6 +469,7 @@ vk::Buffer BufferCache::UploadCopies(Buffer& buffer, std::span<vk::BufferCopy> c
 	if (copies.empty()) {
 		return nullptr;
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::BufferUploadBytes, total_size);
 
 	const auto source = UploadMemory(m_graphics, m_scheduler, m_staging_buffer, total_size, 4,
 	    vk::BufferUsageFlagBits::eTransferSrc, [&](uint8_t* mapped, uint64_t base_offset) {
@@ -658,6 +661,7 @@ void BufferCache::RunGarbageCollector() {
 		} else {
 			m_memory_tracker.UntrackMemory(buffer.CpuAddress(), buffer.Size());
 			DeleteBuffer(id);
+			Common::DebugCounters::Add(Common::DebugCounters::Counter::BufferEvictions);
 		}
 		return ++retire_count == limit;
 	});

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

@@ -2,6 +2,7 @@
 
 #include "common/alignment.h"
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -262,14 +263,8 @@ void TextureCache::DeleteImage(ImageId id) {
 		EXIT("TextureCache: deleting a GPU-modified image without resolving its contents\n");
 	}
 	m_download_images.erase(id);
-	if (image->info.HasMetadata()) {
-		const auto metadata = m_surface_metas.find(image->info.metadata.range.address);
-		if (metadata != m_surface_metas.end() &&
-		    image->info.metadata.kind == ImageMetadataKind::Htile &&
-		    metadata->second.type == MetaDataInfo::Type::HTile) {
-			// A later binding may have reused this address for another metadata type.
-			m_surface_metas.erase(metadata);
-		}
+	if (image->info.metadata.kind == ImageMetadataKind::Htile) {
+		m_surface_metas.erase(image->info.metadata.range.address);
 	}
 	UnregisterImage(id);
 	if (m_scheduler.Active()) {
@@ -954,6 +949,7 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		const auto linear = m_tiler.Detile(source.Handle(), source_offset, info);
 		const auto copies = info.BufferCopies(linear.offset);
 		destination.Upload(copies, linear.buffer, linear.offset, linear.size);
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureUploadBytes, linear.size);
 		return;
 	}
 
@@ -999,6 +995,7 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		copies[0].imageExtent       = info.extent;
 	}
 	destination.Upload(copies, linear.buffer, linear.offset, linear.size);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureUploadBytes, linear.size);
 }
 
 void TextureCache::InitializeImage(ImageId id) {
@@ -1039,7 +1036,7 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 		std::scoped_lock lock {m_lock};
 		auto& image         = m_slot_images[id];
 		image.info.metadata = desc.info.metadata;
-		// Native color metadata must not retain a reused HTile/CMask/FMask clear flag.
+		// Native color metadata must not retain a reused HTile clear flag.
 		m_surface_metas.erase(range.address);
 		if (range.size == 0 || desc.info.resources.levels != 1) {
 			return;
@@ -1417,11 +1414,8 @@ vk::ImageView TextureCache::FindDepthTarget(ImageId id, const ImageDesc& desc) {
 	image.info.metadata = desc.info.metadata;
 	if (desc.info.HasMetadata()) {
 		const auto [metadata, inserted] = m_surface_metas.try_emplace(desc.info.metadata.range.address);
-		if (inserted) {
-			metadata->second.type = MetaDataInfo::Type::HTile;
-			if (image.info.htile_initial_clear) {
-				metadata->second.clear_mask.set();
-			}
+		if (inserted && image.info.htile_initial_clear) {
+			metadata->second.clear_mask.set();
 		}
 	}
 	RefreshImage(id);
@@ -1774,6 +1768,7 @@ bool TextureCache::DownloadImageMemory(ImageId id) {
 	download.Flush(offset, range.size);
 
 	DownloadImage(image, download, offset, range.size, image.info.resources.levels);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureDownloadBytes, range.size);
 	vk::BufferMemoryBarrier barrier {};
 	barrier.srcAccessMask = vk::AccessFlagBits::eMemoryWrite | vk::AccessFlagBits::eTransferWrite |
 	                        vk::AccessFlagBits::eShaderWrite;
@@ -1943,6 +1938,7 @@ void TextureCache::RunGarbageCollector() {
 				}
 			}
 			FreeImage(id);
+			Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureEvictions);
 			if (m_total_used_memory < m_critical_gc_memory && aggressive) {
 				deletions >>= 2;
 				aggressive = false;

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

@@ -81,9 +81,6 @@ public:
 
 private:
 	struct MetaDataInfo {
-		enum class Type : uint8_t { CMask, FMask, HTile };
-
-		Type type;
 		// PS5 DB_DEPTH_VIEW encodes 13-bit base and last array-slice indices.
 		std::bitset<8192> clear_mask;
 	};

src/graphics/host_gpu/renderer/colorRenderTarget.cpp

@@ -72,7 +72,16 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		return;
 	}
 	const auto samples = render_sample_count(rt.attrib.num_fragments);
-	if (samples == 0 || rt.attrib.num_samples != rt.attrib.num_fragments) {
+	// TODO(PPSA24156): Replace this 2x MSAA approximation with proper 4-coverage/2-fragment EQAA.
+	const bool approximate_eqaa = rt.attrib.num_samples == 2 && rt.attrib.num_fragments == 1;
+	if (approximate_eqaa) {
+		static std::atomic_bool logged = false;
+		if (!logged.exchange(true, std::memory_order_relaxed)) {
+			Log::WriteToConsoleAndLog(
+			    "Warning: approximating 4-coverage/2-fragment EQAA with 2x MSAA.\n");
+		}
+	}
+	if (samples == 0 || (rt.attrib.num_samples != rt.attrib.num_fragments && !approximate_eqaa)) {
 		EXIT("unsupported render-target sample configuration: samples=%u fragments=%u\n",
 		     rt.attrib.num_samples, rt.attrib.num_fragments);
 	}
@@ -220,8 +229,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 	}
 	const bool has_dcc        = rt.info.dcc_compression_enable && rt.dcc_addr.addr != 0;
 	const bool has_cmask      = !rt.info.dcc_compression_enable && rt.info.cmask_fast_clear_enable &&
-	                            rt.cmask.addr != 0 && samples == 1 &&
-	                            !rt.info.fmask_compression_enable &&
+	                            rt.cmask.addr != 0 &&
 	                            !rt.attrib3.write_vrs_rate_hint_to_cmask;
 	if (has_dcc || has_cmask) {
 		TileSizeAlign metadata_size {};
@@ -233,7 +241,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		} else {
 			(void)TileGetCmaskSize(width, height, layers, levels, metadata_size);
 		}
-		// DCC owns the clear when both planes are enabled; single-sample CMASK stays expanded.
+		// DCC owns the clear when both planes are enabled.
 		desc.info.metadata.kind = has_dcc ? ImageMetadataKind::Dcc : ImageMetadataKind::Cmask;
 		desc.info.metadata.range = {has_dcc ? rt.dcc_addr.addr : rt.cmask.addr, metadata_size.size};
 		desc.info.metadata.clear_word =

src/graphics/host_gpu/renderer/commandScheduler.cpp

@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/commandScheduler.h"
 
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/logging/log.h"
 #include "graphics/host_gpu/graphicContext.h"
 
@@ -373,6 +374,7 @@ uint64_t CommandScheduler::Submit(SubmitInfo submit) {
 
 		result = graphics.queue.submit(1, &submit_info, nullptr);
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::QueueSubmits);
 
 	if (result != vk::Result::eSuccess) {
 		ReportVulkanFatal("vkQueueSubmit", result, tick, m_command.m_debug_op,

src/graphics/host_gpu/renderer/debug.cpp

@@ -294,16 +294,6 @@ static void RtCheck(const HW::RenderTarget& rt) {
 		EXIT_NOT_IMPLEMENTED(rt.dcc.overwrite_combiner_disable != false);
 		// EXIT_NOT_IMPLEMENTED(rt.dcc.data_write_on_dcc_clear_to_reg != false);
 		EXIT_NOT_IMPLEMENTED(rt.dcc.dcc_clear_key_enable != false);
-		if (rt.cmask.addr != 0x0000000000000000 || rt.fmask.addr != 0x0000000000000000 ||
-		    rt.dcc_addr.addr != 0x0000000000000000) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("RenderTarget: temporary: ignoring PS5 metadata addresses cmask=0x%016" PRIx64
-				     " fmask=0x%016" PRIx64 " dcc=0x%016" PRIx64 "\n",
-				     rt.cmask.addr, rt.fmask.addr, rt.dcc_addr.addr);
-				logged = true;
-			}
-		}
 	}
 }
 

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

@@ -353,8 +353,12 @@ void Image::CopyImage(Image& source) {
 	std::vector<vk::ImageCopy> copies;
 	copies.reserve(levels);
 	for (uint32_t level = 0; level < levels; level++) {
-		const auto width  = std::max(source.backing.extent.width >> level, 1u);
-		const auto height = std::max(source.backing.extent.height >> level, 1u);
+		// Overlapping cache images can differ in size; copying the source extent into a smaller
+		// destination writes out of bounds and loses the device.
+		const auto width  = std::min(std::max(source.backing.extent.width >> level, 1u),
+		                             std::max(backing.extent.width >> level, 1u));
+		const auto height = std::min(std::max(source.backing.extent.height >> level, 1u),
+		                             std::max(backing.extent.height >> level, 1u));
 		const auto depth  = std::max(base_depth >> level, 1u);
 		const auto [source_layers, destination_layers] = SanitizeCopyLayers(source, *this, depth);
 		vk::ImageCopy copy {};

src/graphics/host_gpu/renderer/masterSemaphore.cpp

@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/masterSemaphore.h"
 
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "graphics/host_gpu/graphicContext.h"
 
 namespace Libs::Graphics {
@@ -49,6 +50,8 @@ void MasterSemaphore::Wait(uint64_t tick) {
 	wait_info.pSemaphores    = &m_semaphore;
 	wait_info.pValues        = &tick;
 
+	Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::GpuWaits,
+	                                         Common::DebugCounters::Counter::GpuWaitNs);
 	const auto result = m_graphics.device.waitSemaphores(&wait_info, UINT64_MAX);
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 	Refresh();

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

@@ -764,7 +764,10 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 					size = std::min({size, uint64_t {256} * 1024 * 1024, limit});
 				}
 			}
-			size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
+			// Finite GPU descriptors can span sparse PRT holes; clamp only oversized/sentinel ranges.
+			if (size >= (uint64_t {1} << 30u) || descriptor.NumRecords() == UINT32_MAX) {
+				size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
+			}
 			source = {address, size, cache.FindBuffer(address, size)};
 		}
 	}

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

@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
@@ -24,6 +25,7 @@
 #include <array>
 #include <atomic>
 #include <cctype>
+#include <chrono>
 #include <cstdio>
 #include <cstring>
 #include <fmt/format.h>
@@ -273,6 +275,7 @@ struct PipelineCache::ProgramCache {
 		ShaderRecompiler::IR::ResourceSnapshot       resources;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		std::vector<Permutation>                    permutations;
+		bool                                       skip_dispatch = false;
 	};
 
 	struct ProgramKeyHash {
@@ -299,6 +302,9 @@ struct PipelineCache::ProgramCache {
 	                               ShaderRecompiler::IR::ResourceSpecialization specialization,
 	                               uint32_t push_data_start_dword,
 	                               std::span<const uint32_t> function_code) {
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::ShadersCompiled,
+		                                         Common::DebugCounters::Counter::ShaderCompileNs,
+		                                         Common::DebugCounters::Peak::ShaderCompileNs);
 		auto result = ShaderRecompiler::CompileProgram(std::move(translated), options,
 		                                               specialization, push_data_start_dword);
 		if (!ValidateShaderSpirv(options.dump_label, options.shader_hash, result.spirv,
@@ -355,6 +361,7 @@ struct PipelineCache::ProgramCache {
 		bool source_changed = false;
 		if (entry != programs.end()) {
 			auto& cached = entry->second;
+			if (cached.skip_dispatch) return {};
 			Permutation* permutation = nullptr;
 			if (cached.call_source) {
 				source_changed = cached.RefreshCallSource(runtime);
@@ -414,10 +421,18 @@ struct PipelineCache::ProgramCache {
 		DumpShaderOriginal(stage_name, options.shader_hash, params.code);
 		ShaderRecompiler::ShaderSource source;
 		if (entry == programs.end()) source = ShaderRecompiler::PrepareShaderSource(params.code, options);
+		if (source.unsupported_call_pc != UINT32_MAX) {
+			PipelineCacheLog("!!! TEMPORARY COMPUTE DISPATCH SKIP !!! shader=0x{:016x} pc=0x{:08x} "
+			                 "reason=unsupported GPU-selected callable target; rendering is incomplete",
+			                 params.hash, source.unsupported_call_pc);
+			programs.try_emplace(lookup_key, ShaderRecompiler::IR::ResourcePlan{}, nullptr)
+			    .first->second.skip_dispatch = true;
+			return {};
+		}
 		auto* current = entry == programs.end() ? &source : entry->second.call_source.get();
 		ShaderRecompiler::TranslateResult translated;
 		if (current != nullptr) {
-			if (current->call) {
+			if (!current->calls.empty()) {
 				if (entry == programs.end()) ShaderRecompiler::RefreshShaderSource(*current, runtime);
 				translated = ShaderRecompiler::TranslateProgram(current->linked->decoded, options);
 				translated.program.source_reads = current->reads;
@@ -428,7 +443,7 @@ struct PipelineCache::ProgramCache {
 			translated = ShaderRecompiler::TranslateProgram(params.code, options);
 		}
 		if (entry == programs.end()) {
-			auto retained = source.call ? std::make_unique<ShaderRecompiler::ShaderSource>(std::move(source)) : nullptr;
+			auto retained = !source.calls.empty() ? std::make_unique<ShaderRecompiler::ShaderSource>(std::move(source)) : nullptr;
 			entry = programs.try_emplace(lookup_key,
 			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program), std::move(retained)).first;
 			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
@@ -575,8 +590,8 @@ void PipelineCache::InitializeDriverCache() {
 		return;
 	}
 	if (!initial_data.empty()) {
-		PipelineCacheLog("Vulkan pipeline cache: loaded {} bytes from {}", initial_data.size(),
-		                 path);
+		PipelineCacheLog("Vulkan pipeline cache: loaded {:.2f} MB from {}",
+		                 static_cast<double>(initial_data.size()) / 1'000'000.0, path);
 	} else {
 		PipelineCacheLog("Vulkan pipeline cache: initialized empty");
 	}
@@ -616,8 +631,8 @@ void PipelineCache::Save() {
 	}
 	if (result != vk::Result::eSuccess || size == 0 ||
 	    size > std::numeric_limits<uint32_t>::max()) {
-		PipelineCacheLog("Vulkan pipeline cache: save failed ({}, {} bytes)",
-		                 vk::to_string(result), size);
+		PipelineCacheLog("Vulkan pipeline cache: save failed ({}, {:.2f} MB)",
+		                 vk::to_string(result), static_cast<double>(size) / 1'000'000.0);
 		return;
 	}
 	payload.resize(size);
@@ -649,7 +664,8 @@ void PipelineCache::Save() {
 		                 Common::PathToString(m_driver_cache_path));
 		return;
 	}
-	PipelineCacheLog("Vulkan pipeline cache: saved {} bytes to {}", payload.size(),
+	PipelineCacheLog("Vulkan pipeline cache: saved {:.2f} MB to {}",
+	                 static_cast<double>(payload.size()) / 1'000'000.0,
 	                 Common::PathToString(m_driver_cache_path));
 	m_driver_cache_dirty = false;
 }
@@ -993,8 +1009,13 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 
 	auto cached = std::make_unique<Pipeline>();
 	LogPipelineTrace("CreatePipelineInternal begin", vs_id, ps_id);
-	CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vertex_info,
-	                       ps_input_info, programs, static_params, m_driver_cache);
+	{
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::PipelinesCreated,
+		                                         Common::DebugCounters::Counter::PipelineCreateNs,
+		                                         Common::DebugCounters::Peak::PipelineCreateNs);
+		CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vertex_info,
+		                       ps_input_info, programs, static_params, m_driver_cache);
+	}
 	LogPipelineTrace("CreatePipelineInternal done", vs_id, ps_id);
 
 	EXIT_NOT_IMPLEMENTED(cached->pipeline == nullptr);
@@ -1024,7 +1045,13 @@ PipelineCache::GetComputePipeline(const ShaderComputeInputInfo& input_info,
 	}
 
 	auto cached = std::make_unique<Pipeline>();
-	CreatePipelineInternal(m_graphics, *cached, input_info, compute_program.module, m_driver_cache);
+	{
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::PipelinesCreated,
+		                                         Common::DebugCounters::Counter::PipelineCreateNs,
+		                                         Common::DebugCounters::Peak::PipelineCreateNs);
+		CreatePipelineInternal(m_graphics, *cached, input_info, compute_program.module,
+		                       m_driver_cache);
+	}
 
 	EXIT_NOT_IMPLEMENTED(cached->pipeline == nullptr);
 	EXIT_NOT_IMPLEMENTED(cached->pipeline_layout == nullptr);

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

@@ -196,21 +196,27 @@ static void CreateDescriptorLayout(GraphicContext& graphics, PipelineCache::Pipe
 	const auto& limits = graphics.GetPhysicalDeviceProperties().limits;
 	uint64_t descriptor_count = 0;
 	uint64_t storage_buffers = 0;
-	std::array<uint64_t, 32> stage_buffers {}, stage_resources {};
+	uint64_t samplers = 0;
+	std::array<uint64_t, 32> stage_buffers {}, stage_resources {}, stage_samplers {};
 	for (const auto& binding: bindings) {
 		descriptor_count += binding.descriptorCount;
 		const bool storage = binding.descriptorType == vk::DescriptorType::eStorageBuffer;
+		const bool sampler = binding.descriptorType == vk::DescriptorType::eSampler;
 		if (storage) storage_buffers += binding.descriptorCount;
+		if (sampler) samplers += binding.descriptorCount;
 		for (auto stages = static_cast<uint32_t>(binding.stageFlags); stages != 0; stages &= stages - 1u) {
 			const auto stage = std::countr_zero(stages);
 			if (storage) stage_buffers[stage] += binding.descriptorCount;
-			if (binding.descriptorType != vk::DescriptorType::eSampler)
+			if (sampler) stage_samplers[stage] += binding.descriptorCount;
+			else
 				stage_resources[stage] += binding.descriptorCount;
 			if (stage_buffers[stage] > limits.maxPerStageDescriptorStorageBuffers ||
-			    stage_resources[stage] > limits.maxPerStageResources) {
+			    stage_resources[stage] > limits.maxPerStageResources ||
+			    stage_samplers[stage] > limits.maxPerStageDescriptorSamplers) {
 				EXIT("native descriptor stage %u exceeds device limits: buffers=%" PRIu64
-				     "/%u resources=%" PRIu64 "/%u\n", stage, stage_buffers[stage],
-				     limits.maxPerStageDescriptorStorageBuffers, stage_resources[stage], limits.maxPerStageResources);
+				     "/%u resources=%" PRIu64 "/%u samplers=%" PRIu64 "/%u\n", stage, stage_buffers[stage],
+				     limits.maxPerStageDescriptorStorageBuffers, stage_resources[stage], limits.maxPerStageResources,
+				     stage_samplers[stage], limits.maxPerStageDescriptorSamplers);
 			}
 		}
 	}
@@ -218,6 +224,10 @@ static void CreateDescriptorLayout(GraphicContext& graphics, PipelineCache::Pipe
 		EXIT("native storage-buffer descriptors exceed device limit: %" PRIu64 "/%u\n",
 		     storage_buffers, limits.maxDescriptorSetStorageBuffers);
 	}
+	if (samplers > limits.maxDescriptorSetSamplers) {
+		EXIT("native sampler descriptors exceed device limit: %" PRIu64 "/%u\n",
+		     samplers, limits.maxDescriptorSetSamplers);
+	}
 	pipeline.uses_push_descriptors = descriptor_count <= graphics.max_push_descriptors;
 
 	vk::DescriptorSetLayoutCreateInfo create {};

src/graphics/host_gpu/renderer/render.h

@@ -201,11 +201,11 @@ private:
 	void                      BindImage(ImageId id, bool storage);
 	void                      BindRenderTarget(ImageId id);
 	void                      ResetBindings();
-	void                     UpdateComputeMetaClears(const ShaderComputeInputInfo& input,
-	                                                 const CommandBuffer&          buffer);
-	[[nodiscard]] bool TryConsumeComputeImageClear(const ShaderComputeInputInfo& input,
-	                                              CommandBuffer& command, uint32_t group_x,
-	                                              uint32_t group_y, uint32_t group_z, uint32_t mode);
+	[[nodiscard]] bool        TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
+	                                                     const CommandBuffer&          buffer);
+	[[nodiscard]] bool TryConsumeComputeFill(const ShaderComputeInputInfo& input,
+	                                        CommandBuffer& command, uint32_t group_x,
+	                                        uint32_t group_y, uint32_t group_z, uint32_t mode);
 
 	RenderContext&                        m_context;
 	GraphicsBindings                     m_graphics_bindings;

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -1,5 +1,6 @@
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
@@ -50,8 +51,8 @@ static bool FillSourcesDisjoint(std::span<const ShaderRecompiler::IR::Descriptor
 	return true;
 }
 
-void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input,
-                                            const CommandBuffer&          buffer) {
+bool RenderExecutor::TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
+                                                const CommandBuffer&          buffer) {
 	const auto& program   = *input.stage.program;
 	const auto& resources = *input.stage.resources;
 	if (resources.buffers.size() != program.info.buffers.size()) {
@@ -61,9 +62,10 @@ void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input
 	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
 		const auto& resource   = program.info.buffers[i];
 		const auto  descriptor = DecodeNativeDescriptor<ShaderBufferResource>(resources.buffers[i]);
-		// Reads and transforms of metadata do not establish a clear state.
+		// A metadata resource that is also read is not proven to be a full overwrite. Execute it
+		// conservatively instead of replacing the dispatch with a coarse full-surface clear.
 		if ((!resource.written || resource.read) && cache.IsMeta(descriptor.Base48())) {
-			return;
+			return false;
 		}
 	}
 
@@ -73,10 +75,13 @@ void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input
 			if (resource.written) {
 				const auto descriptor =
 				    DecodeNativeDescriptor<ShaderBufferResource>(resources.buffers[i]);
-				(void)cache.ClearMeta(descriptor.Base48());
+				if (cache.ClearMeta(descriptor.Base48())) {
+					return true;
+				}
 			}
 		}
 	}
+	return false;
 }
 
 bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t group_x,
@@ -127,9 +132,9 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 	return true;
 }
 
-bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& input,
-                                                CommandBuffer& command, uint32_t group_x,
-                                                uint32_t group_y, uint32_t group_z, uint32_t mode) {
+bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
+                                          CommandBuffer& command, uint32_t group_x,
+                                          uint32_t group_y, uint32_t group_z, uint32_t mode) {
 	const auto& program   = *input.stage.program;
 	const auto& resources = *input.stage.resources;
 	const auto& fill      = resources.uniform_fill;
@@ -185,14 +190,26 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
 	                              size)) {
 		return false;
 	}
-	if (!cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear)) {
-		return false;
+	const bool image_clear =
+	    cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear);
+	if (!image_clear) {
+		// CPU fills invalidate overlapping cached images. Preserve the native dispatch for
+		// partial or ambiguous image backings, including pooled metadata aliases.
+		if ((descriptor.Base48() & 3u) != 0 ||
+		    !m_context.IsMapped(descriptor.Base48(), size) ||
+		    cache.IsRegionRegistered(descriptor.Base48(), size)) {
+			return false;
+		}
+		// Metadata need not have an image backing yet. Keep proven fills in guest memory
+		// when CPU-owned, so color-clear discovery can read them without a GPU download.
+		m_context.GetBufferCache().FillBuffer(descriptor.Base48(), size, packed_clear, false);
 	}
 	static std::atomic<uint32_t> logged_clears {0};
 	if (logged_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
-		LOGF("GraphicsRenderDispatchDirect: compute image clear shader=0x%016" PRIx64
+		LOGF("GraphicsRenderDispatchDirect: compute %s shader=0x%016" PRIx64
 		     " addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
-		     input.stage.program->shader_hash, descriptor.Base48(), size, packed_clear);
+		     image_clear ? "image clear" : "buffer fill", input.stage.program->shader_hash,
+		     descriptor.Base48(), size, packed_clear);
 	}
 	return true;
 }
@@ -296,6 +313,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 	const auto compute_program =
 	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
+	if (!compute_program) return;
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
@@ -304,14 +322,12 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 
 	const auto& program   = *input_info.stage.program;
 	const auto& resources = *input_info.stage.resources;
-	if (resources.specialization_reads.empty()) {
-		// Cached metadata addresses can be reused as ordinary buffers; preserve their guest writes.
-		UpdateComputeMetaClears(input_info, buffer);
-		if (TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
-		                               thread_group_z, mode)) {
-			ResetBindings();
-			return;
-		}
+	if (resources.specialization_reads.empty() &&
+	    (TryConsumeComputeMetaClear(input_info, buffer) ||
+	     TryConsumeComputeFill(input_info, buffer, thread_group_x, thread_group_y,
+	                           thread_group_z, mode))) {
+		ResetBindings();
+		return;
 	}
 
 	if (use_thread_dimensions) {
@@ -432,6 +448,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 
 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
 	vk_buffer.dispatch(thread_group_x, thread_group_y, thread_group_z);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::Dispatches);
 
 	// The removed host fence also ordered read-only dispatches before later writers.
 	ShaderAccessBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
@@ -462,6 +479,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.indirect_args_address = args_addr;
 	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
 	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
+	if (!compute_program) return;
 	buffer.EndRendering();
 	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
 	auto& bindings = m_compute_bindings;
@@ -501,6 +519,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	                          1, &barrier, 0, nullptr, 0, nullptr);
 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
 	vk_buffer.dispatchIndirect(args_buffer->Handle(), args_offset);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::Dispatches);
 	ShaderAccessBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
 	ResetBindings();
 }

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -2,6 +2,7 @@
 
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
@@ -1233,6 +1234,8 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	} else {
 		EmitDrawPrimitives(ucfg, vk_buffer, draw, emit);
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::Draws);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::DrawInstances, draw.instance_count);
 
 	if (!draw.IsIndexed()) {
 		SetDrawDebugPhase(buffer, submit_id, draw, 0x600u);

src/graphics/host_gpu/vma.cpp

@@ -97,6 +97,24 @@ uint64_t GraphicContext::GetDeviceMemoryUsage() const {
 	return usage;
 }
 
+uint32_t GraphicContext::GetHeapUsage(std::span<HeapUsage, VK_MAX_MEMORY_HEAPS> heaps) const {
+	if (allocator == nullptr) {
+		return 0;
+	}
+	VmaBudget budgets[VK_MAX_MEMORY_HEAPS] {};
+	vmaGetHeapBudgets(allocator, budgets);
+	const auto count = physical_device_memory_properties.memoryHeapCount;
+	for (uint32_t heap = 0; heap < count; heap++) {
+		const auto& properties = physical_device_memory_properties.memoryHeaps[heap];
+		heaps[heap] = {.usage        = budgets[heap].usage,
+		               .budget       = budgets[heap].budget,
+		               .allocations  = budgets[heap].statistics.allocationCount,
+		               .device_local = static_cast<bool>(properties.flags &
+		                                                 vk::MemoryHeapFlagBits::eDeviceLocal)};
+	}
+	return count;
+}
+
 uint64_t GraphicContext::GetTotalMemoryBudget() const {
 	if (allocator == nullptr) {
 		return 0;

src/graphics/presentation/performanceOverlay.cpp

@@ -0,0 +1,747 @@
+#include "graphics/presentation/performanceOverlay.h"
+
+#include "common/debugCounters.h"
+#include "common/emulatorConfig.h"
+#include "common/hostStats.h"
+#include "graphics/host_gpu/graphicContext.h"
+#include "imgui.h"
+#include "kernel/memory.h"
+
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <chrono>
+#include <condition_variable>
+#include <cstdarg>
+#include <cstdio>
+#include <mutex>
+#include <optional>
+#include <stop_token>
+#include <thread>
+
+namespace Libs::Graphics {
+
+namespace {
+
+using Clock   = std::chrono::steady_clock;
+using Counter = Common::DebugCounters::Counter;
+using Peak    = Common::DebugCounters::Peak;
+
+constexpr auto   REFRESH_INTERVAL  = std::chrono::seconds(1);
+constexpr size_t FRAME_HISTORY     = 120;
+constexpr float  FRAME_GRAPH_MS    = 50.0f;
+constexpr float  FRAME_TIME_GOOD   = 17.5f;
+constexpr float  FRAME_TIME_MEDIUM = 34.0f;
+constexpr size_t COUNTER_COUNT     = static_cast<size_t>(Counter::Count);
+
+// Tables as wide as their columns; the default fills the window, which an auto-resizing
+// window then never shrinks.
+constexpr ImGuiTableFlags TABLE_FLAGS =
+    ImGuiTableFlags_SizingFixedFit | ImGuiTableFlags_NoHostExtendX;
+
+constexpr ImU32 GOOD    = IM_COL32(96, 220, 120, 255);
+constexpr ImU32 MEDIUM  = IM_COL32(230, 195, 60, 255);
+constexpr ImU32 BAD     = IM_COL32(235, 80, 70, 255);
+constexpr ImU32 NEUTRAL = IM_COL32(160, 160, 160, 255);
+
+struct FrameWindow {
+	std::array<float, FRAME_HISTORY> intervals_ms {};
+	size_t                           next         = 0;
+	size_t                           count        = 0;
+	uint32_t                         flips        = 0;
+	float                            max_ms       = 0.0f;
+	uint32_t                         guest_width  = 0;
+	uint32_t                         guest_height = 0;
+	Clock::time_point                last_flip    = {};
+	Clock::time_point                window_start = {};
+};
+
+// Shown values, recomputed once per refresh interval.
+struct Snapshot {
+	double                              fps          = 0.0;
+	double                              frame_ms     = 0.0;
+	double                              max_frame_ms = 0.0;
+	double                              low_fps      = 0.0; // From the 99th percentile.
+	double                              stalled_s    = 0.0;
+	uint32_t                            guest_width  = 0;
+	uint32_t                            guest_height = 0;
+	std::array<double, COUNTER_COUNT>   rates {};
+	std::array<uint64_t, COUNTER_COUNT> totals {};
+	uint64_t                            shader_compile_peak_ns  = 0;
+	uint64_t                            pipeline_create_peak_ns = 0;
+	Common::HostStats                   host;
+	std::optional<uint64_t>             device_memory;
+	uint64_t                            device_budget = 0;
+	std::array<GraphicContext::HeapUsage, VK_MAX_MEMORY_HEAPS> heaps {};
+	uint32_t                                                   heap_count = 0;
+	std::array<float, FRAME_HISTORY>                           history {};
+	size_t                                                     history_count = 0;
+	Clock::time_point                                          sample_time {};
+	LibKernel::Memory::DebugStats                              guest_memory;
+};
+
+std::mutex            g_frame_mutex;
+FrameWindow           g_frames;
+std::atomic<uint64_t> g_generation {0};
+std::atomic<bool>     g_details {false};
+std::atomic<uint64_t> g_revision {0};
+
+std::mutex                  g_host_mutex;
+Common::HostStats           g_host_stats;
+uint64_t                    g_host_generation = 0;
+std::condition_variable_any g_sampler_wait;
+std::jthread                g_sampler;
+
+void SetOverlayEnabled(bool enabled) {
+	{
+		std::scoped_lock lock(g_host_mutex);
+		if (enabled) {
+			std::scoped_lock frame_lock(g_frame_mutex);
+			g_frames = {};
+			g_generation.fetch_add(1, std::memory_order_release);
+		}
+		Common::DebugCounters::SetEnabled(enabled);
+	}
+	g_sampler_wait.notify_all();
+}
+
+bool GuestStalled(Clock::time_point now) {
+	std::scoped_lock lock(g_frame_mutex);
+	return now - g_frames.last_flip >= REFRESH_INTERVAL;
+}
+
+// Host statistics can block for milliseconds (IOKit, thread enumeration), so sample them off
+// the present path. While the guest stops flipping (loading screens), ask for one repaint per
+// interval.
+void SamplerThread(const std::stop_token& stop) {
+	std::optional<Common::HostStatsSampler> sampler;
+	uint64_t                                sampled_generation = 0;
+	std::unique_lock                        lock(g_host_mutex);
+	while (!stop.stop_requested()) {
+		g_sampler_wait.wait(lock, stop, [] { return Common::DebugCounters::Enabled(); });
+		if (stop.stop_requested()) {
+			break;
+		}
+		const auto generation = g_generation.load(std::memory_order_acquire);
+		lock.unlock();
+		if (!sampler || generation != sampled_generation) {
+			sampler.emplace();
+			sampled_generation = generation;
+		}
+		auto stats = sampler->Sample();
+		lock.lock();
+		g_host_stats      = std::move(stats);
+		g_host_generation = generation;
+		if (Common::DebugCounters::Enabled() && GuestStalled(Clock::now())) {
+			g_revision.fetch_add(1, std::memory_order_acq_rel);
+		}
+		g_sampler_wait.wait_for(lock, stop, REFRESH_INTERVAL, [generation] {
+			return !Common::DebugCounters::Enabled() ||
+			       generation != g_generation.load(std::memory_order_acquire);
+		});
+	}
+}
+
+void EnsureSampler() {
+	static std::once_flag once;
+	std::call_once(once, [] { g_sampler = std::jthread(SamplerThread); });
+}
+
+ImU32 FrameTimeColor(double ms) {
+	return ms <= FRAME_TIME_GOOD ? GOOD : ms <= FRAME_TIME_MEDIUM ? MEDIUM : BAD;
+}
+
+// Green below `warn`, yellow below `bad`, red above.
+ImU32 LevelColor(double value, double warn, double bad) {
+	return value < warn ? GOOD : value < bad ? MEDIUM : BAD;
+}
+
+ImVec4 Color(ImU32 color) {
+	return ImGui::ColorConvertU32ToFloat4(color);
+}
+
+std::string Bytes(uint64_t bytes) {
+	constexpr double GIB = 1024.0 * 1024.0 * 1024.0;
+	constexpr double MIB = 1024.0 * 1024.0;
+	char             buffer[32];
+	if (static_cast<double>(bytes) >= GIB) {
+		std::snprintf(buffer, sizeof(buffer), "%.2f GB", static_cast<double>(bytes) / GIB);
+	} else {
+		std::snprintf(buffer, sizeof(buffer), "%.0f MB", static_cast<double>(bytes) / MIB);
+	}
+	return buffer;
+}
+
+std::string Bytes(std::optional<uint64_t> bytes) {
+	return bytes ? Bytes(*bytes) : "n/a";
+}
+
+std::string Throughput(double bytes_per_second) {
+	char buffer[32];
+	std::snprintf(buffer, sizeof(buffer), "%.1f MB/s", bytes_per_second / (1024.0 * 1024.0));
+	return buffer;
+}
+
+std::string Percent(std::optional<double> value) {
+	if (!value) {
+		return "n/a";
+	}
+	char buffer[16];
+	std::snprintf(buffer, sizeof(buffer), "%.0f%%", *value);
+	return buffer;
+}
+
+void Row(const char* label, const std::string& value, std::optional<ImU32> color = std::nullopt) {
+	ImGui::TableNextRow();
+	ImGui::TableNextColumn();
+	ImGui::TextDisabled("%s", label);
+	ImGui::TableNextColumn();
+	if (color) {
+		ImGui::TextColored(Color(*color), "%s", value.c_str());
+	} else {
+		ImGui::TextUnformatted(value.c_str());
+	}
+}
+
+[[gnu::format(printf, 1, 2)]] std::string Fmt(const char* format, ...) {
+	char    buffer[128];
+	va_list args;
+	va_start(args, format);
+	std::vsnprintf(buffer, sizeof(buffer), format, args);
+	va_end(args);
+	return buffer;
+}
+
+bool BeginSection(const char* title) {
+	ImGui::SeparatorText(title);
+	return ImGui::BeginTable(title, 2, TABLE_FLAGS);
+}
+
+void DrawFrameGraph(const std::array<float, FRAME_HISTORY>& history, size_t count, float width,
+                    float height) {
+	const ImVec2 origin = ImGui::GetCursorScreenPos();
+	auto*        draw   = ImGui::GetWindowDrawList();
+	draw->AddRectFilled(origin, {origin.x + width, origin.y + height}, IM_COL32(0, 0, 0, 90));
+	const auto y_of = [&](float ms) {
+		return origin.y + height - std::min(ms, FRAME_GRAPH_MS) / FRAME_GRAPH_MS * height;
+	};
+	for (const float reference: {1000.0f / 60.0f, 1000.0f / 30.0f}) {
+		draw->AddLine({origin.x, y_of(reference)}, {origin.x + width, y_of(reference)},
+		              IM_COL32(255, 255, 255, 45));
+	}
+	const float bar = width / static_cast<float>(FRAME_HISTORY);
+	for (size_t i = 0; i < count; i++) {
+		const float ms = history[i];
+		const float x  = origin.x + width - static_cast<float>(count - i) * bar;
+		draw->AddRectFilled({x, y_of(ms)}, {x + std::max(bar - 1.0f, 1.0f), origin.y + height},
+		                    FrameTimeColor(ms));
+	}
+	ImGui::Dummy({width, height});
+}
+
+const char* PressureName(Common::MemoryPressure pressure) {
+	switch (pressure) {
+		case Common::MemoryPressure::Normal: return "Normal";
+		case Common::MemoryPressure::Warning: return "Warning";
+		case Common::MemoryPressure::Critical: return "Critical";
+		case Common::MemoryPressure::Unknown: break;
+	}
+	return "n/a";
+}
+
+ImU32 PressureColor(Common::MemoryPressure pressure) {
+	switch (pressure) {
+		case Common::MemoryPressure::Normal: return GOOD;
+		case Common::MemoryPressure::Warning: return MEDIUM;
+		case Common::MemoryPressure::Critical: return BAD;
+		case Common::MemoryPressure::Unknown: break;
+	}
+	return NEUTRAL;
+}
+
+const char* ThermalName(Common::ThermalPressure thermal) {
+	switch (thermal) {
+		case Common::ThermalPressure::Nominal: return "Nominal";
+		case Common::ThermalPressure::Moderate: return "Moderate";
+		case Common::ThermalPressure::Heavy: return "Heavy (throttling)";
+		case Common::ThermalPressure::Critical: return "Critical (throttling)";
+		case Common::ThermalPressure::Unknown: break;
+	}
+	return "n/a";
+}
+
+ImU32 ThermalColor(Common::ThermalPressure thermal) {
+	switch (thermal) {
+		case Common::ThermalPressure::Nominal: return GOOD;
+		case Common::ThermalPressure::Moderate: return MEDIUM;
+		case Common::ThermalPressure::Heavy:
+		case Common::ThermalPressure::Critical: return BAD;
+		case Common::ThermalPressure::Unknown: break;
+	}
+	return NEUTRAL;
+}
+
+const char* PresentModeName(Config::PresentMode mode) {
+	switch (mode) {
+		case Config::PresentMode::Fifo: return "FIFO";
+		case Config::PresentMode::Mailbox: return "Mailbox";
+		case Config::PresentMode::Immediate: return "Immediate";
+	}
+	return "?";
+}
+
+double Rate(const Snapshot& shown, Counter counter) {
+	return shown.rates[static_cast<size_t>(counter)];
+}
+
+uint64_t Total(const Snapshot& shown, Counter counter) {
+	return shown.totals[static_cast<size_t>(counter)];
+}
+
+void TransferRow(const Snapshot& shown, const char* label, Counter upload, Counter download) {
+	Row(label,
+	    Throughput(Rate(shown, upload)) + " up / " + Throughput(Rate(shown, download)) + " down");
+}
+
+void GpuActivityRows(const Snapshot& shown) {
+	const auto draws = Rate(shown, Counter::Draws);
+	Row("Draw calls", Fmt("%.0f/frame  %.0f/s", draws / std::max(shown.fps, 1.0), draws));
+	// Includes present pacing waits, so a high value alone is not a stall.
+	Row("CPU waiting on GPU",
+	    Fmt("%.0f ms/s  (%.0f waits/s)", Rate(shown, Counter::GpuWaitNs) / 1e6,
+	        Rate(shown, Counter::GpuWaits)));
+}
+
+void CompilationRows(const Snapshot& shown) {
+	const auto row = [&](const char* label, Counter count_counter, Counter duration_counter,
+	                     uint64_t peak_ns) {
+		const auto count = Total(shown, count_counter);
+		Row(label,
+		    Fmt("%llu  (avg %.1f ms, peak %.1f ms)", static_cast<unsigned long long>(count),
+		        count != 0 ? static_cast<double>(Total(shown, duration_counter)) / 1e6 /
+		                         static_cast<double>(count)
+		                   : 0.0,
+		        static_cast<double>(peak_ns) / 1e6),
+		    Rate(shown, count_counter) > 0.0 ? MEDIUM : GOOD);
+	};
+	row("Shader recompiler", Counter::ShadersCompiled, Counter::ShaderCompileNs,
+	    shown.shader_compile_peak_ns);
+	row("Pipeline compile", Counter::PipelinesCreated, Counter::PipelineCreateNs,
+	    shown.pipeline_create_peak_ns);
+}
+
+void Refresh(Snapshot& shown, const GraphicContext& graphics, Clock::time_point now,
+             uint64_t generation) {
+	const double seconds = shown.sample_time == Clock::time_point {}
+	                           ? 0.0
+	                           : std::chrono::duration<double>(now - shown.sample_time).count();
+	for (size_t i = 0; i < COUNTER_COUNT; i++) {
+		const auto value = Common::DebugCounters::Get(static_cast<Counter>(i));
+		shown.rates[i] =
+		    seconds > 0.0 ? static_cast<double>(value - shown.totals[i]) / seconds : 0.0;
+		shown.totals[i] = value;
+	}
+	shown.shader_compile_peak_ns  = Common::DebugCounters::Get(Peak::ShaderCompileNs);
+	shown.pipeline_create_peak_ns = Common::DebugCounters::Get(Peak::PipelineCreateNs);
+	shown.sample_time             = now;
+	{
+		std::scoped_lock lock(g_host_mutex);
+		if (g_host_generation == generation) {
+			shown.host = g_host_stats;
+		}
+	}
+	shown.heap_count = graphics.GetHeapUsage(shown.heaps);
+	shown.device_memory =
+	    graphics.CanReportMemoryUsage() ? std::optional<uint64_t>(0) : std::nullopt;
+	shown.device_budget = 0;
+	const bool discrete =
+	    graphics.GetPhysicalDeviceProperties().deviceType == vk::PhysicalDeviceType::eDiscreteGpu;
+	for (uint32_t i = 0; i < shown.heap_count; i++) {
+		const auto& heap = shown.heaps[i];
+		if (shown.device_memory && (!discrete || heap.device_local)) {
+			*shown.device_memory += heap.usage;
+			shown.device_budget += heap.budget;
+		}
+	}
+	if (g_details.load(std::memory_order_acquire)) {
+		shown.guest_memory = LibKernel::Memory::GetDebugStats();
+	}
+}
+
+void DrawCompact(const Snapshot& shown) {
+	if (!ImGui::BeginTable("##performance", 2, TABLE_FLAGS)) {
+		return;
+	}
+	const auto& host = shown.host;
+	if (shown.max_frame_ms > 0.0) {
+		Row("Worst frame", Fmt("%.1f ms", shown.max_frame_ms), FrameTimeColor(shown.max_frame_ms));
+	}
+	Row("CPU", host.process_cpu_percent
+	               ? Fmt("%.0f%%  of %u threads", *host.process_cpu_percent, host.logical_cores)
+	               : std::string("n/a"));
+	Row("GPU", Percent(host.gpu_percent));
+	GpuActivityRows(shown);
+	CompilationRows(shown);
+	if (shown.device_memory) {
+		Row("GPU memory", Bytes(*shown.device_memory) + " / " + Bytes(shown.device_budget));
+	}
+	TransferRow(shown, "Texture transfers", Counter::TextureUploadBytes,
+	            Counter::TextureDownloadBytes);
+	TransferRow(shown, "Buffer transfers", Counter::BufferUploadBytes,
+	            Counter::BufferDownloadBytes);
+	Row("Emulator RAM", Bytes(host.process_memory));
+	if (host.system_memory_used && host.system_memory_total) {
+		Row("System RAM",
+		    Bytes(*host.system_memory_used) + " / " + Bytes(*host.system_memory_total));
+	}
+	ImGui::EndTable();
+}
+
+void DrawFramesAndCpu(const Snapshot& shown, vk::Extent2D extent) {
+	const auto& host = shown.host;
+	if (BeginSection("Frames")) {
+		if (shown.low_fps > 0.0) {
+			Row("1% low", Fmt("%.1f FPS", shown.low_fps), FrameTimeColor(1000.0 / shown.low_fps));
+		}
+		Row("Worst frame", Fmt("%.1f ms", shown.max_frame_ms), FrameTimeColor(shown.max_frame_ms));
+		Row("Guest output", shown.guest_width != 0
+		                        ? Fmt("%ux%u", shown.guest_width, shown.guest_height)
+		                        : std::string("n/a"));
+		Row("Window", Fmt("%ux%u  %s  %u Hz", extent.width, extent.height,
+		                  PresentModeName(Config::GetPresentMode()), Config::GetVblankFrequency()));
+		Row("Flips", Fmt("%.0f/s requested", Rate(shown, Counter::FlipsRequested)));
+		const auto rejected = Rate(shown, Counter::FlipsRejected);
+		Row("Flip queue full", Fmt("%.0f/s", rejected), rejected > 0.0 ? BAD : GOOD);
+		const auto late = Rate(shown, Counter::LateVblanks);
+		Row("Late vblanks", Fmt("%.0f/s", late), LevelColor(late, 1.0, 10.0));
+		Row("Swapchain rebuilds",
+		    Fmt("%llu  (dropped %llu)",
+		        static_cast<unsigned long long>(Total(shown, Counter::SwapchainRecreations)),
+		        static_cast<unsigned long long>(Total(shown, Counter::DroppedPresents))));
+		ImGui::EndTable();
+	}
+	if (BeginSection("CPU")) {
+		Row("Emulator",
+		    host.process_cpu_percent
+		        ? Fmt("%.0f%% of %u threads  (%.1f cores)", *host.process_cpu_percent,
+		              host.logical_cores, *host.process_cpu_percent * host.logical_cores / 100.0)
+		        : std::string("n/a"));
+		Row("Machine", Percent(host.system_cpu_percent),
+		    host.system_cpu_percent ? std::optional(LevelColor(*host.system_cpu_percent, 70, 90))
+		                            : std::nullopt);
+		if (host.thread_count) {
+			Row("Host threads", Fmt("%u", *host.thread_count));
+		}
+		for (const auto& thread: host.busiest_threads) {
+			Row(Fmt("  %.24s", thread.name.c_str()).c_str(), Fmt("%.0f%%", thread.cpu_percent),
+			    LevelColor(thread.cpu_percent, 70, 95));
+		}
+		if (host.context_switches_per_s) {
+			Row("Context switches", Fmt("%.0f/s", *host.context_switches_per_s));
+		}
+		if (host.page_faults_per_s) {
+			Row("Page faults", Fmt("%.0f/s  (%.0f from disk)", *host.page_faults_per_s,
+			                       host.page_ins_per_s.value_or(0.0)));
+		}
+		Row("Thermal", ThermalName(host.thermal_pressure), ThermalColor(host.thermal_pressure));
+		const auto emulated = Rate(shown, Counter::EmulatedInstructions);
+		Row("Emulated instrs", Fmt("%.0f/s", emulated), LevelColor(emulated, 1000, 100000));
+		const auto unresolved = Rate(shown, Counter::UnresolvedImportCalls);
+		Row("Missing HLE calls",
+		    Fmt("%.0f/s  (%llu total)", unresolved,
+		        static_cast<unsigned long long>(Total(shown, Counter::UnresolvedImportCalls))),
+		    unresolved > 0.0 ? MEDIUM : GOOD);
+		ImGui::EndTable();
+	}
+}
+
+void DrawGpu(const Snapshot& shown, const GraphicContext& graphics) {
+	const auto& host = shown.host;
+	if (BeginSection("GPU")) {
+		Row("Device", host.gpu_model.empty()
+		                  ? std::string(graphics.GetPhysicalDeviceProperties().deviceName.data())
+		              : host.gpu_cores
+		                  ? Fmt("%s (%u cores)", host.gpu_model.c_str(), *host.gpu_cores)
+		                  : host.gpu_model);
+		Row("Utilization", Percent(host.gpu_percent),
+		    host.gpu_percent ? std::optional(LevelColor(*host.gpu_percent, 80, 95)) : std::nullopt);
+		if (host.gpu_renderer_percent || host.gpu_tiler_percent) {
+			Row("Renderer / tiler",
+			    Percent(host.gpu_renderer_percent) + " / " + Percent(host.gpu_tiler_percent));
+		}
+		if (host.gpu_memory_in_use) {
+			Row("Driver memory", Bytes(*host.gpu_memory_in_use) + " in use / " +
+			                         Bytes(host.gpu_memory_allocated) + " allocated");
+		}
+		if (host.gpu_recoveries) {
+			Row("GPU resets", Fmt("%llu", static_cast<unsigned long long>(*host.gpu_recoveries)),
+			    *host.gpu_recoveries > 0 ? BAD : GOOD);
+		}
+		const double fps = std::max(shown.fps, 1.0);
+		GpuActivityRows(shown);
+		Row("Instances", Fmt("%.0f/frame", Rate(shown, Counter::DrawInstances) / fps));
+		Row("Indirect draws", Fmt("%.0f/s", Rate(shown, Counter::IndirectDraws)));
+		Row("Dispatches", Fmt("%.0f/frame  %.0f/s", Rate(shown, Counter::Dispatches) / fps,
+		                      Rate(shown, Counter::Dispatches)));
+		Row("Queue submits", Fmt("%.0f/s", Rate(shown, Counter::QueueSubmits)));
+		const auto faults = Rate(shown, Counter::GpuFaults);
+		Row("Write-tracking faults", Fmt("%.0f/s", faults), LevelColor(faults, 20000, 100000));
+		CompilationRows(shown);
+		ImGui::EndTable();
+	}
+	if (BeginSection("GPU memory")) {
+		if (shown.device_memory) {
+			const double share = shown.device_budget != 0
+			                         ? static_cast<double>(*shown.device_memory) /
+			                               static_cast<double>(shown.device_budget) * 100.0
+			                         : 0.0;
+			Row("Vulkan budget", Bytes(*shown.device_memory) + " / " + Bytes(shown.device_budget),
+			    LevelColor(share, 80, 95));
+		}
+		for (size_t heap = 0; heap < shown.heap_count; heap++) {
+			const auto& usage = shown.heaps[heap];
+			Row(Fmt("  Heap %zu%s", heap, usage.device_local ? " (device)" : "").c_str(),
+			    Bytes(usage.usage) + " / " + Bytes(usage.budget) +
+			        Fmt("  %u allocs", usage.allocations));
+		}
+		TransferRow(shown, "Texture transfers", Counter::TextureUploadBytes,
+		            Counter::TextureDownloadBytes);
+		Row("Texture evictions", Fmt("%.0f/s", Rate(shown, Counter::TextureEvictions)));
+		TransferRow(shown, "Buffer transfers", Counter::BufferUploadBytes,
+		            Counter::BufferDownloadBytes);
+		Row("Buffer evictions", Fmt("%.0f/s", Rate(shown, Counter::BufferEvictions)));
+		ImGui::EndTable();
+	}
+}
+
+void DrawMemoryAndIo(const Snapshot& shown) {
+	const auto& host  = shown.host;
+	const auto& guest = shown.guest_memory;
+	if (BeginSection("PS5 memory")) {
+		const auto direct_used = guest.direct_allocated + guest.pooled_allocated;
+		Row("Direct", Bytes(direct_used) + " / " + Bytes(guest.direct_total),
+		    LevelColor(guest.direct_total != 0 ? static_cast<double>(direct_used) /
+		                                             static_cast<double>(guest.direct_total) * 100.0
+		                                       : 0.0,
+		               85, 97));
+		Row("  Mapped", Bytes(guest.direct_mapped));
+		Row("  Pool",
+		    Bytes(guest.pooled_allocated) + " (" + Bytes(guest.pool_committed) + " committed)");
+		Row("Flexible", Bytes(guest.flexible_used) + " / " + Bytes(guest.flexible_total));
+		Row("Extended", Bytes(guest.automatic_allocated));
+		// Past the pool size the guest sees no page-table entries available.
+		const auto entries = std::max(guest.cpu_page_entries, guest.gpu_page_entries);
+		Row("Page table (2 MB)",
+		    Fmt("CPU %llu  GPU %llu  / %llu",
+		        static_cast<unsigned long long>(guest.cpu_page_entries),
+		        static_cast<unsigned long long>(guest.gpu_page_entries),
+		        static_cast<unsigned long long>(guest.page_entries_total)),
+		    entries > guest.page_entries_total            ? BAD
+		    : entries * 10 > guest.page_entries_total * 9 ? MEDIUM
+		                                                  : GOOD);
+		ImGui::EndTable();
+	}
+	if (BeginSection("Emulator process")) {
+		Row("Footprint", Bytes(host.process_memory));
+		Row("Resident", Bytes(host.process_resident));
+		if (host.process_compressed) {
+			Row("Compressed / swapped", Bytes(*host.process_compressed),
+			    *host.process_compressed > (1ull << 30u) ? MEDIUM : GOOD);
+		}
+		Row("File-backed", Bytes(host.process_file_backed));
+		Row("Address space", Bytes(host.process_virtual));
+		ImGui::EndTable();
+	}
+	if (BeginSection("System memory")) {
+		if (host.system_memory_used && host.system_memory_total) {
+			Row("Used", Bytes(*host.system_memory_used) + " / " + Bytes(*host.system_memory_total));
+		}
+		if (host.system_wired) {
+			Row("Wired", Bytes(*host.system_wired));
+		}
+		if (host.system_compressed) {
+			Row("Compressed", Bytes(*host.system_compressed));
+		}
+		Row("Cached files", Bytes(host.system_cached));
+		Row("Free", Bytes(host.system_free));
+		if (host.swap_used) {
+			Row("Swap", Bytes(*host.swap_used) + " / " + Bytes(host.swap_total),
+			    *host.swap_used > (4ull << 30u) ? MEDIUM : GOOD);
+		}
+		Row("Pressure", PressureName(host.memory_pressure), PressureColor(host.memory_pressure));
+		ImGui::EndTable();
+	}
+	if (BeginSection("I/O & audio")) {
+		Row("Disk read / write", Throughput(host.disk_read_per_s.value_or(0.0)) + " / " +
+		                             Throughput(host.disk_write_per_s.value_or(0.0)));
+		Row("Guest file read / write", Throughput(Rate(shown, Counter::GuestFileReadBytes)) +
+		                                   " / " +
+		                                   Throughput(Rate(shown, Counter::GuestFileWriteBytes)));
+		const auto underruns = Rate(shown, Counter::AudioUnderruns);
+		Row("Audio underruns",
+		    Fmt("%.0f/s  (%llu total)", underruns,
+		        static_cast<unsigned long long>(Total(shown, Counter::AudioUnderruns))),
+		    underruns > 0.0 ? BAD : GOOD);
+		ImGui::EndTable();
+	}
+}
+
+} // namespace
+
+void PerformanceOverlayRecordFlip(uint32_t guest_width, uint32_t guest_height) noexcept {
+	if (!Common::DebugCounters::Enabled()) {
+		return;
+	}
+	std::scoped_lock lock(g_frame_mutex);
+	const auto       now = Clock::now();
+	if (g_frames.window_start == Clock::time_point {}) {
+		g_frames.window_start = now;
+	}
+	auto& frames = g_frames;
+	if (frames.last_flip != Clock::time_point {}) {
+		const float ms = std::chrono::duration<float, std::milli>(now - frames.last_flip).count();
+		frames.intervals_ms[frames.next] = ms;
+		frames.next                      = (frames.next + 1) % FRAME_HISTORY;
+		frames.count                     = std::min(frames.count + 1, FRAME_HISTORY);
+		frames.max_ms                    = std::max(frames.max_ms, ms);
+	}
+	frames.last_flip    = now;
+	frames.guest_width  = guest_width;
+	frames.guest_height = guest_height;
+	frames.flips++;
+}
+
+bool PerformanceOverlayEnabled() noexcept {
+	return Common::DebugCounters::Enabled();
+}
+
+void TogglePerformanceOverlay() noexcept {
+	SetOverlayEnabled(!Common::DebugCounters::Enabled());
+	g_revision.fetch_add(1, std::memory_order_acq_rel);
+}
+
+void TogglePerformanceOverlayDetails() noexcept {
+	if (Common::DebugCounters::Enabled()) {
+		g_details.store(!g_details.load(std::memory_order_acquire), std::memory_order_release);
+	} else {
+		g_details.store(true, std::memory_order_release);
+		SetOverlayEnabled(true);
+	}
+	g_revision.fetch_add(1, std::memory_order_acq_rel);
+}
+
+uint64_t PerformanceOverlayRevision() noexcept {
+	return g_revision.load(std::memory_order_acquire);
+}
+
+float DrawPerformanceOverlay(const GraphicContext& graphics, vk::Extent2D extent) {
+	if (!PerformanceOverlayEnabled()) {
+		return 0.0f;
+	}
+	EnsureSampler();
+
+	static Snapshot shown;
+	static uint64_t shown_generation = 0;
+	static bool     shown_details    = false;
+	const auto      generation       = g_generation.load(std::memory_order_acquire);
+	const bool      details          = g_details.load(std::memory_order_acquire);
+	const auto      now              = Clock::now();
+	if (generation != shown_generation) {
+		shown            = {};
+		shown_generation = generation;
+	}
+	if (shown.sample_time == Clock::time_point {} || now - shown.sample_time >= REFRESH_INTERVAL ||
+	    details != shown_details) {
+		{
+			std::scoped_lock lock(g_frame_mutex);
+			auto&            frames = g_frames;
+			shown.history_count     = frames.count;
+			for (size_t i = 0; i < frames.count; i++) {
+				shown.history[i] =
+				    frames.intervals_ms[(frames.next + FRAME_HISTORY - frames.count + i) %
+				                        FRAME_HISTORY];
+			}
+			if (frames.window_start == Clock::time_point {}) {
+				frames.window_start = now;
+			}
+			const double seconds = std::chrono::duration<double>(now - frames.window_start).count();
+			shown.fps            = seconds > 0.0 ? frames.flips / seconds : 0.0;
+			shown.frame_ms       = frames.flips > 0 ? seconds * 1000.0 / frames.flips : 0.0;
+			shown.max_frame_ms   = frames.max_ms;
+			shown.stalled_s   = frames.last_flip == Clock::time_point {}
+			                        ? 0.0
+			                        : std::chrono::duration<double>(now - frames.last_flip).count();
+			shown.guest_width = frames.guest_width;
+			shown.guest_height  = frames.guest_height;
+			frames.flips        = 0;
+			frames.max_ms       = 0.0f;
+			frames.window_start = now;
+		}
+		shown_details = details;
+		if (shown.history_count > 0) {
+			auto       sorted     = shown.history;
+			const auto end        = sorted.begin() + shown.history_count;
+			const auto percentile = sorted.begin() + shown.history_count * 99 / 100;
+			std::nth_element(sorted.begin(), percentile, end);
+			shown.low_fps = *percentile > 0.0f ? 1000.0 / *percentile : 0.0;
+		}
+		Refresh(shown, graphics, now, generation);
+	}
+
+	const float scale = std::clamp(std::min(static_cast<float>(extent.width) / 1920.0f,
+	                                        static_cast<float>(extent.height) / 1080.0f),
+	                               0.75f, 1.5f);
+	ImGui::SetNextWindowPos({static_cast<float>(extent.width) - 12.0f * scale, 12.0f * scale},
+	                        ImGuiCond_Always, {1.0f, 0.0f});
+	ImGui::SetNextWindowBgAlpha(details ? 0.82f : 0.72f);
+	constexpr ImGuiWindowFlags FLAGS =
+	    ImGuiWindowFlags_NoDecoration | ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoNav |
+	    ImGuiWindowFlags_NoMove | ImGuiWindowFlags_NoSavedSettings |
+	    ImGuiWindowFlags_NoFocusOnAppearing | ImGuiWindowFlags_NoBringToFrontOnFocus |
+	    ImGuiWindowFlags_AlwaysAutoResize;
+	ImGui::PushFont(nullptr, (details ? 13.0f : 15.0f) * scale);
+	ImGui::PushStyleVar(ImGuiStyleVar_WindowRounding, 8.0f * scale);
+	ImGui::PushStyleVar(ImGuiStyleVar_WindowPadding, {10.0f * scale, 8.0f * scale});
+	ImGui::PushStyleVar(ImGuiStyleVar_ItemSpacing, {8.0f * scale, 2.0f * scale});
+	ImGui::PushStyleVar(ImGuiStyleVar_CellPadding, {4.0f * scale, 1.0f * scale});
+	if (ImGui::Begin("##kyty_performance_overlay", nullptr, FLAGS)) {
+		if (shown.fps > 0.0) {
+			ImGui::TextColored(Color(FrameTimeColor(shown.frame_ms)), "%.1f FPS   %.1f ms",
+			                   shown.fps, shown.frame_ms);
+		} else if (shown.stalled_s >= 1.0) {
+			ImGui::TextColored(Color(BAD), "No frames for %.0f s", shown.stalled_s);
+		} else {
+			ImGui::TextDisabled("Waiting for frames");
+		}
+		DrawFrameGraph(shown.history, shown.history_count, (details ? 330.0f : 260.0f) * scale,
+		               46.0f * scale);
+		if (details) {
+			// Three columns keep the detailed panel within a 720p window.
+			if (ImGui::BeginTable("##details", 3, TABLE_FLAGS)) {
+				ImGui::TableNextRow();
+				ImGui::TableNextColumn();
+				DrawFramesAndCpu(shown, extent);
+				ImGui::TableNextColumn();
+				DrawGpu(shown, graphics);
+				ImGui::TableNextColumn();
+				DrawMemoryAndIo(shown);
+				ImGui::EndTable();
+			}
+		} else {
+			DrawCompact(shown);
+		}
+#if defined(__APPLE__)
+		ImGui::TextDisabled(details ? "Cmd+Shift+P compact   Cmd+P hide"
+		                            : "Cmd+Shift+P details   Cmd+P hide");
+#else
+		ImGui::TextDisabled(details ? "Shift+F2 compact   F2 hide" : "Shift+F2 details   F2 hide");
+#endif
+	}
+	const float bottom = ImGui::GetWindowPos().y + ImGui::GetWindowSize().y;
+	ImGui::End();
+	ImGui::PopStyleVar(4);
+	ImGui::PopFont();
+	return bottom;
+}
+
+} // namespace Libs::Graphics

src/graphics/presentation/performanceOverlay.h

@@ -0,0 +1,30 @@
+#ifndef EMULATOR_SRC_GRAPHICS_PRESENTATION_PERFORMANCEOVERLAY_H_
+#define EMULATOR_SRC_GRAPHICS_PRESENTATION_PERFORMANCEOVERLAY_H_
+
+#include "graphics/host_gpu/vulkanCommon.h"
+
+#include <cstdint>
+
+namespace Libs::Graphics {
+
+struct GraphicContext;
+
+// Called for each main-bus guest flip, with the guest's video-out size.
+void PerformanceOverlayRecordFlip(uint32_t guest_width, uint32_t guest_height) noexcept;
+
+[[nodiscard]] bool PerformanceOverlayEnabled() noexcept;
+void               TogglePerformanceOverlay() noexcept;
+// Switches between the compact and the detailed panel, showing the panel if it was hidden.
+void TogglePerformanceOverlayDetails() noexcept;
+
+// Changes when the panel needs a repaint without a guest flip (toggled, or the guest stalled).
+[[nodiscard]] uint64_t PerformanceOverlayRevision() noexcept;
+
+// Must run between ImGui::NewFrame() and ImGui::Render(), on the present thread while the
+// graphics subsystem is alive. Draws in the top-right corner and returns the panel's bottom
+// edge, or 0 when hidden.
+float DrawPerformanceOverlay(const GraphicContext& graphics, vk::Extent2D extent);
+
+} // namespace Libs::Graphics
+
+#endif // EMULATOR_SRC_GRAPHICS_PRESENTATION_PERFORMANCEOVERLAY_H_

src/graphics/presentation/systemOverlay.cpp

@@ -1,7 +1,5 @@
 #include "graphics/presentation/systemOverlay.h"
 
-#include <SDL3/SDL.h>
-
 #include "common/assert.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
@@ -11,10 +9,11 @@
 #include "gpu_blit_shaders/gpu_imgui_hdr_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/presentation/ddsImage.h"
+#include "graphics/presentation/performanceOverlay.h"
 #include "graphics/presentation/window/hostInput.h"
 #include "imgui.h"
-#include "imgui_internal.h"
 #include "imgui_impl_vulkan.h"
+#include "imgui_internal.h"
 #include "libs/controller.h"
 #include "libs/dialog.h"
 #include "libs/ime.h"
@@ -22,6 +21,7 @@
 #include "loader/systemContent.h"
 #include "stb_image.h"
 
+#include <SDL3/SDL.h>
 #include <algorithm>
 #include <array>
 #include <atomic>
@@ -712,7 +712,8 @@ SystemOverlayVisualState GetSystemOverlayVisualState() noexcept {
 	return {core.active || dialog.active || system.active || trophy_active || splash ||
 	            (cursor.enabled && cursor.cursor_visible),
 	        core.revision + dialog.revision + system.revision + (trophy_active ? 1u : 0u) +
-	            (splash ? 2u : 0u) + static_cast<decltype(core.revision)>(cursor.revision)};
+	            (splash ? 2u : 0u) + static_cast<decltype(core.revision)>(cursor.revision) +
+	            PerformanceOverlayRevision()};
 }
 
 bool ProcessSystemOverlayInput(const SDL_Event& event) {
@@ -1662,8 +1663,7 @@ struct SystemOverlay::Impl {
 		}
 	}
 
-
-	void DrawTrophyNotification(vk::Extent2D frame_extent, Clock::time_point now) {
+	void DrawTrophyNotification(vk::Extent2D frame_extent, Clock::time_point now, float min_top) {
 		std::erase_if(retired_trophy_images, [](const auto& texture) {
 			if (texture->Status != ImTextureStatus_Destroyed) {
 				return false;
@@ -1731,7 +1731,7 @@ struct SystemOverlay::Impl {
 		const float  alpha    = std::min(fade_in, fade_out);
 		const float  slide    = (1.0f - fade_in) * (width + 24.0f);
 		const ImVec2 top_left {static_cast<float>(frame_extent.width) - 32.0f - width + slide,
-		                       32.0f};
+		                       std::max(32.0f, min_top + 12.0f)};
 		const ImVec2 bottom_right {top_left.x + width, top_left.y + height};
 		auto*        draw = ImGui::GetForegroundDrawList();
 		const auto   fade = [alpha](int r, int g, int b, int a) {
@@ -1799,7 +1799,7 @@ struct SystemOverlay::Impl {
 		const auto now = Clock::now();
 		if (!has_overlay && !g_trophy_notification_active.load(std::memory_order_acquire) &&
 		    !g_splash_active.load(std::memory_order_acquire) &&
-		    !GetHostInputTouchPadVisualState().enabled) {
+		    !GetHostInputTouchPadVisualState().enabled && !PerformanceOverlayEnabled()) {
 			return false;
 		}
 		const auto prepared_session = snapshot.session;
@@ -1830,6 +1830,7 @@ struct SystemOverlay::Impl {
 		ImGui_ImplVulkan_NewFrame();
 		ImGui::NewFrame();
 		DrawSplash(frame_extent);
+		const float performance_bottom = DrawPerformanceOverlay(graphics, frame_extent);
 		GetOverlaySnapshot(&snapshot);
 		if (has_overlay && snapshot.session == prepared_session) {
 			if (snapshot.session.kind == OverlayKind::Dialog) {
@@ -1838,7 +1839,7 @@ struct SystemOverlay::Impl {
 				DrawIme(snapshot.ime, frame_extent);
 			}
 		}
-		DrawTrophyNotification(frame_extent, now);
+		DrawTrophyNotification(frame_extent, now, performance_bottom);
 		DrawTouchPadCursor(GetHostInputTouchPadVisualState());
 		ImGui::Render();
 		extent = frame_extent;

src/graphics/presentation/videoOut.cpp

@@ -3,6 +3,7 @@
 #include "common/abi.h"
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -15,6 +16,7 @@
 #include "graphics/host_gpu/renderer/image/imageInfo.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
+#include "graphics/presentation/performanceOverlay.h"
 #include "graphics/presentation/presenter.h"
 #include "graphics/presentation/renderDoc.h"
 #include "kernel/pthread.h"
@@ -909,6 +911,9 @@ void VideoOutDriver::Impl::PresentThread(std::stop_token token) {
 
 		const auto frame_end = Common::Timer::QueryPerformanceCounter();
 		total_wait += static_cast<int64_t>(period) - static_cast<int64_t>(frame_end - frame_begin);
+		if (total_wait < 0) {
+			Common::DebugCounters::Add(Common::DebugCounters::Counter::LateVblanks);
+		}
 	}
 }
 
@@ -938,6 +943,7 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe
 		m_done_cond_var.SignalAll();
 	}
 	if (cfg.flip_status.flipPendingNum >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::FlipsRejected);
 		return false;
 	}
 	auto& pending = source == FlipRequestSource::GpuEop ? m_requests : m_cpu_requests;
@@ -965,6 +971,7 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe
 
 	pending.push_back(r);
 	request_id = r.id;
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::FlipsRequested);
 
 	cfg.flip_status.flipPendingNum++;
 	cfg.flip_status.submitProcessTimeCounter = r.submit_ptc;
@@ -1243,6 +1250,9 @@ bool FlipQueue::Flip(uint32_t micros) {
 	}
 	if (due) {
 		m_presenter.Present(std::span(layers.data(), count));
+		if (requests[0].cfg->bus == 0) {
+			Graphics::PerformanceOverlayRecordFlip(requests[0].cfg->width, requests[0].cfg->height);
+		}
 	}
 
 	m_mutex.Lock();

src/graphics/presentation/window/hostInput.cpp

@@ -254,8 +254,8 @@ public:
 			}
 
 			const bool reserved = binding.key == SDLK_ESCAPE || binding.key == SDLK_F1 ||
-			                      binding.key == SDLK_F7 || binding.key == SDLK_F8 ||
-			                      binding.key == SDLK_F11;
+			                      binding.key == SDLK_F2 || binding.key == SDLK_F7 ||
+			                      binding.key == SDLK_F8 || binding.key == SDLK_F11;
 			if (binding.control == INVALID_CONTROL || reserved ||
 			    (binding.key == SDLK_UNKNOWN && binding.mouse_button == 0)) {
 				EXIT("Invalid input mapping: %s\n", value.c_str());

src/graphics/presentation/window/swapchain.cpp

@@ -1,5 +1,6 @@
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -10,6 +11,7 @@
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/presentation/performanceOverlay.h"
 #include "graphics/presentation/presenter.h"
 #include "graphics/presentation/systemOverlay.h"
 #include "graphics/presentation/videoOut.h"
@@ -324,6 +326,7 @@ struct Presenter::Impl {
 	void RecoverSwapchain(Swapchain::Status status) {
 		LOGF("Recovering Vulkan swapchain%s\n",
 		     status == Swapchain::Status::SurfaceLost ? " and surface" : "");
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::SwapchainRecreations);
 		swapchain.Recreate(status == Swapchain::Status::SurfaceLost);
 		frames.SetFormat(swapchain.Format());
 	}
@@ -1039,8 +1042,8 @@ void Presenter::Impl::Present() {
 		// present_mutex owns the prepared frames and this scheduler. Only queue submission
 		// is shared with rendering, and it has its own queue_mutex.
 		auto&      command = present_scheduler.BeginCommand();
-		const bool draw_system_overlay =
-		    overlay_visual.active && swapchain.PrepareSystemOverlay();
+		const bool draw_system_overlay = (overlay_visual.active || PerformanceOverlayEnabled()) &&
+		                                 swapchain.PrepareSystemOverlay();
 		swapchain.RecordPresentCommands(command, layers[0].frame, layers[1],
 		                                draw_system_overlay);
 		const auto tick = swapchain.Submit(present_scheduler);
@@ -1060,6 +1063,7 @@ void Presenter::Impl::Present() {
 		return;
 	}
 	LOGF("Vulkan presentation retry exhausted; dropping frame\n");
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::DroppedPresents);
 }
 
 void Presenter::Discard(Frame& frame) {

src/graphics/presentation/window/window.cpp

@@ -1,21 +1,20 @@
 #include "graphics/presentation/window.h"
 
-#include <SDL3/SDL.h>
-
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
+#include "common/stringUtils.h"
 #include "common/systemInfo.h"
 #include "common/threads.h"
 #include "common/timer.h"
-#include "common/stringUtils.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/presentation/performanceOverlay.h"
 #include "graphics/presentation/renderDoc.h"
 #include "graphics/presentation/systemOverlay.h"
 #include "graphics/presentation/window/hostInput.h"
@@ -24,13 +23,14 @@
 #include "libs/controller.h"
 #include "loader/systemContent.h"
 
+#include <SDL3/SDL.h>
 #include <cstdlib>
+#include <filesystem>
 #include <fmt/format.h>
 #include <memory>
 #include <string>
 #include <vector>
 #include <vulkan/vk_platform.h>
-#include <filesystem>
 
 #define STB_IMAGE_IMPLEMENTATION
 #define STBI_NO_SIMD
@@ -459,19 +459,37 @@ void WindowContext::ProcessDisplayEvent(const SDL_DisplayEvent& display) {
 void WindowContext::ProcessEvent(double time_s) {
 	auto& game  = loop;
 	auto* event = &game.event;
-	if ((event->type == SDL_EVENT_KEY_DOWN || event->type == SDL_EVENT_KEY_UP) &&
-	    event->key.key == SDLK_F7) {
-		if (event->type == SDL_EVENT_KEY_DOWN && event->key.repeat == 0) {
-			HostInputToggleMouseToJoystick();
+	if (event->type == SDL_EVENT_KEY_DOWN || event->type == SDL_EVENT_KEY_UP) {
+		const bool pressed = event->type == SDL_EVENT_KEY_DOWN && event->key.repeat == 0;
+		switch (event->key.key) {
+			case SDLK_F7:
+				if (pressed) {
+					HostInputToggleMouseToJoystick();
+				}
+				return;
+			case SDLK_F8:
+				if (pressed) {
+					Controller::CycleKeyboardPlayer();
+				}
+				return;
+#if defined(__APPLE__)
+			case SDLK_P:
+				if ((event->key.mod & SDL_KMOD_GUI) == 0) {
+					break;
+				}
+				[[fallthrough]];
+#endif
+			case SDLK_F2:
+				if (pressed) {
+					if ((event->key.mod & SDL_KMOD_SHIFT) != 0) {
+						TogglePerformanceOverlayDetails();
+					} else {
+						TogglePerformanceOverlay();
+					}
+				}
+				return;
+			default: break;
 		}
-		return;
-	}
-	if ((event->type == SDL_EVENT_KEY_DOWN || event->type == SDL_EVENT_KEY_UP) &&
-	    event->key.key == SDLK_F8) {
-		if (event->type == SDL_EVENT_KEY_DOWN && event->key.repeat == 0) {
-			Controller::CycleKeyboardPlayer();
-		}
-		return;
 	}
 	if (ProcessSystemOverlayInput(*event)) {
 		return;

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -481,18 +481,14 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 	return result;
 }
 
-void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
+void PrepareCallTargets(ShaderSource& source, const CompileOptions& options) {
 	using namespace IR;
+	EXIT_NOT_IMPLEMENTED(options.stage != ShaderType::Compute);
 	for (uint32_t i = 0; i < source.decoded.instructions.size(); ++i) {
 		if (source.decoded.instructions[i].opcode != Decoder::Opcode::S_SWAPPC_B64) continue;
-		EXIT_NOT_IMPLEMENTED(source.call.has_value() || options.stage != ShaderType::Compute);
-		source.call = ShaderSource::Call {.instruction = i};
+		source.calls.push_back({.instruction = i});
 	}
 	EXIT_IF(options.input_info.compute == nullptr);
-	const auto& call = source.decoded.instructions[source.call->instruction];
-	EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
-	                     call.dst.kind != Decoder::OperandKind::Sgpr ||
-	                     call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
 	const auto graph = CFG::BuildGraph(source.decoded);
 	Program query;
 	query.stage = options.stage;
@@ -524,7 +520,11 @@ void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
 		return value;
 	};
 	Register = [&](uint32_t before, uint32_t code) -> Value {
-		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code);
+		std::vector<uint32_t> crossed_calls;
+		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code, crossed_calls);
+		for (const auto crossed: crossed_calls) {
+			std::ranges::find(source.calls, crossed, &ShaderSource::Call::instruction)->preserved_registers.set(code);
+		}
 		if (index == UINT32_MAX) {
 			EXIT_IF(code < query.user_data_base || code - query.user_data_base >= query.user_data_count);
 			return IREmitter(&block).GetUserData(static_cast<ScalarReg>(code));
@@ -557,18 +557,30 @@ void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
 		EXIT_IF(!producer.ready || !producer.values.contains(code));
 		return producer.values.at(code).Resolve();
 	};
-	for (uint32_t i = 0; i < 2u; ++i) {
-		const auto value = Register(source.call->instruction, call.src0.reg + i);
-		EXIT_IF(!ValidateRuntimeValue(query, value, RuntimeValueType::Integer));
-		auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
-		root.SetArg(0, value);
-		source.call->target[i] = Value(&root);
+	for (auto& site: source.calls) {
+		const auto& call = source.decoded.instructions[site.instruction];
+		EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
+		                     call.dst.kind != Decoder::OperandKind::Sgpr ||
+		                     call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
+		for (uint32_t i = 0; i < 2u; ++i) {
+			const auto value = Register(site.instruction, call.src0.reg + i);
+			if (!ValidateRuntimeValue(query, value, RuntimeValueType::Integer)) {
+				EXIT_IF(value.Instruction()->GetOpcode() != ValueOpcode::ReadConstBuffer);
+				source.unsupported_call_pc = call.pc;
+				source.calls.clear();
+				return;
+			}
+			auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
+			root.SetArg(0, value);
+			site.target[i] = Value(&root);
+		}
 	}
 	RewriteToSsa(query.blocks);
 	ConstantPropagationPass(query.blocks, query.wave_size);
 	RemoveIdentities(query);
 	EliminateDeadCode(query.blocks);
-	for (auto& value: source.call->target) value = value.Instruction()->Arg(0).Resolve();
+	for (auto& site: source.calls)
+		for (auto& value: site.target) value = value.Instruction()->Arg(0).Resolve();
 	query.value_storage.clear();
 	for (auto& inst: block) {
 		inst.SetParent(nullptr);
@@ -595,7 +607,7 @@ ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOp
 		Decoder::DecodeProgram(code, source.decoded);
 	}
 	if (source.decoded.has_swap_pc) {
-		PrepareCallTarget(source, options);
+		PrepareCallTargets(source, options);
 		source.code.assign(code.begin(), code.end());
 		source.decoded.code = source.code;
 	}
@@ -603,88 +615,122 @@ ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOp
 }
 
 const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime) {
+	EXIT_IF(source.unsupported_call_pc != UINT32_MAX);
 	auto& reads = source.reads;
 	reads.clear();
 	IR::SrtReadCapture capture(runtime, reads);
 	const auto clean = IR::CleanRuntime(capture.ObservedRuntime());
 	IR::SrtWalker walker(source.call_targets, clean);
-	uint32_t low = 0, high = 0;
-	const auto& call = source.decoded.instructions[source.call->instruction];
-	if (!walker.Evaluate(source.call->target[0], low) || !walker.Evaluate(source.call->target[1], high))
-		EXIT("shader call at pc 0x%08x has an unavailable scalar target", call.pc);
-	// SWAPPC reads the old aliased pair and ignores the target's low two bits.
-	const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
-	EXIT_IF(address == 0u);
+	for (auto& site: source.calls) {
+		uint32_t low = 0, high = 0;
+		if (!walker.Evaluate(site.target[0], low) || !walker.Evaluate(site.target[1], high))
+			EXIT("shader call at pc 0x%08x has an unavailable scalar target",
+			     source.decoded.instructions[site.instruction].pc);
+		// SWAPPC reads the old aliased pair and ignores the target's low two bits.
+		const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
+		EXIT_IF(address == 0u);
+		site.address = address;
+	}
 	const auto query_reads = reads.size();
 	if (source.linked) {
 		auto& linked = *source.linked;
-		const auto function = std::span(linked.code).subspan(source.code.size());
-		if (clean.read_memory(clean.userdata, address, linked.observed_function) &&
-		    std::ranges::equal(function, linked.observed_function)) {
-			return linked.decoded;
+		bool unchanged = true;
+		for (size_t i = 0; i < source.calls.size(); ++i) {
+			const auto& function = linked.functions[i];
+			const auto expected = std::span(linked.code).subspan(function.offset, function.size);
+			const auto preceding = std::span(source.calls).first(i);
+			const auto previous = std::ranges::find(preceding, source.calls[i].address, &ShaderSource::Call::address);
+			if (previous != preceding.end()) {
+				const auto& observed = linked.functions[previous - preceding.begin()];
+				if (std::ranges::equal(expected, std::span(linked.code).subspan(observed.offset, observed.size))) continue;
+				unchanged = false;
+				break;
+			}
+			linked.observed_function.resize(function.size);
+			if (!clean.read_memory(clean.userdata, source.calls[i].address, linked.observed_function) ||
+			    !std::ranges::equal(expected, linked.observed_function)) {
+				unchanged = false;
+				break;
+			}
 		}
+		if (unchanged) return linked.decoded;
 		reads.resize(query_reads);
 	}
-	ShaderSource::Linked linked {.code = source.code,
-	                             .decoded = source.decoded};
+	ShaderSource::Linked linked {.code = source.code, .decoded = source.decoded};
 	const auto& last = source.decoded.instructions.back();
 	EXIT_NOT_IMPLEMENTED(last.opcode != Decoder::Opcode::S_ENDPGM);
-	// Keep the caller's data/footer bytes at their native offsets. The leaf's virtual CFG
-	// PCs follow decoded instructions; its physical code is independently read and owned.
-	const auto function_pc = last.pc + last.word_count * sizeof(uint32_t);
-	for (uint32_t index = 0;;) {
-		// The observed queue helper is a scalar leaf. Scalar encodings need at most one literal.
-		EXIT_IF(index >= 64u || address > UINT64_MAX - (uint64_t {index} + 2u) * sizeof(uint32_t));
-		std::array<uint32_t, 2> words {};
-		if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
-		                       std::span(words).first(1)))
-			EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
-		const auto family = Decoder::GetInstructionFamily(words[0]);
-		EXIT_NOT_IMPLEMENTED(family != Decoder::Family::SOP1 && family != Decoder::Family::SOP2 &&
-		                     family != Decoder::Family::SOPC && family != Decoder::Family::SOPK &&
-		                     family != Decoder::Family::SOPP);
-		Decoder::Instruction inst;
-		Decoder::DecodeInstruction(words, 0u, inst);
-		if (inst.word_count == 2u) {
-			if (!clean.read_memory(clean.userdata, address + uint64_t {index + 1u} * sizeof(uint32_t),
-			                       std::span(words).last(1)))
-				EXIT("shader call at pc 0x%08x cannot read leaf literal", call.pc);
-			inst = {};
-			Decoder::DecodeInstruction(words, 0u, inst);
+	// Keep the caller's data/footer at native offsets. Expand leaf CFGs per call site
+	// before SSA, so return edges neither require a dispatcher nor merge resource roots.
+	uint32_t function_pc = last.pc + last.word_count * sizeof(uint32_t);
+	std::map<uint64_t, std::vector<Decoder::Instruction>> functions;
+	for (size_t i = 0; i < source.calls.size(); ++i) {
+		const auto& site = source.calls[i];
+		const auto& call = source.decoded.instructions[site.instruction];
+		const auto address = source.calls[i].address;
+		auto [entry, inserted] = functions.try_emplace(address);
+		auto& function = entry->second;
+		if (inserted) {
+			for (uint32_t index = 0;;) {
+				std::array<uint32_t, Decoder::MaxInstructionRawWords> words {};
+				EXIT_IF(index > UINT32_MAX / sizeof(uint32_t) - words.size() ||
+				        address > UINT64_MAX - (uint64_t {index} + words.size()) * sizeof(uint32_t));
+				uint32_t count = 1;
+				Decoder::Instruction inst;
+				for (;;) {
+					if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
+					                       std::span(words).first(count)))
+						EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
+					inst = {};
+					Decoder::DecodeInstruction(words, 0u, inst);
+					if (inst.word_count <= count) break;
+					count = inst.word_count;
+				}
+				inst.pc = index * sizeof(uint32_t);
+				if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
+				index += inst.word_count;
+				EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
+				                     inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
+				                     inst.opcode == Decoder::Opcode::S_ENDPGM);
+				function.push_back(inst);
+				if (inst.opcode == Decoder::Opcode::S_SETPC_B64) break;
+			}
 		}
-		inst.pc = index * sizeof(uint32_t);
-		if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
-		linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
-		index += inst.word_count;
-		EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
-		                     inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
-		                     inst.opcode == Decoder::Opcode::S_ENDPGM);
-		EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) || CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
-		const bool returns = inst.opcode == Decoder::Opcode::S_SETPC_B64;
-		if (returns) {
-			EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
-			inst.opcode = Decoder::Opcode::S_BRANCH;
-			inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
-		} else if (Decoder::IsDirectBranch(inst.opcode)) {
-			EXIT_IF(inst.branch_target > UINT32_MAX - function_pc);
-			inst.branch_target += function_pc;
+		const uint32_t function_size = function.back().pc + function.back().word_count * sizeof(uint32_t);
+		EXIT_IF(function_pc > UINT32_MAX - function_size);
+		const auto function_offset = static_cast<uint32_t>(linked.code.size());
+		for (auto inst: function) {
+			EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) ||
+			        CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
+			for (uint32_t code = 0; code < site.preserved_registers.size(); ++code) {
+				if (site.preserved_registers[code] && CFG::MayWriteScalarRegister(inst, code))
+					EXIT("shader call at pc 0x%08x clobbers scalar target dependency s%u", call.pc, code);
+			}
+			linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
+			if (inst.opcode == Decoder::Opcode::S_SETPC_B64) {
+				EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
+				inst.opcode = Decoder::Opcode::S_BRANCH;
+				inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
+			} else if (Decoder::IsDirectBranch(inst.opcode)) {
+				EXIT_IF(inst.branch_target >= function_size);
+				inst.branch_target += function_pc;
+			}
+			inst.pc += function_pc;
+			linked.decoded.instructions.push_back(std::move(inst));
 		}
-		inst.pc += function_pc;
-		linked.decoded.instructions.push_back(std::move(inst));
-		if (returns) break;
-	}
-	const auto function_end = linked.decoded.instructions.back().pc +
-	                          linked.decoded.instructions.back().word_count * sizeof(uint32_t);
-	for (size_t i = source.decoded.instructions.size(); i + 1u < linked.decoded.instructions.size(); ++i) {
-		const auto& inst = linked.decoded.instructions[i];
-		if (Decoder::IsDirectBranch(inst.opcode))
-			EXIT_IF(inst.branch_target < function_pc || inst.branch_target >= function_end);
+		linked.functions.push_back({function_offset, function_size / 4u});
+		linked.decoded.instructions[site.instruction].branch_target = function_pc;
+		function_pc += function_size;
 	}
 	reads.resize(query_reads);
-	reads.emplace_back(address, (linked.code.size() - source.code.size()) * sizeof(uint32_t));
-	linked.decoded.instructions[source.call->instruction].branch_target = function_pc;
+	for (size_t i = 0; i < source.calls.size(); ++i) {
+		const auto preceding = std::span(source.calls).first(i);
+		if (std::ranges::find(preceding, source.calls[i].address, &ShaderSource::Call::address) == preceding.end())
+			reads.emplace_back(source.calls[i].address, linked.functions[i].size * sizeof(uint32_t));
+	}
 	linked.decoded.code = linked.code;
-	linked.observed_function.resize(linked.code.size() - source.code.size());
+	linked.observed_function.resize(std::ranges::max_element(linked.functions, {},
+	    &ShaderSource::Linked::Function::size)->size);
+
 	source.linked = std::move(linked);
 	++source.revision;
 	return source.linked->decoded;

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -8,6 +8,7 @@
 #include "graphics/shader/shader.h"
 
 #include <array>
+#include <bitset>
 #include <optional>
 #include <span>
 #include <vector>
@@ -47,16 +48,21 @@ struct ShaderSource {
 	struct Call {
 		uint32_t instruction;
 		std::array<IR::Value, 2> target;
+		std::bitset<128> preserved_registers;
+		uint64_t address = 0;
 	};
 	struct Linked {
+		struct Function { uint32_t offset; uint32_t size; };
 		std::vector<uint32_t> code;
 		Decoder::Program decoded;
+		std::vector<Function> functions;
 		std::vector<uint32_t> observed_function;
 	};
 	std::vector<uint32_t> code;
 	Decoder::Program decoded;
 	IR::ResourcePlan call_targets;
-	std::optional<Call> call;
+	std::vector<Call> calls;
+	uint32_t unsupported_call_pc = UINT32_MAX;
 	std::optional<Linked> linked;
 	uint64_t revision = 0;
 	std::vector<std::pair<uint64_t, uint64_t>> reads;

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

@@ -160,19 +160,6 @@ uint32_t EmitFPMad32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
 	return sum;
 }
 
-uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
-	const auto min_ab   = EmitMinMaxF32Value(state, a, b, false);
-	const auto min3     = EmitMinMaxF32Value(state, min_ab, c, false);
-	const auto max_ab   = EmitMinMaxF32Value(state, a, b, true);
-	const auto high_min = EmitMinMaxF32Value(state, max_ab, c, false);
-	const auto median   = EmitMinMaxF32Value(state, min_ab, high_min, true);
-	const auto nan_ab   = Binary(state, spv::OpLogicalOr, TypeBool(state),
-	                             EmitClassifyF32(state, a).nan, EmitClassifyF32(state, b).nan);
-	const auto any_nan =
-	    Binary(state, spv::OpLogicalOr, TypeBool(state), nan_ab, EmitClassifyF32(state, c).nan);
-	return Select(state, TypeF32(state), any_nan, min3, median);
-}
-
 uint32_t EmitFindUMsb64(EmitterState& state, uint32_t value) {
 	// GLSL FindUMsb is restricted to 32-bit components.
 	const auto unpacked = Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), value);

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

@@ -100,10 +100,9 @@ uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask) {
 	return EmitClassMaskBits(state, bits, mask, true);
 }
 
-uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
-	const auto lhs_class = EmitClassifyF32(state, lhs);
-	const auto rhs_class = EmitClassifyF32(state, rhs);
-
+static uint32_t EmitMinMaxF32(EmitterState& state, uint32_t lhs, uint32_t rhs,
+                             const F32Class& lhs_class, const F32Class& rhs_class,
+                             bool max_value) {
 	const auto numeric_cond = state.builder.AllocateId();
 	state.builder.AddFunction(max_value ? spv::OpFOrdGreaterThanEqual : spv::OpFOrdLessThan,
 	                          TypeBool(state), numeric_cond, lhs, rhs);
@@ -121,6 +120,34 @@ uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, boo
 	return EmitBitcastU32ToF32(state, result_bits);
 }
 
+uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
+	const auto lhs_class = EmitClassifyF32(state, lhs);
+	const auto rhs_class = EmitClassifyF32(state, rhs);
+	return EmitMinMaxF32(state, lhs, rhs, lhs_class, rhs_class, max_value);
+}
+
+uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	const auto a_class = EmitClassifyF32(state, a);
+	const auto b_class = EmitClassifyF32(state, b);
+	const auto c_class = EmitClassifyF32(state, c);
+	const auto min_ab = EmitMinMaxF32(state, a, b, a_class, b_class, false);
+	const auto min_ab_class = EmitClassifyF32(state, min_ab);
+	const auto min3 = EmitMinMaxF32(state, min_ab, c, min_ab_class, c_class, false);
+	const auto max_ab = EmitMinMaxF32(state, a, b, a_class, b_class, true);
+	const auto max_ab_class = EmitClassifyF32(state, max_ab);
+	const auto max3 = EmitMinMaxF32(state, max_ab, c, max_ab_class, c_class, true);
+	const auto max_bc = EmitMinMaxF32(state, b, c, b_class, c_class, true);
+	const auto max_ac = EmitMinMaxF32(state, a, c, a_class, c_class, true);
+	// RDNA2 compares numerically against MAX3, including ties between signed zeros.
+	const auto a_is_max = Binary(state, spv::OpFOrdEqual, TypeBool(state), max3, a);
+	const auto b_is_max = Binary(state, spv::OpFOrdEqual, TypeBool(state), max3, b);
+	const auto median = Select(state, TypeF32(state), a_is_max, max_bc,
+	                           Select(state, TypeF32(state), b_is_max, max_ac, max_ab));
+	const auto nan_ab = EmitLogicalOrBool(state, a_class.nan, b_class.nan);
+	const auto any_nan = EmitLogicalOrBool(state, nan_ab, c_class.nan);
+	return Select(state, TypeF32(state), any_nan, min3, median);
+}
+
 uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {
 	const auto bits      = state.builder.AllocateId();
 	const auto abs_bits  = state.builder.AllocateId();

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

@@ -655,6 +655,14 @@ void EmitStoreCompletion(EmitterState& state) {
 	                          spv::MemorySemanticsImageMemoryMask));
 }
 
+void EmitDataCacheInvalidate(EmitterState& state) {
+	// RDNA2 GL0/GL1 invalidation makes vector memory visible without a workgroup rendezvous.
+	state.builder.AddFunction(
+	    spv::OpMemoryBarrier, ConstantU32(state, spv::ScopeDevice),
+	    ConstantU32(state, spv::MemorySemanticsAcquireMask | spv::MemorySemanticsUniformMemoryMask |
+	                          spv::MemorySemanticsImageMemoryMask));
+}
+
 void EmitBarrier(EmitterState& state) {
 	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
 	if (!tessellation && ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {

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

@@ -97,10 +97,14 @@ uint32_t GatherMip(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memory
 		              ConstantF32Value(state, 1.0f / 256.0f));
 	};
 	const auto sampler_control = ctx.Arg(inst, 6);
+	const auto mip_filter = Extract(sampler_control, 26, 2);
 	const auto preclamp = Binary(state, spv::OpINotEqual, TypeBool(state),
 	                            Extract(sampler_control, 28, 1), ConstantU32(state, 0));
-	const auto half = ConstantF32Value(state, 0.5f);
 	const auto zero = ZeroF32(state);
+	// Point selects the nearest mip; linear gathers return the lower, primary mip only.
+	const auto rounding = Select(state, TypeF32(state),
+	    Binary(state, spv::OpIEqual, TypeBool(state), mip_filter, ConstantU32(state, 1)),
+	    ConstantF32Value(state, 0.5f), zero);
 	const auto base = Unary(state, spv::OpConvertUToF, TypeF32(state),
 	                        Extract(ctx.Arg(inst, 4), 12, 4));
 	const auto minimum = Binary(state, spv::OpFSub, TypeF32(state),
@@ -108,19 +112,19 @@ uint32_t GatherMip(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memory
 	const auto last = ConstantF32Value(state, static_cast<float>(mip_count - 1u));
 	auto lod = Binary(state, spv::OpFAdd, TypeF32(state),
 	                  AddressF32(ctx, mem, address, layout.lod),
-	                  Select(state, TypeF32(state), preclamp, half, zero));
+	                  Select(state, TypeF32(state), preclamp, rounding, zero));
 	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(
 	    state, lod, LodFixed(ctx.Arg(inst, 5), 0), LodFixed(ctx.Arg(inst, 5), 12));
 	// T# minimum LOD is in physical mip space and applies after the S# view-relative clamp.
 	lod = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(state, lod, minimum);
 	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(state, lod, zero, last);
 	lod = Binary(state, spv::OpFAdd, TypeF32(state), lod,
-	             Select(state, TypeF32(state), preclamp, zero, half));
+	             Select(state, TypeF32(state), preclamp, zero, rounding));
 	// The clamped value is nonnegative, so conversion performs the integer mip rounding.
 	const auto mip = EmitUMin32(state, Unary(state, spv::OpConvertFToU, TypeU32(state), lod),
 	                           ConstantU32(state, mip_count - 1u));
 	const auto mip_none = Binary(state, spv::OpIEqual, TypeBool(state),
-	                            Extract(sampler_control, 26, 2), ConstantU32(state, 0));
+	                            mip_filter, ConstantU32(state, 0));
 	return Select(state, TypeU32(state), mip_none, ConstantU32(state, 0), mip);
 }
 
@@ -284,10 +288,10 @@ uint32_t ResultVector(ValueEmitContext& ctx, uint32_t value,
 }
 
 uint32_t QueryDimensions(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
-                         const IR::Inst& address) {
-	const auto  dimension = ctx.state.program.info.images.at(mem.resource).dimension;
+                         const IR::Inst& address, uint32_t resource, uint32_t array_index) {
+	const auto  dimension = ctx.state.program.info.images.at(resource).dimension;
 	const auto& info      = ImageDimensionInfoFor(dimension);
-	const auto  image     = LoadImageDescriptor(ctx.state, mem.resource);
+	const auto  image     = LoadImageDescriptor(ctx.state, resource, 0u, array_index);
 	const auto  size      = ctx.state.builder.AllocateId();
 	if (info.multisampled != 0u) {
 		ctx.state.builder.AddFunction(spv::OpImageQuerySize,
@@ -689,68 +693,58 @@ auto EmitImageAccess(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t resul
 	}
 	const auto selected = EmitIndirectResourceIndex(
 	    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
-	struct ImageRun {
-		uint32_t first;
-		uint32_t count;
-		uint32_t resource;
-		uint32_t slot_bias;
+	const auto SameType = [&](uint32_t lhs, uint32_t rhs) {
+		const auto& first = state.program.info.images[lhs];
+		const auto& second = state.program.info.images[rhs];
+		return first.dimension == second.dimension && first.cube == second.cube &&
+		       IR::DescriptorBindingForImage(first) == IR::DescriptorBindingForImage(second) &&
+		       (inst.GetOpcode() != IR::ValueOpcode::ImageWrite ||
+		        first.shader_swizzle == second.shader_swizzle) &&
+		       first.mip_count == 1u && second.mip_count == 1u;
 	};
-	std::vector<ImageRun> runs;
-	for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
-		const auto resource = image.indirect_resources[ordinal];
-		const auto& candidate = state.program.info.images[resource];
-		const auto kind = *IR::DescriptorBindingForImage(candidate);
-		if (!runs.empty()) {
-			auto& run = runs.back();
-			const auto& first = state.program.info.images[run.resource];
-			const auto next_slot = run.slot_bias + ordinal;
-			if (candidate.dimension == first.dimension && candidate.cube == first.cube &&
-			    IR::DescriptorBindingForImage(first) == kind &&
-			    (inst.GetOpcode() != IR::ValueOpcode::ImageWrite ||
-			     candidate.shader_swizzle == first.shader_swizzle) &&
-			    candidate.mip_count == 1u && first.mip_count == 1u &&
-			    (ordinal == 1u || candidate.descriptor_index == next_slot)) {
-				// Native roots precede appended children; only the root may have a slot gap.
-				if (ordinal == 1u)
-					run.slot_bias = candidate.descriptor_index - ordinal;
-				++run.count;
-				continue;
-			}
-		}
-		runs.push_back({ordinal, 1u, resource,
-		                candidate.descriptor_index - ordinal});
+	// Descriptor order is unrelated to image type. Share each typed access and
+	// select its descriptor slot independently, including gaps between native roots.
+	struct ImageGroup { uint32_t resource; uint32_t count; };
+	std::vector<ImageGroup> groups;
+	std::vector<uint32_t> ordinal_groups;
+	for (const auto resource: image.indirect_resources) {
+		const auto found = std::ranges::find_if(groups, [&](const ImageGroup& group) {
+			return SameType(resource, group.resource);
+		});
+		ordinal_groups.push_back(static_cast<uint32_t>(found - groups.begin()));
+		if (found == groups.end()) groups.push_back({resource, 1u});
+		else ++found->count;
 	}
-	const auto EmitRun = [&](uint32_t index) {
-		const auto& run = runs[index];
-		if (run.count == 1u) return emit(run.resource, 0u);
-		auto slot = Binary(state, spv::OpIAdd, TypeU32(state), selected,
-		                   ConstantU32(state, run.slot_bias));
-		if (run.first == 0u) {
-			const auto root_slot = image.descriptor_index;
-			if (root_slot != run.slot_bias) {
-				const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
-				                            ConstantU32(state, 0u));
-				const auto root_or_child = state.builder.AllocateId();
-				state.builder.AddFunction(spv::OpSelect, TypeU32(state), root_or_child,
-				                          is_root, ConstantU32(state, root_slot), slot);
-				slot = root_or_child;
-			}
-		}
-		return emit(run.resource, slot);
-	};
-	auto run_index = ConstantU32(state, 0u);
-	for (uint32_t index = 1; index < runs.size(); ++index) {
+	const bool dynamic_slot = std::ranges::any_of(groups, [](const ImageGroup& group) {
+		return group.count > 1u;
+	});
+	auto group_index = ConstantU32(state, 0u);
+	auto slot_bias = ConstantU32(state, image.descriptor_index);
+	auto previous_bias = image.descriptor_index;
+	for (uint32_t ordinal = 1; ordinal < image.indirect_resources.size(); ++ordinal) {
+		const auto& candidate = state.program.info.images[image.indirect_resources[ordinal]];
+		const auto bias = candidate.descriptor_index - ordinal;
+		const bool select_group = ordinal_groups[ordinal] != ordinal_groups[ordinal - 1u];
+		const bool select_slot = dynamic_slot && bias != previous_bias;
+		previous_bias = bias;
+		if (!select_group && !select_slot) continue;
 		const auto at_or_after = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state),
-		                               selected, ConstantU32(state, runs[index].first));
-		const auto next = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpSelect, TypeU32(state), next, at_or_after,
-		                          ConstantU32(state, index), run_index);
-		run_index = next;
+		                               selected, ConstantU32(state, ordinal));
+		if (select_group) group_index = Select(state, TypeU32(state), at_or_after,
+		                                      ConstantU32(state, ordinal_groups[ordinal]), group_index);
+		if (select_slot) slot_bias = Select(state, TypeU32(state), at_or_after,
+		                                   ConstantU32(state, bias), slot_bias);
 	}
-	return runs.size() == 1u
-	           ? EmitRun(0u)
-	           : EmitIndexSwitch(state, run_index, static_cast<uint32_t>(runs.size()),
-	                             result_type, EmitRun);
+	const auto slot = dynamic_slot
+	    ? Binary(state, spv::OpIAdd, TypeU32(state), selected, slot_bias) : 0u;
+	const auto EmitGroup = [&](uint32_t index) {
+		const auto& group = groups[index];
+		return emit(group.resource, group.count == 1u ? 0u : slot);
+	};
+	return groups.size() == 1u
+	           ? EmitGroup(0u)
+	           : EmitIndexSwitch(state, group_index, static_cast<uint32_t>(groups.size()),
+	                             result_type, EmitGroup);
 }
 
 template <typename Emit>
@@ -852,7 +846,10 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto* address = ctx.ImageAddress(inst.Arg(image_info.needs_sampler ? 2 : 1));
 	if (op == IR::ValueOpcode::ImageQueryDimensions) {
 		state.builder.RequireCapability(spv::CapabilityImageQuery);
-		ctx.Define(inst, QueryDimensions(ctx, mem, *address));
+		ctx.Define(inst, EmitImageAccess(ctx, inst, TypeU32Vector(state, 4),
+		    [&](uint32_t resource, uint32_t array_index) {
+			    return QueryDimensions(ctx, mem, *address, resource, array_index);
+		    }));
 		return;
 	}
 	if (image_info.needs_sampler &&

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

@@ -218,6 +218,7 @@ inline constexpr auto EmitTtraceData   = EmitVoid;
 inline constexpr auto EmitInstPrefetch = EmitVoid;
 void                  EmitBarrier(EmitterState& state);
 void                  EmitStoreCompletion(EmitterState& state);
+void                  EmitDataCacheInvalidate(EmitterState& state);
 void                  EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);

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

@@ -307,6 +307,7 @@ void EmitBlock(ValueEmitContext& ctx, const IR::Block* block, EmitInstruction&&
 			ctx.state.lane_half = half;
 			if (half == 0 || (inst.GetOpcode() != IR::ValueOpcode::Barrier &&
 			                  inst.GetOpcode() != IR::ValueOpcode::StoreCompletion &&
+			                  inst.GetOpcode() != IR::ValueOpcode::DataCacheInvalidate &&
 			                  inst.GetOpcode() != IR::ValueOpcode::MeshAllocate)) {
 				emit_instruction(lane, inst);
 			}

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

@@ -8,6 +8,8 @@
 #include <deque>
 #include <list>
 #include <map>
+#include <optional>
+#include <set>
 #include <span>
 
 namespace Libs::Graphics::ShaderRecompiler::CFG {
@@ -1650,6 +1652,32 @@ Graph BuildGraph(const Decoder::Program& program) {
 	}
 	SortUnique(graph.code_table_load_pcs);
 
+	if (!indirect_setpc && std::ranges::any_of(program.instructions, [](const auto& inst) {
+		    return inst.opcode == Opcode::S_SWAPPC_B64 && inst.branch_target != UINT32_MAX;
+		})) {
+		// Linked leaf bodies follow the caller in native PC order. Visit them in
+		// execution order so returning to the caller does not manufacture goto loops.
+		std::vector<bool> visited(graph.blocks.size(), false);
+		std::vector<std::pair<uint32_t, bool>> pending {{graph.entry_block, false}};
+		std::vector<BasicBlock> ordered;
+		ordered.reserve(graph.blocks.size());
+		while (!pending.empty()) {
+			const auto [id, finished] = pending.back();
+			pending.pop_back();
+			if (finished) {
+				ordered.push_back(std::move(graph.blocks[id]));
+				continue;
+			}
+			if (visited[id]) continue;
+			visited[id] = true;
+			pending.emplace_back(id, true);
+			for (const auto successor: graph.blocks[id].successors)
+				pending.emplace_back(successor, false);
+		}
+		std::ranges::reverse(ordered);
+		ApplyBlockOrder(graph, std::move(ordered));
+	}
+
 	RecomputeAnalyses(graph);
 
 	if (indirect_setpc) {
@@ -1681,33 +1709,56 @@ Graph Structurize(const Graph& graph) {
 }
 
 bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code) {
-	uint32_t destination = UINT32_MAX;
-	// Scalar ALU pair widths are not all decoded. The typed producer must confirm this word.
-	return InstructionWritesScalarCode(inst, code) ||
-	       (ScalarOperandCode(inst.dst, destination) && code == destination + 1u);
+	// Vector comparisons and carry results can write a scalar pair in wave64.
+	if (inst.family == Decoder::Family::VOP1 || inst.family == Decoder::Family::VOP2 ||
+	    inst.family == Decoder::Family::VOPC || inst.family == Decoder::Family::VOP3 ||
+	    inst.family == Decoder::Family::VOP3P) {
+		uint32_t destination = 0;
+		for (const auto* operand: {&inst.dst, &inst.dst2})
+			if (ScalarOperandCode(*operand, destination) && code >= destination && code <= destination + 1u)
+				return true;
+	}
+	return InstructionWritesScalarCode(inst, code);
 }
 
 uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
-                              uint32_t before, uint32_t code) {
+                              uint32_t before, uint32_t code, std::vector<uint32_t>& crossed_calls) {
 	EXIT_IF(before >= program.instructions.size());
 	const auto found = std::ranges::find_if(graph.blocks, [before](const auto& block) {
 		return block.inst_begin <= before && before < block.inst_end;
 	});
 	EXIT_IF(found == graph.blocks.end());
-	const auto* block = &*found;
-	for (size_t depth = 0; depth < graph.blocks.size(); ++depth) {
-		for (uint32_t i = before; i > block->inst_begin;) {
-			const auto& inst = program.instructions[--i];
-			EXIT_IF(inst.opcode == Opcode::S_SWAPPC_B64);
-			if (MayWriteScalarRegister(inst, code)) return i;
-		}
-		if (block->id == graph.entry_block) return UINT32_MAX;
-		EXIT_IF(block->predecessors.size() != 1u);
-		block = graph.FindBlock(block->predecessors[0]);
+	std::vector<std::pair<uint32_t, uint32_t>> pending {{found->id, before}};
+	std::set<std::pair<uint32_t, uint32_t>> visited;
+	std::optional<uint32_t> definition;
+	while (!pending.empty()) {
+		const auto point = pending.back();
+		pending.pop_back();
+		if (!visited.insert(point).second) continue;
+		const auto* block = graph.FindBlock(point.first);
 		EXIT_IF(block == nullptr);
-		before = block->inst_end;
+		uint32_t index = UINT32_MAX;
+		for (uint32_t i = point.second; i > block->inst_begin;) {
+			const auto& inst = program.instructions[--i];
+			if (inst.opcode == Opcode::S_SWAPPC_B64) {
+				EXIT_NOT_IMPLEMENTED(MayWriteScalarRegister(inst, code));
+				crossed_calls.push_back(i);
+				continue;
+			}
+			if (MayWriteScalarRegister(inst, code)) { index = i; break; }
+		}
+		if (index != UINT32_MAX || block->id == graph.entry_block) {
+			EXIT_IF(definition && *definition != index);
+			definition = index;
+			continue;
+		}
+		EXIT_IF(block->predecessors.empty());
+		for (const auto predecessor: block->predecessors) {
+			pending.emplace_back(predecessor, graph.blocks[predecessor].inst_end);
+		}
 	}
-	EXIT("scalar shader call source has cyclic reaching definitions");
+	EXIT_IF(!definition);
+	return *definition;
 }
 
 std::string BranchConditionToString(BranchCondition condition) {

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

@@ -115,7 +115,7 @@ struct Graph {
 Graph       BuildGraph(const Decoder::Program& program);
 bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code);
 uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
-                              uint32_t before, uint32_t code);
+                              uint32_t before, uint32_t code, std::vector<uint32_t>& crossed_calls);
 // 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);

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

@@ -70,6 +70,8 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {0x52u, Opcode::BUFFER_ATOMIC_ADD_X2, 2, 32},
     {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
     {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
+    {0x71u, Opcode::BUFFER_GL0_INV, 0, 0},
+    {0x72u, Opcode::BUFFER_GL1_INV, 0, 0},
     {0x80u, Opcode::BUFFER_LOAD_FORMAT_D16_X, 1, 16, false, false, true},
     {0x83u, Opcode::BUFFER_LOAD_FORMAT_D16_XYZW, 2, 16, false, false, true},
     {0x84u, Opcode::BUFFER_STORE_FORMAT_D16_X, 1, 16, false, false, true},
@@ -312,6 +314,9 @@ void DecodeMubuf(uint32_t pc, std::span<const uint32_t> code, uint32_t word_inde
 	if (inst.opcode == Opcode::UNSUPPORTED) {
 		SetUnsupported(inst, Family::MUBUF, opcode, "MUBUF opcode is not implemented");
 	}
+	if (inst.opcode == Opcode::BUFFER_GL0_INV || inst.opcode == Opcode::BUFFER_GL1_INV) {
+		return;
+	}
 
 	DecodeVectorGpr(vdata, inst.dst);
 	if (inst.opcode == Opcode::BUFFER_LOAD_FORMAT_D16_XYZW ||

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

@@ -164,9 +164,23 @@ void DecodeSop1(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	switch (inst.opcode) {
 		case Opcode::S_SWAPPC_B64:
 			inst.branch_target = UINT32_MAX;
-			inst.data_dwords = 2;
+			[[fallthrough]];
+		case Opcode::S_MOV_B64:
+		case Opcode::S_CMOV_B64:
+		case Opcode::S_NOT_B64:
+		case Opcode::S_WQM_B64:
+		case Opcode::S_BREV_B64:
+		case Opcode::S_BITSET0_B64:
+		case Opcode::S_BITSET1_B64:
+		case Opcode::S_AND_SAVEEXEC_B64:
+		case Opcode::S_ORN2_SAVEEXEC_B64:
+		case Opcode::S_QUADMASK_B64:
+		case Opcode::S_ANDN1_SAVEEXEC_B64:
+		case Opcode::S_BITREPLICATE_B64_B32:
+			inst.data_dwords = 2u;
 			break;
 		case Opcode::S_GETPC_B64:
+			inst.data_dwords = 2u;
 			inst.src_count = 0;
 			DecodeScalarDestination(sdst, pc, inst.dst);
 			return;
@@ -206,6 +220,8 @@ void DecodeSop2(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 
 	DecodeScalarDestination(sdst, pc, inst.dst);
 	DecodeBinarySources(pc, code, word_index, inst, ssrc0, ssrc1);
+	if (opcode == 0x0bu || (opcode >= 0x0fu && opcode <= 0x25u && (opcode & 1u)) || opcode == 0x29u)
+		inst.data_dwords = 2u;
 }
 
 void DecodeSopk(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,

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

@@ -566,6 +566,9 @@ std::string InstructionToString(const Instruction& inst) {
 			                                               OperandToString(inst.src0).c_str()));
 		case Opcode::S_BARRIER:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: s_barrier", inst.pc));
+		case Opcode::BUFFER_GL0_INV:
+		case Opcode::BUFFER_GL1_INV:
+			return fmt::format("0x{:08x}: {}", inst.pc, magic_enum::enum_name(inst.opcode));
 		case Opcode::V_NOP:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: v_nop", inst.pc));
 		case Opcode::S_ENDPGM:

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

@@ -599,6 +599,8 @@ enum class Opcode {
 	BUFFER_ATOMIC_XOR,
 	BUFFER_ATOMIC_FMIN,
 	BUFFER_ATOMIC_FMAX,
+	BUFFER_GL0_INV,
+	BUFFER_GL1_INV,
 	FLAT_LOAD_UBYTE,
 	FLAT_LOAD_SBYTE,
 	FLAT_LOAD_USHORT,

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

@@ -940,6 +940,10 @@ void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
+		case Decoder::Opcode::BUFFER_GL0_INV:
+		case Decoder::Opcode::BUFFER_GL1_INV:
+			ir.Emit(IR::ValueOpcode::DataCacheInvalidate);
+			return;
 		case Decoder::Opcode::IMAGE_BVH_INTERSECT_RAY: {
 			const auto result = ir.Emit(IR::ValueOpcode::BvhIntersect,
 			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()},

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

@@ -495,7 +495,6 @@ struct ShaderInfo {
 	static constexpr uint32_t MaxBufferRoots  = 64;
 	static constexpr uint32_t MaxImages       = 64;
 	static constexpr uint32_t MaxSamplers     = 32;
-	static constexpr uint32_t MaxSampledPairs = 64;
 
 	std::vector<BufferResource>      buffers;
 	std::vector<ImageResource>       images;

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

@@ -94,6 +94,7 @@ bool HasSideEffects(ValueOpcode opcode) {
 		case ValueOpcode::SetTessellationAttribute:
 		case ValueOpcode::MeshAllocate:
 		case ValueOpcode::StoreCompletion:
+		case ValueOpcode::DataCacheInvalidate:
 		case ValueOpcode::Barrier: return true;
 		default: return false;
 	}

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

@@ -350,6 +350,7 @@ VALUE_OPCODE(SetAttribute, Void, U32x4, U1)
 
 VALUE_OPCODE(ControlNop, Void)
 VALUE_OPCODE(StoreCompletion, Void)
+VALUE_OPCODE(DataCacheInvalidate, Void)
 VALUE_OPCODE(Barrier, Void)
 VALUE_OPCODE(Sendmsg, Void)
 VALUE_OPCODE(TtraceData, Void)

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

@@ -1029,6 +1029,19 @@ void ConstantPropagationPass(const BlockList& blocks, uint32_t wave_size) {
 		}
 		drain();
 	} while (normalized);
+	// Repeated register reads share a bitcast only within the same SSA block.
+	std::unordered_map<const Inst*, Inst*> bitcasts;
+	for (auto* block: blocks) {
+		bitcasts.clear();
+		for (auto& inst: *block) {
+			if (inst.GetOpcode() != ValueOpcode::BitCastF32U32 &&
+			    inst.GetOpcode() != ValueOpcode::BitCastU32F32) continue;
+			const auto* source = Arg(inst, 0).TryInstruction();
+			if (source == nullptr) continue;
+			const auto [entry, inserted] = bitcasts.try_emplace(source, &inst);
+			if (!inserted) Replace(inst, Value(entry->second));
+		}
+	}
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -30,32 +30,62 @@ void RemoveIdentities(Program& program) {
 }
 
 void EliminateDeadCode(const BlockList& blocks) {
-	std::vector<Inst*> pending;
-	for (auto* block: blocks) {
-		for (auto& inst: *block) {
-			// Retained resource-planning instructions live outside the block list.
-			inst.live = inst.MayHaveSideEffects() || std::ranges::any_of(inst.Uses(),
-			    [](const Use& use) { return use.user->Parent() == nullptr; });
-			if (inst.live) pending.push_back(&inst);
-		}
-	}
-	while (!pending.empty()) {
-		const auto* inst = pending.back();
-		pending.pop_back();
-		for (size_t arg = 0; arg < inst->NumArgs(); ++arg) {
-			auto* input = inst->Arg(arg).TryInstruction();
-			if (input != nullptr && input->Parent() != nullptr && !input->live) {
-				input->live = true;
-				pending.push_back(input);
+	bool changed;
+	do {
+		std::vector<Inst*> pending;
+		for (auto* block: blocks) {
+			for (auto& inst: *block) {
+				// Retained resource-planning instructions live outside the block list.
+				inst.live = inst.MayHaveSideEffects() || std::ranges::any_of(inst.Uses(),
+				    [](const Use& use) { return use.user->Parent() == nullptr; });
+				if (inst.live) pending.push_back(&inst);
 			}
 		}
-	}
-	// Unlink every dead operand before freeing nodes in a dead Phi cycle.
-	for (auto* block: blocks)
-		for (auto& inst: *block)
-			if (!inst.live) inst.Invalidate();
-	for (auto* block: blocks)
-		std::erase_if(block->Instructions(), [](const Inst& inst) { return !inst.live; });
+		while (!pending.empty()) {
+			const auto* inst = pending.back();
+			pending.pop_back();
+			for (size_t arg = 0; arg < inst->NumArgs(); ++arg) {
+				auto* input = inst->Arg(arg).TryInstruction();
+				if (input != nullptr && input->Parent() != nullptr && !input->live) {
+					input->live = true;
+					pending.push_back(input);
+				}
+			}
+		}
+		// Unlink every dead operand before freeing nodes in a dead Phi cycle.
+		for (auto* block: blocks)
+			for (auto& inst: *block)
+				if (!inst.live) inst.Invalidate();
+		for (auto* block: blocks)
+			std::erase_if(block->Instructions(), [](const Inst& inst) { return !inst.live; });
+
+		changed = false;
+		for (auto* block: blocks) {
+			const auto& term = block->terminator;
+			if (term.kind != CFG::TerminatorKind::ConditionalBranch || term.loop_header ||
+			    block->condition.IsImmediate()) continue;
+			for (const bool take_true: {false, true}) {
+				const auto* empty = take_true ? term.false_block : term.true_block;
+				const auto* merge = take_true ? term.true_block : term.false_block;
+				if (empty == nullptr || merge == nullptr || !empty->empty() ||
+				    empty->terminator.loop_header || merge->terminator.loop_header ||
+				    empty->terminator.kind != CFG::TerminatorKind::Branch ||
+				    empty->terminator.true_block != merge ||
+				    std::ranges::any_of(*merge, [](const Inst& inst) {
+					    return inst.GetOpcode() == ValueOpcode::Phi;
+				    })) continue;
+				// Keep the CFG and SSA edges intact; only the empty detour's predicate is dead.
+				for (auto& inst: *block) {
+					if (inst.GetOpcode() == ValueOpcode::Reference &&
+					    inst.Arg(0).Resolve() == block->condition.Resolve())
+						inst.SetArg(0, Value(take_true));
+				}
+				block->condition = Value(take_true);
+				changed = true;
+				break;
+			}
+		}
+	} while (changed);
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -292,11 +292,23 @@ bool NormalizeIndirectBuffer(DescriptorValue& value, bool indexed, const BufferR
 bool IsBoundedDescriptorTable(const ResourcePlan& program,
                               const DescriptorSource::IndirectDescriptor& indirect) {
 	const auto* table = Source(program, indirect.table_source);
-	return indirect.sources.empty() && !indirect.selector &&
+	return indirect.sources.empty() && (!indirect.selector || indirect.table_scalar) &&
 	       (indirect.key_count.IsEmpty() || indirect.table_maximum) && indirect.workgroup_axis == UINT32_MAX &&
 	       table != nullptr && table->dword_count == 4u;
 }
 
+bool CanReadIndexedSelector(const ShaderBufferResource& material, uint32_t offset) {
+	if (material.Type() != 0u || material.SwizzleEnabled() || material.AddTid() ||
+	    material.OutOfBounds() != 0u) return false;
+	if (material.Format() == Prospero::BufferFormat::kInvalid || material.NumRecords() == 0u ||
+	    offset >= material.Stride()) return true;
+	return material.Base48() != 0u &&
+	       ((material.Base48() | material.Stride() | offset) & 3u) == 0u &&
+	       material.NumRecords() <= MaxIndirectDescriptorProbes &&
+	       uint64_t {material.NumRecords() - 1u} * material.Stride() + offset + 4u <=
+	           uint64_t {UINT32_MAX} + 1u;
+}
+
 bool EvaluateIndirectCount(const DescriptorSource::IndirectDescriptor& indirect,
                            const SrtRuntime& runtime, SrtWalker& clean, uint32_t& count) {
 	if (indirect.workgroup_axis != UINT32_MAX) {
@@ -490,6 +502,15 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 			keys.push_back(key);
 			mask &= mask - 1u;
 		}
+	} else if (selector->stride == 0u) {
+		ShaderBufferResource material;
+		if (!DecodeBufferDescriptor(material_value, material) ||
+		    !CanReadIndexedSelector(material, selector->offset)) return false;
+		if (material.Format() != Prospero::BufferFormat::kInvalid && material.NumRecords() != 0u &&
+		    selector->offset < material.Stride() &&
+		    !read_keys(material.Base48(), material.GetSize(), selector->offset,
+		               material.Stride(), material.NumRecords())) return false;
+		keys.push_back(0u); // Every out-of-bounds indexed load returns zero.
 	} else {
 		ShaderBufferResource material;
 		if (!DecodeBufferDescriptor(material_value, material) || selector->stride < 4u ||
@@ -524,29 +545,29 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
                                    ResourceSnapshot&             snapshot,
                                    std::vector<DescriptorValue>& descriptors,
                                    std::vector<Specialization>&  specializations,
-                                   uint32_t maximum_resources, Normalize&& normalize) {
+                                   Normalize&& normalize) {
 	const auto& indirect = *program.descriptor_sources[source_index].indirect_descriptor;
 	const auto  descriptor_bytes = read_dwords * sizeof(uint32_t);
 	const auto& sources = indirect.sources;
 	const auto* selector = indirect.selector ? &*indirect.selector : nullptr;
+	const bool byte_keys = IsBoundedDescriptorTable(program, indirect);
 	auto* selector_keys = selector != nullptr
 	    ? &program.selector_keys[program.selector_groups[source_index]] : nullptr;
 	auto& keys = selector_keys != nullptr ? selector_keys->values : program.material_keys;
+	auto& byte_offsets = program.material_keys;
 	const auto children_begin = descriptors.size();
 	const auto root_resource = specializations[resource_index];
-	const auto intern_candidate = [&](const DescriptorValue& candidate, uint32_t& ordinal) {
-		ordinal = 0u;
-		if (candidate == descriptors[resource_index]) return true;
+	const auto intern_candidate = [&](const DescriptorValue& candidate) {
+		if (candidate == descriptors[resource_index]) return 0u;
 		const auto found = std::find(descriptors.begin() + children_begin, descriptors.end(), candidate);
-		ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
+		const auto ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
 		if (found == descriptors.end()) {
-			if (descriptors.size() >= maximum_resources) return false;
 			descriptors.push_back(candidate);
 			auto child = root_resource;
 			child.indirect_root = resource_index;
 			specializations.push_back(child);
 		}
-		return true;
+		return ordinal;
 	};
 	uint64_t table_base = 0;
 	uint64_t table_size = UINT64_MAX; // Scalar addresses have no buffer descriptor bounds.
@@ -558,7 +579,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		    "indirect {} resource={} source={} table_source={} words={:08x},{:08x},{:08x},{:08x} "
 		    "base={:#x} size={} type={} runtime_stride={} records={} swizzle={} addtid={} oob={} "
 		    "maximum={} offset={} immediate={} record_bytes={} child_count={} child_stride={} "
-		    "child_offset={} indexed_pointer={} candidates={} capacity={} {}",
+		    "child_offset={} indexed_pointer={} candidates={} {}",
 		    reason, resource_index, source_index, indirect.table_source,
 		    table_value.dwords[0], table_value.dwords[1], table_value.dwords[2], table_value.dwords[3],
 		    table_base, table_size, table.Type(), table.Stride(), table.NumRecords(),
@@ -567,7 +588,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		    indirect.table_offset, indirect.table_immediate, indirect.table_record_bytes,
 		    pointer != nullptr ? pointer->count : 1u, pointer != nullptr ? pointer->stride : 0u,
 		    pointer != nullptr ? pointer->offset : 0u, pointer != nullptr && pointer->indexed,
-		    descriptors.size(), maximum_resources, details));
+		    descriptors.size(), details));
 	};
 	const auto read_failed = [&](std::string_view reason, uint32_t row, uint32_t child,
 	                             uint64_t base, uint64_t offset, size_t bytes) {
@@ -583,7 +604,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 	if (sources.empty()) {
 		if (selector == nullptr) keys.clear();
 		DescriptorValue material_value;
-		if ((selector != nullptr && !selector_keys->resolved &&
+		if ((selector != nullptr && (byte_keys || !selector_keys->resolved) &&
 		     !clean.EvaluateDescriptor(selector->source, material_value)) ||
 		    !clean.EvaluateDescriptor(indirect.table_source, table_value)) {
 			return false;
@@ -596,16 +617,40 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		} else {
 			return false;
 		}
-		if (IsBoundedDescriptorTable(program, indirect)) {
+		if (byte_keys && table_size > UINT32_MAX) return failed("table geometry");
+		bool use_selector = selector != nullptr;
+		if (byte_keys && selector != nullptr) {
+			ShaderBufferResource material;
+			if (!DecodeBufferDescriptor(material_value, material)) return false;
+			use_selector = CanReadIndexedSelector(material, selector->offset);
+			if (use_selector && material.Format() != Prospero::BufferFormat::kInvalid &&
+			    material.NumRecords() != 0u && selector->offset < material.Stride()) {
+				const auto begin = material.Base48() + selector->offset;
+				const auto end = begin + uint64_t {material.NumRecords() - 1u} * material.Stride() + 4u;
+				for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
+					const auto& buffer = program.info.buffers[i];
+					if (!buffer.written) continue;
+					ShaderBufferResource written;
+					if (program.descriptor_sources[buffer.source].indirect_descriptor ||
+					    !DecodeBufferDescriptor(snapshot.buffers[i], written) ||
+					    (written.Base48() != 0u && written.GetSize() != 0u &&
+					     begin < written.Base48() + written.GetSize() &&
+					     written.Base48() < end)) {
+						use_selector = false;
+						break;
+					}
+				}
+			}
+		}
+		if (byte_keys && !use_selector) {
 			if (!indirect.table_scalar && table.Format() == Prospero::BufferFormat::kInvalid) {
 				descriptors[resource_index] = {.dword_count = dword_count};
 				return true;
 			}
-			if (table_size > UINT32_MAX ||
-			    (indirect.table_record_bytes != 0u &&
+			if (indirect.table_record_bytes != 0u &&
 			     (table.Type() != 0u || table.SwizzleEnabled() || table.AddTid() ||
 			      table.OutOfBounds() != 0u || (table_base & 3u) != 0u ||
-			      (table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes)))
+			      (table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes))
 				return failed("table geometry");
 			const auto* pointer = indirect.pointer_table ? &*indirect.pointer_table : nullptr;
 			const bool indexed_pointer = pointer != nullptr && pointer->indexed;
@@ -640,10 +685,11 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 				return failed("candidate count", fmt::format("rows={} limit={}", count, MaxIndirectDescriptorProbes));
 			// Affine SMEM offsets wrap in U32 before adding the aligned immediate.
 			// Image heaps retain one bulk snapshot and their absolute byte-offset mapping.
+			auto& heap_words = program.material_keys;
 			if (image_heap) {
-				keys.resize(table_size / 4u);
-				if (!ReadScalarTable(table_base, table_size, 0u, runtime, keys))
-					return read_failed("image heap read", 0u, 0u, table_base, 0u, keys.size() * sizeof(uint32_t));
+				heap_words.resize(table_size / 4u);
+				if (!ReadScalarTable(table_base, table_size, 0u, runtime, heap_words))
+					return read_failed("image heap read", 0u, 0u, table_base, 0u, heap_words.size() * sizeof(uint32_t));
 			}
 			const auto mapping_offset = snapshot.flattened_srt.size();
 			descriptors[resource_index] = {.dword_count = dword_count};
@@ -670,15 +716,14 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 						                     std::span(candidate.dwords).first(read_dwords))) return false;
 					} else {
 						const auto table_word = table_offset / 4u;
-						std::copy_n(keys.begin() + table_word,
-						            std::min<size_t>(read_dwords, keys.size() - table_word),
+						std::copy_n(heap_words.begin() + table_word,
+						            std::min<size_t>(read_dwords, heap_words.size() - table_word),
 						            candidate.dwords.begin());
 					}
 					if (!normalize(candidate))
 						return normalization_failed(table_offset, word, step,
 						                            indirect.table_maximum ? "key" : "probe");
-					uint32_t ordinal;
-					if (!intern_candidate(candidate, ordinal)) return false;
+					const auto ordinal = intern_candidate(candidate);
 					if (ordinal == 0u) continue;
 					const auto mapping_word = image_heap ? table_offset / 4u : word * child_count + child;
 					snapshot.flattened_srt.resize(mapping_offset + 2u + mapping_word, 0u);
@@ -722,21 +767,38 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 				selector_keys->resolved = true;
 			}
 		}
+		if (byte_keys) {
+			// Keep cached selector values independent of each destination table's
+			// U32 arithmetic, and use the same byte keys as the full-heap GPU path.
+			byte_offsets.clear();
+			byte_offsets.reserve(keys.size());
+			for (const auto key: keys) {
+				const auto offset = uint64_t {(key * indirect.table_stride + indirect.table_offset) & ~3u} +
+				                    (indirect.table_immediate & ~3u);
+				byte_offsets.push_back(offset <= UINT32_MAX ? static_cast<uint32_t>(offset) : UINT32_MAX);
+			}
+			std::ranges::sort(byte_offsets);
+			byte_offsets.erase(std::unique(byte_offsets.begin(), byte_offsets.end()), byte_offsets.end());
+		}
 	}
 
 	const auto mapping_offset = snapshot.flattened_srt.size();
-	const auto key_count = sources.empty() ? keys.size() : sources.size();
+	const auto selected_keys = byte_keys ? std::span<const uint32_t>(byte_offsets)
+	                                    : std::span<const uint32_t>(keys);
+	const auto key_count = sources.empty() ? selected_keys.size() : sources.size();
 	snapshot.flattened_srt.resize(mapping_offset + 1u + key_count * 2u);
 	snapshot.flattened_srt[mapping_offset] = static_cast<uint32_t>(key_count);
 	for (uint32_t entry = 0; entry < key_count; ++entry) {
-		const auto key = sources.empty() ? keys[entry] : entry;
+		const auto key = sources.empty() ? selected_keys[entry] : entry;
 		uint64_t table_offset = 0;
 		DescriptorValue candidate;
 		candidate.dword_count = dword_count;
 		if (sources.empty()) {
-			table_offset = uint64_t {(key * indirect.table_stride + indirect.table_offset) & ~3u} +
-			               (indirect.table_immediate & ~3u);
-			if (!ReadScalarTable(table_base, table_size, table_offset, runtime,
+			table_offset = byte_keys ? key
+			    : uint64_t {(key * indirect.table_stride + indirect.table_offset) & ~3u} +
+			          (indirect.table_immediate & ~3u);
+			if (!(byte_keys && key == UINT32_MAX) &&
+			    !ReadScalarTable(table_base, table_size, table_offset, runtime,
 			                     std::span(candidate.dwords).first(read_dwords))) {
 				return read_failed("table descriptor read", key, 0u, table_base, table_offset, descriptor_bytes);
 			}
@@ -746,10 +808,10 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		if (!normalize(candidate))
 			return normalization_failed(table_offset, key, indirect.table_stride, "key");
 		uint32_t ordinal = 0;
-		if (entry == 0) {
+		if (entry == 0 && !byte_keys) {
 			descriptors[resource_index] = candidate;
-		} else if (!intern_candidate(candidate, ordinal)) {
-			return false;
+		} else {
+			ordinal = intern_candidate(candidate);
 		}
 		snapshot.flattened_srt[mapping_offset + 1u + entry * 2u] = key;
 		snapshot.flattened_srt[mapping_offset + 2u + entry * 2u] = ordinal;
@@ -772,9 +834,8 @@ struct SamplerPlan {
 		uint32_t     source;
 		SamplerClass type;
 	};
-	std::array<std::array<uint32_t, 2>, ShaderInfo::MaxSamplers> mapping;
-	std::array<Binding, ShaderInfo::MaxSamplers>                bindings;
-	uint32_t                                                  sampler_count = 0;
+	std::vector<std::array<uint32_t, 2>> mapping;
+	std::vector<Binding>                bindings;
 };
 
 template <typename T, typename Keep>
@@ -913,6 +974,7 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			if (i >= program.info.images.size() && source != nullptr &&
 			    source->indirect_descriptor &&
 			    IsBoundedDescriptorTable(program, *source->indirect_descriptor) &&
+			    specialization.images[base_index].indirect_search_iterations == 0u &&
 			    !source->indirect_descriptor->pointer_table) {
 				const auto& indirect = *source->indirect_descriptor;
 				const size_t mapping = specialization.images[base_index].indirect_mapping_offset;
@@ -997,7 +1059,6 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			    is_1d_or_2d(image_class.dimension);
 			if ((!same_coordinates && !(is_2d(image.dimension) && is_2d(image_class.dimension)) &&
 			     !unfiltered_coordinates) ||
-			    image.mip_count != image_class.mip_count ||
 			    image.conversion_format != image_class.conversion_format ||
 			    (!storage_swizzle && image.shader_swizzle != image_class.shader_swizzle)) {
 				return SpecializationFail(
@@ -1011,11 +1072,9 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 }
 
 bool BuildSamplerPlan(const ShaderInfo& base, SamplerPlan& plan) {
-	if (base.samplers.size() > plan.mapping.size()) {
-		return false;
-	}
-	std::array<uint8_t, ShaderInfo::MaxSamplers> usage {};
-	plan.sampler_count = static_cast<uint32_t>(base.samplers.size());
+	plan.mapping.resize(base.samplers.size());
+	plan.bindings.resize(base.samplers.size());
+	std::vector<uint8_t> usage(base.samplers.size());
 	for (const auto& pair: base.sampled_pairs) {
 		if (pair.image >= base.images.size() || pair.sampler >= base.samplers.size()) {
 			return false;
@@ -1037,8 +1096,8 @@ bool BuildSamplerPlan(const ShaderInfo& base, SamplerPlan& plan) {
 		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;
+			const auto target = first ? index : static_cast<uint32_t>(plan.bindings.size());
+			if (!first) plan.bindings.emplace_back();
 			mapping[type]         = target;
 			plan.bindings[target] = {index, static_cast<SamplerClass>(type)};
 			first                = false;
@@ -1544,9 +1603,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		return true;
 	};
 	const auto resource_failed = [&](std::string_view kind, uint32_t resource, uint32_t source,
-	                                 size_t candidates, uint32_t capacity) {
-		return SpecializationFail(fmt::format("indirect {} resource={} source={} candidates={} capacity={}",
-		    kind, resource, source, candidates, capacity));
+	                                 size_t candidates) {
+		return SpecializationFail(fmt::format("indirect {} resource={} source={} candidates={}",
+		    kind, resource, source, candidates));
 	};
 	snapshot.buffers.resize(program.info.buffers.size());
 	specialization.buffers.assign(program.info.buffers.size(), {});
@@ -1563,12 +1622,12 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 				if (active.empty() || active[base.source]) {
 					if (!MaterializeIndirectDescriptor(
 					        program, base.source, i, 4u, 4u, observed, clean, snapshot,
-					        snapshot.buffers, specialization.buffers, UINT32_MAX,
+					        snapshot.buffers, specialization.buffers,
 					        [&](DescriptorValue& value) {
 						        return NormalizeIndirectBuffer(
 						            value, source->indirect_descriptor->indexed, base);
 					        }))
-						return resource_failed("buffer", i, base.source, snapshot.buffers.size(), UINT32_MAX);
+						return resource_failed("buffer", i, base.source, snapshot.buffers.size());
 				}
 			} else if (!evaluate(base.source, snapshot.buffers[i], base.written)) {
 				return false;
@@ -1627,14 +1686,14 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			const bool bounded_table = IsBoundedDescriptorTable(program, indirect);
 			if (!MaterializeIndirectDescriptor(
 			        program, image.source, i, 8u, image.r128 ? 4u : 8u, observed, clean, snapshot, snapshot.images,
-			        specialization.images, UINT32_MAX, [&](DescriptorValue& value) {
+			        specialization.images, [&](DescriptorValue& value) {
 				        // A broad heap also contains resources for other typed image operations.
 				        if (NullImageDescriptor(value) || !ValidImageDescriptor(value, image.r128) ||
 				            (bounded_table && DescriptorDimension(value, image.dimension) != image.dimension))
 					        value.dwords.fill(0u);
 				        return true;
 			        })) {
-				return resource_failed("image", i, image.source, snapshot.images.size(), UINT32_MAX);
+				return resource_failed("image", i, image.source, snapshot.images.size());
 			}
 		} else {
 			if (!evaluate(image.source, snapshot.images[i], image.written)) {
@@ -1660,9 +1719,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 				snapshot.samplers[i] = {.dword_count = 4u};
 				if ((active.empty() || active[source]) && !MaterializeIndirectDescriptor(
 				        program, source, i, 4u, 4u, observed, clean, snapshot, snapshot.samplers,
-				        specialization.samplers, ShaderInfo::MaxSamplers,
+				        specialization.samplers,
 				        [](DescriptorValue&) { return true; }))
-					return resource_failed("sampler", i, source, snapshot.samplers.size(), ShaderInfo::MaxSamplers);
+					return resource_failed("sampler", i, source, snapshot.samplers.size());
 			} else {
 				if (indirect != nullptr) source = indirect->sources[0];
 				if (!evaluate(source, snapshot.samplers[i])) return false;
@@ -1671,11 +1730,13 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		if (program.info.samplers[base_index].gather_lod) {
 			const auto control = snapshot.samplers[i].dwords[2];
 			const auto filter = (control >> 26u) & 3u;
-			// MipNone always selects the base level. Explicit point gathers currently require
-			// encoded-zero primary and secondary bias; linear primary-mip selection is unsupported.
-			if (filter > 1u || (filter == 1u && (control & 0xfffffu) != 0u)) {
+			const auto perf_mip = (snapshot.samplers[i].dwords[1] >> 24u) & 15u;
+			// MipNone always selects the base level. Point/linear selection requires zero
+			// primary/secondary bias; linear mip snapping is not implemented.
+			if (filter > 2u || (filter != 0u && (control & 0xfffffu) != 0u) ||
+			    (filter == 2u && perf_mip != 0u)) {
 				return SpecializationFail(
-				    "explicit-LOD gather requires mip filtering None or Point with zero LOD biases");
+				    "explicit-LOD gather requires None, unbiased Point, or unbiased Linear without mip snapping");
 			}
 		}
 	}
@@ -1768,8 +1829,8 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 	SamplerPlan sampler_plan;
 	EXIT_IF(!BuildSamplerPlan(program.info, sampler_plan));
 	const auto original_sampler_count = samplers.size();
-	samplers.reserve(sampler_plan.sampler_count);
-	for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {
+	samplers.reserve(sampler_plan.bindings.size());
+	for (uint32_t index = 0; index < sampler_plan.bindings.size(); index++) {
 		const auto& binding = sampler_plan.bindings[index];
 		if (index >= original_sampler_count) {
 			samplers.push_back(samplers[binding.source]);
@@ -1778,7 +1839,7 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 		samplers[index].force_point_filtering = binding.type == SamplerClass::Integer;
 		samplers[index].integer_border        = binding.type != SamplerClass::Float;
 	}
-	for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {
+	for (uint32_t index = 0; index < sampler_plan.bindings.size(); index++) {
 		const auto& binding = sampler_plan.bindings[index];
 		for (auto& candidate: samplers[index].indirect_resources) {
 			candidate = sampler_plan.mapping[candidate][static_cast<uint32_t>(binding.type)];
@@ -1905,6 +1966,7 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 			EXIT_IF(image.indirect_root == memory.resource &&
 			        inst.GetOpcode() != ValueOpcode::ImageSampleRaw &&
 			        inst.GetOpcode() != ValueOpcode::ImageGatherRaw &&
+			        inst.GetOpcode() != ValueOpcode::ImageQueryDimensions &&
 			        inst.GetOpcode() != ValueOpcode::ImageRead &&
 			        inst.GetOpcode() != ValueOpcode::ImageWrite);
 		}

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

@@ -1558,14 +1558,49 @@ private:
 		return read;
 	}
 
+	const Inst* IndexedMaterialRead(const Inst* read, Value active,
+	                               DescriptorSource::IndirectDescriptor::SelectorRead& selector) const {
+		uint32_t lane = 0;
+		if (read != nullptr && read->GetOpcode() == ValueOpcode::CompositeExtractU32x2) {
+			if (!ImmediateU32(read->Arg(1), lane) || lane >= 2u) return nullptr;
+			read = read->Arg(0).Resolve().TryInstruction();
+			if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32x2) return nullptr;
+		} else if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32) {
+			return nullptr;
+		}
+		const auto& memory = m_program.memory_info[read->Flags<MemoryFlags>().index];
+		if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
+		    !memory.idxen || memory.offen || memory.offset > UINT32_MAX - lane * 4u ||
+		    read->Arg(2).Resolve() != Value(0u) || read->Arg(3).Resolve() != Value(0u) ||
+		    !EquivalentValue(m_program, active, read->Arg(4))) return nullptr;
+		selector.offset = memory.offset + lane * 4u;
+		return read;
+	}
+
 	bool MatchUniformizedMaterialKey(Value key, const Inst& image,
 	                                DescriptorSource::IndirectDescriptor& indirect,
-	                                DescriptorSource& material_source) {
+	                                DescriptorSource& material_source, bool indexed) {
 		const auto* selected = UniformizedMaterialValue(key, image);
 		if (selected == nullptr) return false;
 		const auto active = selected->Arg(0).Resolve();
 		DescriptorSource::IndirectDescriptor::SelectorRead selector;
 		const auto* read = SelectorReadValue(selected->Arg(1), active, selector);
+		if (indexed) {
+			read = IndexedMaterialRead(read, active, selector);
+			// This optional domain is checked against the dispatch's writable buffers.
+			// Other stages and unbounded/image writes need the original heap domain.
+			if (read == nullptr || m_program.stage != ShaderType::Compute ||
+			    std::ranges::any_of(m_program.blocks, [](const Block* block) {
+				    return std::ranges::any_of(*block, [](const Inst& inst) {
+					    const auto op = inst.GetOpcode();
+					    return ImageOpcodeInfoOf(op).resource_class == ImageResourceClass::Storage ||
+					           AddressOpcodeInfoOf(op).access == AddressAccess::Write;
+				    });
+			    }) || !MakeRuntimeTableSource(*read, material_source)) return false;
+			selector.source = InternSource(material_source);
+			indirect.selector = std::move(selector);
+			return true;
+		}
 		if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadAddressU32 ||
 		    read->NumArgs() != 4u || !Implies(active, read->Arg(3)))
 			return false;
@@ -2815,7 +2850,7 @@ private:
 				} else result.maximum = std::max(result.maximum, SelectorMaximum(key, {}, 0u, active));
 			}
 		}
-		if (lane != nullptr) result = SelectedFiniteDomain(*lane, handle != nullptr ? lane->Arg(1) : Value {});
+		if (lane != nullptr) result = SelectedSelectorDomain(*lane, handle != nullptr ? lane->Arg(1) : Value {});
 		else if (handle == nullptr) result.maximum = SelectorMaximum(key);
 		if (result.maximum > UINT32_MAX / 4u ||
 		    uint64_t {result.maximum} * indirect.table_stride + indirect.table_offset > UINT32_MAX)
@@ -2971,6 +3006,13 @@ private:
 			// Keep live descriptor payloads, but project descriptor-only SMEM onto
 			// the GPU key using the same plan as other indirect image reads.
 			plan.retain_reads |= table_source.dword_count == 4u;
+		} else if (!sampler && !plan.table_indexed && table_source.dword_count == 4u &&
+		           MatchUniformizedMaterialKey(key, handle, indirect, material_source, true)) {
+			// Preserve the heap's byte-offset key so runtime geometry can choose the
+			// tighter selector domain without changing the GPU's descriptor reads.
+			indirect.table_scalar = true;
+			plan.table_read = plan.reads[0];
+			plan.retain_reads = true;
 		} else if (sampler && table_source.dword_count == 4u && !plan.table_indexed &&
 		           SetExactTableMaximum(key, indirect, &handle)) {
 			indirect.table_scalar = true;
@@ -3020,7 +3062,7 @@ private:
 				}
 			}
 			if (indirect.key_count.IsEmpty() &&
-			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source) &&
+			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source, false) &&
 			    !MatchReversedImageRange(key, indirect)) {
 				const auto maximum = SelectorMaximum(key);
 				if (maximum >= MaxIndirectDescriptorProbes) return false;
@@ -3064,6 +3106,10 @@ private:
 					indirect.selector_mask = selected.selector_mask;
 					indirect.signed_count = selected.signed_count;
 					indirect.workgroup_axis = selected.workgroup_axis;
+					if (selected.table_scalar) {
+						indirect.table_scalar = true;
+						plan.table_read = plan.reads[0];
+					}
 					material_source = m_sources[selected.selector->source];
 					shared_selector = true;
 					break;
@@ -3264,26 +3310,13 @@ private:
 			if (value == nullptr || (selected != nullptr && selected != value)) return false;
 			selected = value;
 		}
-		const auto* read = selected->Arg(1).Resolve().TryInstruction();
-		uint32_t lane = 0;
-		if (read != nullptr && read->GetOpcode() == ValueOpcode::CompositeExtractU32x2) {
-			if (!ImmediateU32(read->Arg(1), lane) || lane >= 2u) return false;
-			read = read->Arg(0).Resolve().TryInstruction();
-			if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32x2) return false;
-		} else if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32) {
-			return false;
-		}
-		const auto& memory = m_program.memory_info[read->Flags<MemoryFlags>().index];
-		uint32_t offset = 1u, scalar = 1u;
-		if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
-		    !memory.idxen || memory.offen || memory.offset > UINT32_MAX - lane * 4u ||
-		    !ImmediateU32(read->Arg(2), offset) || offset != 0u ||
-		    !ImmediateU32(read->Arg(3), scalar) || scalar != 0u ||
-		    !EquivalentValue(m_program, selected->Arg(0), read->Arg(4))) return false;
-		if (!MakeRuntimeTableSource(*read, material_source) ||
+		DescriptorSource::IndirectDescriptor::SelectorRead selector;
+		const auto* read = IndexedMaterialRead(selected->Arg(1).Resolve().TryInstruction(),
+		                                       selected->Arg(0), selector);
+		if (read == nullptr || !MakeRuntimeTableSource(*read, material_source) ||
 		    !BoundedSelectedIndex(read->Arg(1), read->Arg(4), indirect)) return false;
-		indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead{.source = InternSource(material_source),
-		                              .offset = memory.offset + lane * 4u});
+		selector.source = InternSource(material_source);
+		indirect.selector = std::move(selector);
 		return finish();
 	}
 
@@ -3372,10 +3405,12 @@ private:
 		return found ? SelectorDomain {maximum, counted ? count : Value {}} : SelectorDomain {};
 	}
 
-	SelectorDomain SelectedFiniteDomain(const Inst& lane, Value guard) const {
+	SelectorDomain SelectedSelectorDomain(const Inst& lane, Value guard) const {
 		if (guard.IsEmpty() ? !HasActiveLane(lane.Arg(1), lane.Parent())
 		                    : !Implies(guard, lane.Arg(1))) return {};
-		return FiniteSelectorDomain(lane.Arg(0), lane.Arg(1));
+		auto result = FiniteSelectorDomain(lane.Arg(0), lane.Arg(1));
+		if (result.maximum == UINT32_MAX) result.maximum = PossibleU32Bits(lane.Arg(0));
+		return result;
 	}
 
 	uint32_t SelectorMaximum(Value value, SelectorBounds bounds = {}, uint32_t depth = 0,
@@ -3392,7 +3427,7 @@ private:
 		};
 		switch (inst->GetOpcode()) {
 			case ValueOpcode::ReadFirstLane:
-				return SelectedFiniteDomain(*inst, guard).maximum;
+				return SelectedSelectorDomain(*inst, guard).maximum;
 			case ValueOpcode::ReadLane:
 				// ReadLane ignores EXEC; only context-free bounds on every source lane apply.
 				return PossibleU32Bits(inst->Arg(0));
@@ -3823,9 +3858,6 @@ private:
 				return;
 			}
 		}
-		if (m_info.sampled_pairs.size() >= ShaderInfo::MaxSampledPairs) {
-			Fail(pc, "sampled image/sampler pair limit exceeded");
-		}
 		m_info.sampled_pairs.push_back({image, sampler, pc});
 	}
 

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

@@ -509,6 +509,7 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 		case ValueOpcode::ReadConst: info.uses_flattened_srt = true; break;
 		case ValueOpcode::BitwiseXor32: info.has_bitwise_xor = true; break;
 		case ValueOpcode::StoreCompletion: info.subgroup_barrier = true; break;
+		case ValueOpcode::DataCacheInvalidate: info.coherent_buffers = true; break;
 		case ValueOpcode::BvhIntersect: info.bvh = true; break;
 		case ValueOpcode::ConditionRef:
 			info.subgroup_vote |=

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

@@ -159,9 +159,11 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::ULessThan32:
 		case ValueOpcode::ULessThanEqual32:
 		case ValueOpcode::IEqual32:
+		case ValueOpcode::IEqual64:
 		case ValueOpcode::UGreaterThan32:
 		case ValueOpcode::SGreaterThanEqual32:
 		case ValueOpcode::INotEqual32:
+		case ValueOpcode::INotEqual64:
 		case ValueOpcode::LogicalOr:
 		case ValueOpcode::LogicalAnd:
 		case ValueOpcode::LogicalXor:
@@ -1053,6 +1055,11 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::IEqual64:
+		case ValueOpcode::INotEqual64:
+			if (!binary()) return false;
+			result = inst.GetOpcode() == ValueOpcode::IEqual64 ? a == b : a != b;
+			return true;
 		case ValueOpcode::ULessThan32:
 			// No U32 is below zero, including an unknown loop-carried value.
 			if (!Arg(inst, 1, b)) return false;

src/kernel/fileSystem.cpp

@@ -4,6 +4,7 @@
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/dateTime.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/hash.h"
@@ -669,6 +670,7 @@ int64_t KYTY_SYSV_ABI KernelRead(int d, void* buf, size_t nbytes) {
 	                         std::min<uint64_t>(nbytes, remaining));
 	uint32_t bytes_read = 0;
 	file->f.Read(buf, static_cast<uint32_t>(nbytes), &bytes_read);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileReadBytes, bytes_read);
 
 	file->mutex.Unlock();
 
@@ -731,6 +733,7 @@ int64_t KYTY_SYSV_ABI KernelWrite(int d, const void* buf, size_t nbytes) {
 		file->f.Seek(file->f.Size());
 	}
 	file->f.Write(buf, static_cast<uint32_t>(nbytes), &bytes_written);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileWriteBytes, bytes_written);
 	if (file->sync_writes) {
 		file->f.Flush();
 	}
@@ -795,6 +798,7 @@ int64_t KYTY_SYSV_ABI KernelPread(int d, void* buf, size_t nbytes, int64_t offse
 	uint32_t bytes_read = 0;
 	file->f.Seek(offset);
 	file->f.Read(buf, static_cast<uint32_t>(nbytes), &bytes_read);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileReadBytes, bytes_read);
 	file->f.Seek(pos);
 
 	file->mutex.Unlock();
@@ -896,6 +900,7 @@ int64_t KYTY_SYSV_ABI KernelPreadv(int d, const KernelIovec* iov, int iovcnt, in
 		Memory::InvalidateMemory(reinterpret_cast<uint64_t>(buffer.iov_base), count);
 		uint32_t bytes = 0;
 		file->f.Read(buffer.iov_base, count, &bytes);
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileReadBytes, bytes);
 		bytes_read += bytes;
 		remaining -= bytes;
 		if (bytes < count) {
@@ -945,6 +950,7 @@ int64_t KYTY_SYSV_ABI KernelPwrite(int d, const void* buf, size_t nbytes, int64_
 	uint32_t bytes_written = 0;
 	file->f.Seek(file->append ? file->f.Size() : static_cast<uint64_t>(offset));
 	file->f.Write(buf, static_cast<uint32_t>(nbytes), &bytes_written);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileWriteBytes, bytes_written);
 	if (file->sync_writes) {
 		file->f.Flush();
 	}
@@ -1016,6 +1022,7 @@ static int64_t Writev(int d, const KernelIovec* iov, int iovcnt, std::optional<i
 		}
 		uint32_t bytes = 0;
 		file->f.Write(buffer.iov_base, static_cast<uint32_t>(buffer.iov_len), &bytes);
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileWriteBytes, bytes);
 		bytes_written += bytes;
 		if (bytes < buffer.iov_len) {
 			break;
@@ -1319,6 +1326,27 @@ int KYTY_SYSV_ABI KernelFtruncate(int d, int64_t length) {
 	return OK;
 }
 
+int KYTY_SYSV_ABI KernelFsync(int d) {
+	PRINT_NAME();
+
+	if (d < DESCRIPTOR_MIN) {
+		return KERNEL_ERROR_EBADF;
+	}
+
+	auto* file = g_files->GetFile(d);
+	if (file == nullptr || !file->opened) {
+		return KERNEL_ERROR_EBADF;
+	}
+
+	Common::LockGuard lock(file->mutex);
+	if (file->f.IsInvalid() || !file->f.Flush()) {
+		return KERNEL_ERROR_EIO;
+	}
+
+	LOGF("\tFsync: %s\n", Common::PathToString(file->real_name).c_str());
+	return OK;
+}
+
 int KYTY_SYSV_ABI KernelUnlink(const char* path) {
 	PRINT_NAME();
 

src/kernel/fileSystem.h

@@ -72,6 +72,7 @@ int KYTY_SYSV_ABI     KernelStat(const char* path, FileStat* sb);
 int KYTY_SYSV_ABI     KernelFstat(int d, FileStat* sb);
 int KYTY_SYSV_ABI     KernelTruncate(const char* path, int64_t length);
 int KYTY_SYSV_ABI     KernelFtruncate(int d, int64_t length);
+int KYTY_SYSV_ABI     KernelFsync(int d);
 int KYTY_SYSV_ABI     KernelUnlink(const char* path);
 int KYTY_SYSV_ABI     KernelRename(const char* from, const char* to);
 int KYTY_SYSV_ABI     KernelGetdirentries(int fd, char* buf, int nbytes, int64_t* basep);

src/kernel/memory.cpp

@@ -4414,4 +4414,39 @@ int KYTY_SYSV_ABI KernelMemoryPoolGetBlockStats(KernelMemoryPoolBlockStats* outp
 	return OK;
 }
 
+DebugStats GetDebugStats() {
+	DebugStats stats;
+	stats.direct_total       = PhysicalMemory::Size();
+	stats.flexible_total     = FlexibleMemory::Size();
+	stats.pool_committed     = g_memory_pool_committed.load(std::memory_order_relaxed);
+	stats.page_entries_total = PAGE_TABLE_POOL_ENTRIES;
+	if (g_physical_memory != nullptr) {
+		Common::LockGuard lock(g_physical_memory->GetMutex());
+		for (const auto& [start, block]: g_physical_memory->GetPhysicalBlocks()) {
+			switch (block.kind) {
+				case PhysicalMemory::AllocationKind::Direct:
+					stats.direct_allocated += block.size;
+					break;
+				case PhysicalMemory::AllocationKind::Pooled:
+					stats.pooled_allocated += block.size;
+					break;
+				case PhysicalMemory::AllocationKind::Automatic:
+					stats.automatic_allocated += block.size;
+					break;
+			}
+		}
+		for (const auto& mapping: g_physical_memory->GetMappings()) {
+			stats.direct_mapped += mapping.map_size;
+		}
+	}
+	if (g_flexible_memory != nullptr) {
+		stats.flexible_used = stats.flexible_total - g_flexible_memory->Available();
+	}
+	if (g_virtual_ranges != nullptr) {
+		stats.cpu_page_entries = g_virtual_ranges->CountPageTableEntries(false);
+		stats.gpu_page_entries = g_virtual_ranges->CountPageTableEntries(true);
+	}
+	return stats;
+}
+
 } // namespace Libs::LibKernel::Memory

src/kernel/memory.h

@@ -15,6 +15,23 @@ namespace Libs::LibKernel::Memory {
 void Initialize();
 void Shutdown();
 
+// Guest memory usage for the performance panel. Takes only bookkeeping locks, so it never
+// waits on the GPU; call it only while the memory subsystem is initialized.
+struct DebugStats {
+	uint64_t direct_total        = 0;
+	uint64_t direct_allocated    = 0; // sceKernelAllocateDirectMemory.
+	uint64_t pooled_allocated    = 0; // Memory pool expansions.
+	uint64_t automatic_allocated = 0; // Extended memory.
+	uint64_t direct_mapped       = 0;
+	uint64_t flexible_total      = 0;
+	uint64_t flexible_used       = 0;
+	uint64_t pool_committed      = 0;
+	uint64_t cpu_page_entries    = 0; // 2 MiB page-table entries in use.
+	uint64_t gpu_page_entries    = 0;
+	uint64_t page_entries_total  = 0; // Per side.
+};
+[[nodiscard]] DebugStats GetDebugStats();
+
 struct Lifecycle {
 	static constexpr const char* name       = "Memory";
 	static constexpr auto        initialize = Libs::LibKernel::Memory::Initialize;

src/kernel/pthread.cpp

@@ -343,14 +343,21 @@ static bool KernelClockGettimeSpecial(KernelClockid clock_id, KernelTimespec* tp
 #endif
 
 struct PthreadMutexPrivate {
+	struct Waiter {
+		Pthread                 thread = nullptr;
+		std::condition_variable cv;
+		Waiter*                 previous = nullptr;
+		Waiter*                 next     = nullptr;
+	};
 	uint8_t                 reserved[256];
 	std::string             name;
 	std::mutex              m;
-	std::condition_variable cv;
 	Pthread                 owner     = nullptr;
 	uint32_t                count     = 0;
 	int                     type      = 1;
 	int                     pprotocol = PTHREAD_PRIO_NONE;
+	Waiter*                 first_waiter = nullptr;
+	Waiter*                 last_waiter  = nullptr;
 };
 
 struct PthreadMutexattrPrivate {
@@ -1242,6 +1249,20 @@ static int PthreadRwlockInitNamed(PthreadRwlock* rwlock, const PthreadRwlockattr
                                   const char* name);
 static int PthreadCondInitNamed(PthreadCond* cond, const PthreadCondattr* attr, const char* name);
 
+static void NativeMutexRemoveWaiter(PthreadMutexPrivate* mutex,
+                                     PthreadMutexPrivate::Waiter* waiter) {
+	if (waiter->previous != nullptr) {
+		waiter->previous->next = waiter->next;
+	} else {
+		mutex->first_waiter = waiter->next;
+	}
+	if (waiter->next != nullptr) {
+		waiter->next->previous = waiter->previous;
+	} else {
+		mutex->last_waiter = waiter->previous;
+	}
+}
+
 static int NativeMutexLock(PthreadMutexPrivate* mutex, KernelUseconds* timeout_us) {
 	EXIT_IF(mutex == nullptr);
 
@@ -1258,54 +1279,53 @@ static int NativeMutexLock(PthreadMutexPrivate* mutex, KernelUseconds* timeout_u
 			mutex->count++;
 			return OK;
 		}
-		if (timeout_us != nullptr) {
-			if (*timeout_us > 0) {
-				mutex->cv.wait_for(lock, std::chrono::microseconds(*timeout_us));
-			}
-			return ETIMEDOUT;
-		}
 		return EDEADLK;
 	}
 
-	if (timeout_us == nullptr) {
-		while (mutex->owner != nullptr) {
-			mutex->cv.wait_for(lock, std::chrono::microseconds(SIGNAL_APC_POLL_MICROS));
-			if (mutex->owner != nullptr) {
-				lock.unlock();
-				KernelDispatchPendingSignalForCurrentThread();
-				lock.lock();
-			}
-		}
-	} else if (*timeout_us == 0) {
-		if (mutex->owner != nullptr) {
-			return ETIMEDOUT;
-		}
-	} else {
-		const auto deadline =
-		    std::chrono::steady_clock::now() + std::chrono::microseconds(*timeout_us);
-		while (mutex->owner != nullptr) {
-			const auto now = std::chrono::steady_clock::now();
-			if (now >= deadline) {
-				return ETIMEDOUT;
-			}
-
-			const auto remaining = deadline - now;
-			const auto poll      = (remaining < std::chrono::microseconds(SIGNAL_APC_POLL_MICROS)
-			                            ? remaining
-			                            : std::chrono::steady_clock::duration(
-			                                  std::chrono::microseconds(SIGNAL_APC_POLL_MICROS)));
-			mutex->cv.wait_for(lock, poll);
-			if (mutex->owner != nullptr) {
-				lock.unlock();
-				KernelDispatchPendingSignalForCurrentThread();
-				lock.lock();
-			}
-		}
+	if (mutex->owner == nullptr) {
+		mutex->owner = self;
+		mutex->count = 1;
+		return OK;
+	}
+	if (timeout_us != nullptr && *timeout_us == 0) {
+		return ETIMEDOUT;
 	}
 
-	mutex->owner = self;
-	mutex->count = 1;
-	return OK;
+	// Guest mutex waits are not cancellation points, unlike the host condition wait.
+	int previous_cancel_state = 0;
+	pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &previous_cancel_state);
+	PthreadMutexPrivate::Waiter waiter;
+	waiter.thread   = self;
+	waiter.previous = mutex->last_waiter;
+	if (mutex->last_waiter != nullptr) {
+		mutex->last_waiter->next = &waiter;
+	} else {
+		mutex->first_waiter = &waiter;
+	}
+	mutex->last_waiter = &waiter;
+	const auto deadline = timeout_us != nullptr
+	                          ? std::chrono::steady_clock::now() +
+	                                std::chrono::microseconds(*timeout_us)
+	                          : std::chrono::steady_clock::time_point::max();
+	int result = OK;
+	while (mutex->owner != self) {
+		const auto now = std::chrono::steady_clock::now();
+		if (now >= deadline) {
+			NativeMutexRemoveWaiter(mutex, &waiter);
+			result = ETIMEDOUT;
+			break;
+		}
+		waiter.cv.wait_until(lock, std::min(deadline, now +
+		    std::chrono::microseconds(SIGNAL_APC_POLL_MICROS)));
+		if (mutex->owner != self) {
+			lock.unlock();
+			KernelDispatchPendingSignalForCurrentThread();
+			lock.lock();
+		}
+	}
+	lock.unlock();
+	pthread_setcancelstate(previous_cancel_state, nullptr);
+	return result;
 }
 
 static int NativeMutexTrylock(PthreadMutexPrivate* mutex) {
@@ -1358,10 +1378,16 @@ static int NativeMutexUnlock(PthreadMutexPrivate* mutex, uint32_t* recurse = nul
 		return OK;
 	}
 
-	mutex->owner = nullptr;
-	mutex->count = 0;
-	lock.unlock();
-	mutex->cv.notify_one();
+	if (auto* waiter = mutex->first_waiter) {
+		NativeMutexRemoveWaiter(mutex, waiter);
+		// Transfer ownership in FIFO order before waking the selected waiter.
+		mutex->owner = waiter->thread;
+		mutex->count = 1;
+		waiter->cv.notify_one();
+	} else {
+		mutex->owner = nullptr;
+		mutex->count = 0;
+	}
 
 	return OK;
 }

src/launcher/forms/configuration_edit_dialog.ui

@@ -174,12 +174,6 @@
              </widget>
             </item>
             <item row="7" column="1">
-             <widget class="QCheckBox" name="checkBox_sync_raw_image_buffers">
-              <property name="toolTip"><string>Can fix missing textures and incorrect transparency in some games. May reduce performance.</string></property>
-              <property name="text"><string>Accurate texture synchronization</string></property>
-             </widget>
-            </item>
-            <item row="8" column="1">
              <widget class="QCheckBox" name="checkBox_hdr">
               <property name="toolTip"><string>Enable HDR10 output for supported games. Requires HDR enabled on the host display.</string></property>
               <property name="text"><string>Enable HDR display support</string></property>
@@ -242,6 +236,12 @@
               <property name="text"><string>Auto-hide cursor</string></property>
              </widget>
             </item>
+            <item>
+             <widget class="QCheckBox" name="checkBox_perf_overlay">
+              <property name="toolTip"><string>Show FPS, frame times, CPU/GPU usage, RAM and memory pressure in the game window. Press F2 (Cmd+P on macOS) to toggle it, and add Shift for the detailed view.</string></property>
+              <property name="text"><string>Show performance panel</string></property>
+             </widget>
+            </item>
             <item>
              <widget class="QCheckBox" name="checkBox_skip_notice_screen">
               <property name="toolTip"><string>Skip startup logos and notices in supported games.</string></property>
@@ -477,6 +477,25 @@
          </layout>
         </widget>
        </item>
+       <item>
+        <widget class="QGroupBox" name="rendering_compatibility_group">
+         <property name="title"><string>Rendering compatibility</string></property>
+         <layout class="QVBoxLayout" name="renderingCompatibilityLayout">
+          <item>
+           <widget class="QLabel" name="label_rendering_compatibility">
+            <property name="text"><string>These settings may help with rendering bugs at a slight performance cost.</string></property>
+            <property name="wordWrap"><bool>true</bool></property>
+           </widget>
+          </item>
+          <item>
+           <widget class="QCheckBox" name="checkBox_sync_raw_image_buffers">
+            <property name="toolTip"><string>Can fix missing textures and incorrect transparency (Resident Evil 2).</string></property>
+            <property name="text"><string>Accurate texture synchronization</string></property>
+           </widget>
+          </item>
+         </layout>
+        </widget>
+       </item>
        <item>
         <spacer name="debugSpacer">
          <property name="orientation"><enum>Qt::Vertical</enum></property>

src/launcher/include/configuration.h

@@ -130,6 +130,7 @@ public:
 	bool                   tessellation_enabled        = false;
 	bool                   trophy_enabled              = true;
 	int                    trophy_duration_seconds     = 6;
+	bool                   perf_overlay_enabled        = false;
 	bool                   skip_notice_screen          = false;
 	bool                   network_connected           = false;
 	int                    vblank_frequency            = 60;
@@ -168,6 +169,7 @@ public:
 		tessellation_enabled        = other.tessellation_enabled;
 		trophy_enabled              = other.trophy_enabled;
 		trophy_duration_seconds     = other.trophy_duration_seconds;
+		perf_overlay_enabled        = other.perf_overlay_enabled;
 		skip_notice_screen          = other.skip_notice_screen;
 		network_connected           = other.network_connected;
 		vblank_frequency            = other.vblank_frequency;

src/launcher/src/configuration.cpp

@@ -17,6 +17,7 @@ QVariantMap Configuration::GameSettings() const {
 	KYTY_CFG_SET(tessellation_enabled);
 	KYTY_CFG_SET(trophy_enabled);
 	KYTY_CFG_SET(trophy_duration_seconds);
+	KYTY_CFG_SET(perf_overlay_enabled);
 	KYTY_CFG_SET(skip_notice_screen);
 	KYTY_CFG_SET(network_connected);
 	KYTY_CFG_SET(vblank_frequency);
@@ -65,6 +66,7 @@ void Configuration::ReadGameSettingsValues(const Settings& s) {
 	KYTY_CFG_GET(tessellation_enabled);
 	trophy_enabled   = s.value("trophy_enabled", trophy_enabled).toBool();
 	KYTY_CFG_GET(trophy_duration_seconds);
+	KYTY_CFG_GET(perf_overlay_enabled);
 	KYTY_CFG_GET(skip_notice_screen);
 	KYTY_CFG_GET(network_connected);
 	vblank_frequency = s.value("vblank_frequency", vblank_frequency).toInt();

src/launcher/src/configurationEditDialog.cpp

@@ -292,6 +292,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_network_connected->setChecked(info.network_connected);
 	m_ui->checkBox_trophy_notifications->setChecked(info.trophy_enabled);
 	m_ui->spinBox_trophy_duration->setValue(info.trophy_duration_seconds);
+	m_ui->checkBox_perf_overlay->setChecked(info.perf_overlay_enabled);
 	m_ui->spinBox_vblank_frequency->setValue(info.vblank_frequency);
 	m_ui->comboBox_console_language->clear();
 	m_ui->comboBox_console_language->addItems(CONSOLE_LANGUAGE_NAMES);
@@ -484,6 +485,7 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 	info.network_connected         = ui.checkBox_network_connected->isChecked();
 	info.trophy_enabled            = ui.checkBox_trophy_notifications->isChecked();
 	info.trophy_duration_seconds   = ui.spinBox_trophy_duration->value();
+	info.perf_overlay_enabled      = ui.checkBox_perf_overlay->isChecked();
 	info.vblank_frequency          = ui.spinBox_vblank_frequency->value();
 	info.console_language          = ui.comboBox_console_language->currentIndex();
 	info.vulkan_validation_enabled = ui.checkBox_vulkan_validation->isChecked();

src/launcher/src/inputMappingDialog.cpp

@@ -203,7 +203,8 @@ public:
 
 		auto* layout = new QVBoxLayout(this);
 		m_label      = new QLabel(
-		    tr("Press a key or mouse button.\nF1, F7, and F11 are reserved; Esc cancels."), this);
+		    tr("Press a key or mouse button.\nF1, F2, F7, F8, and F11 are reserved; Esc cancels."),
+		    this);
 		m_label->setAlignment(Qt::AlignCenter);
 		layout->addWidget(m_label);
 	}
@@ -219,7 +220,8 @@ protected:
 			reject();
 			return;
 		}
-		if (event->key() == Qt::Key_F1 || event->key() == Qt::Key_F7 ||
+		if (event->key() == Qt::Key_F1 || event->key() == Qt::Key_F2 ||
+		    event->key() == Qt::Key_F7 || event->key() == Qt::Key_F8 ||
 		    event->key() == Qt::Key_F11) {
 			m_label->setText(tr("That key is reserved by the emulator."));
 			return;

src/launcher/src/mainDialog.cpp

@@ -264,6 +264,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	args << "--sync-raw-image-buffers" << BoolArg(info.sync_raw_image_buffers);
 	args << "--trophy-notifications" << BoolArg(info.trophy_enabled);
 	args << "--trophy-duration" << QString::number(info.trophy_duration_seconds);
+	args << "--perf-overlay" << BoolArg(info.perf_overlay_enabled);
 	args << "--skip-notice-screen" << BoolArg(info.skip_notice_screen);
 	args << "--network-connected" << BoolArg(info.network_connected);
 	if (info.tessellation_enabled) {

src/libs/agc.cpp

@@ -670,7 +670,7 @@ int KYTY_SYSV_ABI AgcCreateShader(Shader** dst, void* header, const volatile voi
 	return OK;
 }
 
-int KYTY_SYSV_ABI AgcUnknownGetFusedShaderSize(SizeAlign* dst, const Shader* front,
+int KYTY_SYSV_ABI AgcGetFusedShaderSize(SizeAlign* dst, const Shader* front,
                                                const Shader* back) {
 	PRINT_NAME();
 

src/libs/agc.h

@@ -40,7 +40,7 @@ int KYTY_SYSV_ABI   AgcInit(uint32_t* state, uint32_t ver);
 void* KYTY_SYSV_ABI AgcGetRegisterDefaults2(uint32_t ver);
 void* KYTY_SYSV_ABI AgcGetRegisterDefaults2Internal(uint32_t ver);
 int KYTY_SYSV_ABI   AgcCreateShader(Shader** dst, void* header, const volatile void* code);
-int KYTY_SYSV_ABI   AgcUnknownGetFusedShaderSize(SizeAlign* dst, const Shader* front,
+int KYTY_SYSV_ABI   AgcGetFusedShaderSize(SizeAlign* dst, const Shader* front,
                                                  const Shader* back);
 int KYTY_SYSV_ABI   AgcFuseShaderHalves_0200(Shader* fused_result, const Shader* front,
                                                const Shader* back, void* scratch_mem);

src/libs/audio.cpp

@@ -1,6 +1,7 @@
 #include "libs/audio.h"
 
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/threads.h"
@@ -403,6 +404,9 @@ bool Audio::QueueSdlAudio(PortOut* port, const void* data, bool blocking, float
 			queued = SDL_GetAudioStreamQueued(port->stream);
 		}
 		if (queued < static_cast<int>(prepared_size)) {
+			if (port->queue_primed) {
+				Common::DebugCounters::Add(Common::DebugCounters::Counter::AudioUnderruns);
+			}
 			port->queue_primed = false;
 		}
 	}

src/libs/libAgcDriver.cpp

@@ -18,7 +18,7 @@ LIB_DEFINE(InitAgcDriver_1) {
 	LIB_FUNC("2JtWUUiYBXs", Gen5::AgcGetRegisterDefaults2);
 	LIB_FUNC("wRbq6ZjNop4", Gen5::AgcGetRegisterDefaults2Internal);
 	LIB_FUNC("f3dg2CSgRKY", Gen5::AgcCreateShader);
-	LIB_FUNC("dolOmWH+huQ", Gen5::AgcUnknownGetFusedShaderSize);
+	LIB_FUNC("dolOmWH+huQ", Gen5::AgcGetFusedShaderSize);
 	LIB_FUNC("fd5Bp5tGTgo", Gen5::AgcFuseShaderHalves_0200);
 	LIB_FUNC("nApJjpKNBl4", Gen5::AgcFuseShaderHalves);
 	LIB_FUNC("vcmNN+AAXnY", Gen5::AgcSetCxRegIndirectPatchSetAddress);

src/libs/libKernel.cpp

@@ -1341,14 +1341,6 @@ static int KYTY_SYSV_ABI KernelGetOperationMode(int* mode, int* submode) {
 	return OK;
 }
 
-static int KYTY_SYSV_ABI KernelFsync(int fd) {
-	PRINT_NAME();
-
-	LOGF("\t fd = %d\n", fd);
-
-	return OK;
-}
-
 static void KYTY_SYSV_ABI KernelSync() {
 	PRINT_NAME();
 }
@@ -1986,6 +1978,12 @@ int KYTY_SYSV_ABI ftruncate(int d, int64_t length) {
 	return POSIX_CALL(LibKernel::FileSystem::KernelFtruncate(d, length));
 }
 
+int KYTY_SYSV_ABI fsync(int d) {
+	PRINT_NAME();
+
+	return POSIX_CALL(LibKernel::FileSystem::KernelFsync(d));
+}
+
 int KYTY_SYSV_ABI socket(int family, int type, int protocol) {
 	PRINT_NAME();
 	return Network::Net::Socket(family, type, protocol);
@@ -2174,6 +2172,7 @@ LIB_DEFINE(InitLibKernel_1_Posix) {
 	LIB_FUNC("E6ao34wPw+U", stat);
 	LIB_FUNC("JGMio+21L4c", mkdir);
 	LIB_FUNC("ih4CD9-gghM", Posix::ftruncate);
+	LIB_FUNC("juWbTNM+8hw", Posix::fsync);
 	LIB_FUNC("8nY19bKoiZk", Posix::fcntl);
 	LIB_FUNC("pDuPEf3m4fI", Posix::sem_init);
 	LIB_FUNC("cDW233RAwWo", Posix::sem_destroy);
@@ -3394,7 +3393,7 @@ LIB_DEFINE(InitLibKernel_1) {
 	LIB_FUNC("959qrazPIrg", LibKernel::KernelGetProcParam);
 	LIB_FUNC("tU5e3f9gSiU", LibKernel::KernelIsTrinityMode);
 	LIB_FUNC("NH6xARDOVv8", LibKernel::KernelGetOperationMode);
-	LIB_FUNC("fTx66l5iWIA", LibKernel::KernelFsync);
+	LIB_FUNC("fTx66l5iWIA", FileSystem::KernelFsync);
 	LIB_FUNC("uvT2iYBBnkY", LibKernel::KernelSync);
 	LIB_FUNC("HoLVWNanBBc", LibKernel::getpid);
 	LIB_FUNC("9BcDykPmo1I", LibKernel::get_error_addr);

src/libs/libRazorCpu.cpp

@@ -24,10 +24,15 @@ static int KYTY_SYSV_ABI RazorCpuPushMarker(const char*, uint32_t, uint32_t) {
 	return 0;
 }
 
+static int KYTY_SYSV_ABI RazorCpuNamedSync(const char*) {
+	return 0;
+}
+
 LIB_DEFINE(InitRazorCpu_1) {
 	LIB_FUNC("EboejOQvLL4", LibRazorCpu::RazorCpuIsCapturing);
 	LIB_FUNC("YpkGsMXP3ew", LibRazorCpu::RazorCpuPopMarker);
 	LIB_FUNC("zw+celG7zSI", LibRazorCpu::RazorCpuPushMarker);
+	LIB_FUNC("XCuZoBSVFG8", LibRazorCpu::RazorCpuNamedSync);
 }
 
 } // namespace LibRazorCpu

src/loader/runtimeLinker.cpp

@@ -2,6 +2,7 @@
 
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/hostException.h"
@@ -190,6 +191,7 @@ static uint64_t RegisterStubbedImport(uint32_t index, const Program* program,
 }
 
 static KYTY_SYSV_ABI uint64_t UnresolvedImportStub(uint64_t record_id) {
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::UnresolvedImportCalls);
 	const auto log_index = g_unresolved_stub_call_log_count.fetch_add(1);
 	if (log_index < 1024) {
 		if (record_id < g_stubbed_imports.size()) {
@@ -652,6 +654,7 @@ static bool KytyExceptionHandler(const Common::HostException::ExceptionInfo& exc
 
 	if (info->type == Common::HostException::ExceptionType::IllegalInstruction &&
 	    Loader::X64InstructionEmulator::TryEmulate(info->native_context)) {
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::EmulatedInstructions);
 		return true;
 	}
 
@@ -666,6 +669,7 @@ static bool KytyExceptionHandler(const Common::HostException::ExceptionInfo& exc
 			case CoreAccess::Unknown: return false;
 		}
 		if (Libs::LibKernel::Memory::HandleGpuFault(access, info->access_violation_vaddr)) {
+			Common::DebugCounters::Add(Common::DebugCounters::Counter::GpuFaults);
 			return true;
 		}
 	}

src/main.cpp

@@ -95,6 +95,7 @@ static void PrintUsage() {
 	    "  --trophy-notifications <true|false>   Show trophy unlock toasts and play their sound.\n");
 	::printf(
 	    "  --trophy-duration <seconds>          Trophy toast duration, 5-10. Default: 6.\n");
+	::printf("  --perf-overlay <true|false>           Show the performance panel (F2 or Cmd+P).\n");
 	::printf("  --playgo-hack                       Use the supplied PlayGo stub fallback.\n");
 	::printf(
 	    "  --skip-notice-screen <true|false>    Skip startup logos and notices in supported games.\n"
@@ -486,6 +487,11 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 				::printf("invalid boolean for %s: %s\n", arg.c_str(), value.c_str());
 				return false;
 			}
+		} else if (arg == "--perf-overlay") {
+			if (!ParseBool(value, options.config.perf_overlay_enabled)) {
+				::printf("invalid boolean for %s: %s\n", arg.c_str(), value.c_str());
+				return false;
+			}
 		} else if (arg == "--keymap") {
 			const auto split = value.find('=');
 			if (split == std::string::npos || split == 0 || split + 1 == value.size()) {

tests/KernelFileSystemTests.cpp

@@ -6,15 +6,18 @@
 #include "common/file.h"
 #include "ArchiveTestFixture.h"
 #include "common/logging/log.h"
+#include "common/singleton.h"
 #include "common/subsystems.h"
 #include "common/threads.h"
 #include "common/stringUtils.h"
 #include "graphics/presentation/window/windowInternal.h"
 #include "kernel/eventQueue.h"
 #include "kernel/fileSystem.h"
+#include "kernel/pthread.h"
 #include "libs/errno.h"
 #include "libs/network.h"
 #include "loader/additionalContent.h"
+#include "loader/runtimeLinker.h"
 #include "loader/symbolDatabase.h"
 
 #include <algorithm>
@@ -85,6 +88,148 @@ void Check(bool value, const char *text) {
   }
 }
 
+template <typename Predicate>
+void WaitForMutexTest(Predicate ready) {
+  const auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(5);
+  while (!ready()) {
+    Check(std::chrono::steady_clock::now() < deadline, "mutex waiter made progress");
+    std::this_thread::yield();
+  }
+}
+
+struct MutexWaiter {
+  Libs::LibKernel::PthreadMutex *mutex;
+  Libs::LibKernel::KernelUseconds timeout = 0;
+  bool cancel = false;
+  Libs::LibKernel::Pthread thread = nullptr;
+  std::vector<uint8_t> stack = std::vector<uint8_t>(1024 * 1024);
+  std::atomic<unsigned> parked{0}, resume{0};
+  std::atomic<bool> acquired{false}, release{false}, finished{false};
+  int result = 0;
+};
+
+thread_local MutexWaiter *current_mutex_waiter = nullptr;
+
+void KYTY_SYSV_ABI ParkMutexWaiter(int, void *) {
+  auto &waiter = *current_mutex_waiter;
+  const auto ticket = waiter.parked.fetch_add(1) + 1;
+  WaitForMutexTest([&] { return waiter.resume.load() >= ticket; });
+}
+
+void *KYTY_SYSV_ABI RunMutexWaiter(void *arg) {
+  using namespace Libs::LibKernel;
+  auto &waiter = *static_cast<MutexWaiter *>(arg);
+  current_mutex_waiter = &waiter;
+  waiter.result = waiter.timeout == 0 ? PthreadMutexLock(waiter.mutex)
+                                      : PthreadMutexTimedlock(waiter.mutex, waiter.timeout);
+  if (waiter.result == OK) {
+    waiter.acquired = true;
+    WaitForMutexTest([&] { return waiter.release.load(); });
+    Check(PthreadMutexUnlock(waiter.mutex) == OK, "waiter unlocks owned mutex");
+  }
+  waiter.finished = true;
+  if (waiter.cancel) {
+    PthreadTestcancel();
+    Check(false, "deferred mutex waiter cancellation remains pending");
+  }
+  return nullptr;
+}
+
+void QueueMutexWaiter(MutexWaiter &waiter) {
+  using namespace Libs::LibKernel;
+  PthreadAttr attr = nullptr;
+  Check(PthreadAttrInit(&attr) == OK &&
+            PthreadAttrSetstack(&attr, waiter.stack.data(), waiter.stack.size()) == OK &&
+            PthreadCreate(&waiter.thread, &attr, RunMutexWaiter, &waiter, "MutexTest") == OK &&
+            PthreadAttrDestroy(&attr) == OK,
+        "create guest mutex waiter");
+  // Pending guest signals are dispatched inside the contended lock, after enqueueing.
+  // Holding that callback proves queue order without letting the waiter acquire.
+  PthreadQueuePendingSignal(waiter.thread, 30);
+  WaitForMutexTest([&] { return waiter.parked.load() == 1; });
+}
+
+void TestMutexHandoff() {
+  using namespace Libs::LibKernel;
+  Common::Singleton<Loader::RuntimeLinker>::Instance();
+  Initialize();
+  PthreadInitSelfForMainThread();
+  Loader::SymbolDatabase symbols;
+  Libs::InitLibKernel_1(&symbols);
+  const auto *install_symbol = symbols.FindByNid("WkwEd3N7w0Y", Loader::SymbolType::Func);
+  const auto *remove_symbol = symbols.FindByNid("Qhv5ARAoOEc", Loader::SymbolType::Func);
+  Check(install_symbol != nullptr && remove_symbol != nullptr, "mutex signal exports resolve");
+  const auto install = reinterpret_cast<int(KYTY_SYSV_ABI *)(int, void *)>(install_symbol->vaddr);
+  const auto remove = reinterpret_cast<int(KYTY_SYSV_ABI *)(int)>(remove_symbol->vaddr);
+  Check(install(30, reinterpret_cast<void *>(ParkMutexWaiter)) == OK, "install mutex test signal");
+
+  PthreadMutexattr attr = nullptr;
+  Check(PthreadMutexattrInit(&attr) == OK && PthreadMutexattrSettype(&attr, 2) == OK,
+        "create recursive mutex attributes");
+  for (unsigned timed = 0; timed < 3; ++timed) {
+    PthreadMutex mutex = nullptr;
+    Check(PthreadMutexInit(&mutex, &attr, "MutexHandoff") == OK &&
+              PthreadMutexLock(&mutex) == OK && PthreadMutexLock(&mutex) == OK,
+          "lock recursive mutex twice");
+    std::array<MutexWaiter, 3> waiters{{{&mutex}, {&mutex}, {&mutex}}};
+    waiters[timed].timeout = 5000;
+    for (auto &waiter : waiters) {
+      QueueMutexWaiter(waiter);
+    }
+    waiters[timed].resume = 1;
+    Check(PthreadJoin(waiters[timed].thread, nullptr) == OK &&
+              waiters[timed].result == KERNEL_ERROR_ETIMEDOUT,
+          "timed waiter unlinks from queue head, middle and tail");
+    Check(PthreadMutexUnlock(&mutex) == OK && PthreadMutexTrylock(&mutex) == OK &&
+              PthreadMutexUnlock(&mutex) == OK,
+          "partial recursive unlock retains ownership");
+    Check(PthreadMutexUnlock(&mutex) == OK &&
+              PthreadMutexTrylock(&mutex) == KERNEL_ERROR_EBUSY,
+          "unlock reserves ownership before queued waiter runs");
+    for (auto &waiter : waiters) {
+      if (&waiter == &waiters[timed]) {
+        continue;
+      }
+      waiter.resume = 1;
+      WaitForMutexTest([&] { return waiter.acquired.load(); });
+      Check(PthreadMutexTrylock(&mutex) == KERNEL_ERROR_EBUSY, "FIFO waiter owns mutex");
+      waiter.release = true;
+      WaitForMutexTest([&] { return waiter.finished.load(); });
+      Check(PthreadJoin(waiter.thread, nullptr) == OK && waiter.result == OK,
+            "FIFO waiter completes");
+    }
+    Check(PthreadMutexLock(&mutex) == OK && PthreadMutexUnlock(&mutex) == OK &&
+              PthreadMutexDestroy(&mutex) == OK,
+          "mutex remains usable after timed waiter removal");
+  }
+  Check(PthreadMutexattrDestroy(&attr) == OK, "destroy recursive mutex attributes");
+
+  PthreadMutex mutex = nullptr;
+  Check(PthreadMutexInit(&mutex, nullptr, "MutexCancel") == OK &&
+            PthreadMutexLock(&mutex) == OK &&
+            PthreadMutexTimedlock(&mutex, 5000) == KERNEL_ERROR_EDEADLK,
+        "nonrecursive self timed lock reports deadlock");
+  MutexWaiter waiter{&mutex};
+  waiter.cancel = true;
+  QueueMutexWaiter(waiter);
+  Check(PthreadCancel(waiter.thread) == OK, "cancel queued mutex waiter");
+  PthreadQueuePendingSignal(waiter.thread, 30);
+  waiter.resume = 1;
+  WaitForMutexTest([&] { return waiter.parked.load() == 2; });
+  Check(PthreadMutexUnlock(&mutex) == OK &&
+            PthreadMutexTrylock(&mutex) == KERNEL_ERROR_EBUSY,
+        "canceled waiter retains its queue position");
+  waiter.release = true;
+  waiter.resume = 2;
+  void *result = nullptr;
+  Check(PthreadJoin(waiter.thread, &result) == OK &&
+            result == reinterpret_cast<void *>(-1) && waiter.acquired && waiter.finished,
+        "mutex lock defers cancellation until explicit cancellation point");
+  Check(PthreadMutexLock(&mutex) == OK && PthreadMutexUnlock(&mutex) == OK &&
+            PthreadMutexDestroy(&mutex) == OK && remove(30) == OK,
+        "mutex remains usable after waiter cancellation");
+}
+
 class TempDirectory {
 public:
   TempDirectory() {
@@ -483,6 +628,27 @@ void TestSaveOpenVisibility() {
   Check(FileSystem::KernelClose(truncated) == OK, "close truncated save file");
 }
 
+void TestFsync() {
+  Loader::SymbolDatabase symbols;
+  Libs::InitLibKernel_1(&symbols);
+  const auto* symbol = symbols.Find(
+      {"juWbTNM+8hw", "Posix", 1, "libkernel", 1, 1, Loader::SymbolType::Func});
+  Check(symbol != nullptr, "POSIX fsync export resolves");
+  const auto fsync = reinterpret_cast<int (KYTY_SYSV_ABI *)(int)>(symbol->vaddr);
+  Check(fsync(std::numeric_limits<int>::min()) == -1 &&
+            *Libs::Posix::GetErrorAddr() == Libs::Posix::POSIX_EBADF,
+        "invalid fsync descriptor reports POSIX EBADF");
+
+  constexpr char Path[] = "/savedata0/fsync.dat";
+  const int fd = FileSystem::KernelOpen(Path, 0x602, 0777);
+  Check(fd >= 3 && FileSystem::KernelWrite(fd, "save", 4) == 4 && fsync(fd) == OK &&
+            FileSystem::KernelLseek(fd, 0, 1) == 4,
+        "fsync flushes a writable file without changing its offset");
+  Check(FileSystem::KernelClose(fd) == OK && fsync(fd) == -1 &&
+            *Libs::Posix::GetErrorAddr() == Libs::Posix::POSIX_EBADF,
+        "closed fsync descriptor reports POSIX EBADF");
+}
+
 void TestSaveTruncate() {
   namespace Kernel = Libs::LibKernel;
   Loader::SymbolDatabase symbols;
@@ -973,6 +1139,9 @@ void CheckArchiveMount(const std::filesystem::path &root) {
             std::equal(bytes.begin(), bytes.begin() + 9, payload.end() - 9) &&
             FileSystem::KernelRead(fd, bytes.data(), bytes.size()) == 0,
         "archive reads stop at the member boundary");
+  Check(FileSystem::KernelFsync(fd) == OK &&
+            FileSystem::KernelLseek(fd, 0, 1) == payload.size(),
+        "fsync succeeds on an immutable archive member without changing its offset");
   Check(FileSystem::KernelWrite(fd, bytes.data(), 1) ==
             Libs::LibKernel::KERNEL_ERROR_EBADF &&
             FileSystem::KernelPwrite(fd, bytes.data(), 1, 0) ==
@@ -2236,7 +2405,7 @@ void CheckSocketWakeup() {
 
 } // namespace
 
-int main(int, char**) {
+int main(int argc, char** argv) {
   Common::InitializeThreads();
   Common::Subsystems subsystems;
   subsystems.Initialize<Config::Lifecycle>();
@@ -2245,6 +2414,13 @@ int main(int, char**) {
   Config::Load(options);
   subsystems.Initialize<Log::Lifecycle>();
 
+  TestMutexHandoff();
+  if (argc == 2 && std::string_view(argv[1]) == "--mutex-only") {
+    subsystems.Destroy();
+    std::printf("KernelFileSystemTests: mutex cases passed\n");
+    return 0;
+  }
+
   Check(SDL_InitSubSystem(SDL_INIT_VIDEO), "initialize Vulkan test video");
   auto graphics = std::make_unique<Libs::Graphics::WindowContext>();
   graphics->graphic_ctx.screen_width = 64;
@@ -2271,6 +2447,7 @@ int main(int, char**) {
   CheckAprPaths(temporary.Path());
   FileSystem::Mount(temporary.Path(), "/savedata0");
   TestSaveOpenVisibility();
+  TestFsync();
   TestSaveTruncate();
   TestSaveWritev();
   TestAioBatches();

tests/ResourceTrackingTests.cpp

@@ -593,8 +593,9 @@ void TestExactUnfilteredImageCandidates() {
     second[3] |= 1u << 16u;
     fill();
     plan.info.images[0].mip_mode = ImageMipMode::Dynamic;
-    Check(!MaterializeResources(plan, runtime, snapshot, specialization),
-          "mixed-dimension acceptance discarded mip-count compatibility");
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              specialization.images[0].mip_count == 1u && specialization.images[1].mip_count == 2u,
+          "mixed-dimension candidates lost their individual mip counts");
   }
   // The same image may be merged from several uses; one query keeps the guard strict.
   auto merged = MakeIndirectImageFixture(false, 32u, 0u, 0u, false, ValueOpcode::ImageRead, true);
@@ -889,7 +890,7 @@ void TestNestedPointerImageTable() {
           "nested pointer descriptors lost their row geometry or captured the GPU selector payload");
   }
 
-  constexpr uint32_t table = 0x2000u, objects = 0x4000u, stride = 592u, count = 6u;
+  constexpr uint32_t table = 0x2000u, objects = 0x10000u, stride = 592u, count = 59u;
   LinearTestMemory memory;
   memory.words.resize((objects + count * 0x430u - memory.base) / 4u);
   memory.fail_address = 0x1800u; // The GPU selector's contents are unavailable to the CPU.
@@ -898,7 +899,8 @@ void TestNestedPointerImageTable() {
     const auto object = objects + row * 0x430u;
     *at(table + row * stride) = object;
     const std::array<uint32_t, 8> image{10u + row,
-        uint32_t(Libs::Graphics::Prospero::BufferFormat::k32Float) << 20u, 0u,
+        uint32_t(row == count - 1u ? Libs::Graphics::Prospero::BufferFormat::k32UInt
+                                  : Libs::Graphics::Prospero::BufferFormat::k32Float) << 20u, 0u,
         Libs::Graphics::DstSel(4, 0, 0, 1) | (9u << 28u), 0u, 0u, 0u, 0u};
     std::copy(image.begin(), image.end(), at(object + 0x3d0u));
     const std::array<uint32_t, 4> sampler{row, 0x00fff000u, 0u, 0u};
@@ -917,6 +919,21 @@ void TestNestedPointerImageTable() {
     Check(snapshot.images[row + 1u].dwords[0] == 10u + row &&
               snapshot.samplers[row + 1u].dwords[0] == row,
           "nested pointer table mixed rows or pointee members");
+  auto typed = make();
+  typed->PlanAndTrack();
+  ApplyResourceSpecialization(typed->program, specialization);
+  const auto& samplers = typed->program.info.samplers;
+  Check(samplers.size() == 2u * (count + 1u) && samplers[0].indirect_resources.size() == count + 1u &&
+            samplers[count + 1u].indirect_resources.size() == count + 1u &&
+            snapshot.samplers[0].dwords[1] == 0u,
+        "expanded sampler table lost its null root or numeric-class variants");
+  for (uint32_t index = 0; index <= count; ++index) {
+    const auto& integer = samplers[index + count + 1u];
+    Check(samplers[index].snapshot_index == index && !samplers[index].integer_border &&
+              integer.snapshot_index == index && integer.integer_border && integer.force_point_filtering &&
+              samplers[count + 1u].indirect_resources[index] == index + count + 1u,
+          "large sampler class mapping changed descriptor identity or integer filtering");
+  }
   const auto populated = specialization;
   auto equivalent = make(0u, Address::Equivalent);
   equivalent->PlanAndTrack();
@@ -939,11 +956,11 @@ void TestNestedPointerImageTable() {
             specialization == populated,
         "empty-to-positive pointer table refresh did not restore its specialization");
   *at(table) = objects + 0x430u;
-  *at(objects + 0x430u + 0x3d0u) = 42u;
-  *at(objects + 0x430u + 0x3f8u) = 7u;
+  *at(objects + 0x430u + 0x3d0u) = 10u + count;
+  *at(objects + 0x430u + 0x3f8u) = count;
   Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
             snapshot.images.size() == count && snapshot.samplers.size() == count &&
-            snapshot.images[1].dwords[0] == 42u && snapshot.samplers[1].dwords[0] == 7u,
+            snapshot.images[1].dwords[0] == 10u + count && snapshot.samplers[1].dwords[0] == count,
         "nested descriptor refresh retained stale pointers or failed to deduplicate new pointees");
   *at(table) = objects;
 
@@ -3164,6 +3181,243 @@ void TestUniformizedMaterialImageKeys() {
                "packed selector accepted an unproved OR guard, offset domain, or field extraction");
 }
 
+void TestIndexedMaterialImageKeys() {
+  namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
+  const auto make_program = [](ShaderType stage = ShaderType::Compute, bool shared = false) {
+    Fixture f(stage);
+    auto *entry = f.block, *body = f.AddBlock();
+    auto *later = shared ? f.AddBlock() : nullptr;
+    auto *done = f.AddBlock();
+    entry->AddBranch(body);
+    entry->AddBranch(done);
+    body->AddBranch(shared ? later : done);
+    if (shared) later->AddBranch(done);
+    const auto lane = f.Emit(ValueOpcode::GetBuiltin,
+        {Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(0u)});
+    const auto active = f.Emit(ValueOpcode::INotEqual32, {lane, Value(UINT32_MAX)});
+    entry->condition = f.Emit(ValueOpcode::ConditionRef, {active}, CFG::BranchCondition::ExecNonZero);
+    entry->terminator = {.kind = CFG::TerminatorKind::ConditionalBranch,
+                         .true_block = body, .false_block = done};
+    body->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = shared ? later : done};
+    if (shared) later->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = done};
+    done->terminator.kind = CFG::TerminatorKind::Return;
+    const auto material = f.Buffer({f.UserData(0), f.UserData(1), f.UserData(2), f.UserData(3)});
+    const auto heap = f.Buffer({f.UserData(4), f.UserData(5), f.UserData(6), f.UserData(7)});
+    const auto output = f.Buffer({f.UserData(8), f.UserData(9), f.UserData(10), f.UserData(11)});
+    const auto sampler = f.Sampler({Value(0u), Value(0u), Value(1u << 26u), Value(0u)});
+    const auto address = f.ImageAddress();
+    Value owner_key, shared_image;
+    for (uint32_t index = 0; index < (shared ? 3u : 2u); ++index) {
+      const uint32_t field = index == 1u ? 68u : 64u;
+      auto *use = index == 2u ? later : body;
+      auto key = owner_key;
+      if (index < 2u) {
+        MemoryInfo memory;
+        memory.kind = ResourceKind::Buffer;
+        memory.idxen = true;
+        memory.offset = field;
+        const auto loaded = f.Emit(ValueOpcode::LoadBufferU32,
+            {material, lane, Value(0u), Value(0u), active}, f.AddMemory(memory, field), entry);
+        const auto selected = f.Emit(ValueOpcode::SelectU32, {active, loaded, Value(0xdeadbeefu)}, 0, entry);
+        key = shared && index == 0u
+            ? f.Emit(ValueOpcode::ReadLane, {selected, Value(0u)}, 0, entry)
+            : f.Emit(ValueOpcode::ReadFirstLane, {selected, active}, 0, body);
+        if (shared && index == 0u) {
+          owner_key = key;
+          const auto equal = f.Emit(ValueOpcode::IEqual32, {key, selected}, 0, entry);
+          const auto guard = f.Emit(ValueOpcode::LogicalAnd, {active, equal}, 0, entry);
+          entry->condition = f.Emit(ValueOpcode::ConditionRef, {guard}, CFG::BranchCondition::ExecNonZero, entry);
+        }
+      }
+      const auto offset = f.Emit(ValueOpcode::ShiftLeftLogical32, {key, Value(5u)}, 0, use);
+      std::array<Value, 8> words;
+      for (uint32_t word = 0; word < words.size(); ++word) {
+        MemoryInfo scalar;
+        scalar.kind = ResourceKind::ScalarBuffer;
+        scalar.offset = word * 4u;
+        words[word] = f.Emit(ValueOpcode::ReadConstBuffer, {heap, offset},
+                            f.AddMemory(scalar, 0x100u + field), use);
+      }
+      const auto image_handle = [&] {
+        return f.Emit(ValueOpcode::GetImageResource,
+            {words[0], words[1], words[2], words[3], words[4], words[5], words[6], words[7]}, 0, use);
+      };
+      MemoryInfo gather;
+      gather.kind = ResourceKind::Image;
+      gather.image_dimension = Decoder::ImageDimension::Dim2D;
+      if (index == 0u)
+        f.Emit(ValueOpcode::ImageQueryDimensions, {image_handle(), address}, f.AddMemory(gather, 0x180u), body);
+      const auto image = image_handle();
+      if (index == 2u) shared_image = image;
+      gather.image_sample_flags = Decoder::ImageSampleFlagLod;
+      f.Emit(ValueOpcode::ImageGatherRaw,
+          {image, sampler, address, words[1], words[3], Value(0u), Value(1u << 26u), active},
+          f.AddMemory(gather, 0x200u + field), use);
+    }
+    MemoryInfo store;
+    store.kind = ResourceKind::Buffer;
+    f.Emit(ValueOpcode::StoreBufferU32,
+        {output, Value(0u), Value(0u), Value(0u), Value(1u), Value(true)},
+        f.AddMemory(store, 0x300u), done);
+    f.PlanAndTrack();
+    if (shared) {
+      owner_key.ResolveInstruction()->ReplaceUsesWith(Value(12994u));
+      uint32_t offset = 0u;
+      Check(SrtWalker(f.program, {}).Evaluate(shared_image.ResolveInstruction()->Arg(0), offset) &&
+                offset == 12994u * 32u,
+            "inherited indexed selector changed GPU byte offsets into raw material keys");
+    }
+    return std::move(f.program);
+  };
+  auto program = make_program();
+  auto plan = ExtractResourcePlan(program);
+  Check(plan.info.images.size() == 3u && plan.selector_keys.size() == 2u,
+        "indexed material fields did not share their query/gather selector domain");
+  for (const auto &image : plan.info.images) {
+    const auto &source = *plan.descriptor_sources[image.source].indirect_descriptor;
+    Check(source.table_scalar && source.selector && source.selector->stride == 0u &&
+              source.key_count.IsEmpty() && source.selector_first.IsEmpty(),
+          "indexed material field fell back to an unconditional heap scan");
+  }
+  constexpr uint32_t base = 0x1000u, heap = 0x2000u, records = 14517u;
+  struct Memory {
+    std::vector<uint32_t> words = std::vector<uint32_t>((heap - base + records * 32u) / 4u);
+    uint32_t selectors = 0, descriptors = 0;
+    bool narrow = true, unreadable = false;
+  } memory;
+  const auto put = [&](uint32_t key, uint32_t mips) {
+    auto *words = memory.words.data() + (heap - base + key * 32u) / 4u;
+    words[0] = 0x100u + key;
+    words[1] = static_cast<uint32_t>(Libs::Graphics::Prospero::BufferFormat::k16UNorm) << 20u;
+    words[3] = (9u << 28u) | ((mips - 1u) << 16u) | Libs::Graphics::DstSel(4, 5, 6, 7);
+  };
+  for (uint32_t key = 0; key < 64u; ++key) put(key, 1u);
+  put(12994u, 8u);
+  put(12996u, 8u);
+  for (uint32_t row = 0; row < 10u; ++row) {
+    memory.words[(row * 96u + 64u) / 4u] = 12994u;
+    memory.words[(row * 96u + 68u) / 4u] = 12996u;
+  }
+  std::array<uint32_t, 12> user_data{base, 96u << 16u, 10u, 0x5204u,
+      heap, 32u << 16u, records, 0x5204u, 0x90000u, 4u << 16u, 16u, 0x5204u};
+  const auto read = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+    auto &m = *static_cast<Memory *>(data);
+    if (address < base || address - base + words.size_bytes() > m.words.size() * 4u) return false;
+    if (address < heap) {
+      ++m.selectors;
+      if (m.unreadable) return false;
+    } else {
+      ++m.descriptors;
+      const auto key = (address - heap) / 32u;
+      if (m.narrow && (words.size() != 8u || (key != 0u && key != 12994u && key != 12996u))) return false;
+    }
+    std::copy_n(m.words.begin() + (address - base) / 4u, words.size(), words.begin());
+    return true;
+  };
+  const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+                           .read_specialization_memory = read};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            memory.selectors == 2u && memory.descriptors == 6u && snapshot.images.size() == 9u,
+        "indexed selector scanned the broad heap or did not share its captured keys");
+  ApplyResourceSpecialization(program, specialization);
+  Check(program.info.images[0].indirect_resources.size() == 3u,
+        "indirect dimension query lost its specialized candidates");
+  for (uint32_t root = 0; root < plan.info.images.size(); ++root) {
+    const auto &image = specialization.images[root];
+    const auto offset = image.indirect_mapping_offset;
+    Check(snapshot.images[root].dwords[0] == 0u && image.indirect_search_iterations == 2u &&
+              snapshot.flattened_srt[offset] == 2u && snapshot.flattened_srt[offset + 1u] == 0u &&
+              snapshot.flattened_srt[offset + 3u] == (root == 2u ? 12996u : 12994u) * 32u,
+          "byte-key selector lost native OOB zero, sorted offsets, or the null default");
+  }
+  Check(std::ranges::count_if(specialization.images, [](const auto &image) {
+          return image.mip_count == 8u;
+        }) == 2u, "indirect gathers lost candidate-specific mip counts");
+  for (const uint32_t field : {64u, 68u})
+    Check(std::ranges::find(snapshot.specialization_reads,
+              std::pair<uint64_t, uint64_t>{base + field, 868u}) != snapshot.specialization_reads.end(),
+          "indexed selector batch escaped writable-alias tracking");
+  auto shared_program = make_program(ShaderType::Compute, true);
+  auto shared = ExtractResourcePlan(shared_program);
+  Check(MaterializeResources(shared, runtime, snapshot, specialization) && snapshot.images.size() == 9u,
+        "inherited indexed selector lost its bounded material domain");
+  memory.words[64u / 4u] = 12996u;
+  memory.selectors = 0;
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            memory.selectors == 2u && snapshot.images.size() == 11u,
+        "changed material rows reused stale selector keys");
+  memory.words[64u / 4u] = 12994u;
+  memory.unreadable = true;
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
+        "unreadable eligible selector silently fell back to the heap");
+  memory.narrow = false;
+  const auto original = user_data;
+  for (const uint32_t variant : {0u, 1u, 2u, 3u, 4u, 5u, 6u, 7u}) {
+    user_data = original;
+    if (variant == 0u) user_data[1] |= 1u << 31u;
+    if (variant == 1u) user_data[3] |= 1u << 23u;
+    if (variant == 2u) user_data[3] |= 1u << 28u;
+    if (variant == 3u) user_data[0] += 1u;
+    if (variant == 4u) user_data[2] = MaxIndirectDescriptorProbes + 1u;
+    if (variant == 5u) user_data[8] = base + 68u;
+    if (variant == 6u) user_data[1] = 66u << 16u;
+    if (variant == 7u) user_data[0] = 0u;
+    memory.selectors = 0;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              memory.selectors == 0u && snapshot.images.size() > 100u,
+          "unsupported geometry or writable selector lost the original heap domain");
+  }
+  user_data = original;
+  auto graphics_program = make_program(ShaderType::Pixel);
+  auto graphics = ExtractResourcePlan(graphics_program);
+  Check(MaterializeResources(graphics, runtime, snapshot, specialization) && memory.selectors == 0u,
+        "graphics material lookup acquired unproved selector read dependencies");
+  memory.unreadable = false;
+  for (uint32_t row = 0; row < 10u; ++row)
+    memory.words[(row * 96u + 64u) / 4u] = 0x08000000u; // U32 key*32 wraps to zero.
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) && snapshot.images.size() == 7u,
+        "indexed selector lost wrapped scalar descriptor offsets");
+  for (auto &source : plan.descriptor_sources) {
+    if (!source.indirect_descriptor) continue;
+    source.indirect_descriptor->table_offset = 0xffffffe0u;
+    source.indirect_descriptor->table_immediate = 32u;
+  }
+  for (uint32_t row = 0; row < 10u; ++row) {
+    memory.words[(row * 96u + 64u) / 4u] = row == 0u ? 1u : 0u;
+    memory.words[(row * 96u + 68u) / 4u] = 0u;
+  }
+  memory.selectors = memory.descriptors = 0u;
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            memory.selectors == 2u && memory.descriptors == 2u && snapshot.images.size() == 5u,
+        "overflowed scalar immediate read a descriptor or lost valid wrapped keys");
+  for (uint32_t root = 0; root < 2u; ++root) {
+    const auto offset = specialization.images[root].indirect_mapping_offset;
+    Check(snapshot.images[root].dwords[0] == 0u && snapshot.flattened_srt[offset] == 2u &&
+              snapshot.flattened_srt[offset + 1u] == 32u && snapshot.flattened_srt[offset + 3u] == UINT32_MAX &&
+              snapshot.flattened_srt[offset + 4u] == 0u,
+          "invalid byte key selected a real texture instead of the null root");
+  }
+  for (auto &source : plan.descriptor_sources) {
+    if (!source.indirect_descriptor) continue;
+    source.indirect_descriptor->table_offset = source.indirect_descriptor->table_immediate = 0u;
+  }
+  memory.narrow = true;
+  memory.unreadable = true;
+  for (const uint32_t variant : {0u, 1u, 2u, 3u}) {
+    user_data = original;
+    if (variant == 0u) user_data[2] = 0u;
+    if (variant == 1u) user_data[1] = 0u;
+    if (variant == 2u) user_data[3] = 0u;
+    if (variant == 3u) user_data[1] = 64u << 16u;
+    memory.selectors = 0;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) && memory.selectors == 0u &&
+              snapshot.images.size() == 6u,
+          "native zero/OOB indexed load read its unavailable payload");
+  }
+}
+
 void TestImageDescriptorFields() {
   constexpr std::array<std::pair<uint32_t, uint32_t>, 5> reserved{
       {{1u, 0x20000000u}, {2u, 0x70003000u}, {4u, 0xe000e000u},
@@ -4929,6 +5183,38 @@ void TestImagesSamplersAndAliases() {
         "buffer/image descriptor alias was not linked");
 }
 
+void TestIndependentImageSamplerPairs() {
+  Fixture fixture(ShaderType::Pixel);
+  const auto address = fixture.ImageAddress();
+  MemoryInfo memory;
+  memory.kind = ResourceKind::Image;
+  memory.image_dimension = Decoder::ImageDimension::Dim2D;
+  for (uint32_t use = 0; use < 144; ++use) {
+    const auto pair = use % 72;
+    const auto pc = use * 4;
+    const auto image = fixture.Image(
+        {Value(pair / 8), Value(0u), Value(0u), Value(0u), Value(0u),
+         Value(0u), Value(0u), Value(0u)}, pc);
+    const auto sampler = fixture.Sampler(
+        {Value(pair % 8), Value(0u), Value(0u), Value(0u)}, pc);
+    fixture.Emit(ValueOpcode::ImageSampleRaw,
+                 {image, sampler, address, Value(true)},
+                 fixture.AddMemory(memory, pc));
+  }
+  fixture.PlanAndTrack();
+
+  Check(fixture.program.info.images.size() == 9 &&
+            fixture.program.info.samplers.size() == 8 &&
+            fixture.program.info.sampled_pairs.size() == 72,
+        "distinct image/sampler combinations were limited or duplicated");
+  for (uint32_t index = 0; index < 72; ++index) {
+    const auto &pair = fixture.program.info.sampled_pairs[index];
+    Check(pair.image == index / 8 && pair.sampler == index % 8 &&
+              pair.first_use_pc == index * 4,
+          "sampled pair lost its independent resources or first use");
+  }
+}
+
 void TestSampleAdjustSamplerScratch() {
   Fixture fixture(ShaderType::Pixel);
   const auto active = fixture.Emit(
@@ -5145,19 +5431,24 @@ void TestGatherLodSamplerValidation() {
           Libs::Graphics::Prospero::BufferFormat::k32Float) << 20u;
       user_data[3] = Libs::Graphics::DstSel(4, 5, 6, 7) |
           (static_cast<uint32_t>(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
-      user_data[9] = 0xfff000u;
       constexpr std::array sampler_controls{
           0u, 0x4001u, 1u << 26u, (1u << 26u) | 256u,
-          (1u << 26u) | (1u << 14u), 2u << 26u, 3u << 26u};
+          (1u << 26u) | (1u << 14u), 2u << 26u, (2u << 26u) | 256u,
+          (2u << 26u) | (1u << 14u), 0x0a500000u, 0x1a500000u, 3u << 26u};
       for (const auto control : sampler_controls) {
-        user_data[10] = control;
-        ResourceSnapshot snapshot;
-        ResourceSpecialization specialization;
-        const bool supported = !explicit_lod || (control >> 26u) == 0 ||
-                               control == (1u << 26u);
-        Check(MaterializeResources(plan, {.user_data = user_data}, snapshot,
-                                   specialization) == supported,
-              "explicit gather accepted an unsupported sampler or rejected a valid one");
+        for (const uint32_t perf_mip : {0u, 1u}) {
+          user_data[9] = 0xfff000u | (perf_mip << 24u);
+          user_data[10] = control;
+          ResourceSnapshot snapshot;
+          ResourceSpecialization specialization;
+          const auto filter = (control >> 26u) & 3u;
+          const bool supported = !explicit_lod || filter == 0 ||
+              (filter <= 2 && (control & 0xfffffu) == 0 &&
+               (filter != 2 || perf_mip == 0));
+          Check(MaterializeResources(plan, {.user_data = user_data}, snapshot,
+                                     specialization) == supported,
+                "explicit gather accepted an unsupported sampler or rejected a valid one");
+        }
       }
     }
   }
@@ -5376,6 +5667,65 @@ void TestSrtFlatteningAndRuntimeMemoization() {
         "flattened typed SRT reads were not collected");
 }
 
+void TestScalarDataReadControlFlow() {
+  namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
+  for (const auto comparison : {ValueOpcode::Void, ValueOpcode::IEqual64,
+                                ValueOpcode::INotEqual64}) {
+    Fixture fixture;
+    const auto low = fixture.UserData(0);
+    const auto high = fixture.UserData(1);
+    const auto address = fixture.Address(low, high);
+    auto* use = fixture.block;
+    if (comparison != ValueOpcode::Void) {
+      const auto pointer = fixture.Emit(ValueOpcode::CompositeConstructU64, {low, high});
+      auto enabled = fixture.Emit(comparison, {pointer, Value(uint64_t{0})});
+      if (comparison == ValueOpcode::IEqual64)
+        enabled = fixture.Emit(ValueOpcode::LogicalNot, {enabled});
+      const auto lane = fixture.Emit(ValueOpcode::GetBuiltin,
+          {Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(0u)});
+      const auto varying = fixture.Emit(ValueOpcode::INotEqual32, {lane, Value(0u)});
+      fixture.block->condition = fixture.Emit(ValueOpcode::LogicalAnd, {varying, enabled});
+      use = fixture.AddBlock();
+      auto* done = fixture.AddBlock();
+      fixture.block->AddBranch(use);
+      fixture.block->AddBranch(done);
+      use->AddBranch(done);
+      fixture.block->terminator = {.kind = CFG::TerminatorKind::ConditionalBranch,
+          .true_block = use, .false_block = done};
+      use->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = done};
+      done->terminator.kind = CFG::TerminatorKind::Return;
+    }
+    MemoryInfo scalar;
+    scalar.kind = ResourceKind::ScalarAddress;
+    fixture.Emit(ValueOpcode::LoadAddressU32,
+        {address, Value(0u), Value(0u), Value(true)}, fixture.AddMemory(scalar, 4), use);
+    fixture.PlanAndTrack();
+    auto plan = ExtractResourcePlan(fixture.program);
+    Check(plan.srt_reads.size() == 1 && !fixture.program.info.uses_dma &&
+              bool(plan.srt_code) == (comparison == ValueOpcode::Void),
+          "scalar data lost flattening or conditional reads entered the unconditional JIT");
+    uint32_t reads = 0;
+    std::array<uint32_t, 2> user_data {};
+    SrtRuntime runtime {.user_data = user_data,
+        .read_memory = +[](void* data, uint64_t address, std::span<uint32_t> words) {
+          ++*static_cast<uint32_t*>(data);
+          if (address == 0) return false;
+          words[0] = 42;
+          return true;
+        }, .userdata = &reads};
+    ResourceSnapshot snapshot;
+    ResourceSpecialization specialization;
+    for (const uint64_t pointer : {uint64_t{0}, uint64_t{0x1000}, uint64_t{1} << 32u}) {
+      if (pointer == 0 && comparison == ValueOpcode::Void) continue;
+      user_data = {static_cast<uint32_t>(pointer), static_cast<uint32_t>(pointer >> 32u)};
+      reads = 0;
+      Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+                reads == (pointer != 0) && snapshot.flattened_srt[0] == (pointer ? 42u : 0u),
+            "scalar data refresh ignored a mixed EXEC/64-bit null-pointer branch");
+    }
+  }
+}
+
 void TestDynamicSrtReadRemainsExplicit() {
   Fixture fixture;
   const auto base =
@@ -7030,6 +7380,83 @@ void TestFiniteImagePhiCycle() {
   }
 }
 
+void TestPackedDescriptorWaterfall() {
+  for (const uint32_t variant : {0u, 1u, 2u}) {
+    Fixture f(ShaderType::Pixel);
+    const auto packed = f.Emit(ValueOpcode::GetInterpolationParameter,
+                               {Value(1u), Value(1u), Value(2u)});
+    const auto active = f.Emit(ValueOpcode::ULessThan32,
+                               {f.Emit(ValueOpcode::LaneId), Value(16u)});
+    const auto image_key = f.Emit(ValueOpcode::BitwiseAnd32, {packed, Value(65535u)});
+    const auto sampler_key = variant == 2u ? packed :
+        f.Emit(ValueOpcode::ShiftRightLogical32, {packed, Value(16u)});
+    auto selected_active = active;
+    const auto descriptor = [&](Value key, uint32_t first, uint32_t count, uint32_t shift) {
+      const auto selected = f.Emit(ValueOpcode::ReadFirstLane, {key, active});
+      selected_active = f.Emit(ValueOpcode::LogicalAnd, {selected_active,
+          f.Emit(ValueOpcode::IEqual32, {selected, key})});
+      const auto offset = f.Emit(ValueOpcode::ShiftLeftLogical32, {selected, Value(shift)});
+      const auto table = f.Buffer({f.UserData(first), f.UserData(first + 1u),
+                                   f.UserData(first + 2u), f.UserData(first + 3u)});
+      std::array<Value, 8> words;
+      for (uint32_t word = 0; word < count; ++word)
+        words[word] = f.Emit(ValueOpcode::ReadConstBuffer, {table, offset},
+            f.AddMemory({.kind = ResourceKind::ScalarBuffer, .offset = word * 4u}, 0x58u));
+      return words;
+    };
+    const auto image = f.Image(descriptor(image_key, 0u, 8u, 5u));
+    const auto sampler_words = descriptor(sampler_key, 4u, 4u, 4u);
+    const auto sampler = f.Sampler({sampler_words[0], sampler_words[1],
+                                   sampler_words[2], sampler_words[3]});
+    f.Emit(ValueOpcode::ImageSampleRaw,
+        {image, sampler, f.ImageAddress(), variant == 1u ? Value(true) : selected_active},
+        f.AddMemory({.kind = ResourceKind::Image,
+                     .image_dimension = Decoder::ImageDimension::Dim2D}, 0x6cu));
+    if (variant != 0u) {
+      CheckFatal([&] { f.PlanAndTrack(); }, "not a valid runtime value",
+                 "packed waterfall accepted widened EXEC or an unbounded sampler key");
+      continue;
+    }
+    f.PlanAndTrack();
+    const auto plan = ExtractResourcePlan(f.program);
+    for (const auto source : {plan.info.images[0].source, plan.info.samplers[0].source}) {
+      const auto& value = plan.descriptor_sources[source];
+      const auto& table = value.indirect_descriptor;
+      Check(table && !table->selector && table->table_scalar &&
+                table->table_stride == value.dword_count * 4u &&
+                (value.dword_count != 4u || table->table_maximum == 65535u),
+            "packed waterfall lost its static sampler bit bound or captured the varying key");
+    }
+    LinearTestMemory memory;
+    const std::array<uint32_t, 8> user_data{
+        uint32_t(memory.base) + 256u, 32u << 16u, 2u, 0x5204u,
+        uint32_t(memory.base) + 512u, 16u << 16u, 59u, 0x5204u};
+    for (uint32_t row = 0; row < 2u; ++row) {
+      const auto word = (256u + row * 32u) / 4u;
+      memory.words[word] = 0x1000u + row * 0x1000u;
+      memory.words[word + 1u] = uint32_t(Libs::Graphics::Prospero::BufferFormat::k32Float) << 20u;
+      memory.words[word + 3u] = Libs::Graphics::DstSel(4, 0, 0, 1) | (9u << 28u);
+    }
+    for (uint32_t row = 0; row < 59u; ++row) {
+      const auto word = (512u + row * 16u) / 4u;
+      memory.words[word] = row;
+      memory.words[word + 1u] = 0x00fff000u;
+      memory.words[word + 2u] = (row & 1u) ? 0x0a500000u : 0u;
+    }
+    const SrtRuntime runtime{.user_data = user_data, .read_memory = ReadLinearTestMemory,
+        .userdata = &memory, .read_specialization_memory = ReadLinearTestMemory};
+    ResourceSnapshot snapshot;
+    ResourceSpecialization specialization;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              snapshot.images.size() == 3u && snapshot.samplers.size() == 60u,
+          "packed waterfall ignored the two-image/59-sampler heap bounds or lost the null roots");
+    Check(std::ranges::all_of(snapshot.specialization_reads, [&](const auto& span) {
+      return (span.first >= memory.base + 256u && span.first + span.second <= memory.base + 320u) ||
+          (span.first >= memory.base + 512u && span.first + span.second <= memory.base + 512u + 59u * 16u);
+    }), "packed waterfall materialization read outside the descriptor heaps");
+  }
+}
+
 void TestFiniteImageWaterfall() {
   namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
   enum class Variant { Valid, Staggered, Cycle, DifferentMask, MixedWord, DifferentSelect, WiderExec, BadLeaf };
@@ -8255,6 +8682,7 @@ int main() {
     Run("scalar/vector alias", TestScalarAndVectorBufferAlias);
     Run("runtime unsigned min", TestRuntimeUnsignedMinDescriptor);
     Run("images and samplers", TestImagesSamplersAndAliases);
+    Run("independent image/sampler pairs", TestIndependentImageSamplerPairs);
     Run("SampleAdjust sampler scratch", TestSampleAdjustSamplerScratch);
     Run("FMASK load specialization", TestFmaskLoadSpecialization);
     Run("gather LOD sampler validation", TestGatherLodSamplerValidation);
@@ -8269,6 +8697,7 @@ int main() {
     Run("guarded direct image table", TestGuardedDirectImageTable);
     Run("bounded compute image loop", TestBoundedComputeImageLoop);
     Run("uniformized material image keys", TestUniformizedMaterialImageKeys);
+    Run("indexed material image keys", TestIndexedMaterialImageKeys);
     Run("image descriptor fields", TestImageDescriptorFields);
     Run("draw-uniform scalar image", TestUniformScalarBufferImage);
     Run("compact scalar image table", TestCompactScalarImageTable);
@@ -8278,6 +8707,7 @@ int main() {
     Run("shared append selector bounds", TestSharedAppendSelectorBounds);
     Run("descending image range", TestDescendingImageRange);
     Run("SRT runtime", TestSrtFlatteningAndRuntimeMemoization);
+    Run("scalar data control flow", TestScalarDataReadControlFlow);
     Run("dynamic SRT", TestDynamicSrtReadRemainsExplicit);
     Run("writable descriptor phi", TestWritableDescriptorPhi);
     Run("bounded formatted buffer store", TestBoundedFormattedBufferStore);
@@ -8292,6 +8722,7 @@ int main() {
     Run("packed scalar buffer selection", TestPackedScalarBufferSelection);
     Run("conditional sampler phi", TestConditionalSamplerPhi);
     Run("finite image phi cycle", TestFiniteImagePhiCycle);
+    Run("packed descriptor waterfall", TestPackedDescriptorWaterfall);
     Run("finite image waterfall", TestFiniteImageWaterfall);
     Run("finite inline sampler", TestFiniteInlineSampler);
     Run("finite image bit scan sentinel", TestFiniteImageBitScanSentinel);

tests/shaderCfgTests.cpp

@@ -45,6 +45,7 @@
 #include <span>
 #include <sstream>
 #include <string>
+#include <tuple>
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
@@ -1967,6 +1968,59 @@ void TestNggVertexEntryState() {
   }
 }
 
+void TestNativeDataCacheInvalidation() {
+  using namespace ShaderRecompiler;
+  // Captured f5edf70dbb1f6057:0x194 and 0x1a4, each followed by S_WAITCNT.
+  const std::array cache_ops = {0xe1c40000u, 0u, 0xbf8c0000u,
+                               0xe1c80000u, 0u, 0xbf8c0000u, EncodeSopp(0x01)};
+  Decoder::Program decoded;
+  Decoder::DecodeProgram(cache_ops, decoded);
+  for (const uint32_t index : {0u, 2u}) {
+    const auto &inst = decoded.instructions[index];
+    Check(inst.opcode == (index == 0 ? Decoder::Opcode::BUFFER_GL0_INV
+                                    : Decoder::Opcode::BUFFER_GL1_INV) &&
+              inst.word_count == 2 && inst.src_count == 0 && inst.data_dwords == 0 &&
+              inst.dst.kind == Decoder::OperandKind::Unknown,
+          "RDNA2 cache invalidation acquired a buffer operand or lost its encoding");
+  }
+
+  auto options = MakeCompileOptions(ShaderType::Compute);
+  options.dump_ir = true;
+  ShaderComputeInputInfo compute{};
+  compute.threads_num[1] = compute.threads_num[2] = 1;
+  options.input_info.compute = &compute;
+  for (const uint32_t wave_size : {32u, 64u}) {
+    compute.threads_num[0] = compute.wave_size = options.wave_size = wave_size;
+    const auto minimal = RecompileForTest(cache_ops, options);
+    CheckSpirvBinaryValidates(minimal.spirv);
+    Check(minimal.program.info.buffers.empty(),
+          "cache invalidation introduced a phantom buffer binding");
+    const auto source = DisassembleSpirvBinary(minimal.spirv);
+    const auto first = source.find("OpMemoryBarrier %uint_1 %uint_2114");
+    const auto second = source.find("OpMemoryBarrier %uint_1 %uint_2114", first + 1);
+    Check(first != std::string::npos && second != std::string::npos &&
+              source.find("OpMemoryBarrier", second + 1) == std::string::npos &&
+              source.find("OpControlBarrier") == std::string::npos,
+          "cache invalidation must fence device memory once per wave without convergence");
+
+    const std::array reload = {
+        EncodeMubuf0(0x1c, 0, false), EncodeMubuf1(0, 0, 0),
+        0xe1c40000u, 0u, 0xbf8c0000u,
+        EncodeMubuf0(0x0c, 0, false), EncodeMubuf1(1, 0, 0),
+        EncodeMubuf0(0x1c, 4, false), EncodeMubuf1(1, 0, 0), EncodeSopp(0x01),
+    };
+    const auto result = RecompileForTest(reload, options);
+    CheckSpirvBinaryValidates(result.spirv);
+    const auto store = result.ir_dump.find("StoreBufferU32");
+    const auto fence = result.ir_dump.find("DataCacheInvalidate");
+    const auto load = result.ir_dump.find("LoadBufferU32");
+    Check(store < fence && fence < load && load != std::string::npos &&
+              result.program.info.coherent_buffers &&
+              DisassembleSpirvBinary(result.spirv).find(" Coherent") != std::string::npos,
+          "plain buffer reload lost its cache fence order or coherent visibility");
+  }
+}
+
 void TestNativeStoreCompletion() {
   using namespace ShaderRecompiler;
   const std::array register_wait = {EncodeSopk(0x17, 4, static_cast<int16_t>(0x9234)), EncodeSopp(0x01)};
@@ -10751,9 +10805,14 @@ void TestMemoryFedScalarLeafCall() {
   };
   IR::SrtRuntime runtime{.read_memory = read, .userdata = &memory, .read_specialization_memory = read};
   const auto options = MakeCompileOptions(ShaderType::Compute);
+  const uint32_t gpu_callable[] = {EncodeVop1(0x02, 4, 256),
+      EncodeSmem0(0x09, 14, 0), 4u << 25u, EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
+  const auto skipped = PrepareShaderSource(gpu_callable, options);
+  Check(skipped.unsupported_call_pc == 12u && skipped.calls.empty(),
+        "GPU-buffer callable did not report its unsupported call site");
   auto source = PrepareShaderSource(shader, options);
-  Check(source.call && source.decoded.has_swap_pc &&
-            source.decoded.instructions[source.call->instruction].branch_target == UINT32_MAX,
+  Check(source.calls.size() == 1 && source.decoded.has_swap_pc &&
+            source.decoded.instructions[source.calls.front().instruction].branch_target == UINT32_MAX,
         "memory-fed SWAPPC did not retain an unresolved target");
 #if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
   ExpectFatal([&] { (void)TranslateProgram(source.decoded, options); },
@@ -10761,7 +10820,7 @@ void TestMemoryFedScalarLeafCall() {
 #endif
   const auto& linked = RefreshShaderSource(source, runtime);
   Check(linked.has_swap_pc && source.revision == 1 && linked.instructions.back().branch_target == 0x40 &&
-            linked.instructions[source.call->instruction].branch_target == 0x44 &&
+            linked.instructions[source.calls.front().instruction].branch_target == 0x44 &&
             source.reads.back() == std::pair<uint64_t, uint64_t>{base + 0x100, 12},
         "aliased call target, exact leaf range, or continuation was lost");
   const auto* storage = linked.instructions.data();
@@ -10846,14 +10905,14 @@ void TestMemoryFedScalarLeafCall() {
   memory.words[96] = EncodeSMovB32(14, 129);
   ExpectFatal([&] { (void)RefreshShaderSource(source, runtime); },
               "leaf clobbering its return pair was linked");
-  // AND_B64 writes the high half despite the decoder's generic scalar width.
+  // AND_B64 must replace the preceding high-half definition.
   const uint32_t high_writer[] = {
       EncodeSMovB32(15, 129), EncodeSop2(0x0f, 14, 128, 128),
       EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
   auto pair_source = PrepareShaderSource(high_writer, options);
   IR::SrtWalker pair_walker(pair_source.call_targets, runtime);
   uint32_t high = 1;
-  Check(pair_walker.Evaluate(pair_source.call->target[1], high) && high == 0,
+  Check(pair_walker.Evaluate(pair_source.calls.front().target[1], high) && high == 0,
         "scalar pair producer used a stale preceding high-half definition");
   const uint32_t joined[] = {
       EncodeSopp(0x04, 2), EncodeSMovB32(14, 129), EncodeSopp(0x02, 1),
@@ -10862,7 +10921,7 @@ void TestMemoryFedScalarLeafCall() {
               "ambiguous reaching definitions were accepted for a call target");
   const uint32_t multiple[] = {EncodeSop1(0x21, 14, 14), EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
   ExpectFatal([&] { (void)PrepareShaderSource(multiple, options); },
-              "call metadata bypassed the multiple-call rejection");
+              "a native return pair was accepted as a later call target");
 #endif
 }
 
@@ -10909,7 +10968,7 @@ void TestFusedShaderHandoffPreservesRegisters() {
   for (const auto handoff: {EncodeSop1(0x20, 0, 6), 0xbefd2106u}) {
     front[2] = handoff; // SETPC or captured SWAPPC with NULL destination
     auto source = PrepareShaderSource(front, options);
-    Check(!source.decoded.has_swap_pc && !source.call,
+    Check(!source.decoded.has_swap_pc && source.calls.empty(),
           "fused SETPC handoff entered scalar call preparation");
     auto translated = TranslateProgram(source.decoded, options);
     uint32_t allocations = 0;
@@ -12911,6 +12970,54 @@ void TestLateConstantPropagation() {
         "an ungrounded Phi cycle acquired a constant");
 }
 
+void TestRepeatedBitcastPropagation() {
+  using namespace ShaderRecompiler::IR;
+  Block entry, left, right;
+  IREmitter e(&entry), l(&left), r(&right);
+  const auto raw = e.GetUserData(static_cast<ScalarReg>(0));
+  const auto other = e.GetUserData(static_cast<ScalarReg>(1));
+  const auto fp = e.Emit(ValueOpcode::ConvertF32U32, {raw});
+  for (const auto [opcode, source, distinct] : {
+       std::tuple{ValueOpcode::BitCastF32U32, Value(raw), Value(other)},
+       std::tuple{ValueOpcode::BitCastU32F32, fp,
+                  e.Emit(ValueOpcode::ConvertF32U32, {other})}}) {
+    const auto first = e.Emit(opcode, {source});
+    const auto duplicate = e.Emit(opcode, {source});
+    const auto separate = e.Emit(opcode, {distinct});
+    const auto left_cast = l.Emit(opcode, {source});
+    const auto left_duplicate = l.Emit(opcode, {source});
+    const auto right_cast = r.Emit(opcode, {source});
+    const auto use = opcode == ValueOpcode::BitCastF32U32
+        ? e.Emit(ValueOpcode::FPIsNan32, {duplicate})
+        : e.Emit(ValueOpcode::IEqual32, {duplicate, Value(0u)});
+    ConstantPropagationPass({&entry, &left, &right});
+    Check(duplicate.Resolve() == first && left_duplicate.Resolve() == left_cast &&
+              use.Instruction()->Arg(0).Resolve() == first,
+          "repeated same-block casts were not shared with their earlier definition");
+    Check(separate.Resolve() == separate && left_cast.Resolve() == left_cast &&
+              right_cast.Resolve() == right_cast && left_cast != first && right_cast != left_cast,
+          "cast reuse merged distinct values or crossed block boundaries");
+  }
+  const auto as_float = e.Emit(ValueOpcode::BitCastF32U32, {raw});
+  const auto raw_roundtrip = e.Emit(ValueOpcode::BitCastU32F32, {as_float});
+  const auto as_bits = e.Emit(ValueOpcode::BitCastU32F32, {fp});
+  const auto float_roundtrip = e.Emit(ValueOpcode::BitCastF32U32, {as_bits});
+  ConstantPropagationPass({&entry});
+  Check(raw_roundtrip.Resolve() == raw && float_roundtrip.Resolve() == fp,
+        "sharing casts prevented inverse casts from folding");
+  for (const uint32_t bits : {0u, 0x80000000u, 1u, 0x80000001u, 0x7f800000u,
+                             0xff800000u, 0x7fc12345u, 0x7f812345u}) {
+    const auto constant_float = e.Emit(ValueOpcode::BitCastF32U32, {Value(bits)});
+    const auto constant_bits = e.Emit(ValueOpcode::BitCastU32F32,
+                                      {Value::F32(std::bit_cast<float>(bits))});
+    ConstantPropagationPass({&entry});
+    Check(constant_float.Resolve().IsImmediate() &&
+              std::bit_cast<uint32_t>(constant_float.Resolve().F32Value()) == bits &&
+              constant_bits.Resolve() == Value(bits),
+          "constant bitcast folding changed signed zero, subnormal, infinity or NaN bits");
+  }
+}
+
 void TestMaskedShiftConstantPropagation() {
   using namespace ShaderRecompiler::IR;
   for (const uint32_t count : {0u, 1u, 8u, 31u, 32u, 33u, 40u, 63u, 64u, UINT32_MAX}) {
@@ -15300,6 +15407,58 @@ void TestScalarMemoryLoadCrossesIntoVcc() {
   }
 }
 
+void TestEmptyBranchScalarReadDce() {
+  using namespace ShaderRecompiler::IR;
+  using ShaderRecompiler::CFG::TerminatorKind;
+  for (const bool reverse : {false, true}) {
+    for (uint32_t mode = 0; mode < 4; ++mode) {
+      Block entry, detour, merge;
+      const BlockList blocks{&entry, &detour, &merge};
+      entry.AddBranch(&detour);
+      entry.AddBranch(&merge);
+      detour.AddBranch(&merge);
+      entry.terminator.kind = TerminatorKind::ConditionalBranch;
+      entry.terminator.true_block = reverse ? &merge : &detour;
+      entry.terminator.false_block = reverse ? &detour : &merge;
+      entry.terminator.merge_block = &merge;
+      detour.terminator.kind = TerminatorKind::Branch;
+      detour.terminator.true_block = &merge;
+      IREmitter ir(&entry);
+      const auto address = ir.Emit(ValueOpcode::GetAddressResource,
+                                   {Value(0x1000u), Value(0u)});
+      const auto read = ir.Emit(ValueOpcode::LoadAddressU32,
+                                {address, Value(24u), Value(0u), Value(true)});
+      const auto predicate = ir.INotEqual(U32(read), U32(Value(0u)));
+      entry.condition = ir.Emit(ValueOpcode::ConditionRef, {predicate});
+      ir.Emit(ValueOpcode::Reference, {entry.condition});
+      ir.Emit(ValueOpcode::Barrier);
+      detour.AppendNewInst(ValueOpcode::ControlNop, {});
+      if (mode == 1) detour.AppendNewInst(ValueOpcode::Barrier, {});
+      if (mode == 2) {
+        auto& phi = merge.AppendNewInst(ValueOpcode::Phi, {});
+        phi.SetFlags(Type::U32);
+        phi.AddPhiOperand(&entry, Value(1u));
+        phi.AddPhiOperand(&detour, Value(2u));
+        merge.AppendNewInst(ValueOpcode::ReferenceU32, {Value(&phi)});
+      }
+      if (mode == 3) merge.AppendNewInst(ValueOpcode::ReferenceU32, {read});
+      EliminateDeadCode(blocks);
+      Check(entry.condition.IsImmediate() == (mode == 0 || mode == 3),
+            "empty branch cleanup changed side effects or a live Phi");
+      const auto contains = [&](ValueOpcode op) {
+        return std::ranges::any_of(entry, [=](const Inst& inst) {
+          return inst.GetOpcode() == op;
+        });
+      };
+      Check(contains(ValueOpcode::LoadAddressU32) == (mode != 0),
+            "empty branch cleanup lost a shared read or retained a dead read");
+      Check(contains(ValueOpcode::Barrier) && entry.ImmSuccessors().size() == 2 &&
+                merge.ImmPredecessors().size() == 2,
+            "empty branch cleanup changed earlier side effects or CFG edges");
+    }
+  }
+}
+
 void TestScalarMemoryUnusedTailDce() {
   const uint32_t shader[] = {
       EncodeSmem0(0x0c, 16, 0),
@@ -16523,6 +16682,7 @@ int main() {
   TestNewShaderDecoderArchitecture();
   TestImageAddressOperands();
   TestSopkCompareImmediateExtension();
+  TestNativeDataCacheInvalidation();
   TestNativeStoreCompletion();
   TestNewShaderRecompilerCapturedVopcSdwaCmpxClass();
   TestVopcCmpxClassF16Decoder();
@@ -16616,6 +16776,7 @@ int main() {
   TestValuePhiValidation();
   TestLaneMaskProjection();
   TestLateConstantPropagation();
+  TestRepeatedBitcastPropagation();
   TestMaskedShiftConstantPropagation();
   TestU64ShiftConstantPropagation();
   TestU64RegisterPairRoundTrip();
@@ -16647,6 +16808,7 @@ int main() {
   TestSrtWalkerRealSBufferTranslation();
   TestScalarMemorySourcesCapturedBeforeWrites();
   TestScalarMemoryLoadCrossesIntoVcc();
+  TestEmptyBranchScalarReadDce();
   TestScalarMemoryUnusedTailDce();
   TestResourceTrackingRealDensePatching();
   TestDirectTranslationResetsAnalysisState();