guest_gpu: remove memory-unmap submission deadlock + remove legacy agc buffering

Extracted from 3db2b3c5c5e1a26a861df7ebcacd9ccb8c484420 in KytyPS5/KytyPS5#147.

.github/workflows/build.yml

@@ -83,12 +83,12 @@ jobs:
       - name: Build
         shell: cmd
         run: |
-          cmake --build _Build/windows --target launcher virtual_memory_allocation_tests --parallel
+          cmake --build _Build/windows --target launcher audio_out2_port_tests virtual_memory_allocation_tests --parallel
 
       - name: Test
         shell: cmd
         run: |
-          ctest --test-dir _Build/windows --output-on-failure -R "^virtual_memory_allocation$"
+          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|virtual_memory_allocation)$"
 
       - name: Install
         shell: cmd
@@ -159,14 +159,14 @@ jobs:
         shell: bash
         run: |
           cmake --build _Build/macos \
-            --target launcher virtual_memory_allocation_tests \
+            --target launcher audio_out2_port_tests virtual_memory_allocation_tests \
             --parallel
 
       - name: Test
         shell: bash
         run: |
           ctest --test-dir _Build/macos --output-on-failure \
-            -R '^virtual_memory_allocation$'
+            -R '^(audio_out2_port|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash
@@ -297,14 +297,14 @@ jobs:
         run: |
           cmake --build _Build/linux \
             --target launcher page_manager_tests memory_tracker_tests \
-              virtual_memory_allocation_tests \
+              audio_out2_port_tests virtual_memory_allocation_tests \
             --parallel
 
       - name: Test
         shell: bash
         run: |
           ctest --test-dir _Build/linux --output-on-failure \
-            -R '^(page_manager|memory_tracker|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|page_manager|memory_tracker|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash

README.md

@@ -229,6 +229,17 @@ codesign --force --sign - _Build/macos/install/libMoltenVK.dylib
 
 Release archives already include a signed `libMoltenVK.dylib`.
 
+### Regression tests
+
+Build every regression executable and run the registered tests with:
+
+```powershell
+cmake --build _Build/windows --target kyty_tests
+ctest --test-dir _Build/windows --output-on-failure
+```
+
+Use `_Build/linux` instead of `_Build/windows` for a Linux build.
+
 ### Visual Studio Code
 
 A ready-made Visual Studio Code setup is included in [`.vscode`](.vscode). It configures CMake

src/CMakeLists.txt

@@ -345,6 +345,11 @@ add_executable(page_manager_tests EXCLUDE_FROM_ALL
 )
 target_include_directories(page_manager_tests PRIVATE ${inc_headers})
 
+add_executable(bit_array_tests EXCLUDE_FROM_ALL
+	../tests/BitArrayTests.cpp
+)
+target_include_directories(bit_array_tests PRIVATE ${inc_headers})
+
 add_executable(memory_tracker_tests EXCLUDE_FROM_ALL
 	../tests/MemoryTrackerTests.cpp
 	graphics/host_gpu/pageManager.cpp
@@ -394,6 +399,14 @@ add_executable(resource_mutex_tests EXCLUDE_FROM_ALL
 target_link_libraries(resource_mutex_tests common)
 target_include_directories(resource_mutex_tests PRIVATE ${inc_headers})
 
+add_executable(audio_out2_port_tests EXCLUDE_FROM_ALL
+	../tests/AudioOut2PortTests.cpp
+	libs/libAudio2.cpp
+	loader/timer.cpp
+)
+target_link_libraries(audio_out2_port_tests common fmt::fmt)
+target_include_directories(audio_out2_port_tests PRIVATE ${inc_headers})
+
 add_executable(event_queue_lifetime_tests EXCLUDE_FROM_ALL
 	../tests/EventQueueLifetimeTests.cpp
 	kernel/eventQueue.cpp
@@ -444,12 +457,18 @@ if(NOT KYTY_CLANG_CL)
 endif()
 
 if(BUILD_TESTING)
+	add_test(NAME shader_cfg COMMAND $<TARGET_FILE:shader_cfg_tests>)
 	add_test(NAME scalar_provenance COMMAND $<TARGET_FILE:scalar_provenance_tests>)
 	add_test(NAME image_page_table COMMAND $<TARGET_FILE:image_page_table_tests>)
 	add_test(NAME memory_tracker COMMAND $<TARGET_FILE:memory_tracker_tests>)
 	add_test(NAME page_manager COMMAND $<TARGET_FILE:page_manager_tests>)
+	add_test(NAME bit_array COMMAND $<TARGET_FILE:bit_array_tests>)
+	add_test(NAME shader_vertex_metadata COMMAND $<TARGET_FILE:shader_vertex_metadata_tests>)
+	add_test(NAME shader_stage_runtime COMMAND $<TARGET_FILE:shader_stage_runtime_tests>)
+	add_test(NAME resource_tracking COMMAND $<TARGET_FILE:resource_tracking_tests>)
 	add_test(NAME resource_mutex COMMAND $<TARGET_FILE:resource_mutex_tests>)
 	add_test(NAME event_queue_lifetime COMMAND $<TARGET_FILE:event_queue_lifetime_tests>)
+	add_test(NAME audio_out2_port COMMAND $<TARGET_FILE:audio_out2_port_tests>)
 	add_test(NAME shader_recompiler_compute COMMAND $<TARGET_FILE:shader_recompiler_compute_tests>)
 	add_test(NAME virtual_memory_allocation
 		COMMAND $<TARGET_FILE:virtual_memory_allocation_tests>)
@@ -482,6 +501,22 @@ if(BUILD_TESTING)
 		add_test(NAME buffer_cache_ranges
 			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-range-only)
 	endif()
+
+	add_custom_target(kyty_tests DEPENDS
+		shader_cfg_tests
+		scalar_provenance_tests
+		image_page_table_tests
+		memory_tracker_tests
+		page_manager_tests
+		bit_array_tests
+		shader_vertex_metadata_tests
+		shader_stage_runtime_tests
+		resource_tracking_tests
+		resource_mutex_tests
+		event_queue_lifetime_tests
+		shader_recompiler_compute_tests
+		virtual_memory_allocation_tests
+	)
 endif()
 
 
@@ -515,7 +550,12 @@ set(KYTY_EMULATOR_MAP_LINK_PATH "${CMAKE_CURRENT_BINARY_DIR}/${KYTY_EMULATOR_MAP
 set(KYTY_EMULATOR_PDB_LINK_PATH "${CMAKE_CURRENT_BINARY_DIR}/kyty_emulator.pdb")
 
 if(KYTY_CLANG_CL)
-	set_target_properties(kyty_emulator PROPERTIES LINK_FLAGS "/DYNAMICBASE:NO /DEBUG:FULL /PDB:${KYTY_EMULATOR_PDB_LINK_PATH} /lldmap:${KYTY_EMULATOR_MAP_LINK_PATH}")
+	target_link_options(kyty_emulator PRIVATE
+		"/DYNAMICBASE:NO"
+		"/DEBUG:FULL"
+		"/PDB:${KYTY_EMULATOR_PDB_LINK_PATH}"
+		"/lldmap:${KYTY_EMULATOR_MAP_LINK_PATH}"
+	)
 	add_custom_command(TARGET kyty_emulator POST_BUILD COMMAND ${CMAKE_COMMAND} -E copy_if_different "${KYTY_THIRD_PARTY_DIR}/winpthread/bin/libwinpthread-1.dll" $<TARGET_FILE_DIR:kyty_emulator>/libwinpthread-1.dll)
 elseif(WIN32 OR LINUX)
 	set_target_properties(kyty_emulator PROPERTIES LINK_FLAGS "${KYTY_LD_OPTIONS} -Wl,-Map=${KYTY_EMULATOR_MAP_LINK_PATH}")

src/common/bitArray.h

@@ -0,0 +1,260 @@
+#ifndef EMULATOR_SRC_COMMON_BITARRAY_H_
+#define EMULATOR_SRC_COMMON_BITARRAY_H_
+
+#include <array>
+#include <bit>
+#include <cstddef>
+#include <cstdint>
+#include <iterator>
+#include <utility>
+
+namespace Common {
+
+template <size_t N>
+class BitArray final {
+	static_assert(N != 0, "BitArray size must be nonzero");
+	static_assert(N % 64 == 0, "BitArray size must be a multiple of 64 bits");
+
+	static constexpr size_t BITS_PER_WORD = 64;
+	static constexpr size_t WORD_COUNT    = N / BITS_PER_WORD;
+
+public:
+	using Range = std::pair<size_t, size_t>;
+
+	class Iterator final {
+	public:
+		using iterator_category = std::forward_iterator_tag;
+		using value_type        = Range;
+		using difference_type   = std::ptrdiff_t;
+		using pointer           = const Range*;
+		using reference         = const Range&;
+
+		Iterator(const BitArray& bits, size_t start)
+		    : m_bits(bits), m_range(bits.FirstRangeFrom(start)) {}
+
+		Iterator& operator++() {
+			m_range = m_bits.FirstRangeFrom(m_range.second);
+			return *this;
+		}
+
+		[[nodiscard]] bool operator==(const Iterator& other) const {
+			return &m_bits == &other.m_bits && m_range == other.m_range;
+		}
+
+		[[nodiscard]] bool operator!=(const Iterator& other) const { return !(*this == other); }
+
+		[[nodiscard]] reference operator*() const { return m_range; }
+		[[nodiscard]] pointer   operator->() const { return &m_range; }
+
+	private:
+		const BitArray& m_bits;
+		Range           m_range;
+	};
+
+	using const_iterator = Iterator;
+
+	constexpr BitArray() = default;
+
+	constexpr BitArray(const BitArray& other, size_t start, size_t end) {
+		if (start >= end || end > N) {
+			return;
+		}
+
+		const auto first_word = start / BITS_PER_WORD;
+		const auto last_word  = (end - 1) / BITS_PER_WORD;
+		const auto start_bit  = start % BITS_PER_WORD;
+		const auto end_bit    = (end - 1) % BITS_PER_WORD;
+		const auto start_mask = ~uint64_t {0} << start_bit;
+		const auto end_mask =
+		    end_bit == BITS_PER_WORD - 1 ? ~uint64_t {0} : (uint64_t {1} << (end_bit + 1)) - 1;
+
+		if (first_word == last_word) {
+			m_data[first_word] = other.m_data[first_word] & start_mask & end_mask;
+			return;
+		}
+
+		m_data[first_word] = other.m_data[first_word] & start_mask;
+		for (auto word = first_word + 1; word < last_word; word++) {
+			m_data[word] = other.m_data[word];
+		}
+		m_data[last_word] = other.m_data[last_word] & end_mask;
+	}
+
+	[[nodiscard]] constexpr bool Get(size_t index) const {
+		return (m_data[index / BITS_PER_WORD] & (uint64_t {1} << (index % BITS_PER_WORD))) != 0;
+	}
+
+	constexpr void Set(size_t index) {
+		m_data[index / BITS_PER_WORD] |= uint64_t {1} << (index % BITS_PER_WORD);
+	}
+
+	constexpr void Unset(size_t index) {
+		m_data[index / BITS_PER_WORD] &= ~(uint64_t {1} << (index % BITS_PER_WORD));
+	}
+
+	constexpr void SetRange(size_t start, size_t end) {
+		if (start >= end || end > N) {
+			return;
+		}
+
+		const auto first_word = start / BITS_PER_WORD;
+		const auto last_word  = (end - 1) / BITS_PER_WORD;
+		const auto start_bit  = start % BITS_PER_WORD;
+		const auto end_bit    = (end - 1) % BITS_PER_WORD;
+		const auto start_mask = ~uint64_t {0} << start_bit;
+		const auto end_mask =
+		    end_bit == BITS_PER_WORD - 1 ? ~uint64_t {0} : (uint64_t {1} << (end_bit + 1)) - 1;
+
+		if (first_word == last_word) {
+			m_data[first_word] |= start_mask & end_mask;
+			return;
+		}
+
+		m_data[first_word] |= start_mask;
+		for (auto word = first_word + 1; word < last_word; word++) {
+			m_data[word] = ~uint64_t {0};
+		}
+		m_data[last_word] |= end_mask;
+	}
+
+	constexpr void UnsetRange(size_t start, size_t end) {
+		if (start >= end || end > N) {
+			return;
+		}
+
+		const auto first_word = start / BITS_PER_WORD;
+		const auto last_word  = (end - 1) / BITS_PER_WORD;
+		const auto start_bit  = start % BITS_PER_WORD;
+		const auto end_bit    = (end - 1) % BITS_PER_WORD;
+		const auto start_mask = (uint64_t {1} << start_bit) - 1;
+		const auto end_mask =
+		    end_bit == BITS_PER_WORD - 1 ? uint64_t {0} : ~((uint64_t {1} << (end_bit + 1)) - 1);
+
+		if (first_word == last_word) {
+			m_data[first_word] &= start_mask | end_mask;
+			return;
+		}
+
+		m_data[first_word] &= start_mask;
+		for (auto word = first_word + 1; word < last_word; word++) {
+			m_data[word] = 0;
+		}
+		m_data[last_word] &= end_mask;
+	}
+
+	constexpr void Clear() { m_data.fill(0); }
+	constexpr void Fill() { m_data.fill(~uint64_t {0}); }
+
+	[[nodiscard]] constexpr bool None() const {
+		uint64_t combined = 0;
+		for (const auto word: m_data) {
+			combined |= word;
+		}
+		return combined == 0;
+	}
+
+	[[nodiscard]] constexpr bool Any() const { return !None(); }
+
+	[[nodiscard]] constexpr Range FirstRangeFrom(size_t start) const {
+		if (start >= N) {
+			return {N, N};
+		}
+
+		auto word_index = start / BITS_PER_WORD;
+		auto word       = m_data[word_index] & (~uint64_t {0} << (start % BITS_PER_WORD));
+		while (word == 0) {
+			word_index++;
+			if (word_index == WORD_COUNT) {
+				return {N, N};
+			}
+			word = m_data[word_index];
+		}
+
+		const auto first     = word_index * BITS_PER_WORD + std::countr_zero(word);
+		const auto first_bit = first % BITS_PER_WORD;
+		const auto first_ones =
+		    static_cast<size_t>(std::countr_one(m_data[word_index] >> first_bit));
+		if (first_bit + first_ones < BITS_PER_WORD) {
+			return {first, first + first_ones};
+		}
+
+		for (word_index++; word_index < WORD_COUNT; word_index++) {
+			word = m_data[word_index];
+			if (word != ~uint64_t {0}) {
+				return {first, word_index * BITS_PER_WORD + std::countr_one(word)};
+			}
+		}
+		return {first, N};
+	}
+
+	[[nodiscard]] constexpr Range FirstRange() const { return FirstRangeFrom(0); }
+
+	[[nodiscard]] constexpr Range LastRangeFrom(size_t end) const {
+		if (end == 0) {
+			return {0, 0};
+		}
+		if (end > N) {
+			end = N;
+		}
+
+		auto       word_index = (end - 1) / BITS_PER_WORD;
+		const auto end_bit    = (end - 1) % BITS_PER_WORD;
+		const auto end_mask =
+		    end_bit == BITS_PER_WORD - 1 ? ~uint64_t {0} : (uint64_t {1} << (end_bit + 1)) - 1;
+		auto word = m_data[word_index] & end_mask;
+		while (word == 0) {
+			if (word_index == 0) {
+				return {0, 0};
+			}
+			word = m_data[--word_index];
+		}
+
+		const auto empty_bits = static_cast<size_t>(std::countl_zero(word));
+		const auto ones       = static_cast<size_t>(std::countl_one(word << empty_bits));
+		const auto last       = (word_index + 1) * BITS_PER_WORD - empty_bits;
+		if (empty_bits + ones < BITS_PER_WORD) {
+			return {last - ones, last};
+		}
+
+		while (word_index != 0) {
+			word = m_data[--word_index];
+			if (word != ~uint64_t {0}) {
+				return {(word_index + 1) * BITS_PER_WORD - std::countl_one(word), last};
+			}
+		}
+		return {0, last};
+	}
+
+	[[nodiscard]] constexpr Range LastRange() const { return LastRangeFrom(N); }
+
+	[[nodiscard]] const_iterator begin() const { return Iterator(*this, 0); }
+	[[nodiscard]] const_iterator end() const { return Iterator(*this, N); }
+
+	constexpr BitArray& operator^=(const BitArray& other) {
+		for (size_t word = 0; word < WORD_COUNT; word++) {
+			m_data[word] ^= other.m_data[word];
+		}
+		return *this;
+	}
+
+	[[nodiscard]] constexpr BitArray operator^(const BitArray& other) const {
+		auto result = *this;
+		result ^= other;
+		return result;
+	}
+
+	[[nodiscard]] constexpr BitArray operator~() const {
+		auto result = *this;
+		for (auto& word: result.m_data) {
+			word = ~word;
+		}
+		return result;
+	}
+
+private:
+	std::array<uint64_t, WORD_COUNT> m_data {};
+};
+
+} // namespace Common
+
+#endif // EMULATOR_SRC_COMMON_BITARRAY_H_

src/graphics/guest_gpu/graphicsRun.cpp

@@ -32,11 +32,10 @@
 
 namespace Libs::Graphics {
 
-static thread_local CommandProcessor* g_current_processor      = nullptr;
-static thread_local Pm4Execution*     g_current_execution      = nullptr;
-static thread_local uint32_t          g_submission_pause_depth = 0;
-static thread_local bool              g_gpu_mutex_owned        = false;
-static thread_local bool              g_gpu_thread             = false;
+static thread_local CommandProcessor* g_current_processor = nullptr;
+static thread_local Pm4Execution*     g_current_execution = nullptr;
+static thread_local bool              g_gpu_mutex_owned   = false;
+static thread_local bool              g_gpu_thread        = false;
 
 class GpuMutexLock final {
 public:
@@ -98,8 +97,6 @@ public:
 	                   bool trigger_agc_interrupt_on_done);
 	void SubmitFlipPreparation(uint64_t request_id);
 	void Done();
-	void PauseSubmissions();
-	void ResumeSubmissions();
 	void Shutdown();
 	[[nodiscard]] bool IsStopping();
 	void               SendCommand(Common::UniqueFunction<void>&& command);
@@ -407,7 +404,7 @@ void CommandProcessor::WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw
                                  uint32_t write_control) {
 	const uint32_t dst_sel      = ((write_control >> 30u) & 0x1u) | ((write_control >> 7u) & 0x1eu);
 	const uint32_t cache_policy = (write_control >> 25u) & 0x3u;
-	const uint32_t increment     = (write_control >> 16u) & 0x1u;
+	const uint32_t increment    = (write_control >> 16u) & 0x1u;
 	const uint32_t write_confirm = (write_control >> 20u) & 0x1u;
 
 	switch (dst_sel) {
@@ -696,26 +693,6 @@ bool GpuState::Process(Submission& submission) {
 	return complete;
 }
 
-void GpuState::PauseSubmissions() {
-	if (g_gpu_mutex_owned) {
-		EXIT("GPU submissions are already paused by this thread\n");
-	}
-	g_gpu_mutex_owned = true;
-	m_submission_mutex.Lock();
-	if (!IsGpuThread()) {
-		WaitLocked();
-	}
-	m_renderer.GetCommandScheduler().DrainPriorityOperations();
-}
-
-void GpuState::ResumeSubmissions() {
-	if (!g_gpu_mutex_owned) {
-		EXIT("GPU submissions resumed without an active pause\n");
-	}
-	m_submission_mutex.Unlock();
-	g_gpu_mutex_owned = false;
-}
-
 Pm4ProcessResult CommandProcessor::Process(Pm4Execution& execution, uint32_t* buffer,
                                            uint32_t size_dw) {
 	KYTY_PROFILER_BLOCK("CommandProcessor::Process");
@@ -1693,32 +1670,6 @@ int Gpu::GetFrameNum() const {
 	return m_state->GetFrameNum();
 }
 
-void Gpu::PauseSubmissions() {
-	m_state->PauseSubmissions();
-}
-
-void Gpu::ResumeSubmissions() {
-	m_state->ResumeSubmissions();
-}
-
-Gpu::SubmissionLock::SubmissionLock(Gpu& gpu): m_gpu(gpu) {
-	if (g_current_processor != nullptr || g_submission_pause_depth == UINT32_MAX) {
-		EXIT("cannot acquire GPU submission lock in the current state\n");
-	}
-	if (g_submission_pause_depth++ == 0) {
-		m_gpu.PauseSubmissions();
-	}
-}
-
-Gpu::SubmissionLock::~SubmissionLock() {
-	if (g_submission_pause_depth == 0) {
-		EXIT("GPU submission lock released without ownership\n");
-	}
-	if (--g_submission_pause_depth == 0) {
-		m_gpu.ResumeSubmissions();
-	}
-}
-
 bool Gpu::IsCommandProcessorThread() noexcept {
 	return g_current_processor != nullptr;
 }
@@ -1727,12 +1678,4 @@ CommandProcessor* Gpu::CurrentCommandProcessor() noexcept {
 	return g_current_processor;
 }
 
-bool Gpu::SubmissionLockHeld() noexcept {
-	return g_submission_pause_depth != 0;
-}
-
-bool Gpu::MutexHeld() noexcept {
-	return g_gpu_mutex_owned;
-}
-
 } // namespace Libs::Graphics

src/graphics/guest_gpu/graphicsRun.h

@@ -35,25 +35,8 @@ public:
 
 	[[nodiscard]] static bool              IsCommandProcessorThread() noexcept;
 	[[nodiscard]] static CommandProcessor* CurrentCommandProcessor() noexcept;
-	[[nodiscard]] static bool              SubmissionLockHeld() noexcept;
-	[[nodiscard]] static bool              MutexHeld() noexcept;
-
-	class SubmissionLock final {
-	public:
-		explicit SubmissionLock(Gpu& gpu);
-		~SubmissionLock();
-		KYTY_CLASS_NO_COPY(SubmissionLock);
-
-	private:
-		Gpu& m_gpu;
-	};
 
 private:
-	friend class SubmissionLock;
-
-	void PauseSubmissions();
-	void ResumeSubmissions();
-
 	std::unique_ptr<GpuState> m_state;
 };
 } // namespace Libs::Graphics

src/graphics/host_gpu/memoryTracker.cpp

@@ -6,13 +6,7 @@ namespace Libs::Graphics {
 
 static_assert(std::atomic<void*>::is_always_lock_free);
 
-MemoryTracker::MemoryTracker(PageManager& page_manager, PageWatchMode gpu_watch_mode)
-    : m_page_manager(page_manager), m_gpu_watch_mode(gpu_watch_mode) {
-	switch (m_gpu_watch_mode) {
-		case PageWatchMode::Write:
-		case PageWatchMode::ReadWrite: break;
-		default: EXIT("unsupported memory tracker GPU page-watch mode\n");
-	}
+MemoryTracker::MemoryTracker(PageManager& page_manager): m_page_manager(page_manager) {
 	m_regions = std::make_unique<std::atomic<RegionManager*>[]>(REGION_COUNT);
 	for (size_t i = 0; i < REGION_COUNT; i++) {
 		m_regions[i].store(nullptr, std::memory_order_relaxed);
@@ -21,6 +15,7 @@ MemoryTracker::MemoryTracker(PageManager& page_manager, PageWatchMode gpu_watch_
 
 MemoryTracker::~MemoryTracker() = default;
 
+#if KYTY_BUILD == KYTY_BUILD_DEBUG
 void MemoryTracker::ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
                                           const char* operation) const noexcept {
 	if (vaddr == 0 || size == 0 || size > UINT64_MAX - vaddr ||
@@ -58,6 +53,7 @@ void MemoryTracker::ValidateGpuDirtyOwnership(const RangeSet& dirty, uint64_t va
 		}
 	}
 }
+#endif
 
 void MemoryTracker::ValidateRange(uint64_t vaddr, uint64_t size) {
 	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
@@ -104,34 +100,25 @@ void MemoryTracker::MarkRegionAsCpuModified(uint64_t vaddr, uint64_t size) {
 	std::lock_guard access(m_access_mutex);
 	Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		std::scoped_lock lock(manager->lock);
-		const auto       changed =
-		    manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
-		manager->ApplyProtection(changed, false);
+		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
 }
 
 void MemoryTracker::MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	std::lock_guard access(m_access_mutex);
-	Iterate<true>(vaddr, size, [this](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+	Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		std::scoped_lock lock(manager->lock);
-		const auto       changed =
-		    manager->ChangeState<DirtySource::Gpu, true>(manager->GetCpuAddr() + offset, bytes);
-		manager->ApplyGpuProtection(changed, true, m_gpu_watch_mode);
+		manager->ChangeState<DirtySource::Gpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
 }
 
 void MemoryTracker::UnmarkRegionAsGpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	std::lock_guard access(m_access_mutex);
-	Iterate<true>(vaddr, size, [this](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+	Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		std::scoped_lock lock(manager->lock);
-		if (!manager->IsFullyModified<DirtySource::Gpu>(offset, bytes)) {
-			EXIT("cannot clear partially GPU-dirty tracking range\n");
-		}
-		const auto changed =
-		    manager->ChangeState<DirtySource::Gpu, false>(manager->GetCpuAddr() + offset, bytes);
-		manager->ApplyGpuProtection(changed, false, m_gpu_watch_mode);
+		manager->ChangeState<DirtySource::Gpu, false>(manager->GetCpuAddr() + offset, bytes);
 	});
 }
 
@@ -154,9 +141,7 @@ void MemoryTracker::UntrackMemoryLocked(uint64_t vaddr, uint64_t size) {
 		EXIT("cannot untrack GPU-dirty memory\n");
 	}
 	Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-		const auto changed =
-		    manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
-		manager->ApplyProtection(changed, false);
+		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
 	locks.clear();
 }

src/graphics/host_gpu/memoryTracker.h

@@ -18,8 +18,7 @@ namespace Libs::Graphics {
 
 class MemoryTracker final {
 public:
-	explicit MemoryTracker(PageManager&  page_manager,
-	                       PageWatchMode gpu_watch_mode = PageWatchMode::ReadWrite);
+	explicit MemoryTracker(PageManager& page_manager);
 	~MemoryTracker();
 
 	KYTY_CLASS_NO_COPY(MemoryTracker);
@@ -56,9 +55,8 @@ public:
 			}
 			Iterate<false>(vaddr, size,
 			               [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-				               const auto changed = manager->ChangeState<DirtySource::Cpu, true>(
+				               manager->ChangeState<DirtySource::Cpu, true>(
 				                   manager->GetCpuAddr() + offset, bytes);
-				               manager->ApplyProtection(changed, false);
 			               });
 			return false;
 		};
@@ -71,10 +69,15 @@ public:
 			EXIT("memory invalidation retained GPU-owned pages\n");
 		}
 	}
+#if KYTY_BUILD == KYTY_BUILD_DEBUG
 	void ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
 	                           const char* operation) const noexcept;
 	void ValidateGpuDirtyOwnership(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
 	                               const char* operation);
+#else
+	void ValidateGpuDirtyPages(const RangeSet&, uint64_t, uint64_t, const char*) const noexcept {}
+	void ValidateGpuDirtyOwnership(const RangeSet&, uint64_t, uint64_t, const char*) {}
+#endif
 
 	template <bool clear, typename Preflight, typename Func>
 	void ForEachDownloadRange(uint64_t vaddr, uint64_t size, Preflight&& preflight, Func&& func) {
@@ -104,10 +107,8 @@ public:
 			Iterate<false>(vaddr, size,
 			               [&](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 				               const auto address = manager->GetCpuAddr() + offset;
-				               const auto changed =
-				                   manager->template ForEachModifiedRange<DirtySource::Gpu, true>(
-				                       address, bytes, [](uint64_t, uint64_t) noexcept {});
-				               manager->ApplyGpuProtection(changed, false, m_gpu_watch_mode);
+				               manager->template ForEachModifiedRange<DirtySource::Gpu, true>(
+				                   address, bytes, [](uint64_t, uint64_t) noexcept {});
 			               });
 		}
 	}
@@ -137,13 +138,12 @@ public:
 		});
 		upload_func();
 		if (is_written) {
-			Iterate<false>(
-			    vaddr, size, [this](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-				    const auto changed = manager->template ChangeState<DirtySource::Gpu, true>(
-				        manager->GetCpuAddr() + offset, bytes);
-				    manager->ApplyGpuProtection(changed, true, m_gpu_watch_mode);
-				    manager->lock.unlock();
-			    });
+			Iterate<false>(vaddr, size,
+			               [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+				               manager->template ChangeState<DirtySource::Gpu, true>(
+				                   manager->GetCpuAddr() + offset, bytes);
+				               manager->lock.unlock();
+			               });
 		}
 		s_upload_owner = previous_upload_owner;
 	}
@@ -197,7 +197,6 @@ private:
 	std::mutex                                     m_region_mutex;
 	std::mutex                                     m_access_mutex;
 	PageManager&                                   m_page_manager;
-	PageWatchMode                                  m_gpu_watch_mode = PageWatchMode::ReadWrite;
 };
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/pageManager.cpp

@@ -9,7 +9,6 @@
 #include <cstdarg>
 #include <cstdio>
 #include <cstdlib>
-#include <limits>
 #include <memory>
 #include <mutex>
 #include <vector>
@@ -22,13 +21,9 @@
 #undef min
 #undef max
 #elif defined(__APPLE__)
-#include <pthread.h>
-#include <sys/mman.h>
 #include <unistd.h>
 #else
 #include <execinfo.h>
-#include <sys/mman.h>
-#include <sys/syscall.h>
 #include <unistd.h>
 #endif
 
@@ -41,9 +36,8 @@ constexpr uint64_t ADDRESS_SIZE = TRACKER_ADDRESS_SIZE;
 constexpr uint64_t REGION_COUNT = ADDRESS_SIZE / REGION_SIZE;
 
 #if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
-// The tracker reuses Win32 memory-protection tags as internal page-state values (on
-// Windows they come from <windows.h> and are what VirtualQuery returns). Mirror the
-// canonical Win32 numeric values so the shared state-machine logic is identical.
+// The tracker reuses Win32 memory-protection tags as internal page-state values.
+// Mirror their canonical numeric values so the shared state-machine logic is identical.
 constexpr uint32_t PAGE_NOACCESS  = 0x01;
 constexpr uint32_t PAGE_READONLY  = 0x02;
 constexpr uint32_t PAGE_READWRITE = 0x04;
@@ -53,8 +47,6 @@ constexpr uint64_t REGION_PAGES = REGION_SIZE / PAGE_SIZE;
 constexpr uint32_t NO_ACCESS_PROTECTION  = PAGE_NOACCESS;
 constexpr uint32_t READ_ONLY_PROTECTION  = PAGE_READONLY;
 constexpr uint32_t READ_WRITE_PROTECTION = PAGE_READWRITE;
-// Zero is the unknown protection sentinel.
-constexpr uint32_t UNKNOWN_PROTECTION = 0;
 
 [[noreturn]] void FailFast(const char* reason = nullptr) noexcept {
 	std::fputs("PageManager fail-fast: ", stderr);
@@ -102,25 +94,6 @@ Common::VirtualMemory::Mode ToMemoryMode(uint32_t protection) {
 	}
 }
 
-uint32_t CurrentThread() noexcept {
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-	return GetCurrentThreadId();
-#elif defined(__APPLE__)
-	return static_cast<uint32_t>(pthread_mach_thread_np(pthread_self()));
-#elif defined(__linux__)
-	static thread_local const uint32_t tid = [] {
-		const auto raw = static_cast<uint32_t>(::syscall(SYS_gettid));
-		if (raw == 0) {
-			FailFast("gettid returned the reserved zero owner token");
-		}
-		return raw;
-	}();
-	return tid;
-#else
-	FailFast("page tracking thread identity is unsupported on this platform");
-#endif
-}
-
 class SpinGuard final {
 public:
 	explicit SpinGuard(std::atomic_flag& lock): m_lock(lock) {
@@ -154,38 +127,58 @@ uint64_t PageEnd(uint64_t vaddr, uint64_t size) {
 
 struct PageManager::Impl {
 	struct PageState {
-		std::atomic_flag lock                = ATOMIC_FLAG_INIT;
-		uint32_t         write_watchers      = 0;
-		uint32_t         access_watchers     = 0;
-		uint32_t         original_protection = 0;
-		uint32_t         backing_writer      = 0;
-		// Shadow the protection applied through Protect().
-		uint32_t current_protection = UNKNOWN_PROTECTION;
-		bool     resolving          = false;
-	};
+		uint8_t write_watchers  : 7 = 0;
+		uint8_t access_watchers : 1 = 0;
 
-	struct Region {
-		std::array<PageState, REGION_PAGES> pages;
-	};
+		[[nodiscard]] uint32_t Perms() const noexcept {
+			if (access_watchers != 0) {
+				return NO_ACCESS_PROTECTION;
+			}
+			if (write_watchers != 0) {
+				return READ_ONLY_PROTECTION;
+			}
+			return READ_WRITE_PROTECTION;
+		}
 
-	class PageRangeGuard final {
-	public:
-		explicit PageRangeGuard(std::span<PageState*> pages): m_pages(pages) {
-			for (auto* page: m_pages) {
-				while (page->lock.test_and_set(std::memory_order_acquire)) {
-					std::atomic_signal_fence(std::memory_order_seq_cst);
+		template <int delta, bool is_read>
+		uint32_t AddDelta(uint64_t address) {
+			static_assert(delta >= -1 && delta <= 1);
+			if constexpr (is_read) {
+				if constexpr (delta == 1) {
+					if (access_watchers != 0) {
+						Fatal("read-watcher overflow at 0x%016" PRIx64, address);
+					}
+					return ++access_watchers;
+				} else if constexpr (delta == -1) {
+					if (access_watchers == 0) {
+						Fatal("read-watcher underflow at 0x%016" PRIx64, address);
+					}
+					return --access_watchers;
+				} else {
+					return access_watchers;
+				}
+			} else {
+				if constexpr (delta == 1) {
+					if (write_watchers == 0x7f) {
+						Fatal("write-watcher overflow at 0x%016" PRIx64, address);
+					}
+					return ++write_watchers;
+				} else if constexpr (delta == -1) {
+					if (write_watchers == 0) {
+						Fatal("write-watcher underflow at 0x%016" PRIx64, address);
+					}
+					return --write_watchers;
+				} else {
+					return write_watchers;
 				}
 			}
 		}
-		~PageRangeGuard() {
-			for (auto it = m_pages.rbegin(); it != m_pages.rend(); ++it) {
-				(*it)->lock.clear(std::memory_order_release);
-			}
-		}
-		KYTY_CLASS_NO_COPY(PageRangeGuard);
+	};
+	static_assert(sizeof(PageState) == 1);
 
-	private:
-		std::span<PageState*> m_pages;
+	struct Region {
+		std::atomic_flag                    lock = ATOMIC_FLAG_INIT;
+		std::array<PageState, REGION_PAGES> pages;
 	};
 
 	Impl() {
@@ -215,10 +208,9 @@ struct PageManager::Impl {
 
 	~Impl() {
 		for (const auto& region: region_storage) {
+			SpinGuard lock(region->lock);
 			for (auto& page: region->pages) {
-				SpinGuard lock(page.lock);
-				if (page.write_watchers != 0 || page.access_watchers != 0 ||
-				    page.backing_writer != 0 || page.resolving) {
+				if (page.write_watchers != 0 || page.access_watchers != 0) {
 					FailFast("PageManager destroyed with live page state");
 				}
 			}
@@ -246,56 +238,81 @@ struct PageManager::Impl {
 		return ptr;
 	}
 
-	PageState& GetPage(Region& region, uint64_t vaddr) const {
-		return region.pages[(vaddr % REGION_SIZE) / PAGE_SIZE];
-	}
-
-	static uint32_t WatcherProtection(const PageState& page) {
-		if (page.access_watchers != 0) {
-			return NO_ACCESS_PROTECTION;
-		}
-		if (page.write_watchers != 0) {
-			return READ_ONLY_PROTECTION;
-		}
-		return page.original_protection;
-	}
-
-	static void InitializeProtection(std::span<PageState*> pages) {
-		for (auto* page: pages) {
-			page->original_protection = READ_WRITE_PROTECTION;
-			page->current_protection  = READ_WRITE_PROTECTION;
-		}
-	}
-
-	void ProtectRange(std::span<PageState*> pages, uint64_t vaddr, uint32_t protection,
-	                  std::span<const uint32_t> expected_old) noexcept {
-		const auto size = pages.size() * PAGE_SIZE;
-		if (pages.size() != expected_old.size()) {
-			FailFast("protection range state size mismatch");
-		}
-		for (size_t i = 0; i < pages.size(); i++) {
-			const auto actual = pages[i]->current_protection;
-			if (actual != UNKNOWN_PROTECTION && actual != expected_old[i]) {
-				Fatal("invalid protection transition at 0x%016" PRIx64 ", old=0x%08" PRIx32
-				      ", expected=0x%08" PRIx32 ", new=0x%08" PRIx32,
-				      vaddr + i * PAGE_SIZE, actual, expected_old[i], protection);
-			}
-		}
+	void Protect(uint64_t vaddr, uint64_t size, uint32_t protection) noexcept {
 		if (!Libs::LibKernel::Memory::ProtectGuestHostMemory(vaddr, size,
 		                                                     ToMemoryMode(protection))) {
 			Fatal("address-space protection failed at 0x%016" PRIx64 ", new=0x%08" PRIx32, vaddr,
 			      protection);
 		}
-		for (auto* page: pages) {
-			page->current_protection = protection;
-		}
 	}
 
-	void Protect(PageState& page, uint64_t vaddr, uint32_t protection,
-	             uint32_t expected_old) noexcept {
-		PageState* pages[]    = {&page};
-		uint32_t   expected[] = {expected_old};
-		ProtectRange(pages, vaddr, protection, expected);
+	template <bool track, bool is_read, bool masked>
+	void UpdateRegionWatchers(Region& region, uint64_t base_addr, size_t first, size_t last,
+	                          const RegionBits* mask = nullptr) {
+		SpinGuard lock(region.lock);
+		auto      perms                 = region.pages[first].Perms();
+		uint64_t  range_begin           = 0;
+		uint64_t  range_bytes           = 0;
+		uint64_t  potential_range_bytes = 0;
+
+		const auto release_pending = [&] {
+			if (range_bytes != 0) {
+				Protect(base_addr + range_begin * PAGE_SIZE, range_bytes, perms);
+				range_bytes           = 0;
+				potential_range_bytes = 0;
+			}
+		};
+
+		for (size_t page_index = first; page_index < last; page_index++) {
+			auto&      page    = region.pages[page_index];
+			const auto address = base_addr + page_index * PAGE_SIZE;
+			const bool update  = !masked || mask->Get(page_index);
+
+			const auto old_perms = page.Perms();
+			const auto new_count = update ? page.AddDelta<track ? 1 : -1, is_read>(address)
+			                              : page.AddDelta<0, is_read>(address);
+			const auto new_perms = page.Perms();
+
+			if (new_perms != perms) [[unlikely]] {
+				release_pending();
+				perms = new_perms;
+			} else if (range_bytes != 0) {
+				potential_range_bytes += PAGE_SIZE;
+			}
+
+			if (!update) {
+				continue;
+			}
+
+			const bool watcher_edge = (track && new_count == 1) || (!track && new_count == 0);
+			if (watcher_edge && old_perms != new_perms) {
+				if (range_bytes == 0) {
+					range_begin           = page_index;
+					potential_range_bytes = PAGE_SIZE;
+				}
+				range_bytes = potential_range_bytes;
+			}
+		}
+
+		release_pending();
+	}
+
+	template <bool track, bool is_read>
+	void UpdatePageWatchers(uint64_t vaddr, uint64_t size) {
+		const auto begin = PageStart(vaddr);
+		const auto end   = PageEnd(vaddr, size);
+		for (auto chunk_begin = begin; chunk_begin < end;) {
+			const auto chunk_end   = std::min(end, (chunk_begin / REGION_SIZE + 1) * REGION_SIZE);
+			const auto region_base = chunk_begin / REGION_SIZE * REGION_SIZE;
+			auto*      region = track ? GetOrCreateRegion(chunk_begin) : FindRegion(chunk_begin);
+			if (region == nullptr) {
+				Fatal("untracking unknown page 0x%016" PRIx64, chunk_begin);
+			}
+			const auto first = static_cast<size_t>((chunk_begin - region_base) / PAGE_SIZE);
+			const auto last  = static_cast<size_t>((chunk_end - region_base) / PAGE_SIZE);
+			UpdateRegionWatchers<track, is_read, false>(*region, region_base, first, last);
+			chunk_begin = chunk_end;
+		}
 	}
 
 	std::unique_ptr<std::atomic<Region*>[]> regions;
@@ -313,212 +330,48 @@ uint64_t PageManager::GetPageSize() const {
 	return PAGE_SIZE;
 }
 
-void PageManager::UpdatePageWatchers(bool track, uint64_t vaddr, uint64_t size,
-                                     PageWatchMode mode) {
-	if (mode != PageWatchMode::Write && mode != PageWatchMode::ReadWrite) {
-		Fatal("invalid watcher mode");
-	}
-	const auto begin = PageStart(vaddr);
-	const auto end   = PageEnd(vaddr, size);
-	for (auto chunk_begin = begin; chunk_begin < end;) {
-		const auto chunk_end = std::min(end, (chunk_begin / REGION_SIZE + 1) * REGION_SIZE);
-		auto*      region =
-		    track ? m_impl->GetOrCreateRegion(chunk_begin) : m_impl->FindRegion(chunk_begin);
-		if (region == nullptr) {
-			Fatal("untracking unknown page 0x%016" PRIx64, chunk_begin);
-		}
-
-		const auto page_count = static_cast<size_t>((chunk_end - chunk_begin) / PAGE_SIZE);
-		std::vector<Impl::PageState*> pages;
-		pages.reserve(page_count);
-		for (auto address = chunk_begin; address < chunk_end; address += PAGE_SIZE) {
-			pages.push_back(&m_impl->GetPage(*region, address));
-		}
-		Impl::PageRangeGuard lock(pages);
-
-		std::vector<uint8_t> first_watchers(page_count);
-		for (size_t i = 0; i < page_count; i++) {
-			auto&      page    = *pages[i];
-			const auto address = chunk_begin + i * PAGE_SIZE;
-			if (page.resolving && track) {
-				FailFast("new page watcher raced active fault resolution");
-			}
-			auto& watchers =
-			    (mode == PageWatchMode::ReadWrite ? page.access_watchers : page.write_watchers);
-			if (track) {
-				if (watchers == std::numeric_limits<uint32_t>::max()) {
-					Fatal("watcher overflow at 0x%016" PRIx64, address);
-				}
-				first_watchers[i] = page.write_watchers == 0 && page.access_watchers == 0;
-			} else {
-				if (watchers == 0) {
-					Fatal("watcher underflow at 0x%016" PRIx64, address);
-				}
-				if (page.backing_writer != 0 && page.backing_writer != CurrentThread()) {
-					Fatal("backing write ownership changed at 0x%016" PRIx64, address);
-				}
-			}
-		}
-
-		if (track) {
-			for (size_t first = 0; first < page_count;) {
-				while (first < page_count && first_watchers[first] == 0) {
-					first++;
-				}
-				auto last = first;
-				while (last < page_count && first_watchers[last] != 0) {
-					last++;
-				}
-				if (first != last) {
-					Impl::InitializeProtection(std::span {pages}.subspan(first, last - first));
-				}
-				first = last;
-			}
-		}
-
-		std::vector<uint32_t> old_protections(page_count);
-		std::vector<uint32_t> new_protections(page_count);
-		std::vector<uint8_t>  transitions(page_count);
-		for (size_t i = 0; i < page_count; i++) {
-			auto& page = *pages[i];
-			auto& watchers =
-			    (mode == PageWatchMode::ReadWrite ? page.access_watchers : page.write_watchers);
-			const auto old_protection = Impl::WatcherProtection(page);
-			if (track) {
-				watchers++;
-			} else {
-				watchers--;
-			}
-			const auto new_protection = Impl::WatcherProtection(page);
-			old_protections[i]        = old_protection;
-			new_protections[i]        = new_protection;
-			if (new_protection != old_protection && (track || page.backing_writer == 0)) {
-				transitions[i] = 1;
-			}
-		}
-
-		for (size_t first = 0; first < page_count;) {
-			while (first < page_count && transitions[first] == 0) {
-				first++;
-			}
-			if (first == page_count) {
-				break;
-			}
-			const auto protection = new_protections[first];
-			auto       current    = first + 1;
-			auto       last       = current;
-			for (; current < page_count && new_protections[current] == protection; current++) {
-				if (old_protections[current] != new_protections[current] &&
-				    transitions[current] == 0) {
-					break;
-				}
-				if (transitions[current] != 0) {
-					last = current + 1;
-				}
-			}
-			m_impl->ProtectRange(std::span {pages}.subspan(first, last - first),
-			                     chunk_begin + first * PAGE_SIZE, protection,
-			                     std::span {old_protections}.subspan(first, last - first));
-			first = current;
-		}
-
-		for (auto* page: pages) {
-			if (!track && page->backing_writer == 0 && page->write_watchers == 0 &&
-			    page->access_watchers == 0) {
-				page->original_protection = 0;
-			}
-		}
-		chunk_begin = chunk_end;
-	}
+template <bool track>
+void PageManager::UpdatePageWatchers(uint64_t vaddr, uint64_t size) {
+	m_impl->UpdatePageWatchers<track, false>(vaddr, size);
 }
 
+template void PageManager::UpdatePageWatchers<true>(uint64_t, uint64_t);
+template void PageManager::UpdatePageWatchers<false>(uint64_t, uint64_t);
+
+template <bool track, bool is_read>
+void PageManager::UpdatePageWatchersForRegion(uint64_t base_addr, RegionBits& mask) {
+	if (base_addr % REGION_SIZE != 0 || base_addr >= ADDRESS_SIZE ||
+	    REGION_SIZE > ADDRESS_SIZE - base_addr) {
+		Fatal("invalid tracking region base 0x%016" PRIx64, base_addr);
+	}
+
+	const auto start_range = mask.FirstRange();
+	const auto end_range   = mask.LastRange();
+	if (start_range.first == REGION_PAGES) {
+		FailFast("empty region watcher mask");
+	}
+	const auto first = start_range.first;
+	const auto last  = end_range.second;
+	if (start_range.second == end_range.second) {
+		m_impl->UpdatePageWatchers<track, is_read>(base_addr + first * PAGE_SIZE,
+		                                           (last - first) * PAGE_SIZE);
+		return;
+	}
+
+	auto* region = track ? m_impl->GetOrCreateRegion(base_addr) : m_impl->FindRegion(base_addr);
+	if (region == nullptr) {
+		Fatal("untracking unknown region 0x%016" PRIx64, base_addr);
+	}
+	m_impl->UpdateRegionWatchers<track, is_read, true>(*region, base_addr, first, last, &mask);
+}
+
+template void PageManager::UpdatePageWatchersForRegion<true, true>(uint64_t, RegionBits&);
+template void PageManager::UpdatePageWatchersForRegion<true, false>(uint64_t, RegionBits&);
+template void PageManager::UpdatePageWatchersForRegion<false, true>(uint64_t, RegionBits&);
+template void PageManager::UpdatePageWatchersForRegion<false, false>(uint64_t, RegionBits&);
+
 void PageManager::OnGpuMap(uint64_t, uint64_t) {}
 
 void PageManager::OnGpuUnmap(uint64_t, uint64_t) {}
 
-PageManager::BackingWrite::BackingWrite(PageManager& manager, uint64_t vaddr,
-                                        uint64_t size) noexcept
-    : m_manager(manager), m_vaddr(vaddr), m_size(size) {
-	m_manager.BeginBackingWrite(vaddr, size);
-}
-
-PageManager::BackingWrite::~BackingWrite() {
-	m_manager.EndBackingWrite(m_vaddr, m_size);
-}
-
-std::vector<std::unique_ptr<PageManager::BackingWrite>>
-PageManager::ReserveBackingWrites(std::span<const RangeSet::Range> ranges) {
-	if (ranges.empty()) {
-		Fatal("cannot reserve empty backing-write ranges");
-	}
-	std::vector<std::unique_ptr<BackingWrite>> writes;
-	writes.reserve(ranges.size());
-	uint64_t begin = 0;
-	uint64_t end   = 0;
-	for (const auto& range: ranges) {
-		if (range.address == 0 || range.size == 0 || range.size > UINT64_MAX - range.address ||
-		    range.address + range.size > UINT64_MAX - (PAGE_SIZE - 1)) {
-			Fatal("invalid backing-write range");
-		}
-		const auto page_begin = PageStart(range.address);
-		const auto page_end   = PageStart(range.address + range.size + PAGE_SIZE - 1);
-		if (begin != 0 && page_begin > end) {
-			writes.push_back(std::make_unique<BackingWrite>(*this, begin, end - begin));
-			begin = 0;
-		}
-		if (begin == 0) {
-			begin = page_begin;
-			end   = page_end;
-		} else {
-			end = std::max(end, page_end);
-		}
-	}
-	writes.push_back(std::make_unique<BackingWrite>(*this, begin, end - begin));
-	return writes;
-}
-
-void PageManager::BeginBackingWrite(uint64_t vaddr, uint64_t size) noexcept {
-	const auto end    = PageEnd(vaddr, size);
-	const auto writer = CurrentThread();
-	for (auto address = PageStart(vaddr); address < end; address += PAGE_SIZE) {
-		auto* region = m_impl->FindRegion(address);
-		if (region == nullptr) {
-			Fatal("backing write reserves an unknown page at 0x%016" PRIx64, address);
-		}
-		auto&     page = m_impl->GetPage(*region, address);
-		SpinGuard lock(page.lock);
-		if (page.resolving || page.backing_writer != 0 || page.access_watchers == 0) {
-			Fatal("backing write races page resolution at 0x%016" PRIx64, address);
-		}
-		page.resolving      = true;
-		page.backing_writer = writer;
-	}
-}
-
-void PageManager::EndBackingWrite(uint64_t vaddr, uint64_t size) noexcept {
-	const auto end    = PageEnd(vaddr, size);
-	const auto writer = CurrentThread();
-	for (auto address = PageStart(vaddr); address < end; address += PAGE_SIZE) {
-		auto* region = m_impl->FindRegion(address);
-		if (region == nullptr) {
-			FailFast("backing write ended for an unknown page");
-		}
-		auto&     page = m_impl->GetPage(*region, address);
-		SpinGuard lock(page.lock);
-		if (!page.resolving || page.backing_writer != writer) {
-			FailFast("backing write ended without matching owner and resolving state");
-		}
-		const auto old_protection = NO_ACCESS_PROTECTION;
-		const auto new_protection = Impl::WatcherProtection(page);
-		if (new_protection != old_protection) {
-			m_impl->Protect(page, address, new_protection, old_protection);
-		}
-		if (page.write_watchers == 0 && page.access_watchers == 0) {
-			page.original_protection = 0;
-		}
-		page.backing_writer = 0;
-		page.resolving      = false;
-	}
-}
-
 } // namespace Libs::Graphics

src/graphics/host_gpu/pageManager.h

@@ -2,31 +2,16 @@
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_PAGEMANAGER_H_
 
 #include "common/common.h"
-#include "graphics/host_gpu/rangeSet.h"
+#include "graphics/host_gpu/regionDefinitions.h"
 
 #include <memory>
-#include <span>
-#include <vector>
 
 namespace Libs::Graphics {
 
 enum class PageFaultAccess { Read, Write, Execute, Unknown };
-enum class PageWatchMode { Write, ReadWrite };
 
 class PageManager final {
 public:
-	class BackingWrite final {
-	public:
-		BackingWrite(PageManager& manager, uint64_t vaddr, uint64_t size) noexcept;
-		~BackingWrite();
-		KYTY_CLASS_NO_COPY(BackingWrite);
-
-	private:
-		PageManager& m_manager;
-		uint64_t     m_vaddr = 0;
-		uint64_t     m_size  = 0;
-	};
-
 	PageManager();
 	// The owner must stop all PageManager callers before destruction.
 	~PageManager();
@@ -35,18 +20,14 @@ public:
 
 	[[nodiscard]] uint64_t GetPageSize() const;
 
-	void UpdatePageWatchers(bool track, uint64_t vaddr, uint64_t size,
-	                        PageWatchMode mode = PageWatchMode::Write);
+	template <bool track>
+	void UpdatePageWatchers(uint64_t vaddr, uint64_t size);
+	template <bool track, bool is_read = false>
+	void UpdatePageWatchersForRegion(uint64_t base_addr, RegionBits& mask);
 	void OnGpuMap(uint64_t vaddr, uint64_t size);
 	void OnGpuUnmap(uint64_t vaddr, uint64_t size);
 
-	[[nodiscard]] std::vector<std::unique_ptr<BackingWrite>>
-	ReserveBackingWrites(std::span<const RangeSet::Range> ranges);
-
 private:
-	void BeginBackingWrite(uint64_t vaddr, uint64_t size) noexcept;
-	void EndBackingWrite(uint64_t vaddr, uint64_t size) noexcept;
-
 	struct Impl;
 	std::unique_ptr<Impl> m_impl;
 };

src/graphics/host_gpu/regionDefinitions.h

@@ -1,10 +1,9 @@
 #ifndef EMULATOR_SRC_GRAPHICS_HOST_GPU_REGIONDEFINITIONS_H_
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_REGIONDEFINITIONS_H_
 
+#include "common/bitArray.h"
 #include "common/common.h"
 
-#include <bitset>
-
 namespace Libs::Graphics {
 
 constexpr uint64_t TRACKER_PAGE_SIZE    = 4ull * 1024ull;
@@ -13,7 +12,8 @@ constexpr uint64_t TRACKER_ADDRESS_SIZE = 1ull << 40u;
 constexpr size_t   TRACKER_REGION_PAGES = TRACKER_REGION_SIZE / TRACKER_PAGE_SIZE;
 
 enum class DirtySource { Cpu, Gpu };
-using RegionBits = std::bitset<TRACKER_REGION_PAGES>;
+using RegionBits = Common::BitArray<TRACKER_REGION_PAGES>;
+static_assert(sizeof(RegionBits) == TRACKER_REGION_PAGES / 8);
 
 } // namespace Libs::Graphics
 

src/graphics/host_gpu/regionManager.h

@@ -76,8 +76,9 @@ public:
 		if (m_cpu_addr % TRACKER_REGION_SIZE != 0) {
 			EXIT("invalid region tracking manager construction\n");
 		}
-		m_cpu_dirty.set();
-		m_writable.set();
+		m_cpu_dirty.Fill();
+		m_writable.Fill();
+		m_readable.Fill();
 	}
 
 	KYTY_CLASS_NO_COPY(RegionManager);
@@ -87,106 +88,81 @@ public:
 	[[nodiscard]] bool IsModified(uint64_t offset, uint64_t size) const {
 		const auto [start, end] = GetPageRange(m_cpu_addr + offset, size);
 		const auto& bits        = GetBits<source>();
-		for (auto page = start; page < end; page++) {
-			if (bits.test(page)) {
-				return true;
-			}
-		}
-		return false;
-	}
-
-	template <DirtySource source>
-	[[nodiscard]] bool IsFullyModified(uint64_t offset, uint64_t size) const {
-		const auto [start, end] = GetPageRange(m_cpu_addr + offset, size);
-		const auto& bits        = GetBits<source>();
-		for (auto page = start; page < end; page++) {
-			if (!bits.test(page)) {
-				return false;
-			}
-		}
-		return true;
+		return RegionBits(bits, start, end).Any();
 	}
 
 	template <DirtySource source, bool enable>
-	RegionBits ChangeState(uint64_t vaddr, uint64_t size) {
+	void ChangeState(uint64_t vaddr, uint64_t size) {
 		const auto [start, end] = GetPageRange(vaddr, size);
 		if constexpr (source == DirtySource::Cpu && enable) {
-			for (auto page = start; page < end; page++) {
-				if (m_gpu_dirty.test(page)) {
-					EXIT("CPU dirty state conflicts with GPU dirty state\n");
-				}
+			if (RegionBits(m_gpu_dirty, start, end).Any()) {
+				EXIT("CPU dirty state conflicts with GPU dirty state\n");
 			}
 		}
 		if constexpr (source == DirtySource::Gpu && enable) {
-			for (auto page = start; page < end; page++) {
-				if (m_cpu_dirty.test(page)) {
-					EXIT("GPU dirty state conflicts with CPU dirty state\n");
-				}
+			if (RegionBits(m_cpu_dirty, start, end).Any()) {
+				EXIT("GPU dirty state conflicts with CPU dirty state\n");
 			}
 		}
-		auto& bits    = GetBits<source>();
-		auto  changed = bits;
-		for (auto page = start; page < end; page++) {
-			bits.set(page, enable);
+		auto& bits = GetBits<source>();
+		if constexpr (enable) {
+			bits.SetRange(start, end);
+		} else {
+			bits.UnsetRange(start, end);
 		}
-		changed ^= bits;
 		if constexpr (source == DirtySource::Cpu) {
-			changed    = m_cpu_dirty ^ m_writable;
-			m_writable = m_cpu_dirty;
+			UpdateCpuProtection<!enable>();
+		} else {
+			UpdateGpuProtection<enable>();
 		}
-		return changed;
 	}
 
 	template <DirtySource source, bool clear, typename Func>
-	RegionBits ForEachModifiedRange(uint64_t vaddr, uint64_t size, Func&& func) {
+	void ForEachModifiedRange(uint64_t vaddr, uint64_t size, Func&& func) {
 		const auto [start, end] = GetPageRange(vaddr, size);
-		auto mask               = GetBits<source>();
-		for (auto page = 0u; page < start; page++) {
-			mask.reset(page);
-		}
-		for (auto page = end; page < TRACKER_REGION_PAGES; page++) {
-			mask.reset(page);
-		}
+		RegionBits mask(GetBits<source>(), start, end);
 		if constexpr (clear) {
-			auto& bits = GetBits<source>();
-			for (auto page = start; page < end; page++) {
-				if (mask.test(page)) {
-					bits.reset(page);
-				}
-			}
+			GetBits<source>().UnsetRange(start, end);
 		}
 		if constexpr (source == DirtySource::Cpu && clear) {
-			auto changed = m_cpu_dirty ^ m_writable;
-			m_writable   = m_cpu_dirty;
-			ApplyProtection(changed, true);
+			UpdateCpuProtection<true>();
 			ForEachRange(mask, std::forward<Func>(func));
-			return changed;
+			return;
+		}
+		if constexpr (source == DirtySource::Gpu && clear) {
+			UpdateGpuProtection<false>();
 		}
 		ForEachRange(mask, std::forward<Func>(func));
-		if constexpr (clear) {
-			return mask;
-		}
-		return {};
-	}
-
-	void ApplyProtection(const RegionBits& changed, bool track) {
-		ForEachRange(changed, [this, track](uint64_t vaddr, uint64_t size) {
-			m_page_manager.UpdatePageWatchers(track, vaddr, size);
-		});
-	}
-
-	void ApplyGpuProtection(const RegionBits& changed, bool track, PageWatchMode mode) {
-		if (mode != PageWatchMode::Write && mode != PageWatchMode::ReadWrite) {
-			EXIT("unsupported GPU page-watch mode\n");
-		}
-		ForEachRange(changed, [this, track, mode](uint64_t vaddr, uint64_t size) {
-			m_page_manager.UpdatePageWatchers(track, vaddr, size, mode);
-		});
 	}
 
 	TrackingSpinLock lock;
 
 private:
+	template <bool track>
+	void UpdateCpuProtection() {
+		auto mask  = m_cpu_dirty ^ m_writable;
+		m_writable = m_cpu_dirty;
+		if (mask.None()) {
+			return;
+		}
+		m_page_manager.UpdatePageWatchersForRegion<track>(m_cpu_addr, mask);
+	}
+
+	template <bool track>
+	void UpdateGpuProtection() {
+		auto readable = ~m_gpu_dirty;
+		auto mask     = readable ^ m_readable;
+		m_readable    = readable;
+		if (mask.None()) {
+			return;
+		}
+		if constexpr (track) {
+			m_page_manager.UpdatePageWatchersForRegion<true, true>(m_cpu_addr, mask);
+		} else {
+			m_page_manager.UpdatePageWatchersForRegion<false, true>(m_cpu_addr, mask);
+		}
+	}
+
 	template <DirtySource source>
 	RegionBits& GetBits() {
 		if constexpr (source == DirtySource::Cpu) {
@@ -217,18 +193,8 @@ private:
 
 	template <typename Func>
 	void ForEachRange(const RegionBits& bits, Func&& func) const {
-		size_t page = 0;
-		while (page < TRACKER_REGION_PAGES) {
-			while (page < TRACKER_REGION_PAGES && !bits.test(page)) {
-				page++;
-			}
-			const auto start = page;
-			while (page < TRACKER_REGION_PAGES && bits.test(page)) {
-				page++;
-			}
-			if (start != page) {
-				func(m_cpu_addr + start * TRACKER_PAGE_SIZE, (page - start) * TRACKER_PAGE_SIZE);
-			}
+		for (const auto [start, end]: bits) {
+			func(m_cpu_addr + start * TRACKER_PAGE_SIZE, (end - start) * TRACKER_PAGE_SIZE);
 		}
 	}
 
@@ -237,6 +203,7 @@ private:
 	RegionBits   m_cpu_dirty;
 	RegionBits   m_gpu_dirty;
 	RegionBits   m_writable;
+	RegionBits   m_readable;
 };
 
 } // namespace Libs::Graphics

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

@@ -123,12 +123,6 @@ struct BufferCache::RetiredBuffer {
 	std::shared_ptr<Buffer> owner;
 };
 
-struct BufferCache::PendingBackingPublication {
-	uint64_t address = 0;
-	uint64_t size    = 0;
-	uint64_t tick    = 0;
-};
-
 std::pair<uint64_t, uint64_t> BufferCache::DownloadEnvelope(const DownloadCopy& copy) {
 	if (copy.owner == nullptr || copy.size == 0 || copy.source_offset > copy.owner->Size() ||
 	    copy.size > copy.owner->Size() - copy.source_offset) {
@@ -200,34 +194,26 @@ void BufferCache::QueueGarbageDownload(std::span<const DownloadCopy> copies, Ret
 	if (copies.empty()) {
 		return;
 	}
-	auto       downloads = RecordDownloads(copies);
-	const auto tick      = m_scheduler.CurrentTick();
-	BeginBackingPublication(retire.address, retire.size, tick);
-	m_scheduler.DeferOperation([this, downloads = std::move(downloads), retire = std::move(retire),
-	                            tick]() mutable {
-		PublishDownloads(downloads);
-		{
-			FaultSafeCacheLock lock(this, m_mutex);
-			if (m_memory_tracker.IsRegionGpuModified(retire.address, retire.size)) {
-				m_memory_tracker.ForEachDownloadRange<true>(
-				    retire.address, retire.size,
-				    [&](uint64_t address, uint64_t size) noexcept {
-					    m_memory_tracker.ValidateGpuDirtyPages(m_gpu_modified_ranges, address, size,
-					                                           "asynchronous garbage retirement");
-				    },
-				    [](uint64_t, uint64_t) noexcept {});
-			}
-			for (const auto& range: downloads) {
-				m_gpu_modified_ranges.Subtract(range.address, range.size);
-			}
-			if (m_memory_tracker.IsRegionGpuModified(retire.address, retire.size) ||
-			    !m_gpu_modified_ranges.Intersections(retire.address, retire.size).empty()) {
-				EXIT("BufferCache: asynchronous garbage collection retained GPU ownership\n");
-			}
-			m_memory_tracker.UntrackMemory(retire.address, retire.size);
-		}
-		CompleteBackingPublication(retire.address, retire.size, tick);
-	});
+	auto downloads = RecordDownloads(copies);
+	m_scheduler.DeferOperation(
+	    [this, downloads = std::move(downloads), retire = std::move(retire)]() mutable {
+		    PublishDownloads(downloads);
+		    {
+			    FaultSafeCacheLock lock(this, m_mutex);
+			    for (const auto& range: downloads) {
+				    m_gpu_modified_ranges.Subtract(range.address, range.size);
+			    }
+			    // ForEachDownloadRange reports full tracker pages, and every exact GPU-owned
+			    // interval on those pages was downloaded and removed. Clearing the original
+			    // query therefore cannot orphan a dirty sibling on an edge page.
+			    m_memory_tracker.UnmarkRegionAsGpuModified(retire.address, retire.size);
+			    if (m_memory_tracker.IsRegionGpuModified(retire.address, retire.size) ||
+			        !m_gpu_modified_ranges.Intersections(retire.address, retire.size).empty()) {
+				    EXIT("BufferCache: asynchronous garbage collection retained GPU ownership\n");
+			    }
+			    m_memory_tracker.UntrackMemory(retire.address, retire.size);
+		    }
+	    });
 }
 
 BufferCache::BufferCache(GraphicContext& graphics, CommandScheduler& scheduler,
@@ -240,8 +226,7 @@ BufferCache::BufferCache(GraphicContext& graphics, CommandScheduler& scheduler,
       m_stream_buffer(graphics, scheduler, MemoryUsage::Stream, 64 * MiB),
       m_download_buffer(graphics, scheduler, MemoryUsage::Download, 32 * MiB),
       m_device_buffer(graphics, scheduler, MemoryUsage::DeviceLocal, 128 * MiB),
-      m_page_manager(page_manager), m_texture_cache(texture_cache),
-      m_resource_mutex(resource_mutex) {
+      m_texture_cache(texture_cache), m_resource_mutex(resource_mutex) {
 	std::memset(m_gds_buffer.Mapped().data(), 0, static_cast<size_t>(m_gds_buffer.Size()));
 	m_gds_buffer.Flush(0, m_gds_buffer.Size());
 	if (!m_graphics.CanReportMemoryUsage()) {
@@ -262,9 +247,6 @@ BufferCache::~BufferCache() {
 	if (!m_gpu_modified_ranges.Empty()) {
 		EXIT("BufferCache: destroyed with pending GPU-modified ranges\n");
 	}
-	if (!m_pending_backing_publications.empty()) {
-		EXIT("BufferCache: destroyed with pending backing publications\n");
-	}
 	for (const auto& [vaddr, cached]: m_buffers) {
 		(void)vaddr;
 		if (m_memory_tracker.IsRegionGpuModified(cached->vaddr, cached->size)) {
@@ -274,68 +256,6 @@ BufferCache::~BufferCache() {
 	m_buffers.clear();
 }
 
-bool BufferCache::SynchronizeBacking(uint64_t vaddr, uint64_t size) {
-	bool waited = false;
-	for (;;) {
-		uint64_t   tick       = 0;
-		const auto page_begin = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-		const auto page_end   = (vaddr + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
-		CacheRange affected {.address = page_begin, .size = page_end - page_begin};
-		{
-			FaultSafeCacheLock lock(this, m_mutex);
-			bool               changed = true;
-			while (changed) {
-				changed = false;
-				for (const auto& [address, cached]: m_buffers) {
-					const CacheRange previous = affected;
-					if (ResolveOverlap(affected, {address, cached->size}) &&
-					    (previous.address != affected.address || previous.size != affected.size)) {
-						changed = true;
-					}
-				}
-			}
-		}
-		{
-			std::lock_guard lock(m_publication_mutex);
-			for (const auto& publication: m_pending_backing_publications) {
-				if (publication.address < affected.address + affected.size &&
-				    affected.address < publication.address + publication.size) {
-					tick = std::max(tick, publication.tick);
-				}
-			}
-		}
-		if (tick == 0) {
-			return waited;
-		}
-		waited = true;
-		m_scheduler.Wait(tick);
-		m_scheduler.WaitPriorityOperations(tick);
-	}
-}
-
-void BufferCache::RefreshInvalidatedRanges(CommandBuffer& command, CachedBuffer& cached,
-                                           uint64_t vaddr, uint64_t size, bool upload) {
-	const auto invalidated = m_image_invalidated_ranges.Intersections(vaddr, size);
-	if (upload) {
-		std::array<uint8_t, 64 * 1024> bytes;
-		for (const auto& range: invalidated) {
-			for (uint64_t copied = 0; copied < range.size;) {
-				const auto chunk = std::min<uint64_t>(range.size - copied, bytes.size());
-				if (!Libs::LibKernel::Memory::TryReadBacking(range.address + copied, bytes.data(),
-				                                             chunk)) {
-					EXIT("BufferCache: failed to refresh an invalidated image alias\n");
-				}
-				Upload(command, *cached.buffer, cached.buffer->Offset(range.address + copied),
-				       bytes.data(), chunk);
-				copied += chunk;
-			}
-		}
-	}
-	if (!invalidated.empty()) {
-		m_image_invalidated_ranges.Subtract(vaddr, size);
-	}
-}
-
 StreamBuffer& BufferCache::GetUtilityBuffer(MemoryUsage usage) noexcept {
 	switch (usage) {
 		case MemoryUsage::Upload: return m_staging_buffer;
@@ -364,7 +284,6 @@ void BufferCache::InvalidateMemory(uint64_t vaddr, uint64_t size) {
 	    size > TRACKER_ADDRESS_SIZE - vaddr) {
 		EXIT("BufferCache: invalid memory-invalidation range\n");
 	}
-	(void)SynchronizeBacking(vaddr, size);
 	if (!HasPageOverlap(vaddr, size)) {
 		return;
 	}
@@ -373,7 +292,6 @@ void BufferCache::InvalidateMemory(uint64_t vaddr, uint64_t size) {
 }
 
 void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size) {
-	(void)SynchronizeBacking(vaddr, size);
 	std::vector<DownloadCopy> copies;
 	{
 		FaultSafeCacheLock lock(this, m_mutex);
@@ -413,16 +331,11 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size) {
 	PublishDownloads(downloads);
 	{
 		FaultSafeCacheLock lock(this, m_mutex);
-		m_memory_tracker.ForEachDownloadRange<true>(
-		    vaddr, size,
-		    [&](uint64_t address, uint64_t bytes) noexcept {
-			    m_memory_tracker.ValidateGpuDirtyPages(m_gpu_modified_ranges, address, bytes,
-			                                           "memory invalidation completion");
-		    },
-		    [](uint64_t, uint64_t) noexcept {});
 		for (const auto& range: downloads) {
 			m_gpu_modified_ranges.Subtract(range.address, range.size);
 		}
+		// The enumeration above covered whole dirty pages and every exact interval on them.
+		m_memory_tracker.UnmarkRegionAsGpuModified(vaddr, size);
 	}
 }
 
@@ -430,13 +343,9 @@ void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 	if (vaddr == 0 || size == 0 || size > UINT64_MAX - vaddr) {
 		EXIT("BufferCache: invalid unmap range\n");
 	}
-	(void)SynchronizeBacking(vaddr, size);
-
-	std::vector<DownloadCopy>                               copies;
-	std::vector<RangeSet::Range>                            dirty_ranges;
-	std::vector<std::pair<uint64_t, uint64_t>>              modified_buffers;
-	std::vector<std::unique_ptr<PageManager::BackingWrite>> backing_writes;
-	std::vector<std::pair<uint64_t, uint64_t>>              retired_buffers;
+	std::vector<DownloadCopy>                  copies;
+	std::vector<std::pair<uint64_t, uint64_t>> modified_buffers;
+	std::vector<std::pair<uint64_t, uint64_t>> retired_buffers;
 	{
 		FaultSafeCacheLock lock(this, m_mutex);
 		for (const auto& [begin, cached]: m_buffers) {
@@ -453,12 +362,8 @@ void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 			if (dirty.empty()) {
 				EXIT("BufferCache: GPU-modified buffer has no dirty ranges\n");
 			}
-			dirty_ranges.insert(dirty_ranges.end(), dirty.begin(), dirty.end());
 			modified_buffers.emplace_back(begin, cached->size);
 		}
-		if (!dirty_ranges.empty()) {
-			backing_writes = m_page_manager.ReserveBackingWrites(dirty_ranges);
-		}
 		for (const auto& [begin, bytes]: modified_buffers) {
 			auto owner = m_buffers.find(begin);
 			if (owner == m_buffers.end() || owner->second->size != bytes) {
@@ -488,24 +393,11 @@ void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 		// command stream before removing such backing.
 		m_scheduler.FinishCurrent();
 	}
-	backing_writes.clear();
-
 	{
 		FaultSafeCacheLock lock(this, m_mutex);
-		for (const auto& [begin, bytes]: modified_buffers) {
-			if (!m_memory_tracker.IsRegionGpuModified(begin, bytes)) {
-				continue;
-			}
-			m_memory_tracker.ForEachDownloadRange<true>(
-			    begin, bytes,
-			    [&](uint64_t address, uint64_t download_size) noexcept {
-				    m_memory_tracker.ValidateGpuDirtyPages(m_gpu_modified_ranges, address,
-				                                           download_size, "unmap retirement");
-			    },
-			    [](uint64_t, uint64_t) noexcept {});
-		}
 		for (const auto& [begin, bytes]: modified_buffers) {
 			m_gpu_modified_ranges.Subtract(begin, bytes);
+			m_memory_tracker.UnmarkRegionAsGpuModified(begin, bytes);
 		}
 		for (const auto& [begin, bytes]: retired_buffers) {
 			m_memory_tracker.MarkRegionAsCpuModified(begin, bytes);
@@ -513,7 +405,6 @@ void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 		if (!m_gpu_modified_ranges.Intersections(vaddr, size).empty()) {
 			EXIT("BufferCache: unmap retained dirty byte ranges\n");
 		}
-		m_image_invalidated_ranges.Subtract(vaddr, size);
 		m_memory_tracker.UntrackMemory(vaddr, size);
 		for (auto it = m_buffers.begin(); it != m_buffers.end();) {
 			if (vaddr < it->first + it->second->size && it->first < vaddr + size) {
@@ -607,20 +498,29 @@ BufferBinding BufferCache::ObtainBuffer(CommandBuffer& command, uint64_t vaddr,
 		EXIT("BufferCache: buffer request requires a recording command buffer\n");
 	}
 	std::lock_guard transaction(m_resource_mutex);
-	(void)SynchronizeBacking(vaddr, size);
 
 	if (is_read && !is_written && size <= CACHING_PAGE_SIZE &&
 	    !m_memory_tracker.IsRegionGpuModified(vaddr, size) &&
 	    m_memory_tracker.IsRegionCpuModified(vaddr, size)) {
-		std::vector<uint8_t> data(size);
-		if (Libs::LibKernel::Memory::TryReadBacking(vaddr, data.data(), size)) {
-			return UploadTransient(data.data(), size, 16);
+		const auto alignment = std::max<uint64_t>(
+		    m_graphics.physical_device_properties.limits.minUniformBufferOffsetAlignment, 1);
+		if (auto [mapped, offset] = m_stream_buffer.Map(size, alignment, false);
+		    mapped != nullptr) {
+			if (Libs::LibKernel::Memory::TryReadBacking(vaddr, mapped, size)) {
+				m_stream_buffer.Commit();
+				return {{}, m_stream_buffer.Handle(), offset};
+			}
+		} else {
+			auto owner = std::make_shared<Buffer>(m_graphics, m_scheduler, MemoryUsage::Upload, 0,
+			                                      AllFlags, size);
+			if (Libs::LibKernel::Memory::TryReadBacking(vaddr, owner->Mapped().data(), size)) {
+				owner->Flush(0, size);
+				return {owner, owner->Handle(), 0};
+			}
 		}
 	}
 
-	if (is_formatted && is_read && !is_written) {
-		(void)m_texture_cache.SynchronizeImageToBuffer(vaddr, size);
-	} else if (is_formatted && is_written) {
+	if (is_formatted && is_written) {
 		(void)m_texture_cache.InvalidateMemoryFromGPU(vaddr, size, true);
 	}
 
@@ -636,10 +536,12 @@ BufferBinding BufferCache::ObtainBuffer(CommandBuffer& command, uint64_t vaddr,
 			           reinterpret_cast<const void*>(address), bytes);
 		    }
 	    });
-	RefreshInvalidatedRanges(command, cached, vaddr, size, is_read);
 	if (is_written) {
 		m_gpu_modified_ranges.Add(vaddr, size);
 	}
+	if (is_formatted && is_read && !is_written) {
+		(void)SynchronizeBufferFromImage(*cached.buffer, vaddr, size);
+	}
 	return {cached.buffer, cached.buffer->Handle(), cached.buffer->Offset(vaddr)};
 }
 
@@ -664,7 +566,6 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 	    size > TRACKER_ADDRESS_SIZE - vaddr) {
 		EXIT("BufferCache: invalid image source\n");
 	}
-	(void)SynchronizeBacking(vaddr, size);
 	auto find_owner = [&]() {
 		auto owner = m_buffers.upper_bound(vaddr);
 		if (owner == m_buffers.begin()) {
@@ -679,13 +580,12 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 		const bool         cpu_modified = m_memory_tracker.IsRegionCpuModified(vaddr, size);
 		const bool         gpu_modified = m_memory_tracker.IsRegionGpuModified(vaddr, size);
 		const auto         dirty        = m_gpu_modified_ranges.Intersections(vaddr, size);
-		const bool invalidated = !m_image_invalidated_ranges.Intersections(vaddr, size).empty();
-		const bool requested_gpu_owned = !dirty.empty();
+		const bool         has_dirty_buffer_source = !dirty.empty();
 		m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
 		                                           "image source");
 
 		auto owner = find_owner();
-		if (requested_gpu_owned && owner == m_buffers.end()) {
+		if (has_dirty_buffer_source && owner == m_buffers.end()) {
 			CacheRange merged {.address = AlignDown(vaddr),
 			                   .size    = AlignUp(vaddr + size) - AlignDown(vaddr)};
 			using Iterator = decltype(m_buffers.begin());
@@ -735,42 +635,32 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 				EXIT("BufferCache: merged image source does not contain the requested range\n");
 			}
 		}
-		if (owner != m_buffers.end() && !cpu_modified && !invalidated &&
-		    (!gpu_modified || requested_gpu_owned)) {
-			DiscardGpuDirtyBytesLocked(vaddr, size, "image source transfer");
+		if (owner != m_buffers.end() && !cpu_modified &&
+		    (!gpu_modified || has_dirty_buffer_source)) {
 			owner->second->tick_accessed_last = m_gc_tick;
-			return {owner->second->buffer.get(), owner->second->buffer->Offset(vaddr),
-			        requested_gpu_owned};
+			return {owner->second->buffer.get(), owner->second->buffer->Offset(vaddr)};
 		}
-		if (requested_gpu_owned && owner == m_buffers.end()) {
+		if (has_dirty_buffer_source && owner == m_buffers.end()) {
 			EXIT("BufferCache: GPU-dirty image source could not resolve its native owner\n");
 		}
 	}
 
-	// Direct-memory backing remains readable while PageManager protects the guest mapping. The
-	// fallback exists for plain host mappings used by standalone renderer tests and is deliberately
-	// performed outside the cache lock so a page fault cannot recurse into BufferCache.
-	const auto stage_address = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-	const auto stage_end     = (vaddr + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
-	const auto stage_size    = stage_end - stage_address;
-	(void)SynchronizeBacking(stage_address, stage_size);
-	std::vector<uint8_t> bytes(stage_size);
-	if (!Libs::LibKernel::Memory::TryReadBacking(stage_address, bytes.data(), stage_size)) {
+	auto [staging, stage_offset] = m_staging_buffer.Map(size, 16);
+	if (staging == nullptr || !Libs::LibKernel::Memory::TryReadBacking(vaddr, staging, size)) {
 		EXIT("BufferCache: failed to read mapped guest image backing\n");
 	}
+	m_staging_buffer.Commit();
 
 	FaultSafeCacheLock lock(this, m_mutex);
-	const auto         dirty       = m_gpu_modified_ranges.Intersections(vaddr, size);
-	const bool         invalidated = !m_image_invalidated_ranges.Intersections(vaddr, size).empty();
-	const bool         requested_gpu_owned = !dirty.empty();
-	auto               owner               = find_owner();
-	if (requested_gpu_owned && owner == m_buffers.end()) {
+	const auto         dirty                   = m_gpu_modified_ranges.Intersections(vaddr, size);
+	const bool         has_dirty_buffer_source = !dirty.empty();
+	auto               owner                   = find_owner();
+	if (has_dirty_buffer_source && owner == m_buffers.end()) {
 		EXIT("BufferCache: GPU-dirty image source lost its native owner\n");
 	}
-	const auto stage_offset = m_staging_buffer.Copy(bytes.data(), stage_size, 16);
-	if (owner == m_buffers.end() || invalidated ||
-	    (m_memory_tracker.IsRegionGpuModified(vaddr, size) && !requested_gpu_owned)) {
-		return {&m_staging_buffer, stage_offset + vaddr - stage_address, false};
+	if (owner == m_buffers.end() ||
+	    (m_memory_tracker.IsRegionGpuModified(vaddr, size) && !has_dirty_buffer_source)) {
+		return {&m_staging_buffer, stage_offset};
 	}
 
 	auto& cached              = *owner->second;
@@ -784,42 +674,17 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 	    [&]() noexcept {
 		    for (const auto& [address, upload_size]: uploads) {
 			    cached.buffer->CopyFrom(
-			        m_scheduler.Current(), m_staging_buffer, stage_offset + address - stage_address,
+			        m_scheduler.Current(), m_staging_buffer, stage_offset + address - vaddr,
 			        cached.buffer->Offset(address), upload_size, vk::AccessFlagBits::eHostWrite);
 		    }
 	    });
-	DiscardGpuDirtyBytesLocked(vaddr, size, "staged image source transfer");
-	return {cached.buffer.get(), cached.buffer->Offset(vaddr), requested_gpu_owned};
-}
-
-void BufferCache::DiscardGpuDirtyBytesLocked(uint64_t vaddr, uint64_t size, const char* operation) {
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size, operation);
-	m_gpu_modified_ranges.Subtract(vaddr, size);
-	const auto page_begin = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-	const auto page_end   = (vaddr + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
-	for (auto page = page_begin; page < page_end; page += TRACKER_PAGE_SIZE) {
-		if (m_gpu_modified_ranges.Intersections(page, TRACKER_PAGE_SIZE).empty() &&
-		    m_memory_tracker.IsRegionGpuModified(page, TRACKER_PAGE_SIZE)) {
-			m_memory_tracker.UnmarkRegionAsGpuModified(page, TRACKER_PAGE_SIZE);
-		}
-	}
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size, operation);
-}
-
-void BufferCache::DiscardGpuDirtyBytes(uint64_t vaddr, uint64_t size) {
-	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
-	    size > TRACKER_ADDRESS_SIZE - vaddr) {
-		EXIT("BufferCache: invalid dirty-byte discard range\n");
-	}
-	FaultSafeCacheLock lock(this, m_mutex);
-	DiscardGpuDirtyBytesLocked(vaddr, size, "image output supersession");
+	return {cached.buffer.get(), cached.buffer->Offset(vaddr)};
 }
 
 void BufferCache::WriteHostMemory(uint64_t vaddr, std::span<const uint8_t> data) {
 	if (vaddr == 0 || data.empty() || data.size() > UINT64_MAX - vaddr) {
 		EXIT("BufferCache: invalid host DMA write\n");
 	}
-	(void)SynchronizeBacking(vaddr, data.size());
 	Libs::LibKernel::Memory::WriteBacking(vaddr, data.data(), data.size());
 
 	FaultSafeCacheLock lock(this, m_mutex);
@@ -835,45 +700,6 @@ void BufferCache::WriteHostMemory(uint64_t vaddr, std::span<const uint8_t> data)
 		       data.data() + begin - vaddr, range_end - begin);
 		cached->tick_accessed_last = m_gc_tick;
 	}
-	m_image_invalidated_ranges.Subtract(vaddr, data.size());
-}
-
-std::pair<std::shared_ptr<Buffer>, uint64_t> BufferCache::ObtainBufferForImageWrite(uint64_t vaddr,
-                                                                                    uint64_t size) {
-	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
-	    size > TRACKER_ADDRESS_SIZE - vaddr) {
-		EXIT("BufferCache: invalid image destination\n");
-	}
-	const auto stage_address = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-	const auto stage_end     = (vaddr + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
-	const auto stage_size    = stage_end - stage_address;
-	(void)SynchronizeBacking(stage_address, stage_size);
-	std::vector<uint8_t> bytes(stage_size);
-	if (!Libs::LibKernel::Memory::TryReadBacking(stage_address, bytes.data(), stage_size)) {
-		EXIT("BufferCache: failed to preserve guest bytes around an image mirror\n");
-	}
-	FaultSafeCacheLock lock(this, m_mutex);
-	auto&              cached = GetOrCreateBuffer(m_scheduler.Current(), vaddr, size);
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
-	                                           "image destination");
-	if (!m_gpu_modified_ranges.Intersections(vaddr, size).empty()) {
-		EXIT("BufferCache: image destination aliases GPU-owned buffer bytes\n");
-	}
-	const auto stage_offset = m_staging_buffer.Copy(bytes.data(), stage_size, 16);
-	std::vector<std::pair<uint64_t, uint64_t>> uploads;
-	m_memory_tracker.ForEachUploadRange(
-	    vaddr, size, false,
-	    [&](uint64_t address, uint64_t upload_size) noexcept {
-		    uploads.emplace_back(address, upload_size);
-	    },
-	    [&]() noexcept {
-		    for (const auto& [address, upload_size]: uploads) {
-			    cached.buffer->CopyFrom(
-			        m_scheduler.Current(), m_staging_buffer, stage_offset + address - stage_address,
-			        cached.buffer->Offset(address), upload_size, vk::AccessFlagBits::eHostWrite);
-		    }
-	    });
-	return {cached.buffer, cached.buffer->Offset(vaddr)};
 }
 
 void BufferCache::FillBuffer(uint64_t vaddr, uint64_t size, uint32_t value, bool is_gds) {
@@ -933,17 +759,10 @@ void BufferCache::CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t si
 	}
 	if (src_memory || dst_memory) {
 		std::lock_guard transaction(m_resource_mutex);
-		if (src_memory) {
-			(void)SynchronizeBacking(src_vaddr, size);
-		}
-		const auto src_region =
+		const auto      src_region =
 		    src_memory ? m_texture_cache.QueryRegion(src_vaddr, size) : TextureCache::RegionInfo {};
 		const auto dst_region =
 		    dst_memory ? m_texture_cache.QueryRegion(dst_vaddr, size) : TextureCache::RegionInfo {};
-		if (src_memory && src_region.gpu_image_bytes &&
-		    !m_texture_cache.SynchronizeImageToBuffer(src_vaddr, size)) {
-			EXIT("BufferCache: GPU copy source image could not be synchronized\n");
-		}
 		if (src_memory && dst_memory && !HasGpuDirtyBytes(src_vaddr, size) &&
 		    !HasGpuDirtyBytes(dst_vaddr, size) && !src_region.gpu_image_bytes &&
 		    !dst_region.gpu_image_bytes) {
@@ -965,7 +784,7 @@ void BufferCache::CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t si
 	}
 
 	auto& command = m_scheduler.Current();
-	auto  src = src_memory ? ObtainBuffer(command, src_vaddr, size, false, true)
+	auto  src = src_memory ? ObtainBuffer(command, src_vaddr, size, false, true, true)
 	                       : BufferBinding {.buffer = m_gds_buffer.Handle(), .offset = src_vaddr};
 	auto  dst = dst_memory ? ObtainBuffer(command, dst_vaddr, size, true, false, true)
 	                       : BufferBinding {.buffer = m_gds_buffer.Handle(), .offset = dst_vaddr};
@@ -1018,75 +837,6 @@ bool BufferCache::IsRegionCpuModified(uint64_t vaddr, uint64_t size) {
 	return m_memory_tracker.IsRegionCpuModified(vaddr, size);
 }
 
-void BufferCache::InvalidateImageAliases(uint64_t vaddr, uint64_t size) {
-	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
-	    size > TRACKER_ADDRESS_SIZE - vaddr) {
-		EXIT("BufferCache: invalid image-alias invalidation\n");
-	}
-	FaultSafeCacheLock lock(this, m_mutex);
-	const auto         end = vaddr + size;
-	for (const auto& [address, cached]: m_buffers) {
-		const auto cached_end = address + cached->size;
-		const auto begin      = std::max(vaddr, address);
-		const auto range_end  = std::min(end, cached_end);
-		if (begin >= range_end) {
-			continue;
-		}
-		const auto bytes = range_end - begin;
-		if (!m_gpu_modified_ranges.Intersections(begin, bytes).empty()) {
-			EXIT("BufferCache: image ownership overlaps exact dirty buffer bytes\n");
-		}
-		m_image_invalidated_ranges.Add(begin, bytes);
-	}
-}
-
-void BufferCache::BeginBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick) {
-	if (vaddr == 0 || size == 0 || tick == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
-	    size > TRACKER_ADDRESS_SIZE - vaddr) {
-		EXIT("BufferCache: invalid pending backing publication\n");
-	}
-	std::lock_guard lock(m_publication_mutex);
-	m_pending_backing_publications.push_back({vaddr, size, tick});
-}
-
-void BufferCache::CompleteBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick) {
-	std::lock_guard lock(m_publication_mutex);
-	const auto      publication =
-	    std::ranges::find_if(m_pending_backing_publications, [&](const auto& pending) {
-		    return pending.address == vaddr && pending.size == size && pending.tick == tick;
-	    });
-	if (publication == m_pending_backing_publications.end()) {
-		EXIT("BufferCache: completed an unknown backing publication\n");
-	}
-	m_pending_backing_publications.erase(publication);
-}
-
-void BufferCache::PublishImageBuffer(uint64_t vaddr, uint64_t size) {
-	FaultSafeCacheLock lock(this, m_mutex);
-	auto               owner = m_buffers.end();
-	for (auto it = m_buffers.begin(); it != m_buffers.end(); ++it) {
-		if (!PageOverlaps(vaddr, size, it->second->vaddr, it->second->size)) {
-			continue;
-		}
-		if (owner != m_buffers.end() || !it->second->buffer->IsInBounds(vaddr, size)) {
-			EXIT("BufferCache: image destination aliases a non-containing cached buffer\n");
-		}
-		owner = it;
-	}
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
-	                                           "image destination publication");
-	if (owner == m_buffers.end() || m_memory_tracker.IsRegionCpuModified(vaddr, size) ||
-	    !m_gpu_modified_ranges.Intersections(vaddr, size).empty()) {
-		EXIT("BufferCache: image destination requires clean buffer ownership\n");
-	}
-	m_memory_tracker.MarkRegionAsGpuModified(vaddr, size);
-	m_gpu_modified_ranges.Add(vaddr, size);
-	m_image_invalidated_ranges.Subtract(vaddr, size);
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
-	                                           "published image destination");
-	owner->second->tick_accessed_last = m_gc_tick;
-}
-
 void BufferCache::RunGarbageCollector() {
 	std::lock_guard transaction(m_resource_mutex);
 	const auto      tick = m_gc_tick++;
@@ -1162,7 +912,6 @@ void BufferCache::RunGarbageCollector() {
 		if (!m_memory_tracker.IsRegionGpuModified(retire.address, retire.size)) {
 			m_memory_tracker.UntrackMemory(retire.address, retire.size);
 		}
-		m_image_invalidated_ranges.Subtract(retire.address, retire.size);
 		if (retire.size > m_total_used_memory) {
 			EXIT("BufferCache: allocation accounting underflow\n");
 		}

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

@@ -10,7 +10,6 @@
 
 #include <map>
 #include <memory>
-#include <mutex>
 #include <span>
 #include <utility>
 #include <vector>
@@ -30,9 +29,8 @@ struct BufferBinding {
 };
 
 struct ImageBufferSource {
-	Buffer*  buffer    = nullptr;
-	uint64_t offset    = 0;
-	bool     gpu_owned = false;
+	Buffer*  buffer = nullptr;
+	uint64_t offset = 0;
 };
 
 class BufferCache {
@@ -60,9 +58,6 @@ public:
 	                                            uint64_t alignment);
 	[[nodiscard]] std::shared_ptr<Buffer> ObtainNullBuffer();
 	[[nodiscard]] ImageBufferSource       ObtainBufferForImage(uint64_t vaddr, uint64_t size);
-	[[nodiscard]] std::pair<std::shared_ptr<Buffer>, uint64_t>
-	     ObtainBufferForImageWrite(uint64_t vaddr, uint64_t size);
-	void DiscardGpuDirtyBytes(uint64_t vaddr, uint64_t size);
 	void FillBuffer(uint64_t vaddr, uint64_t size, uint32_t value, bool is_gds = false);
 	void CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t size, bool dst_gds = false,
 	                bool src_gds = false);
@@ -70,11 +65,6 @@ public:
 	[[nodiscard]] bool HasGpuDirtyBytes(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionCpuModified(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionGpuModified(uint64_t vaddr, uint64_t size);
-	void               InvalidateImageAliases(uint64_t vaddr, uint64_t size);
-	void               BeginBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick);
-	void               CompleteBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick);
-	[[nodiscard]] bool SynchronizeBacking(uint64_t vaddr, uint64_t size);
-	void               PublishImageBuffer(uint64_t vaddr, uint64_t size);
 	void               RunGarbageCollector();
 
 private:
@@ -88,7 +78,6 @@ private:
 	struct DownloadCopy;
 	struct DownloadRange;
 	struct RetiredBuffer;
-	struct PendingBackingPublication;
 	static constexpr uint64_t               DOWNLOAD_ALIGNMENT = 64;
 	[[nodiscard]] static uint64_t           AlignDown(uint64_t value) noexcept;
 	[[nodiscard]] static uint64_t           AlignUp(uint64_t value);
@@ -103,12 +92,10 @@ private:
 	            const void* source, uint64_t size);
 	[[nodiscard]] CachedBuffer& GetOrCreateBuffer(CommandBuffer& command, uint64_t vaddr,
 	                                              uint64_t size);
+	[[nodiscard]] bool SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size);
 	[[nodiscard]] std::vector<DownloadRange> RecordDownloads(std::span<const DownloadCopy> copies);
 	void PublishDownloads(std::span<const DownloadRange> downloads);
 	void QueueGarbageDownload(std::span<const DownloadCopy> copies, RetiredBuffer retire);
-	void RefreshInvalidatedRanges(CommandBuffer& command, CachedBuffer& cached, uint64_t vaddr,
-	                              uint64_t size, bool upload);
-	void DiscardGpuDirtyBytesLocked(uint64_t vaddr, uint64_t size, const char* operation);
 	void WriteHostMemory(uint64_t vaddr, std::span<const uint8_t> data);
 
 	GraphicContext&                                   m_graphics;
@@ -118,15 +105,11 @@ private:
 	std::shared_ptr<Buffer>                           m_null_buffer;
 	std::map<uint64_t, std::unique_ptr<CachedBuffer>> m_buffers;
 	RangeSet                                          m_gpu_modified_ranges;
-	RangeSet                                          m_image_invalidated_ranges;
-	std::mutex                                        m_publication_mutex;
-	std::vector<PendingBackingPublication>            m_pending_backing_publications;
 	MemoryTracker                                     m_memory_tracker;
 	StreamBuffer                                      m_staging_buffer;
 	StreamBuffer                                      m_stream_buffer;
 	StreamBuffer                                      m_download_buffer;
 	StreamBuffer                                      m_device_buffer;
-	PageManager&                                      m_page_manager;
 	TextureCache&                                     m_texture_cache;
 	ResourceMutex&                                    m_resource_mutex;
 	uint64_t                                          m_total_used_memory = 0;

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

@@ -7,7 +7,8 @@
 namespace Libs::Graphics {
 
 GpuResourceManager::GpuResourceManager(GraphicContext& graphics, CommandScheduler& scheduler)
-    : m_buffer_cache(graphics, scheduler, m_page_manager, m_texture_cache, m_resource_mutex),
+    : m_scheduler(scheduler),
+      m_buffer_cache(graphics, scheduler, m_page_manager, m_texture_cache, m_resource_mutex),
       m_texture_cache(graphics, scheduler, m_page_manager, m_buffer_cache, m_resource_mutex) {}
 
 GpuResourceManager::~GpuResourceManager() = default;
@@ -101,7 +102,22 @@ void GpuResourceManager::MapMemory(uint64_t vaddr, uint64_t size) {
 }
 
 void GpuResourceManager::UnmapMemory(uint64_t vaddr, uint64_t size) {
+	if (CommandScheduler::InDeferredOperation()) {
+		EXIT("unsupported memory unmap from an asynchronous GPU completion, "
+		     "addr=0x%016" PRIx64 " size=0x%016" PRIx64 "\n",
+		     vaddr, size);
+	}
+	if (m_resource_mutex.IsOwnedByCurrentThread()) {
+		EXIT("unsupported memory unmap from a pre-owned resource transaction, "
+		     "addr=0x%016" PRIx64 " size=0x%016" PRIx64 "\n",
+		     vaddr, size);
+	}
 	const auto unmap = [this, vaddr, size] {
+		if (m_scheduler.Active()) {
+			const auto tick = m_scheduler.CurrentTick();
+			m_scheduler.FinishCurrent();
+			m_scheduler.WaitPriorityOperations(tick);
+		}
 		m_buffer_cache.UnmapMemory(vaddr, size);
 		m_texture_cache.UnmapMemory(vaddr, size);
 		m_page_manager.OnGpuUnmap(vaddr, size);
@@ -109,13 +125,9 @@ void GpuResourceManager::UnmapMemory(uint64_t vaddr, uint64_t size) {
 		m_mapped_ranges.Subtract(vaddr, size);
 	};
 	if (m_gpu == nullptr) {
-		if (m_resource_mutex.IsOwnedByCurrentThread()) {
-			EXIT("cannot synchronously unmap from a resource transaction\n");
-		}
 		unmap();
 		return;
 	}
-	Gpu::SubmissionLock submissions(*m_gpu);
 	m_gpu->SendCommandSync(unmap);
 }
 

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

@@ -36,6 +36,7 @@ public:
 private:
 	PageManager               m_page_manager;
 	ResourceMutex             m_resource_mutex;
+	CommandScheduler&         m_scheduler;
 	BufferCache               m_buffer_cache;
 	TextureCache              m_texture_cache;
 	mutable std::shared_mutex m_mapped_ranges_mutex;

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

@@ -219,7 +219,7 @@ private:
 		typename CoarseTable::PageRange   coarse_range {};
 		typename TrackingTable::PageRange tracking_range {};
 		if (!CoarseTable::TryGetPageRange(address, size, coarse_range) ||
-		    !TrackingTable::TryGetPageRange(address, size, tracking_range)) {
+		    (!strict_bytes && !TrackingTable::TryGetPageRange(address, size, tracking_range))) {
 			return {};
 		}
 		MembershipList candidates;
@@ -230,8 +230,8 @@ private:
 		}
 		std::vector<OwnerT> result;
 		for (const Registration* registration: candidates) {
-			if ((!strict_bytes || Overlaps(registration->ranges, address, size)) &&
-			    HasTrackingMembership(registration, tracking_range) &&
+			if ((strict_bytes ? Overlaps(registration->ranges, address, size)
+			                  : HasTrackingMembership(registration, tracking_range)) &&
 			    predicate(registration->owner)) {
 				result.push_back(registration->owner);
 			}

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

@@ -323,7 +323,7 @@ void TextureCache::TrackImage(ImageId id) {
 	if (!image.IsTracked()) {
 		image.track_addr     = image_begin;
 		image.track_addr_end = image_end;
-		m_page_manager.UpdatePageWatchers(true, image_begin, image.info.data.size);
+		m_page_manager.UpdatePageWatchers<true>(image_begin, image.info.data.size);
 		return;
 	}
 	if (image_begin < image.track_addr) {
@@ -348,7 +348,7 @@ void TextureCache::TrackImageHead(ImageId id) {
 	}
 	const auto size  = image.track_addr - image_begin;
 	image.track_addr = image_begin;
-	m_page_manager.UpdatePageWatchers(true, image_begin, size);
+	m_page_manager.UpdatePageWatchers<true>(image_begin, size);
 }
 
 void TextureCache::TrackImageTail(ImageId id) {
@@ -366,7 +366,7 @@ void TextureCache::TrackImageTail(ImageId id) {
 	const auto address   = image.track_addr_end;
 	const auto size      = image_end - address;
 	image.track_addr_end = image_end;
-	m_page_manager.UpdatePageWatchers(true, address, size);
+	m_page_manager.UpdatePageWatchers<true>(address, size);
 }
 
 void TextureCache::UntrackImage(ImageId id) {
@@ -379,7 +379,7 @@ void TextureCache::UntrackImage(ImageId id) {
 	image.track_addr     = 0;
 	image.track_addr_end = 0;
 	if (size != 0) {
-		m_page_manager.UpdatePageWatchers(false, address, size);
+		m_page_manager.UpdatePageWatchers<false>(address, size);
 	}
 }
 
@@ -400,7 +400,7 @@ void TextureCache::UntrackImageHead(ImageId id) {
 		UntrackImage(id);
 	}
 	if (size != 0) {
-		m_page_manager.UpdatePageWatchers(false, begin, size);
+		m_page_manager.UpdatePageWatchers<false>(begin, size);
 	}
 }
 
@@ -421,7 +421,7 @@ void TextureCache::UntrackImageTail(ImageId id) {
 		UntrackImage(id);
 	}
 	if (size != 0) {
-		m_page_manager.UpdatePageWatchers(false, address, size);
+		m_page_manager.UpdatePageWatchers<false>(address, size);
 	}
 }
 
@@ -867,22 +867,28 @@ TextureCache::BuildColorTransfer(const Image& image, BindingType binding,
 			case BindingType::Texture: break;
 			case BindingType::Storage: owner = "StorageTextureCache"; break;
 			case BindingType::RenderTarget:
+				if (info.resources.layers == 0 || info.data.size % info.resources.layers != 0 ||
+				    info.samples != 1 || image.backing.samples != 1) {
+					EXIT("TextureCache: invalid color-attachment upload\n");
+				}
+				format           = ImageOps::RenderTargetTransferFormat(info.bytes_per_block);
+				allow_depth_tile = false;
+				plan.swap_bgra16 = info.bgra16;
+				owner            = "RenderTarget";
+				break;
 			case BindingType::VideoOut:
 				if (info.resources.layers == 0 || info.data.size % info.resources.layers != 0 ||
 				    info.samples != 1 || image.backing.samples != 1 ||
-				    (binding == BindingType::VideoOut &&
-				     info.metadata.compression != VideoOutCompression::Uncompressed)) {
+				    info.metadata.compression != VideoOutCompression::Uncompressed) {
 					EXIT("TextureCache: invalid color-attachment upload\n");
 				}
-				format           = binding == BindingType::RenderTarget
-				                       ? ImageOps::RenderTargetTransferFormat(info.bytes_per_block)
-				                       : info.guest_format;
+				format           = info.guest_format;
 				layers           = info.resources.layers;
 				volume           = false;
 				layered          = layers > 1;
 				allow_depth_tile = false;
 				plan.swap_bgra16 = info.bgra16;
-				owner = binding == BindingType::RenderTarget ? "RenderTarget" : "VideoOut";
+				owner            = "VideoOut";
 				break;
 			case BindingType::DepthTarget: return plan;
 		}
@@ -1038,25 +1044,20 @@ void TextureCache::InitializeImage(ImageId id, const ImageDesc& desc) {
 	if (image.info.samples > 1) {
 		return;
 	}
-	bool       data_gpu_owned = false;
-	bool       data_imported  = false;
-	const bool upload         = image.IsBufferModified() || image.IsCpuDirty();
+	bool       data_imported = false;
+	const bool upload        = image.IsBufferModified() || image.IsCpuDirty();
 	if (upload) {
 		const auto source =
 		    m_buffer_cache.ObtainBufferForImage(image.info.data.address, image.info.data.size);
 		if (source.buffer == nullptr) {
 			EXIT("TextureCache: failed to obtain image upload source\n");
 		}
-		data_gpu_owned |= source.gpu_owned;
 		data_imported = true;
 		UploadImage(image, desc, *source.buffer, source.offset);
 	}
 	if (data_imported) {
 		image.ClearBufferModified();
 	}
-	if (data_gpu_owned) {
-		image.MarkGpuModified();
-	}
 	if (image.IsCpuDirty()) {
 		image.RefreshComplete();
 	}
@@ -1131,8 +1132,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 	}
 
 	ImageId result {};
-	bool    replacement_buffer = false;
-	bool    inserted_new       = false;
+	bool    inserted_new = false;
 	{
 		std::lock_guard            transaction(m_resource_mutex);
 		CacheLock                  lock(*this, m_lock);
@@ -1159,8 +1159,8 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 				if (owner == nullptr) {
 					continue;
 				}
-				const auto merged_info = result ? ResolveImage(result).info : desc.info;
-				const auto overlap     = ResolveOverlap(merged_info, desc.type, candidate, result);
+				const auto& merged_info = result ? ResolveImage(result).info : desc.info;
+				const auto  overlap     = ResolveOverlap(merged_info, desc.type, candidate, result);
 				if (overlap.image) {
 					result     = overlap.image;
 					view_mip   = overlap.mip;
@@ -1174,23 +1174,15 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 			if (exact_format && resolved.info.pixel_format != desc.info.pixel_format) {
 				result = {};
 			} else if (resolved.info.resources < desc.info.resources) {
-				ImageDesc refresh {
-				    .info = resolved.info, .view_info = {}, .type = UploadBinding(resolved)};
-				RefreshImage(result, refresh);
-				if (resolved.IsGpuModified() && !SynchronizeImageToBuffer(result)) {
-					EXIT("TextureCache: cannot preserve an unsupported replacement image\n");
-				}
-				replacement_buffer = resolved.IsBufferModified();
-				DeleteImage(result);
-				result = {};
+				result = ExpandImage(desc.info, result);
 			}
 		}
 		if (!result) {
 			result         = InsertImage(desc.info);
 			inserted_new   = true;
 			auto& inserted = ResolveImage(result);
-			if (replacement_buffer || m_buffer_cache.HasGpuDirtyBytes(inserted.info.data.address,
-			                                                          inserted.info.data.size)) {
+			if (m_buffer_cache.HasGpuDirtyBytes(inserted.info.data.address,
+			                                    inserted.info.data.size)) {
 				inserted.MarkBufferModified();
 			}
 		}
@@ -1360,11 +1352,6 @@ void TextureCache::CommitGpuWrite(Image& image) {
 	if (image.depth_id || image.backing.image == nullptr) {
 		EXIT("TextureCache: stencil association cannot own image contents\n");
 	}
-	const auto range = image.info.data;
-	if (m_buffer_cache.HasGpuDirtyBytes(range.address, range.size)) {
-		m_buffer_cache.DiscardGpuDirtyBytes(range.address, range.size);
-	}
-	m_buffer_cache.InvalidateImageAliases(range.address, range.size);
 	image.ClearBufferModified();
 	if (image.IsCpuDirty()) {
 		image.RefreshComplete();
@@ -1429,9 +1416,6 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 			return false;
 		}
 	}
-	if (m_buffer_cache.HasGpuDirtyBytes(address, size)) {
-		m_buffer_cache.DiscardGpuDirtyBytes(address, size);
-	}
 	if (image.IsBufferModified() || image.IsCpuDirty()) {
 		ImageDesc refresh {.info = image.info, .view_info = {}, .type = UploadBinding(image)};
 		InitializeImage(selected, refresh);
@@ -1551,11 +1535,14 @@ void TextureCache::DownloadDepth(Image& image, Buffer& destination, uint64_t des
 }
 
 void TextureCache::DownloadImageData(Image& image, Buffer& destination, uint64_t destination_offset,
-                                     DownloadPlan plan) {
+                                     uint64_t destination_size, DownloadPlan plan) {
 	if (!plan.valid) {
 		EXIT("TextureCache: invalid image download plan\n");
 	}
 	if (plan.depth) {
+		if (destination_size != image.info.data.size) {
+			EXIT("TextureCache: partial depth image download is unsupported\n");
+		}
 		DownloadDepth(image, destination, destination_offset);
 		return;
 	}
@@ -1565,22 +1552,118 @@ void TextureCache::DownloadImageData(Image& image, Buffer& destination, uint64_t
 	                                         : TileManager::ColorTransform::None;
 	if (!color.tiled) {
 		if (transform == TileManager::ColorTransform::SwapBgra16) {
-			auto linear = m_tiler->GetScratchBuffer(image.info.data.size);
+			auto linear = m_tiler->GetScratchBuffer(destination_size);
 			image.Download(color.regions, linear.buffer, 0, linear.size);
 			m_tiler->SwapBgra16(linear,
-			                    {destination.Handle(), destination_offset, image.info.data.size});
+			                    {destination.Handle(), destination_offset, destination_size});
 			return;
 		}
 		for (auto& copy: color.regions) {
 			copy.bufferOffset += destination_offset;
 		}
-		image.Download(color.regions, destination.Handle(), destination_offset,
-		               image.info.data.size);
+		image.Download(color.regions, destination.Handle(), destination_offset, destination_size);
 		return;
 	}
 
 	m_tiler->TileImage(image, color.regions, destination.Handle(), destination_offset,
-	                   image.info.data.size, image.info.data.size, color.tiles, transform);
+	                   destination_size, destination_size, color.tiles, transform);
+}
+
+bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size) {
+	CacheLock            lock(m_texture_cache, m_texture_cache.m_lock);
+	std::vector<ImageId> matches;
+	for (const auto id: m_texture_cache.FindImagesInRegion(vaddr, size, false)) {
+		auto owner = m_texture_cache.ResolveOwner(id);
+		if (owner == nullptr || owner->info.data.address != vaddr) {
+			continue;
+		}
+		if (owner->depth_id) {
+			owner = m_texture_cache.ResolveOwner(owner->depth_id);
+		}
+		if (owner != nullptr && owner->SafeToDownload()) {
+			matches.push_back(id);
+		}
+	}
+
+	ImageId selected {};
+	if (matches.size() == 1) {
+		selected = matches.front();
+	} else {
+		for (const auto id: matches) {
+			const auto& image = m_texture_cache.ResolveImage(id);
+			if (image.info.data.size == size) {
+				selected = id;
+				break;
+			}
+		}
+	}
+	if (!selected) {
+		return false;
+	}
+	if (const auto owner = m_texture_cache.ResolveOwner(selected);
+	    owner != nullptr && owner->depth_id) {
+		selected = owner->depth_id;
+	}
+
+	auto& image = m_texture_cache.ResolveImage(selected);
+	if (!buffer.IsInBounds(image.info.data.address, 1)) {
+		return false;
+	}
+	const auto buf_offset = buffer.Offset(image.info.data.address);
+	const auto available  = buffer.Size() - buf_offset;
+	uint32_t   levels     = 0;
+	uint64_t   copy_size  = 0;
+	if (image.info.IsVolume()) {
+		// Volume mips contain strided block slices, so a mip's linear span cannot prove that
+		// every retained slice fits. Keep volume synchronization whole-image only.
+		if (!buffer.IsInBounds(image.info.data.address, image.info.data.size)) {
+			return false;
+		}
+		levels    = image.info.resources.levels;
+		copy_size = image.info.data.size;
+	} else {
+		for (; levels < image.info.resources.levels; ++levels) {
+			const auto& mip = image.info.mip_layout[levels];
+			if (mip.size == 0 || mip.offset > available || mip.size > available - mip.offset) {
+				break;
+			}
+			copy_size = std::max(copy_size, mip.offset + mip.size);
+		}
+	}
+	if (copy_size == 0) {
+		return false;
+	}
+	auto plan = m_texture_cache.BuildDownload(image);
+	if (!plan.valid) {
+		return false;
+	}
+	if (plan.depth && copy_size != image.info.data.size) {
+		return false;
+	}
+	if (!plan.depth && levels < image.info.resources.levels) {
+		auto& color = plan.color;
+		std::erase_if(color.regions, [levels](const vk::BufferImageCopy& region) {
+			return region.imageSubresource.mipLevel >= levels;
+		});
+		if (color.regions.empty()) {
+			return false;
+		}
+		if (color.tiled) {
+			const auto binding = m_texture_cache.UploadBinding(image);
+			const auto format =
+			    binding == TextureCache::BindingType::RenderTarget
+			        ? ImageOps::RenderTargetTransferFormat(image.info.bytes_per_block)
+			        : image.info.guest_format;
+			color.tiles.clear();
+			if (!TextureBuildGpuTileInfos(copy_size, color.regions, color.layout, format,
+			                              image.info.TransferLayers(), levels, color.tiles)) {
+				return false;
+			}
+		}
+	}
+	m_texture_cache.DownloadImageData(image, buffer, buf_offset, copy_size, std::move(plan));
+	m_texture_cache.RetainImage(m_scheduler.Current(), selected);
+	return true;
 }
 
 std::pair<uint8_t*, uint64_t> TextureCache::MapDownload(uint64_t size, uint64_t alignment) {
@@ -1612,12 +1695,9 @@ void TextureCache::QueueDownload(GuestRange range, StreamBuffer& download, uint8
 	m_scheduler.Current().Handle().pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
 	                                               vk::PipelineStageFlagBits::eHost, {}, 0, nullptr,
 	                                               1, &barrier, 0, nullptr);
-	const auto tick = m_scheduler.CurrentTick();
-	m_buffer_cache.BeginBackingPublication(range.address, range.size, tick);
-	m_scheduler.DeferPriorityOperation([this, &download, range, mapped, offset, tick] {
+	m_scheduler.DeferPriorityOperation([&download, range, mapped, offset] {
 		download.Invalidate(offset, range.size);
 		LibKernel::Memory::WriteBacking(range.address, mapped, range.size);
-		m_buffer_cache.CompleteBackingPublication(range.address, range.size, tick);
 	});
 }
 
@@ -1638,7 +1718,7 @@ bool TextureCache::TryDownloadImage(ImageId id) {
 	}
 	download.Flush(offset, range.size);
 
-	DownloadImageData(image, download, offset, std::move(plan));
+	DownloadImageData(image, download, offset, range.size, std::move(plan));
 
 	QueueDownload(range, download, mapped, offset);
 	return true;
@@ -1652,62 +1732,6 @@ void TextureCache::DownloadImage(ImageId id) {
 	m_scheduler.DrainPriorityOperations();
 }
 
-bool TextureCache::SynchronizeImageToBuffer(ImageId id) {
-	auto& image = ResolveImage(id);
-	if (image.depth_id) {
-		return true;
-	}
-	auto plan = BuildDownload(image);
-	if (!plan.valid) {
-		return false;
-	}
-	const auto range = image.info.data;
-	if (image.IsCpuDirty()) {
-		RefreshImage(id,
-		             ImageDesc {.info = image.info, .view_info = {}, .type = UploadBinding(image)});
-	}
-	if (!image.IsGpuModified()) {
-		return true;
-	}
-	if (image.IsDefinitelyCpuDirty() || image.IsBufferModified()) {
-		EXIT("TextureCache: image mirror source is not native-current\n");
-	}
-	auto [destination, offset] =
-	    m_buffer_cache.ObtainBufferForImageWrite(range.address, range.size);
-	if (destination == nullptr) {
-		EXIT("TextureCache: failed to allocate image mirror\n");
-	}
-	DownloadImageData(image, *destination, offset, std::move(plan));
-	m_scheduler.Current().RetainResourceUntilFence(destination);
-	m_buffer_cache.PublishImageBuffer(range.address, range.size);
-	image.MarkBufferModified();
-	RetainImage(m_scheduler.Current(), id);
-	ClearGpuModified(id);
-	return true;
-}
-
-bool TextureCache::SynchronizeImageToBuffer(uint64_t address, uint64_t size) {
-	if (!GuestRange {address, size}.Valid()) {
-		return false;
-	}
-	CacheLock lock(*this, m_lock);
-	ImageId   selected {};
-	for (const auto id: FindImagesInRegion(address, size, true)) {
-		auto owner = ResolveOwner(id);
-		if (owner == nullptr || !owner->GpuOverlaps(address, size)) {
-			continue;
-		}
-		if (selected) {
-			EXIT("TextureCache: ambiguous image-to-buffer synchronization\n");
-		}
-		selected = id;
-	}
-	if (!selected) {
-		return false;
-	}
-	return SynchronizeImageToBuffer(selected);
-}
-
 bool TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size,
                                            bool formatted_buffer_write) {
 	if (!GuestRange {address, size}.Valid()) {

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

@@ -66,7 +66,6 @@ public:
 	[[nodiscard]] bool ClearImageFromBuffer(CommandBuffer& command, uint64_t address, uint64_t size,
 	                                        uint32_t packed_clear);
 	void               InvalidateMemory(uint64_t address, uint64_t size);
-	[[nodiscard]] bool SynchronizeImageToBuffer(uint64_t address, uint64_t size);
 	[[nodiscard]] bool InvalidateMemoryFromGPU(uint64_t address, uint64_t size,
 	                                           bool formatted_buffer_write = false);
 	[[nodiscard]] RegionInfo QueryRegion(uint64_t address, uint64_t size);
@@ -140,7 +139,7 @@ private:
 	[[nodiscard]] DownloadPlan      BuildDownload(const Image& image) const;
 	void UploadImage(Image& image, const ImageDesc& desc, Buffer& source, uint64_t source_offset);
 	void DownloadImageData(Image& image, Buffer& destination, uint64_t destination_offset,
-	                       DownloadPlan plan);
+	                       uint64_t destination_size, DownloadPlan plan);
 	void DownloadDepth(Image& image, Buffer& destination, uint64_t destination_offset);
 	void CommitGpuWrite(Image& image);
 	void PrepareImageCopy(Image& image);
@@ -155,7 +154,6 @@ private:
 	void InvalidateCpuAliases(uint64_t address, uint64_t size);
 	void ClearGpuModified(ImageId id);
 
-	[[nodiscard]] bool                          SynchronizeImageToBuffer(ImageId id);
 	void                                        DownloadImage(ImageId id);
 	[[nodiscard]] bool                          TryDownloadImage(ImageId id);
 	[[nodiscard]] std::pair<uint8_t*, uint64_t> MapDownload(uint64_t size, uint64_t alignment);
@@ -184,6 +182,7 @@ private:
 	bool     m_readback_linear_images = false;
 
 	friend struct TextureCacheTestAccess;
+	friend class BufferCache;
 	friend class RenderExecutor;
 };
 

src/graphics/host_gpu/renderer/colorRenderTarget.cpp

@@ -119,7 +119,18 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	uint32_t   pitch  = 0;
 	uint64_t   size   = 0;
 	bool       tile   = false;
-	const bool standard64 =
+	const bool volume = rt.attrib3.dimension == 2;
+	if (rt.attrib3.dimension != 1 && !volume) {
+		EXIT("unsupported render-target dimension: %u\n", rt.attrib3.dimension);
+	}
+	if (!volume && rt.attrib3.depth != 0) {
+		EXIT("2D render target has nonzero depth: %u\n", rt.attrib3.depth);
+	}
+	if (volume && samples != 1) {
+		EXIT("multisampled 3D render targets are unsupported\n");
+	}
+	const uint32_t depth = volume ? rt.attrib3.depth + 1u : 1u;
+	const bool     standard64 =
 	    rt.attrib3.tile_mode == Prospero::GpuEnumValue(Prospero::TileMode::kStandard64KB);
 
 	switch (rt.attrib3.tile_mode) {
@@ -145,6 +156,7 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	if (bytes_per_element == 0) {
 		EXIT("render-target format has no valid element size\n");
 	}
+	const auto transfer_format = ImageOps::RenderTargetTransferFormat(bytes_per_element);
 	if (standard64 &&
 	    (rt.attrib3.dimension != 1 || rt.attrib3.depth != 0 || levels != 1 ||
 	     rt.view.current_mip_level != 0 || view.base_layer != 0 || view.image_layers != 1 ||
@@ -165,10 +177,14 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	if (rt.pitch.pitch_div8_minus1 != 0) {
 		pitch = (rt.pitch.pitch_div8_minus1 + 1u) << 3u;
 	} else if (tile) {
-		pitch = standard64
-		            ? TileGetTexturePitch(Prospero::GpuEnumValue(Prospero::BufferFormat::k32Float),
-		                                  width, levels, rt.attrib3.tile_mode)
-		            : TileGetRenderTargetPitch(width, bytes_per_element, rt.attrib.num_fragments);
+		if (volume) {
+			pitch = TileGetTexturePitch(transfer_format, width, levels, rt.attrib3.tile_mode);
+		} else if (standard64) {
+			pitch = TileGetTexturePitch(Prospero::GpuEnumValue(Prospero::BufferFormat::k32Float),
+			                            width, levels, rt.attrib3.tile_mode);
+		} else {
+			pitch = TileGetRenderTargetPitch(width, bytes_per_element, rt.attrib.num_fragments);
+		}
 		if (pitch == 0) {
 			EXIT("unsupported render-target pitch: width=%u bytes=%u\n", width, bytes_per_element);
 		}
@@ -176,9 +192,19 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 		pitch = width;
 	}
 
-	TileSizeOffset mip_sizes[16] {};
-	TilePaddedSize mip_padded[16] {};
-	if (tile) {
+	TileSizeOffset   mip_sizes[16] {};
+	TilePaddedSize   mip_padded[16] {};
+	TileVolumeLayout volume_layout {};
+	uint64_t         backing_size = 0;
+	if (volume) {
+		if (!tile || !TileGetTextureVolumeLayout(transfer_format, width, height, depth, levels,
+		                                         rt.attrib3.tile_mode, volume_layout)) {
+			EXIT("unsupported 3D render-target layout: %ux%ux%u levels=%u tile=%u\n", width, height,
+			     depth, levels, rt.attrib3.tile_mode);
+		}
+		size         = volume_layout.block_slice_size;
+		backing_size = volume_layout.total_size;
+	} else if (tile) {
 		TileSizeAlign layout {};
 		bool          valid_layout = false;
 		if (standard64) {
@@ -203,12 +229,6 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 			mip_sizes[0]  = {static_cast<uint32_t>(size), 0, 0, 0, 0, 0};
 			mip_padded[0] = {pitch, height};
 		}
-		if (rt.slice.slice_div64_minus1 != 0 &&
-		    (static_cast<uint64_t>(rt.slice.slice_div64_minus1) + 1u) * 64u != size) {
-			EXIT("render-target slice span mismatch: encoded=0x%016" PRIx64 " derived=0x%016" PRIx64
-			     "\n",
-			     (static_cast<uint64_t>(rt.slice.slice_div64_minus1) + 1u) * 64u, size);
-		}
 	} else {
 		size = static_cast<uint64_t>(pitch) * height * bytes_per_element * samples;
 		if (size > UINT32_MAX) {
@@ -217,40 +237,66 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 		mip_sizes[0]  = {static_cast<uint32_t>(size), 0, 0, 0, 0, 0};
 		mip_padded[0] = {pitch, height};
 	}
-	if (size == 0 || size > UINT64_MAX / view.image_layers) {
+	if (rt.slice.slice_div64_minus1 != 0 &&
+	    (static_cast<uint64_t>(rt.slice.slice_div64_minus1) + 1u) * 64u != size) {
+		EXIT("render-target slice span mismatch: encoded=0x%016" PRIx64 " derived=0x%016" PRIx64
+		     "\n",
+		     (static_cast<uint64_t>(rt.slice.slice_div64_minus1) + 1u) * 64u, size);
+	}
+	if (size == 0 || (!volume && size > UINT64_MAX / view.image_layers)) {
 		EXIT("render-target memory footprint is invalid\n");
 	}
-	const auto backing_size = size * view.image_layers;
+	if (!volume) {
+		backing_size = size * view.image_layers;
+	}
+	if (backing_size == 0) {
+		EXIT("render-target backing is empty\n");
+	}
 	if (backing_size > TRACKER_ADDRESS_SIZE - rt.base.addr) {
 		EXIT("render-target backing range is invalid\n");
 	}
 
 	const vk::Extent2D view_extent = {std::max(width >> rt.view.current_mip_level, 1u),
 	                                  std::max(height >> rt.view.current_mip_level, 1u)};
+	const uint32_t     view_depth  = std::max(depth >> rt.view.current_mip_level, 1u);
+	if (volume &&
+	    (view.base_layer >= view_depth || view.layer_count > view_depth - view.base_layer)) {
+		EXIT("3D render-target view exceeds mip depth: base=%u count=%u depth=%u mip=%u\n",
+		     view.base_layer, view.layer_count, view_depth, rt.view.current_mip_level);
+	}
 
 	auto decision_log_id = g_render_color_log_count.fetch_add(1);
 	if (decision_log_id < 128) {
 		LOGF("RenderColorTarget: slot=%" PRIu32 " addr=0x%010" PRIx64 " size=0x%016" PRIx64
-		     " extent=%ux%u view_mip=%u view_extent=%ux%u levels=%u pitch=%u"
+		     " extent=%ux%ux%u view_mip=%u view_extent=%ux%u levels=%u pitch=%u"
 		     " fmt=0x%08" PRIx32 " nfmt=0x%08" PRIx32 " order=0x%08" PRIx32 " samples=%u tile=%s\n",
-		     rt_slot, rt.base.addr, backing_size, width, height, rt.view.current_mip_level,
+		     rt_slot, rt.base.addr, backing_size, width, height, depth, rt.view.current_mip_level,
 		     view_extent.width, view_extent.height, levels, pitch, rt.info.format,
 		     rt.info.channel_type, rt.info.channel_order, samples, tile ? "tiled" : "linear");
 	}
 
 	TextureCache::ImageDesc desc {};
-	desc.type                 = TextureCache::BindingType::RenderTarget;
-	desc.info.data            = {rt.base.addr, backing_size};
-	desc.info.pixel_format    = target_format.format;
-	desc.info.guest_format    = ImageOps::RenderTargetTransferFormat(bytes_per_element);
-	desc.info.type            = Prospero::ImageType::kColor2D;
-	desc.info.extent          = {width, height, 1};
-	desc.info.resources       = {levels, view.image_layers};
-	desc.info.pitch           = pitch;
+	desc.type              = TextureCache::BindingType::RenderTarget;
+	desc.info.data         = {rt.base.addr, backing_size};
+	desc.info.pixel_format = target_format.format;
+	desc.info.guest_format = transfer_format;
+	desc.info.type         = volume ? Prospero::ImageType::kColor3D : Prospero::ImageType::kColor2D;
+	desc.info.extent       = {width, height, depth};
+	desc.info.resources    = {levels, volume ? 1u : view.image_layers};
+	desc.info.pitch        = pitch;
 	desc.info.bytes_per_block = bytes_per_element;
 	desc.info.samples         = samples;
 	desc.info.tile_mode       = rt.attrib3.tile_mode;
 	for (uint32_t level = 0; level < levels; level++) {
+		if (volume) {
+			desc.info.mip_layout[level] = {
+			    volume_layout.level_offsets[level],
+			    volume_layout.level_sizes[level],
+			    volume_layout.level_widths[level],
+			    volume_layout.level_heights[level],
+			};
+			continue;
+		}
 		const auto level_offset =
 		    mip_sizes[level].src_size != 0 ? mip_sizes[level].src_offset : mip_sizes[level].offset;
 		const auto level_size =

src/graphics/host_gpu/renderer/debug.cpp

@@ -359,30 +359,12 @@ static void RtCheck(const HW::RenderTarget& rt) {
 				logged = true;
 			}
 		}
-		if (rt.attrib3.depth != 0x00000000) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("RenderTarget: temporary: ignoring PS5 color target depth_minus1=0x%08" PRIx32
-				     "\n",
-				     rt.attrib3.depth);
-				logged = true;
-			}
-		}
 		if (!RenderIsColorTileMode(rt.attrib3.tile_mode)) {
 			EXIT("unknown PS5 render-target tile mode: 0x%08" PRIx32 "\n", rt.attrib3.tile_mode);
 		}
 		if (!RenderIsColorDimension(rt.attrib3.dimension)) {
 			EXIT("unknown PS5 render-target dimension: 0x%08" PRIx32 "\n", rt.attrib3.dimension);
 		}
-		if (rt.attrib3.dimension != 0x00000001) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("RenderTarget: temporary: using 2D fallback for PS5 color "
-				     "dimension=0x%08" PRIx32 "\n",
-				     rt.attrib3.dimension);
-				logged = true;
-			}
-		}
 		if (!rt.attrib3.cmask_pipe_aligned) {
 			static bool logged = false;
 			if (!logged) {

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

@@ -105,6 +105,10 @@ Image::Barriers Image::GetBarriers(vk::ImageLayout                      destinat
                                    std::optional<ImageSubresourceRange> range) {
 	auto& state              = backing.state;
 	auto& subresource_states = backing.subresource_states;
+	if (range && info.IsVolume()) {
+		range->base_layer  = 0;
+		range->layer_count = 1;
+	}
 
 	const bool partial =
 	    range && (range->base_level != 0 || range->level_count != info.resources.levels ||

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

@@ -113,6 +113,13 @@ inline void ValidateStorageColorView(vk::Format image_format, vk::Format view_fo
 
 [[nodiscard]] inline bool
 IsSupportedStorageImageResource(const ShaderRecompiler::IR::ImageResource& resource) noexcept {
+	const bool supported_mip =
+	    (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None &&
+	     resource.mip_levels == 1u) ||
+	    (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage &&
+	     resource.mip_levels > 0u &&
+	     resource.mip_levels <= ShaderRecompiler::IR::ImageResource::MaxMipLevels &&
+	     !resource.read && !resource.atomic);
 	return (resource.kind == ShaderRecompiler::IR::ResourceKind::StorageImage ||
 	        resource.kind == ShaderRecompiler::IR::ResourceKind::StorageImageUint) &&
 	       (resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim1D ||
@@ -120,7 +127,7 @@ IsSupportedStorageImageResource(const ShaderRecompiler::IR::ImageResource& resou
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2D ||
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim3D ||
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DArray) &&
-	       resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None && resource.written &&
+	       supported_mip && resource.written &&
 	       (!resource.atomic ||
 	        (resource.kind == ShaderRecompiler::IR::ResourceKind::StorageImageUint &&
 	         resource.read)) &&

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

@@ -94,10 +94,11 @@ vk::DescriptorBufferInfo BufferInfo(const BufferView& view) {
 
 } // namespace
 
-vk::DescriptorImageInfo DescriptorCache::MakeImageInfo(const TextureBinding& texture) {
-	EXIT_IF(!texture.image_id || texture.image_view == nullptr ||
-	        texture.layout == vk::ImageLayout::eUndefined);
-	return {nullptr, texture.image_view, texture.layout};
+vk::DescriptorImageInfo DescriptorCache::MakeImageInfo(const TextureBinding& texture,
+                                                       uint32_t              mip) {
+	const auto view = texture.mip_views.empty() ? texture.image_view : texture.mip_views.at(mip);
+	EXIT_IF(!texture.image_id || view == nullptr || texture.layout == vk::ImageLayout::eUndefined);
+	return {nullptr, view, texture.layout};
 }
 
 DescriptorCache::~DescriptorCache() {
@@ -154,7 +155,8 @@ void DescriptorCache::CreatePool() {
 	     MaxSets * (ShaderRecompiler::IR::ShaderInfo::MaxBuffers +
 	                ShaderRecompiler::IR::ShaderInfo::MaxAddresses + 3u)},
 	    {vk::DescriptorType::eSampledImage, MaxSets * ShaderRecompiler::IR::ShaderInfo::MaxImages},
-	    {vk::DescriptorType::eStorageImage, MaxSets * ShaderRecompiler::IR::ShaderInfo::MaxImages},
+	    {vk::DescriptorType::eStorageImage, MaxSets * ShaderRecompiler::IR::ShaderInfo::MaxImages *
+	                                            ShaderRecompiler::IR::ImageResource::MaxMipLevels},
 	    {vk::DescriptorType::eSampler, MaxSets * ShaderRecompiler::IR::ShaderInfo::MaxSamplers},
 	};
 	vk::DescriptorPoolCreateInfo info {};
@@ -229,8 +231,10 @@ VulkanDescriptorSet& DescriptorCache::GetDescriptor(Stage
 	auto* set = Allocate(stage, program);
 	EXIT_NOT_IMPLEMENTED(set == nullptr);
 
-	const auto descriptor_count = program.info.buffers.size() + program.info.images.size() +
-	                              program.info.samplers.size() + program.info.addresses.size() + 3u;
+	uint32_t descriptor_count = 0;
+	for (const auto& binding: program.bindings.descriptors) {
+		descriptor_count += DescriptorCount(binding);
+	}
 	std::vector<vk::DescriptorBufferInfo> buffer_infos;
 	std::vector<vk::DescriptorImageInfo>  image_infos;
 	std::vector<vk::WriteDescriptorSet>   writes;
@@ -238,6 +242,7 @@ VulkanDescriptorSet& DescriptorCache::GetDescriptor(Stage
 	image_infos.reserve(descriptor_count);
 	writes.reserve(program.bindings.descriptors.size());
 
+	std::vector<uint32_t> image_mips(program.info.images.size());
 	for (const auto& binding: program.bindings.descriptors) {
 		vk::WriteDescriptorSet write {};
 		write.sType             = vk::StructureType::eWriteDescriptorSet;
@@ -273,7 +278,11 @@ VulkanDescriptorSet& DescriptorCache::GetDescriptor(Stage
 			default: {
 				for (const auto resource: binding.resources) {
 					const auto& texture = data.images.at(resource);
-					image_infos.push_back(MakeImageInfo(texture));
+					const auto  mip     = program.info.images.at(resource).mip_mode ==
+					                              ShaderRecompiler::IR::ImageMipMode::DynamicStorage
+					                          ? image_mips.at(resource)++
+					                          : 0u;
+					image_infos.push_back(MakeImageInfo(texture, mip));
 				}
 				break;
 			}

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

@@ -47,10 +47,11 @@ public:
 	enum class Stage { Unknown, Vertex, Pixel, Compute };
 
 	struct TextureBinding {
-		ImageId                 image_id;
-		vk::ImageView           image_view = nullptr;
-		TextureCache::ImageDesc desc;
-		vk::ImageLayout         layout = vk::ImageLayout::eUndefined;
+		ImageId                    image_id;
+		vk::ImageView              image_view = nullptr;
+		TextureCache::ImageDesc    desc;
+		vk::ImageLayout            layout = vk::ImageLayout::eUndefined;
+		std::vector<vk::ImageView> mip_views;
 	};
 
 	struct NativeDescriptors {
@@ -94,7 +95,7 @@ private:
 		int                next_free_pool = -1;
 	};
 
-	static vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture);
+	static vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture, uint32_t mip = 0);
 	void                           CreatePool();
 	VulkanDescriptorSet* Allocate(Stage stage, const ShaderRecompiler::IR::Program& program);
 	vk::DescriptorSetLayout

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

@@ -391,9 +391,16 @@ static bool IsSupportedStorageTextureDescriptor(const ShaderRecompiler::IR::Imag
 	const bool supported_swizzle =
 	    IsValidImageSwizzle(swizzle) &&
 	    (swizzle == DstSel(4, 5, 6, 7) || !resource.read || resource.atomic);
+	const auto base_level = static_cast<uint32_t>(descriptor.BaseLevel());
+	const auto last_level = static_cast<uint32_t>(descriptor.LastLevel());
+	const auto mip_levels = last_level >= base_level ? last_level - base_level + 1u : 0u;
+	const bool dynamic_mip =
+	    resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage;
 	const bool supported_mip_view = descriptor.BaseLevel() == 0 || is_1d || is_2d;
 	return (is_1d || is_1d_array || is_2d || is_2d_array || is_3d) && supported_tile &&
-	       supported_mip_view && descriptor.BaseLevel() == descriptor.LastLevel() &&
+	       supported_mip_view && mip_levels != 0u &&
+	       ((dynamic_mip && mip_levels == resource.mip_levels) ||
+	        (!dynamic_mip && descriptor.BaseLevel() == descriptor.LastLevel())) &&
 	       descriptor.LastLevel() <= descriptor.MaxMip() && descriptor.MinLod() == 0 &&
 	       supported_swizzle && descriptor.BCSwizzle() == 0 && !descriptor.MsaaDepth();
 }
@@ -618,14 +625,15 @@ RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   reso
 	const bool multisampled = IsMultisampledTexture(type);
 	const auto levels       = multisampled ? 1u : static_cast<uint32_t>(descriptor.MaxMip()) + 1u;
 	const auto tile         = descriptor.TileMode();
-	const bool depth_tile = tile == Prospero::GpuEnumValue(Prospero::TileMode::kDepth);
+	const bool depth_tile   = tile == Prospero::GpuEnumValue(Prospero::TileMode::kDepth);
 	const bool msaa_tile =
 	    depth_tile || tile == Prospero::GpuEnumValue(Prospero::TileMode::kRenderTarget);
 	const bool msaa_array = type == Prospero::ImageType::kColor2DMsaaArray;
 	if ((!multisampled && (base_level > last_level || last_level >= levels)) ||
 	    (multisampled &&
 	     (base_level != 0 || last_level == 0 || last_level > 3 ||
-	      descriptor.MaxMip() != last_level || !msaa_tile || (descriptor.MsaaDepth() && !depth_tile) ||
+	      descriptor.MaxMip() != last_level || !msaa_tile ||
+	      (descriptor.MsaaDepth() && !depth_tile) ||
 	      (!msaa_array && (descriptor.Depth() != 0 || descriptor.BaseArray5() != 0))))) {
 		EXIT("unsupported texture mip view: base=%u last=%u levels=%u max=%u type=%u tile=%u "
 		     "kind=%u dimension=%u mip_mode=%u read=%d written=%d "
@@ -634,11 +642,15 @@ RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   reso
 		     static_cast<uint32_t>(resource.kind), static_cast<uint32_t>(resource.dimension),
 		     static_cast<uint32_t>(resource.mip_mode), resource.read, resource.written,
 		     descriptor.fields[0], descriptor.fields[1], descriptor.fields[2], descriptor.fields[3],
-		     descriptor.fields[4], descriptor.fields[5], descriptor.fields[6], descriptor.fields[7]);
+		     descriptor.fields[4], descriptor.fields[5], descriptor.fields[6],
+		     descriptor.fields[7]);
 	}
 	const auto samples = multisampled ? 1u << last_level : 1u;
 	const auto view_levels =
-	    multisampled ? 1u : static_cast<uint32_t>(last_level - base_level) + 1u;
+	    multisampled ||
+	            (storage && resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage)
+	        ? 1u
+	        : static_cast<uint32_t>(last_level - base_level) + 1u;
 	const auto depth  = static_cast<uint32_t>(descriptor.Depth()) + 1u;
 	const auto format = descriptor.Format();
 	const bool sampled_numeric_class =
@@ -661,8 +673,8 @@ RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   reso
 	TileSizeAlign size {};
 	if (multisampled) {
 		const auto bytes = Prospero::NumBytesPerElement(format);
-		pitch = depth_tile ? TileGetDepthPitch(width, bytes, last_level)
-		                   : TileGetRenderTargetPitch(width, bytes, last_level);
+		pitch            = depth_tile ? TileGetDepthPitch(width, bytes, last_level)
+		                              : TileGetRenderTargetPitch(width, bytes, last_level);
 		if (pitch == 0 || !TileGetRenderTargetSize(width, height, pitch, bytes, size, last_level) ||
 		    size.size > UINT32_MAX / image_layers) {
 			EXIT("unsupported multisample texture layout\n");
@@ -679,8 +691,8 @@ RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   reso
 		ValidateStorageTexture(resource, descriptor, size.size);
 	}
 
-	const auto              pixel_format = TextureGetFormat(format);
-	const auto              storage_view_format =
+	const auto pixel_format = TextureGetFormat(format);
+	const auto storage_view_format =
 	    storage && format == Prospero::GpuEnumValue(Prospero::BufferFormat::k32SInt)
 	        ? vk::Format::eR32Uint
 	        : SrgbStorageViewFormat(pixel_format);
@@ -893,7 +905,17 @@ void RenderExecutor::RebindImages(CommandBuffer&                     buffer,
 		}
 		auto& binding      = images[i];
 		binding.image_view = texture_cache.FindTexture(binding.image_id, binding.desc);
-		auto&      image   = texture_cache.GetImage(binding.image_id);
+		auto& image        = texture_cache.GetImage(binding.image_id);
+		binding.mip_views.clear();
+		if (program.info.images[i].mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage) {
+			binding.mip_views.reserve(program.info.images[i].mip_levels);
+			for (uint32_t mip = 0; mip < program.info.images[i].mip_levels; mip++) {
+				auto view = binding.desc.view_info;
+				view.base_level += mip;
+				view.level_count = 1;
+				binding.mip_views.push_back(mip == 0 ? binding.image_view : image.FindView(view));
+			}
+		}
 		const bool storage = binding.desc.type == TextureCache::BindingType::Storage;
 		image.usage.storage |= storage;
 		image.usage.texture |= !storage;
@@ -971,7 +993,11 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 		auto&       image   = m_context.GetTextureCache().GetImage(descriptors.images[i].image_id);
 		auto&       binding = descriptors.images[i];
 		const auto& view    = binding.desc.view_info;
-		const ImageSubresourceRange range {view.base_level, view.level_count, view.base_layer,
+		const auto  level_count =
+		    program.info.images[i].mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage
+		        ? program.info.images[i].mip_levels
+		        : view.level_count;
+		const ImageSubresourceRange range {view.base_level, level_count, view.base_layer,
 		                                   view.layer_count};
 		const bool storage = binding.desc.type == TextureCache::BindingType::Storage;
 		if (image.info.data.Empty()) {

src/graphics/presentation/window/vulkanWindow.cpp

@@ -287,6 +287,10 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 			LOGF("shaderStorageImageReadWithoutFormat is not supported\n");
 			skip_device = true;
 		}
+		if (features12.shaderStorageImageArrayNonUniformIndexing != VK_TRUE) {
+			LOGF("shaderStorageImageArrayNonUniformIndexing is not supported\n");
+			skip_device = true;
+		}
 
 		if (device_features2.features.shaderImageGatherExtended != VK_TRUE) {
 			LOGF("shaderImageGatherExtended is not supported\n");
@@ -514,6 +518,7 @@ static vk::Device VulkanCreateDevice(vk::PhysicalDevice physical_device, const V
 	features12.sType                    = vk::StructureType::ePhysicalDeviceVulkan12Features;
 	features12.pNext                    = &depth_clip_control;
 	features12.samplerMirrorClampToEdge = VK_TRUE;
+	features12.shaderStorageImageArrayNonUniformIndexing = VK_TRUE;
 
 	vk::PhysicalDeviceSubgroupSizeControlFeatures subgroup_size_control {};
 	subgroup_size_control.sType = vk::StructureType::ePhysicalDeviceSubgroupSizeControlFeatures;

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

@@ -871,19 +871,19 @@ std::vector<uint32_t> DominatedBlocks(const Graph& graph, uint32_t header,
 	return blocks;
 }
 
-uint32_t AppendSyntheticMergeBlock(Graph& graph, uint32_t old_merge) {
-	const auto* merge = graph.FindBlock(old_merge);
+uint32_t AppendSyntheticBranchBlock(Graph& graph, uint32_t target) {
+	const auto* target_block = graph.FindBlock(target);
 
 	BasicBlock block;
 	block.id                    = static_cast<uint32_t>(graph.blocks.size());
-	block.start_pc              = merge != nullptr ? merge->start_pc : 0u;
+	block.start_pc              = target_block != nullptr ? target_block->start_pc : 0u;
 	block.end_pc                = block.start_pc;
-	block.inst_begin            = merge != nullptr ? merge->inst_begin : 0u;
+	block.inst_begin            = target_block != nullptr ? target_block->inst_begin : 0u;
 	block.inst_end              = block.inst_begin;
-	block.successors            = {old_merge};
+	block.successors            = {target};
 	block.terminator.kind       = TerminatorKind::Branch;
 	block.terminator.condition  = BranchCondition::Always;
-	block.terminator.true_block = old_merge;
+	block.terminator.true_block = target;
 	graph.blocks.push_back(std::move(block));
 	return graph.blocks.back().id;
 }
@@ -897,15 +897,50 @@ bool IsSyntheticMergeForwarder(const Graph& graph, uint32_t block_id, uint32_t m
 	       block->terminator.true_block == merge;
 }
 
-bool IsInsideLoopConstruct(const Graph& graph, const NaturalLoop& loop, uint32_t block_id) {
-	return block_id != UINT32_MAX && block_id != loop.merge && block_id != loop.continue_block &&
-	       graph.Dominates(loop.header, block_id) &&
-	       (loop.merge == UINT32_MAX || !graph.Dominates(loop.merge, block_id));
+const NaturalLoop* FindInnermostContainingLoop(const Graph& graph, uint32_t block_id) {
+	const NaturalLoop* innermost = nullptr;
+	for (const auto& loop: graph.natural_loops) {
+		if (Contains(loop.body_blocks, block_id) &&
+		    (innermost == nullptr || loop.body_blocks.size() < innermost->body_blocks.size())) {
+			innermost = &loop;
+		}
+	}
+	return innermost;
 }
 
-bool SelectionMergeLeavesContainingLoop(const Graph& graph, uint32_t header, uint32_t merge) {
+bool IsInsideLoopConstruct(const Graph& graph, const NaturalLoop& loop, uint32_t block_id) {
+	return block_id != UINT32_MAX && block_id != loop.merge && block_id != loop.continue_block &&
+	       graph.Dominates(loop.header, block_id) && !graph.Dominates(loop.merge, block_id);
+}
+
+bool IsInnermostLoopControlConditional(const Graph& graph, const BasicBlock& block) {
+	if (block.terminator.kind != TerminatorKind::ConditionalBranch) {
+		return false;
+	}
+	const auto* loop = FindInnermostContainingLoop(graph, block.id);
+	if (loop == nullptr || loop->merge == UINT32_MAX || loop->continue_block == UINT32_MAX) {
+		return false;
+	}
+	const auto true_target  = block.terminator.true_block;
+	const auto false_target = block.terminator.false_block;
+	if (block.id == loop->continue_block) {
+		const auto is_repeat_target = [&](uint32_t target) {
+			return target == loop->header || target == loop->merge;
+		};
+		return is_repeat_target(true_target) && is_repeat_target(false_target);
+	}
+	const auto is_control_target = [&](uint32_t target) {
+		return target == loop->merge || target == loop->continue_block;
+	};
+	return (is_control_target(true_target) &&
+	        (is_control_target(false_target) ||
+	         IsInsideLoopConstruct(graph, *loop, false_target))) ||
+	       (is_control_target(false_target) && IsInsideLoopConstruct(graph, *loop, true_target));
+}
+
+bool MergeLeavesContainingLoop(const Graph& graph, uint32_t header, uint32_t merge) {
 	for (const auto& loop: graph.natural_loops) {
-		if (IsInsideLoopConstruct(graph, loop, header) &&
+		if (loop.header != header && IsInsideLoopConstruct(graph, loop, header) &&
 		    !IsInsideLoopConstruct(graph, loop, merge)) {
 			return true;
 		}
@@ -913,6 +948,80 @@ bool SelectionMergeLeavesContainingLoop(const Graph& graph, uint32_t header, uin
 	return false;
 }
 
+bool CanonicalizeNaturalLoops(Graph& graph, std::string* error) {
+	const auto rewrite_budget = graph.blocks.size() * 2u + 16u;
+	for (size_t rewrite = 0; rewrite < rewrite_budget; rewrite++) {
+		bool changed = false;
+		for (const auto& loop: graph.natural_loops) {
+			std::vector<uint32_t> latches;
+			for (const auto& edge: graph.back_edges) {
+				if (edge.to == loop.header) {
+					AddUnique(latches, edge.from);
+				}
+			}
+			if (latches.size() <= 1u) {
+				continue;
+			}
+
+			const auto continue_block = AppendSyntheticBranchBlock(graph, loop.header);
+			for (auto latch: latches) {
+				auto* block = graph.FindBlock(latch);
+				if (block != nullptr) {
+					ReplaceValue(block->successors, loop.header, continue_block);
+					ReplaceTerminatorTarget(block->terminator, loop.header, continue_block);
+				}
+			}
+			RebuildPredecessors(graph);
+			RecomputeAnalyses(graph);
+			changed = true;
+			break;
+		}
+		if (changed) {
+			continue;
+		}
+
+		for (const auto& loop: graph.natural_loops) {
+			const auto* header                 = graph.FindBlock(loop.header);
+			const auto  is_loop_control_target = [&](uint32_t target) {
+				return target == loop.merge || target == loop.continue_block;
+			};
+			if (header == nullptr || header->terminator.kind != TerminatorKind::ConditionalBranch ||
+			    is_loop_control_target(header->terminator.true_block) ||
+			    is_loop_control_target(header->terminator.false_block) ||
+			    !Contains(loop.body_blocks, header->terminator.true_block) ||
+			    !Contains(loop.body_blocks, header->terminator.false_block)) {
+				continue;
+			}
+
+			const auto old_header   = loop.header;
+			const auto predecessors = header->predecessors;
+			const auto new_header   = AppendSyntheticBranchBlock(graph, old_header);
+			for (auto pred: predecessors) {
+				auto* block = graph.FindBlock(pred);
+				if (block != nullptr) {
+					ReplaceValue(block->successors, old_header, new_header);
+					ReplaceTerminatorTarget(block->terminator, old_header, new_header);
+				}
+			}
+			if (graph.entry_block == old_header) {
+				graph.entry_block = new_header;
+			}
+			MoveBlockBefore(graph, new_header, old_header);
+			RebuildPredecessors(graph);
+			RecomputeAnalyses(graph);
+			changed = true;
+			break;
+		}
+		if (!changed) {
+			return true;
+		}
+	}
+
+	SetFailure(graph, FailureKind::StructuredControlFlow, graph.entry_block,
+	           "CFG loop canonicalization exceeded rewrite budget", error);
+	return false;
+}
+
 bool SplitSharedMergeBlock(Graph& graph, uint32_t merge,
                            const std::vector<uint32_t>& construct_blocks,
                            bool                         force_split = false) {
@@ -948,7 +1057,7 @@ bool SplitSharedMergeBlock(Graph& graph, uint32_t merge,
 		return false;
 	}
 
-	const auto synthetic_merge = AppendSyntheticMergeBlock(graph, merge);
+	const auto synthetic_merge = AppendSyntheticBranchBlock(graph, merge);
 	auto*      synthetic_block = graph.FindBlock(synthetic_merge);
 	if (synthetic_block != nullptr) {
 		synthetic_block->predecessors = predecessors_to_split;
@@ -980,14 +1089,111 @@ bool SplitSharedMergeBlock(Graph& graph, uint32_t merge,
 bool SplitOneLoopMerge(Graph& graph) {
 	const auto& loops = graph.natural_loops;
 	for (const auto& loop: loops) {
-		if (SplitSharedMergeBlock(graph, loop.merge, loop.body_blocks)) {
+		const auto construct_blocks = DominatedBlocks(graph, loop.header, loop.merge);
+		const auto force_split      = MergeLeavesContainingLoop(graph, loop.header, loop.merge);
+		if (SplitSharedMergeBlock(graph, loop.merge, construct_blocks, force_split)) {
 			return true;
 		}
 	}
 	return false;
 }
 
-bool SplitOneSelectionMerge(Graph& graph) {
+std::vector<uint32_t> SelectionRegion(const Graph& graph, const BasicBlock& header,
+                                      uint32_t merge) {
+	std::vector<uint32_t> region;
+	std::vector<uint32_t> pending = {header.terminator.true_block,
+	                                 header.terminator.false_block};
+	while (!pending.empty()) {
+		const auto block_id = pending.back();
+		pending.pop_back();
+		if (block_id == merge || Contains(region, block_id)) {
+			continue;
+		}
+		const auto* block = graph.FindBlock(block_id);
+		if (block == nullptr) {
+			continue;
+		}
+		AddUnique(region, block_id);
+		pending.insert(pending.end(), block->successors.begin(), block->successors.end());
+	}
+	SortUnique(region);
+	return region;
+}
+
+bool DuplicateSelectionRegion(Graph& graph, uint32_t header_id, uint32_t merge,
+                              const std::vector<uint32_t>& region, uint32_t block_budget) {
+	std::vector<uint32_t> cloned_blocks;
+	for (auto block_id: region) {
+		if (!graph.Dominates(header_id, block_id)) {
+			cloned_blocks.push_back(block_id);
+		}
+	}
+	if (cloned_blocks.empty() || graph.FindBlock(header_id) == nullptr || header_id >= merge ||
+	    graph.blocks.size() + cloned_blocks.size() + 1u > block_budget) {
+		return false;
+	}
+
+	const auto first_clone = static_cast<uint32_t>(graph.blocks.size());
+	std::map<uint32_t, uint32_t> clones;
+	for (uint32_t i = 0; i < cloned_blocks.size(); i++) {
+		clones.emplace(cloned_blocks[i], first_clone + i);
+	}
+
+	for (auto block_id: cloned_blocks) {
+		BasicBlock clone = *graph.FindBlock(block_id);
+		clone.id         = clones.at(block_id);
+		clone.predecessors.clear();
+		clone.dominators.clear();
+		clone.post_dominators.clear();
+		graph.blocks.push_back(std::move(clone));
+	}
+
+	const auto remap_block = [&](BasicBlock& block) {
+		const auto remap_target = [&](uint32_t& target) {
+			if (const auto it = clones.find(target); it != clones.end()) {
+				target = it->second;
+			}
+		};
+		for (auto& successor: block.successors) {
+			remap_target(successor);
+		}
+		remap_target(block.terminator.true_block);
+		remap_target(block.terminator.false_block);
+		remap_target(block.terminator.merge_block);
+		remap_target(block.terminator.continue_block);
+		for (auto& target: block.terminator.indirect_targets) {
+			remap_target(target);
+		}
+	};
+	for (auto block_id: region) {
+		const auto owned_id = clones.contains(block_id) ? clones.at(block_id) : block_id;
+		remap_block(*graph.FindBlock(owned_id));
+	}
+
+	const auto private_merge = AppendSyntheticBranchBlock(graph, merge);
+	auto&      header        = *graph.FindBlock(header_id);
+	remap_block(header);
+
+	for (auto block_id: region) {
+		const auto owned_id = clones.contains(block_id) ? clones.at(block_id) : block_id;
+		auto*      block    = graph.FindBlock(owned_id);
+		if (block != nullptr) {
+			ReplaceValue(block->successors, merge, private_merge);
+			ReplaceTerminatorTarget(block->terminator, merge, private_merge);
+		}
+	}
+	ReplaceValue(header.successors, merge, private_merge);
+	ReplaceTerminatorTarget(header.terminator, merge, private_merge);
+
+	for (uint32_t i = 0; i <= cloned_blocks.size(); i++) {
+		MoveBlockBefore(graph, first_clone + i, merge + i);
+	}
+	RebuildPredecessors(graph);
+	RecomputeAnalyses(graph);
+	return true;
+}
+
+bool SplitOneSelectionMerge(Graph& graph, uint32_t block_budget) {
 	std::vector<uint32_t> loop_headers;
 	loop_headers.reserve(graph.natural_loops.size());
 	for (const auto& loop: graph.natural_loops) {
@@ -1001,11 +1207,26 @@ bool SplitOneSelectionMerge(Graph& graph) {
 		    Contains(loop_headers, block_id)) {
 			continue;
 		}
+		if (IsInnermostLoopControlConditional(graph, *block)) {
+			continue;
+		}
 
 		const auto merge = graph.FindNearestCommonPostDominator(block->terminator.true_block,
 		                                                        block->terminator.false_block);
+		if (merge == UINT32_MAX || graph.FindBlock(merge) == nullptr) {
+			continue;
+		}
+		const auto region = SelectionRegion(graph, *block, merge);
+		if (std::any_of(region.begin(), region.end(),
+		                [&](uint32_t member) { return !graph.Dominates(block_id, member); })) {
+			if (graph.natural_loops.empty() &&
+			    DuplicateSelectionRegion(graph, block_id, merge, region, block_budget)) {
+				return true;
+			}
+			continue;
+		}
 		const auto construct_blocks = DominatedBlocks(graph, block_id, merge);
-		const auto force_split      = SelectionMergeLeavesContainingLoop(graph, block_id, merge);
+		const auto force_split      = MergeLeavesContainingLoop(graph, block_id, merge);
 		if (SplitSharedMergeBlock(graph, merge, construct_blocks, force_split)) {
 			return true;
 		}
@@ -1015,10 +1236,12 @@ bool SplitOneSelectionMerge(Graph& graph) {
 
 bool SplitSharedMergeBlocks(Graph& graph, std::string* error) {
 	const auto original_block_count = static_cast<uint32_t>(graph.blocks.size());
-	const auto split_budget =
-	    std::max<uint32_t>(16u, std::min<uint32_t>(128u, original_block_count));
+	const auto split_budget = std::max<uint32_t>(
+	    16u, std::min<uint32_t>(128u, original_block_count * 4u));
+	const auto block_budget = std::max<uint32_t>(
+	    32u, std::min<uint32_t>(512u, original_block_count * 8u));
 	for (uint32_t splits = 0; splits < split_budget; splits++) {
-		if (!SplitOneLoopMerge(graph) && !SplitOneSelectionMerge(graph)) {
+		if (!SplitOneLoopMerge(graph) && !SplitOneSelectionMerge(graph, block_budget)) {
 			return true;
 		}
 		RebuildPredecessors(graph);
@@ -1353,6 +1576,9 @@ bool Structurize(Graph& graph, std::string* error) {
 		return false;
 	}
 
+	if (!CanonicalizeNaturalLoops(graph, error)) {
+		return false;
+	}
 	if (!SplitSharedMergeBlocks(graph, error)) {
 		return false;
 	}
@@ -1395,6 +1621,9 @@ bool Structurize(Graph& graph, std::string* error) {
 		    block.terminator.loop_header) {
 			continue;
 		}
+		if (IsInnermostLoopControlConditional(graph, block)) {
+			continue;
+		}
 
 		const auto merge = graph.FindNearestCommonPostDominator(block.terminator.true_block,
 		                                                        block.terminator.false_block);

src/graphics/shader/recompiler/emitter/SpirvEmitter.cpp

@@ -237,8 +237,14 @@ bool ValidateNativeProgram(const IR::Program& program, std::string* error) {
 		if (!ImageBinding(program.info.images[i], kind)) {
 			return Fail(error, "native shader plan has an invalid image class");
 		}
+		const auto bindings = program.info.images[i].NumBindings();
+		if (bindings == 0 || bindings > IR::ImageResource::MaxMipLevels) {
+			return Fail(error, "native shader plan has an invalid image descriptor count");
+		}
 		present[static_cast<size_t>(kind)] = true;
-		expected[static_cast<size_t>(kind)].push_back(i);
+		for (uint32_t binding = 0; binding < bindings; binding++) {
+			expected[static_cast<size_t>(kind)].push_back(i);
+		}
 	}
 	if (!program.info.samplers.empty()) {
 		Expect(Kind::Samplers, Dense(program.info.samplers.size()));
@@ -316,6 +322,11 @@ bool ValidateNativeProgram(const IR::Program& program, std::string* error) {
 			         inst.memory.image_dimension)) {
 				return Fail(error, "image instruction has an invalid dense resource");
 			}
+			if (inst.op == IR::Opcode::ImageStore &&
+			    ((program.info.images[inst.memory.resource].mip_mode ==
+			      IR::ImageMipMode::DynamicStorage) != inst.memory.image_has_mip)) {
+				return Fail(error, "storage image mip mode does not match the instruction");
+			}
 			const bool address =
 			    inst.op == IR::Opcode::SLoadDword || memory == IR::ResourceKind::Flat ||
 			    memory == IR::ResourceKind::Global || memory == IR::ResourceKind::Scratch;

src/graphics/shader/recompiler/emitter/spirvEmitterAnalysis.cpp

@@ -560,12 +560,21 @@ uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
                                   uint32_t variable_id, uint32_t array_index,
                                   IR::DescriptorBindingKind kind, uint32_t resource,
                                   const char* variable_name) {
+	return DescriptorElementPointerId(state, result_ptr_type, variable_id,
+	                                  ConstantU32(state, array_index), kind, resource,
+	                                  variable_name);
+}
+
+uint32_t DescriptorElementPointerId(EmitterState& state, uint32_t result_ptr_type,
+                                    uint32_t variable_id, uint32_t array_index_id,
+                                    IR::DescriptorBindingKind kind, uint32_t resource,
+                                    const char* variable_name) {
 	if (variable_id == 0) {
 		ExitDescriptorBindingFailure(state, kind, resource, variable_name);
 	}
 	const auto pointer = state.builder.AllocateId();
 	state.builder.AddFunction(
-	    {OpAccessChain, result_ptr_type, pointer, variable_id, ConstantU32(state, array_index)});
+	    {OpAccessChain, result_ptr_type, pointer, variable_id, array_index_id});
 	return pointer;
 }
 

src/graphics/shader/recompiler/emitter/spirvEmitterImageOps.cpp

@@ -33,13 +33,13 @@ uint32_t ConstantImageGatherHorizontalOffsets(EmitterState& state, ImageViewKind
 }
 
 uint32_t LoadStorageImageDescriptorAtIndex(EmitterState& state, uint32_t resource,
-                                           uint32_t array_index, bool uint_image,
+                                           uint32_t array_index_id, bool uint_image,
                                            ImageViewKind view) {
 	const auto  kind        = StorageBindingKind(uint_image, view);
 	const auto& descriptors = state.storage_images[StorageImageIndex(uint_image, view)];
-	const auto  pointer =
-	    DescriptorElementPointer(state, descriptors.pointer_type, descriptors.variable, array_index,
-	                             kind, resource, "storage image descriptor array was not emitted");
+	const auto  pointer     = DescriptorElementPointerId(
+	    state, descriptors.pointer_type, descriptors.variable, array_index_id, kind, resource,
+	    "storage image descriptor array was not emitted");
 	const auto image = state.builder.AllocateId();
 	state.builder.AddFunction({OpLoad, descriptors.image_type, image, pointer});
 	return image;
@@ -168,12 +168,38 @@ void EmitImageStore(EmitterState& state, const IR::Instruction& inst) {
 	const auto view       = StorageImageViewKind(state, inst.memory, uint_image, inst.pc);
 	const auto binding =
 	    ResourceForDescriptor(state, StorageBindingKind(uint_image, view), inst.memory.resource);
-	const auto image = LoadStorageImageDescriptorAtIndex(state, inst.memory.resource,
-	                                                     binding.array_index, uint_image, view);
+	const auto emit_write = [&](uint32_t descriptor_index, bool non_uniform) {
+		if (non_uniform) {
+			state.builder.AddAnnotation({OpDecorate, descriptor_index, DecorationNonUniform});
+		}
+		const auto image = LoadStorageImageDescriptorAtIndex(state, inst.memory.resource,
+		                                                     descriptor_index, uint_image, view);
+		if (non_uniform) {
+			state.builder.AddAnnotation({OpDecorate, image, DecorationNonUniform});
+		}
+		state.builder.AddFunction({OpImageWrite, image, EmitImageCoordU32(state, inst, view),
+		                           uint_image ? EmitImageStoreTexelU32(state, inst)
+		                                      : EmitImageStoreTexelF32(state, inst)});
+	};
+	if (!inst.memory.image_has_mip) {
+		emit_write(ConstantU32(state, binding.array_index), false);
+		return;
+	}
 
+	const auto& resource = state.program.info.images[inst.memory.resource];
+	const auto  mip      = EmitImageMipLodU32(state, inst, inst.src[1], view);
+	const auto  in_range = state.builder.AllocateId();
 	state.builder.AddFunction(
-	    {OpImageWrite, image, EmitImageCoordU32(state, inst, view),
-	     uint_image ? EmitImageStoreTexelU32(state, inst) : EmitImageStoreTexelF32(state, inst)});
+	    {OpULessThan, state.bool_type, in_range, mip, ConstantU32(state, resource.mip_levels)});
+	EmitIfCondition(state, in_range, [&] {
+		auto descriptor_index = mip;
+		if (binding.array_index != 0) {
+			descriptor_index = state.builder.AllocateId();
+			state.builder.AddFunction({OpIAdd, state.uint_type, descriptor_index,
+			                           ConstantU32(state, binding.array_index), mip});
+		}
+		emit_write(descriptor_index, true);
+	});
 }
 
 void EmitImageSampleResult(EmitterState& state, const IR::Instruction& inst, uint32_t sample,

src/graphics/shader/recompiler/emitter/spirvEmitterInternal.h

@@ -23,38 +23,40 @@
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 
 enum : uint32_t {
-	ExecutionModelVertex                     = 0,
-	ExecutionModelFragment                   = 4,
-	ExecutionModelGLCompute                  = 5,
-	ExecutionModeOriginUpperLeft             = 7,
-	ExecutionModeEarlyFragmentTests          = 9,
-	ExecutionModeDepthReplacing              = 12,
-	ExecutionModeLocalSize                   = 17,
-	ExecutionModeDerivativeGroupQuadsKHR     = 5289,
-	AddressingModelLogical                   = 0,
-	MemoryModelGLSL450                       = 1,
-	CapabilityShader                         = 1,
-	CapabilityImageGatherExtended            = 25,
-	CapabilitySampled1D                      = 43,
-	CapabilityImage1D                        = 44,
-	CapabilityImageQuery                     = 50,
-	CapabilityStorageImageReadWithoutFormat  = 55,
-	CapabilityStorageImageWriteWithoutFormat = 56,
-	CapabilityGroupNonUniform                = 61,
-	CapabilityGroupNonUniformBallot          = 64,
-	CapabilityGroupNonUniformShuffle         = 65,
-	CapabilityComputeDerivativeGroupQuadsKHR = 5288,
-	StorageClassUniformConstant              = 0,
-	StorageClassInput                        = 1,
-	StorageClassOutput                       = 3,
-	StorageClassWorkgroup                    = 4,
-	StorageClassFunction                     = 7,
-	StorageClassPushConstant                 = 9,
-	StorageClassImage                        = 11,
-	StorageClassStorageBuffer                = 12,
-	FunctionControlNone                      = 0,
-	SelectionControlNone                     = 0,
-	LoopControlNone                          = 0,
+	ExecutionModelVertex                          = 0,
+	ExecutionModelFragment                        = 4,
+	ExecutionModelGLCompute                       = 5,
+	ExecutionModeOriginUpperLeft                  = 7,
+	ExecutionModeEarlyFragmentTests               = 9,
+	ExecutionModeDepthReplacing                   = 12,
+	ExecutionModeLocalSize                        = 17,
+	ExecutionModeDerivativeGroupQuadsKHR          = 5289,
+	AddressingModelLogical                        = 0,
+	MemoryModelGLSL450                            = 1,
+	CapabilityShader                              = 1,
+	CapabilityImageGatherExtended                 = 25,
+	CapabilitySampled1D                           = 43,
+	CapabilityImage1D                             = 44,
+	CapabilityImageQuery                          = 50,
+	CapabilityStorageImageReadWithoutFormat       = 55,
+	CapabilityStorageImageWriteWithoutFormat      = 56,
+	CapabilityGroupNonUniform                     = 61,
+	CapabilityGroupNonUniformBallot               = 64,
+	CapabilityGroupNonUniformShuffle              = 65,
+	CapabilityShaderNonUniform                    = 5301,
+	CapabilityStorageImageArrayNonUniformIndexing = 5309,
+	CapabilityComputeDerivativeGroupQuadsKHR      = 5288,
+	StorageClassUniformConstant                   = 0,
+	StorageClassInput                             = 1,
+	StorageClassOutput                            = 3,
+	StorageClassWorkgroup                         = 4,
+	StorageClassFunction                          = 7,
+	StorageClassPushConstant                      = 9,
+	StorageClassImage                             = 11,
+	StorageClassStorageBuffer                     = 12,
+	FunctionControlNone                           = 0,
+	SelectionControlNone                          = 0,
+	LoopControlNone                               = 0,
 };
 
 enum : uint32_t {
@@ -67,6 +69,7 @@ enum : uint32_t {
 	DecorationBinding       = 33,
 	DecorationDescriptorSet = 34,
 	DecorationOffset        = 35,
+	DecorationNonUniform    = 5300,
 };
 
 enum : uint32_t {
@@ -679,6 +682,10 @@ uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
                                   uint32_t variable_id, uint32_t array_index,
                                   IR::DescriptorBindingKind kind, uint32_t resource,
                                   const char* variable_name);
+uint32_t DescriptorElementPointerId(EmitterState& state, uint32_t result_ptr_type,
+                                    uint32_t variable_id, uint32_t array_index_id,
+                                    IR::DescriptorBindingKind kind, uint32_t resource,
+                                    const char* variable_name);
 
 ImageViewKind SampledImageViewKind(const EmitterState& state, const IR::MemoryInfo& mem,
                                    uint32_t use_pc);

src/graphics/shader/recompiler/emitter/spirvEmitterModule.cpp

@@ -484,6 +484,13 @@ void EmitHeaderAndTypes(EmitterState& state) {
 	if (state.needs_image_gather_extended) {
 		state.builder.AddCapability({CapabilityImageGatherExtended});
 	}
+	if (std::any_of(
+	        state.program.info.images.begin(), state.program.info.images.end(),
+	        [](const auto& image) { return image.mip_mode == IR::ImageMipMode::DynamicStorage; })) {
+		state.builder.AddCapability({CapabilityShaderNonUniform});
+		state.builder.AddCapability({CapabilityStorageImageArrayNonUniformIndexing});
+		state.builder.AddExtension("SPV_EXT_descriptor_indexing");
+	}
 	if (std::any_of(state.storage_images.begin(),
 	                state.storage_images.begin() + StorageImageViewKindCount,
 	                [](const auto& image) { return image.variable != 0; })) {

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

@@ -269,7 +269,10 @@ bool AllocateBindings(Program& program, const BindingLayoutOptions& options, std
 			}
 			return false;
 		}
-		image_groups[static_cast<size_t>(group - ImageBindingKinds.begin())].push_back(i);
+		auto& resources = image_groups[static_cast<size_t>(group - ImageBindingKinds.begin())];
+		for (uint32_t binding = 0; binding < program.info.images[i].NumBindings(); binding++) {
+			resources.push_back(i);
+		}
 	}
 	for (uint32_t i = 0; i < image_groups.size(); i++) {
 		if (!image_groups[i].empty()) {

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

@@ -71,6 +71,17 @@ bool DescriptorIsCube(const DescriptorValue& descriptor) {
 	       Prospero::ImageType::kCube;
 }
 
+bool DescriptorMipRange(const DescriptorValue& descriptor, uint32_t& count) {
+	const auto base_level = (descriptor.dwords[3] >> 12u) & 0xfu;
+	const auto last_level = (descriptor.dwords[3] >> 16u) & 0xfu;
+	const auto max_mip    = (descriptor.dwords[5] >> 4u) & 0xfu;
+	if (base_level > last_level || last_level > max_mip) {
+		return false;
+	}
+	count = last_level - base_level + 1u;
+	return true;
+}
+
 bool DecodeBufferDescriptor(const DescriptorValue& descriptor, ShaderBufferResource& result) {
 	if (descriptor.dword_count != std::size(result.fields)) {
 		return false;
@@ -196,16 +207,31 @@ bool ValidateResourceSpecialization(const Program& program, const ResourceSnapsh
 			}
 			continue;
 		}
+		if (image.mip_mode == ImageMipMode::DynamicStorage) {
+			uint32_t mip_levels = 0;
+			if (!DescriptorMipRange(descriptor, mip_levels) || mip_levels != image.mip_levels) {
+				if (error != nullptr) {
+					*error = fmt::format(
+					    "image descriptor {} no longer matches specialized storage mip count", i);
+				}
+				return false;
+			}
+		} else if (image.mip_levels != 1u) {
+			if (error != nullptr) {
+				*error = fmt::format("image descriptor {} has invalid non-storage mip count", i);
+			}
+			return false;
+		}
 		const auto dimension = DescriptorDimension(descriptor, image.dimension);
 		if (dimension == Decoder::ImageDimension::Unknown || dimension != image.dimension ||
 		    DescriptorIsCube(descriptor) != image.cube) {
 			if (error != nullptr) {
-				*error = fmt::format(
-				    "image descriptor {} no longer matches specialized dimension: "
-				    "{:08x},{:08x},{:08x},{:08x},{:08x},{:08x},{:08x},{:08x}",
-				    i, descriptor.dwords[0], descriptor.dwords[1], descriptor.dwords[2],
-				    descriptor.dwords[3], descriptor.dwords[4], descriptor.dwords[5],
-				    descriptor.dwords[6], descriptor.dwords[7]);
+				*error =
+				    fmt::format("image descriptor {} no longer matches specialized dimension: "
+				                "{:08x},{:08x},{:08x},{:08x},{:08x},{:08x},{:08x},{:08x}",
+				                i, descriptor.dwords[0], descriptor.dwords[1], descriptor.dwords[2],
+				                descriptor.dwords[3], descriptor.dwords[4], descriptor.dwords[5],
+				                descriptor.dwords[6], descriptor.dwords[7]);
 			}
 			return false;
 		}
@@ -224,10 +250,9 @@ bool ValidateResourceSpecialization(const Program& program, const ResourceSnapsh
 			const bool raw_sint_storage =
 			    storage && format == Prospero::GpuEnumValue(Prospero::BufferFormat::k32SInt) &&
 			    !image.read && !image.atomic;
-			const bool uint_descriptor =
-			    Prospero::IsUintTextureFormat(format) || raw_sint_storage;
-			const auto uint_program = image.kind == ResourceKind::ImageUint ||
-			                          image.kind == ResourceKind::StorageImageUint;
+			const bool uint_descriptor = Prospero::IsUintTextureFormat(format) || raw_sint_storage;
+			const auto uint_program    = image.kind == ResourceKind::ImageUint ||
+			                             image.kind == ResourceKind::StorageImageUint;
 			if (uint_descriptor != uint_program && !(image.atomic && uint_program)) {
 				if (error != nullptr) {
 					*error =
@@ -380,8 +405,9 @@ bool SpecializeResources(Program& program, const ResourceSnapshot& snapshot, std
 		const auto& descriptor = snapshot.images[i];
 		auto&       image      = next.images[i];
 		if (NullImageDescriptor(descriptor)) {
-			image.dimension = Decoder::ImageDimension::Dim2D;
-			image.cube      = false;
+			image.dimension  = Decoder::ImageDimension::Dim2D;
+			image.cube       = false;
+			image.mip_levels = 1;
 			switch (image.kind) {
 				case ResourceKind::ImageUint: image.kind = ResourceKind::Image; break;
 				case ResourceKind::StorageImageUint:
@@ -393,6 +419,16 @@ bool SpecializeResources(Program& program, const ResourceSnapshot& snapshot, std
 			}
 			continue;
 		}
+		if (image.mip_mode == ImageMipMode::DynamicStorage) {
+			if (!DescriptorMipRange(descriptor, image.mip_levels)) {
+				if (error != nullptr) {
+					*error = fmt::format("image descriptor {} has invalid storage mip range", i);
+				}
+				return false;
+			}
+		} else {
+			image.mip_levels = 1;
+		}
 		const auto descriptor_dimension = DescriptorDimension(descriptor, image.dimension);
 		if (descriptor_dimension == Decoder::ImageDimension::Unknown) {
 			if (error != nullptr) {

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

@@ -45,6 +45,8 @@ namespace {
 
 constexpr uint32_t ScalarRegisters = 128;
 constexpr uint32_t VectorRegisters = 256;
+// Clamp X/Y/Z are consecutive three-bit fields; the high bit of each selects a border mode.
+constexpr uint32_t SamplerBorderClampMask = (1u << 2u) | (1u << 5u) | (1u << 8u);
 
 struct ScalarState {
 	std::array<uint32_t, ScalarRegisters> regs          = {};
@@ -647,6 +649,24 @@ private:
 		return AddDescriptor(descriptor);
 	}
 
+	uint32_t AddSamplerDescriptor(const ScalarState& state, uint32_t base) {
+		if (base >= ScalarRegisters || 4u > ScalarRegisters - base) {
+			return ScalarProvenance::Unknown;
+		}
+		DescriptorValue descriptor;
+		descriptor.dword_count = 4;
+		for (uint32_t i = 0; i < 4; i++) {
+			descriptor.dwords[i] = state.regs[base + i];
+		}
+		if (auto d0 = descriptor.dwords[0];
+		    d0 < m_graph.values.size() && m_graph.values[d0].op == ScalarValueOp::Constant &&
+		    (m_graph.values[d0].imm & SamplerBorderClampMask) == 0) {
+			// Without a border clamp, the border color and table index in dword 3 are unused.
+			descriptor.dwords[3] = Constant(0);
+		}
+		return AddDescriptor(descriptor);
+	}
+
 	uint32_t AddFlatAddressDescriptor(const Instruction& inst, const ScalarState& state) {
 		const uint32_t first = FlatStore(inst.op) ? 1u : 0u;
 		if (inst.src_count < first + 2u) {
@@ -704,7 +724,7 @@ private:
 			inst.memory.resource_source = AddDescriptor(state, inst.memory.resource * 4u, 8);
 			if (inst.op == Opcode::ImageSample || inst.op == Opcode::ImageGather4 ||
 			    inst.op == Opcode::ImageGetLod) {
-				inst.memory.sampler_source = AddDescriptor(state, inst.memory.sampler * 4u, 4);
+				inst.memory.sampler_source = AddSamplerDescriptor(state, inst.memory.sampler * 4u);
 			}
 		}
 	}

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

@@ -598,11 +598,14 @@ enum class ImageMipMode { None, DynamicStorage };
 constexpr uint32_t StorageImageIdentitySwizzle = 0x00000facu;
 
 struct ImageResource {
+	static constexpr uint32_t MaxMipLevels = 16;
+
 	uint32_t                source          = 0;
 	uint32_t                first_use_pc    = 0;
 	ResourceKind            kind            = ResourceKind::None;
 	Decoder::ImageDimension dimension       = Decoder::ImageDimension::Unknown;
 	ImageMipMode            mip_mode        = ImageMipMode::None;
+	uint32_t                mip_levels      = 1;
 	uint32_t                storage_swizzle = StorageImageIdentitySwizzle;
 	bool                    read            = false;
 	bool                    written         = false;
@@ -610,6 +613,10 @@ struct ImageResource {
 	bool                    depth_compare   = false;
 	bool                    cube            = false;
 
+	[[nodiscard]] uint32_t NumBindings() const {
+		return mip_mode == ImageMipMode::DynamicStorage ? mip_levels : 1u;
+	}
+
 	bool operator==(const ImageResource& other) const = default;
 };
 

src/graphics/shader/shader.cpp

@@ -1018,6 +1018,7 @@ static void ShaderAppendNativeSpecialization(std::vector<uint32_t>&
 		ids.push_back(static_cast<uint32_t>(image.kind));
 		ids.push_back(static_cast<uint32_t>(image.dimension));
 		ids.push_back(static_cast<uint32_t>(image.mip_mode));
+		ids.push_back(image.mip_levels);
 		ids.push_back(image.storage_swizzle);
 	}
 	ids.push_back(static_cast<uint32_t>(program.info.addresses.size()));

src/kernel/memoryAddressSpace.inc

@@ -1058,6 +1058,76 @@ private:
 		    {HOST_SYSTEM_RESERVED_MIN, HOST_SYSTEM_RESERVED_MAX + 1u},
 		    {HOST_USER_MIN, HOST_USER_MAX + 1u},
 		}};
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
+		std::vector<std::pair<uint64_t, uint64_t>> occupied;
+		FILE*                                      maps = fopen("/proc/self/maps", "r");
+		EXIT_IF(maps == nullptr);
+
+		char line[512];
+		while (fgets(line, sizeof(line), maps) != nullptr) {
+			unsigned long long mapping_start = 0;
+			unsigned long long mapping_end   = 0;
+			if (sscanf(line, "%llx-%llx", &mapping_start, &mapping_end) == 2) {
+				occupied.emplace_back(static_cast<uint64_t>(mapping_start),
+				                      static_cast<uint64_t>(mapping_end));
+			}
+		}
+		fclose(maps);
+
+		auto reserve_range = [this](uint64_t start, uint64_t end) {
+			start = AlignUp(start, PageSize());
+			end   = AlignDown(end, PageSize());
+			if (start == 0 || end <= start) {
+				return;
+			}
+
+			const auto size = end - start;
+			int        flags = MAP_PRIVATE | MAP_ANON | MAP_NORESERVE;
+#if defined(KYTY_LINKED_GUEST_ADDRESS_SPACE)
+			flags |= MAP_FIXED;
+#elif defined(MAP_FIXED_NOREPLACE)
+			flags |= MAP_FIXED_NOREPLACE;
+#endif
+
+			void* ptr = mmap(reinterpret_cast<void*>(start), size, PROT_NONE, flags, -1, 0);
+			if (ptr == MAP_FAILED || reinterpret_cast<uint64_t>(ptr) != start) {
+				if (ptr != MAP_FAILED) {
+					munmap(ptr, size);
+				}
+				return;
+			}
+
+			AddFreeUnlocked(start, size);
+			m_owned.emplace_back(start, size);
+		};
+
+		for (const auto& [region_start, region_end]: regions) {
+			auto current = region_start;
+
+			for (const auto& [mapping_start, mapping_end]: occupied) {
+				if (mapping_end <= current) {
+					continue;
+				}
+				if (mapping_start >= region_end) {
+					break;
+				}
+
+				if (mapping_start > current) {
+					reserve_range(current, std::min(mapping_start, region_end));
+				}
+
+				current = std::max(current, mapping_end);
+				if (current >= region_end) {
+					break;
+				}
+			}
+
+			if (current < region_end) {
+				reserve_range(current, region_end);
+			}
+		}
+#else
 		for (const auto& [start, end]: regions) {
 			int flags = MAP_PRIVATE | MAP_ANON | MAP_NORESERVE;
 #if defined(KYTY_LINKED_GUEST_ADDRESS_SPACE)
@@ -1070,13 +1140,14 @@ private:
 				if (ptr != MAP_FAILED) {
 					munmap(ptr, end - start);
 				}
-				EXIT("failed to reserve guest address space at 0x%016" PRIx64 ", size 0x%016" PRIx64
-				     "\n",
+				EXIT("failed to reserve guest address space at 0x%016" PRIx64
+				     ", size 0x%016" PRIx64 "\n",
 				     start, end - start);
 			}
 			AddFreeUnlocked(start, end - start);
 			m_owned.emplace_back(start, end - start);
 		}
+#endif
 #endif
 	}
 

src/kernel/pthread.cpp

@@ -65,6 +65,10 @@
 
 namespace Libs {
 
+namespace LibcInternalExt {
+void RunThreadAtexitDestructors();
+} // namespace LibcInternalExt
+
 namespace LibKernel {
 
 LIB_NAME("libkernel", "libkernel");
@@ -3347,6 +3351,8 @@ int PthreadGetCurrentPriorityForKernel() {
 static void CleanupThread(void* arg) {
 	auto* thread = static_cast<Pthread>(arg);
 
+	LibcInternalExt::RunThreadAtexitDestructors();
+
 	auto thread_dtors = g_pthread_context->GetThreadDtors();
 
 	if (thread_dtors != nullptr) {
@@ -3880,8 +3886,12 @@ int KYTY_SYSV_ABI KernelGettimeofday(KernelTimeval* tp) {
 	tp->tv_sec  = static_cast<int64_t>(ticks / 1000000);
 	tp->tv_usec = static_cast<int64_t>(ticks % 1000000);
 #else
-	auto dt = Common::DateTime::FromSystemUTC();
-	sec_to_timeval(tp, dt.ToUnix());
+	struct timespec ts {};
+	result = ::clock_gettime(CLOCK_REALTIME, &ts);
+	if (result == 0) {
+		tp->tv_sec  = static_cast<int64_t>(ts.tv_sec);
+		tp->tv_usec = static_cast<int64_t>(ts.tv_nsec / 1000);
+	}
 #endif
 
 	if (result == 0) {

src/libs/agc.cpp

@@ -221,57 +221,6 @@ static RegisterDefaults* get_internal_register_defaults(uint32_t ver) {
 	return get_register_defaults(g_agc_internal_reg_defaults_by_version[index], &storage[index]);
 }
 
-struct PendingGraphicsSegment {
-	uint32_t* start     = nullptr;
-	uint32_t* end       = nullptr;
-	uint32_t* range_end = nullptr;
-};
-
-static std::mutex             g_pending_graphics_segment_mutex;
-static PendingGraphicsSegment g_pending_graphics_segment;
-
-static void track_pending_graphics_segment_after_submit(uint32_t* dcb, uint32_t size_in_dwords) {
-	if (dcb == nullptr || size_in_dwords == 0) {
-		return;
-	}
-
-	auto* segment_start = dcb + size_in_dwords;
-	auto* range_end     = segment_start + 0xfffffu;
-
-	std::lock_guard lock(g_pending_graphics_segment_mutex);
-	g_pending_graphics_segment.start     = segment_start;
-	g_pending_graphics_segment.end       = segment_start;
-	g_pending_graphics_segment.range_end = range_end;
-}
-
-static void track_pending_graphics_allocation(uint32_t* cmd, uint32_t size_dw) {
-	if (cmd == nullptr || size_dw == 0) {
-		return;
-	}
-
-	std::lock_guard lock(g_pending_graphics_segment_mutex);
-	auto*           range_start = g_pending_graphics_segment.start;
-	auto*           range_end   = g_pending_graphics_segment.range_end;
-	if (range_start == nullptr || range_end == nullptr || cmd < range_start || cmd >= range_end) {
-		return;
-	}
-
-	auto* cmd_end = cmd + size_dw;
-	if (cmd > g_pending_graphics_segment.end) {
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1) < 64) {
-			LOGF("\t pending graphics segment: ignoring non-contiguous allocation cmd = "
-			     "0x%016" PRIx64 ", tracked_end = 0x%016" PRIx64 "\n",
-			     reinterpret_cast<uint64_t>(cmd),
-			     reinterpret_cast<uint64_t>(g_pending_graphics_segment.end));
-		}
-		return;
-	}
-	if (cmd_end > g_pending_graphics_segment.end && cmd_end <= range_end) {
-		g_pending_graphics_segment.end = cmd_end;
-	}
-}
-
 struct CommandBuffer {
 	using Callback = KYTY_SYSV_ABI bool (*)(CommandBuffer*, uint32_t, void*);
 
@@ -370,7 +319,6 @@ struct CommandBuffer {
 		}
 		auto* ret_ptr = cursor_up;
 		cursor_up += size_dw;
-		track_pending_graphics_allocation(ret_ptr, size_dw);
 		return ret_ptr;
 	}
 };
@@ -834,6 +782,104 @@ int KYTY_SYSV_ABI GraphicsUnknownFuseShaderHalves(Shader* fused_result, const Sh
 	return OK;
 }
 
+static void merge_shader_register_max_field(ShaderRegister* dst, const ShaderRegister* src,
+                                            uint32_t shift, uint32_t mask) {
+	const auto dst_field = (dst->value >> shift) & mask;
+	const auto src_field = (src->value >> shift) & mask;
+	const auto field     = std::max(dst_field, src_field);
+
+	dst->value &= ~(mask << shift);
+	dst->value |= field << shift;
+}
+
+int KYTY_SYSV_ABI GraphicsUnknownNApJjpKNBl4(Shader* fused_result, const Shader* front,
+                                             const Shader* back, void* scratch_mem) {
+	PRINT_NAME();
+
+	LOGF("\t fused_result = 0x%016" PRIx64 "\n"
+	     "\t front        = 0x%016" PRIx64 "\n"
+	     "\t back         = 0x%016" PRIx64 "\n"
+	     "\t scratch_mem  = 0x%016" PRIx64 "\n",
+	     reinterpret_cast<uint64_t>(fused_result), reinterpret_cast<uint64_t>(front),
+	     reinterpret_cast<uint64_t>(back), reinterpret_cast<uint64_t>(scratch_mem));
+
+	const auto front_type = static_cast<Prospero::ShaderBinaryType>(front->type);
+	const auto is_gs      = front_type == Prospero::ShaderBinaryType::kGsFront;
+	const auto is_hs      = front_type == Prospero::ShaderBinaryType::kHsFront;
+	if ((!is_gs && !is_hs) ||
+	    (is_gs && back->type != static_cast<uint8_t>(Prospero::ShaderBinaryType::kGsBack)) ||
+	    (is_hs && back->type != static_cast<uint8_t>(Prospero::ShaderBinaryType::kHsBack))) {
+		return GRAPHICS5_ERROR_INVALID_SHADER_HALVES;
+	}
+
+	*fused_result      = *back;
+	fused_result->type = static_cast<uint8_t>(is_gs ? Prospero::ShaderBinaryType::kGs
+	                                                : Prospero::ShaderBinaryType::kHs);
+
+	const auto back_stages  = back->specials->vgt_shader_stages_en.value;
+	const auto front_stages = front->specials->vgt_shader_stages_en.value;
+	const auto mismatch_bit = is_gs ? (1u << 22u) : (1u << 21u);
+	if (((front_stages ^ back_stages) & mismatch_bit) != 0) {
+		return GRAPHICS5_ERROR_INVALID_SHADER_HALVES;
+	}
+
+	if (scratch_mem != nullptr) {
+		auto* sh_registers = static_cast<ShaderRegister*>(scratch_mem);
+		memcpy(sh_registers, back->sh_registers,
+		       static_cast<size_t>(back->num_sh_registers) * sizeof(ShaderRegister));
+		fused_result->sh_registers = sh_registers;
+	}
+
+	auto*      fused_regs      = fused_result->sh_registers;
+	const auto fused_reg_count = static_cast<uint32_t>(fused_result->num_sh_registers);
+	const auto front_reg_count = static_cast<uint32_t>(front->num_sh_registers);
+	const auto checksum_offset =
+	    is_gs ? Pm4::SPI_SHADER_PGM_CHKSUM_GS : Pm4::SPI_SHADER_PGM_CHKSUM_HS;
+	const auto* front_checksum0 =
+	    find_shader_register(front->sh_registers, front_reg_count, checksum_offset, 0);
+	const auto* front_checksum1 =
+	    find_shader_register(front->sh_registers, front_reg_count, checksum_offset, 1);
+	auto* fused_checksum0  = find_shader_register(fused_regs, fused_reg_count, checksum_offset, 0);
+	auto* fused_checksum1  = find_shader_register(fused_regs, fused_reg_count, checksum_offset, 1);
+	fused_checksum0->value = front_checksum0->value;
+	fused_checksum1->value = front_checksum1->value;
+
+	const auto  rsrc1_offset = is_gs ? Pm4::SPI_SHADER_PGM_RSRC1_GS : Pm4::SPI_SHADER_PGM_RSRC1_HS;
+	const auto  rsrc2_offset = is_gs ? Pm4::SPI_SHADER_PGM_RSRC2_GS : Pm4::SPI_SHADER_PGM_RSRC2_HS;
+	const auto* front_rsrc1 =
+	    find_shader_register(front->sh_registers, front_reg_count, rsrc1_offset);
+	const auto* front_rsrc2 =
+	    find_shader_register(front->sh_registers, front_reg_count, rsrc2_offset);
+	auto* fused_rsrc1 = find_shader_register(fused_regs, fused_reg_count, rsrc1_offset);
+	auto* fused_rsrc2 = find_shader_register(fused_regs, fused_reg_count, rsrc2_offset);
+
+	merge_shader_register_max_field(fused_rsrc1, front_rsrc1, 0, 0x3fu);
+	merge_shader_register_max_field(fused_rsrc2, front_rsrc2, 28, 0x0fu);
+	if (is_gs) {
+		merge_shader_register_max_field(fused_rsrc1, front_rsrc1, 29, 0x03u);
+		merge_shader_register_max_field(fused_rsrc2, front_rsrc2, 16, 0x03u);
+		fused_rsrc2->value =
+		    (fused_rsrc2->value & 0xf7ffffc1u) | (front_rsrc2->value & 0x0800003eu);
+		fused_rsrc2->value =
+		    (fused_rsrc2->value & 0xfffbffffu) | (front_rsrc2->value & 0x00040000u);
+	} else {
+		merge_shader_register_max_field(fused_rsrc1, front_rsrc1, 28, 0x03u);
+		fused_rsrc2->value =
+		    (fused_rsrc2->value & 0xf7ffffc1u) | (front_rsrc2->value & 0x0800003eu);
+	}
+
+	const auto program_lo_offset = is_gs ? Pm4::SPI_SHADER_PGM_LO_ES : Pm4::SPI_SHADER_PGM_LO_LS;
+	auto*      program_lo = find_shader_register(fused_regs, fused_reg_count, program_lo_offset);
+	const auto address    = reinterpret_cast<uint64_t>(front->code);
+	program_lo->value     = static_cast<uint32_t>(address >> 8u);
+	(program_lo + 1)->value &= 0xffffff00u;
+	(program_lo + 1)->value |= static_cast<uint32_t>((address >> 40u) & 0xffu);
+
+	fused_result->user_data = front->user_data;
+
+	return OK;
+}
+
 static constexpr int GRAPHICS5_ERROR_INVALID_PACKET = static_cast<int>(0x8a6c000cu);
 
 enum class RegIndirectPacket : uint32_t {
@@ -1347,7 +1393,7 @@ int KYTY_SYSV_ABI GraphicsWriteDataPatchSetAddressOrOffset(uint32_t* cmd,
 		return OK;
 	}
 
-	return static_cast<int>(0x8a6c000cu);
+	return GRAPHICS5_ERROR_INVALID_PACKET;
 }
 
 int KYTY_SYSV_ABI GraphicsUnknownJumpPatchSetTarget(uint32_t* cmd, const volatile uint32_t* target,
@@ -1820,7 +1866,6 @@ uint32_t* KYTY_SYSV_ABI GraphicsCbReleaseMem(CommandBuffer* buf, uint8_t action,
 	cmd[5] = static_cast<uint32_t>(packet_data & 0xffffffffu);
 	cmd[6] = static_cast<uint32_t>((packet_data >> 32u) & 0xffffffffu);
 	cmd[7] = interrupt_ctx_id & 0x07ffffffu;
-
 	return cmd;
 }
 
@@ -2380,7 +2425,7 @@ int KYTY_SYSV_ABI GraphicsUnknownIkfdtRIqCE(uint32_t* cmd, uint64_t arg1,
 
 	auto op = (cmd[0] >> 8u) & 0xffu;
 	if (op != Pm4::IT_INDIRECT_BUFFER) {
-		return 0x8a6c000c;
+		return GRAPHICS5_ERROR_INVALID_PACKET;
 	}
 
 	auto vaddr        = reinterpret_cast<uint64_t>(target);
@@ -2800,6 +2845,10 @@ uint32_t KYTY_SYSV_ABI GraphicsAcbCondExecGetSize() {
 	return GraphicsDcbCondExecGetSize();
 }
 
+uint32_t KYTY_SYSV_ABI GraphicsAcbJumpGetSize() {
+	return 0x10u;
+}
+
 uint32_t* KYTY_SYSV_ABI GraphicsAcbWaitRegMem(CommandBuffer* buf, uint8_t size,
                                               uint8_t compare_function, uint8_t cache_policy,
                                               const volatile void* address, uint64_t reference,
@@ -3119,7 +3168,7 @@ int KYTY_SYSV_ABI GraphicsDmaDataPatchSetDstAddressOrOffset(uint32_t* cmd,
 		return OK;
 	}
 
-	return static_cast<int>(0x8a6c000cu);
+	return GRAPHICS5_ERROR_INVALID_PACKET;
 }
 
 int KYTY_SYSV_ABI GraphicsDmaDataPatchSetSrcAddressOrOffsetOrImmediate(
@@ -3133,7 +3182,7 @@ int KYTY_SYSV_ABI GraphicsDmaDataPatchSetSrcAddressOrOffsetOrImmediate(
 		return OK;
 	}
 
-	return static_cast<int>(0x8a6c000cu);
+	return GRAPHICS5_ERROR_INVALID_PACKET;
 }
 
 uint32_t KYTY_SYSV_ABI GraphicsGetPacketSize(uint32_t* packet) {
@@ -3197,6 +3246,15 @@ int KYTY_SYSV_ABI GraphicsSetRangePredication(uint32_t* start, const volatile ui
 	return OK;
 }
 
+int KYTY_SYSV_ABI GraphicsRewindPatchSetRewindState(uint32_t* cmd, uint8_t state) {
+	if (((cmd[0] >> 8u) & 0xffu) != Pm4::IT_REWIND) {
+		return GRAPHICS5_ERROR_INVALID_PACKET;
+	}
+
+	cmd[1] = (cmd[1] & 0x7fffffffu) | (static_cast<uint32_t>(state) << 31u);
+	return OK;
+}
+
 int KYTY_SYSV_ABI GraphicsCondExecPatchSetEnd(uint32_t* cmd, const volatile uint32_t* buffer) {
 	PRINT_NAME();
 
@@ -3205,23 +3263,23 @@ int KYTY_SYSV_ABI GraphicsCondExecPatchSetEnd(uint32_t* cmd, const volatile uint
 	     reinterpret_cast<uint64_t>(cmd), reinterpret_cast<uint64_t>(buffer));
 
 	if (cmd == nullptr || buffer == nullptr) {
-		return static_cast<int>(0x8a6c000cu);
+		return GRAPHICS5_ERROR_INVALID_PACKET;
 	}
 
 	auto op = (cmd[0] >> 8u) & 0xffu;
 	if (op != Pm4::IT_COND_EXEC) {
-		return static_cast<int>(0x8a6c000cu);
+		return GRAPHICS5_ERROR_INVALID_PACKET;
 	}
 
 	auto* packet_end = cmd + 5;
 	auto* range_end  = const_cast<uint32_t*>(reinterpret_cast<const volatile uint32_t*>(buffer));
 	if (range_end < packet_end) {
-		return static_cast<int>(0x8a6c000cu);
+		return GRAPHICS5_ERROR_INVALID_PACKET;
 	}
 
 	auto num_dwords = static_cast<uint64_t>(range_end - packet_end);
 	if (num_dwords > 0x3fffu) {
-		return static_cast<int>(0x8a6c000cu);
+		return GRAPHICS5_ERROR_INVALID_PACKET;
 	}
 
 	cmd[4] = (cmd[4] & ~0x3fffu) | static_cast<uint32_t>(num_dwords);
@@ -3237,12 +3295,12 @@ int KYTY_SYSV_ABI GraphicsCondExecPatchSetCommandAddress(uint32_t*
 	     reinterpret_cast<uint64_t>(cmd), reinterpret_cast<uint64_t>(command));
 
 	if (cmd == nullptr || command == nullptr) {
-		return static_cast<int>(0x8a6c000cu);
+		return GRAPHICS5_ERROR_INVALID_PACKET;
 	}
 
 	auto op = (cmd[0] >> 8u) & 0xffu;
 	if (op != Pm4::IT_COND_EXEC || (reinterpret_cast<uintptr_t>(command) & 0x3u) != 0) {
-		return static_cast<int>(0x8a6c000cu);
+		return GRAPHICS5_ERROR_INVALID_PACKET;
 	}
 
 	auto addr = reinterpret_cast<uint64_t>(command);
@@ -3704,150 +3762,6 @@ static void submit_dcb(uint32_t* dcb, uint32_t size_in_dwords) {
 	EXIT_IF(g_renderer == nullptr);
 	g_renderer->GetGpu().Submit(dcb, size_in_dwords, nullptr, 0,
 	                            !dcb_has_queued_interrupt(dcb, size_in_dwords));
-	Gen5::track_pending_graphics_segment_after_submit(dcb, size_in_dwords);
-}
-
-static std::vector<uint64_t> collect_acb_wait_addresses(const uint32_t* acb,
-                                                        uint32_t        size_in_dwords) {
-	std::vector<uint64_t> addresses;
-
-	for (uint32_t offset = 0; offset < size_in_dwords;) {
-		auto cmd_id = acb[offset];
-		auto len    = KYTY_PM4_LEN(cmd_id);
-		if (len == 0 || len > size_in_dwords - offset) {
-			return addresses;
-		}
-
-		auto op = (cmd_id >> 8u) & 0xffu;
-		if (op == Pm4::IT_NOP && KYTY_PM4_R(cmd_id) == Pm4::R_WAIT_MEM_32 && len >= 7) {
-			auto address = static_cast<uint64_t>(acb[offset + 1]) |
-			               (static_cast<uint64_t>(acb[offset + 2]) << 32u);
-			if (address != 0) {
-				addresses.push_back(address);
-			}
-		} else if (op == Pm4::IT_NOP && KYTY_PM4_R(cmd_id) == Pm4::R_WAIT_MEM_64 && len >= 9) {
-			auto address = static_cast<uint64_t>(acb[offset + 1]) |
-			               (static_cast<uint64_t>(acb[offset + 2]) << 32u);
-			if (address != 0) {
-				addresses.push_back(address);
-			}
-		}
-
-		offset += len;
-	}
-
-	return addresses;
-}
-
-static bool acb_waits_for_address(const std::vector<uint64_t>& wait_addresses,
-                                  uint64_t                     release_address) {
-	for (auto address: wait_addresses) {
-		if (address == release_address) {
-			return true;
-		}
-	}
-
-	return false;
-}
-
-static void flush_pending_graphics_segment_before_acb(const uint32_t* acb,
-                                                      uint32_t        acb_size_in_dwords) {
-	uint32_t* dcb            = nullptr;
-	uint32_t  size_in_dwords = 0;
-	auto      wait_addresses = collect_acb_wait_addresses(acb, acb_size_in_dwords);
-
-	{
-		std::lock_guard lock(Gen5::g_pending_graphics_segment_mutex);
-		if (!wait_addresses.empty() && Gen5::g_pending_graphics_segment.start != nullptr) {
-			auto* scan        = Gen5::g_pending_graphics_segment.start;
-			auto* matched_end = Gen5::g_pending_graphics_segment.start;
-			while (scan < Gen5::g_pending_graphics_segment.end) {
-				auto cmd_id = *scan;
-				if (cmd_id == 0x80000000u) {
-					scan++;
-					continue;
-				}
-				if ((cmd_id & 0xC0000000u) != 0xC0000000u) {
-					break;
-				}
-
-				auto len = KYTY_PM4_LEN(cmd_id);
-				if (len == 0 ||
-				    len > static_cast<uint32_t>(Gen5::g_pending_graphics_segment.end - scan)) {
-					break;
-				}
-
-				if (((cmd_id >> 8u) & 0xffu) == Pm4::IT_NOP &&
-				    KYTY_PM4_R(cmd_id) == Pm4::R_RELEASE_MEM && len >= 7) {
-					auto release_addr =
-					    static_cast<uint64_t>(scan[3]) | (static_cast<uint64_t>(scan[4]) << 32u);
-					if (acb_waits_for_address(wait_addresses, release_addr)) {
-						matched_end = scan + len;
-					}
-				}
-
-				scan += len;
-			}
-
-			if (matched_end > Gen5::g_pending_graphics_segment.start) {
-				Gen5::g_pending_graphics_segment.end = matched_end;
-			}
-		}
-
-		if (Gen5::g_pending_graphics_segment.start != nullptr &&
-		    Gen5::g_pending_graphics_segment.end > Gen5::g_pending_graphics_segment.start) {
-			auto* scan      = Gen5::g_pending_graphics_segment.start;
-			auto* valid_end = Gen5::g_pending_graphics_segment.start;
-			while (scan < Gen5::g_pending_graphics_segment.end) {
-				auto cmd_id = *scan;
-				if (cmd_id == 0x80000000u) {
-					scan++;
-					valid_end = scan;
-					continue;
-				}
-				if ((cmd_id & 0xC0000000u) != 0xC0000000u) {
-					break;
-				}
-
-				auto len = KYTY_PM4_LEN(cmd_id);
-				if (len == 0 ||
-				    len > static_cast<uint32_t>(Gen5::g_pending_graphics_segment.end - scan)) {
-					break;
-				}
-
-				scan += len;
-				valid_end = scan;
-			}
-
-			if (valid_end < Gen5::g_pending_graphics_segment.end) {
-				static std::atomic<uint32_t> log_count {0};
-				if (log_count.fetch_add(1) < 64) {
-					LOGF("\t trimming pending graphics segment: addr = 0x%016" PRIx64
-					     ", old_dw = 0x%08" PRIx32 ", new_dw = 0x%08" PRIx32 "\n",
-					     reinterpret_cast<uint64_t>(Gen5::g_pending_graphics_segment.start),
-					     static_cast<uint32_t>(Gen5::g_pending_graphics_segment.end -
-					                           Gen5::g_pending_graphics_segment.start),
-					     static_cast<uint32_t>(valid_end - Gen5::g_pending_graphics_segment.start));
-				}
-				Gen5::g_pending_graphics_segment.end = valid_end;
-			}
-		}
-
-		if (Gen5::g_pending_graphics_segment.start == nullptr ||
-		    Gen5::g_pending_graphics_segment.end <= Gen5::g_pending_graphics_segment.start) {
-			return;
-		}
-
-		dcb            = Gen5::g_pending_graphics_segment.start;
-		size_in_dwords = static_cast<uint32_t>(Gen5::g_pending_graphics_segment.end -
-		                                       Gen5::g_pending_graphics_segment.start);
-	}
-
-	LOGF("\t flushing pending graphics segment before ACB: addr = 0x%016" PRIx64
-	     ", dw_num = 0x%08" PRIx32 "\n",
-	     reinterpret_cast<uint64_t>(dcb), size_in_dwords);
-
-	submit_dcb(dcb, size_in_dwords);
 }
 
 int KYTY_SYSV_ABI GraphicsDriverSubmitDcb(const Packet* packet) {
@@ -3963,8 +3877,6 @@ static void submit_acb(uint32_t queue, uint32_t* acb, uint32_t size_in_dwords) {
 		LOGF("\t acb[%u] = 0x%08" PRIx32 "\n", i, acb[i]);
 	}
 
-	flush_pending_graphics_segment_before_acb(acb, size_in_dwords);
-
 	GraphicsDbgDumpDcb("a", size_in_dwords, acb);
 
 	const bool trigger_interrupt_on_done = !dcb_has_queued_interrupt(acb, size_in_dwords);

src/libs/agc.h

@@ -38,6 +38,8 @@ int KYTY_SYSV_ABI   GraphicsUnknownGetFusedShaderSize(SizeAlign* dst, const Shad
                                                       const Shader* back);
 int KYTY_SYSV_ABI   GraphicsUnknownFuseShaderHalves(Shader* fused_result, const Shader* front,
                                                     const Shader* back, void* scratch_mem);
+int KYTY_SYSV_ABI   GraphicsUnknownNApJjpKNBl4(Shader* fused_result, const Shader* front,
+                                               const Shader* back, void* scratch_mem);
 int KYTY_SYSV_ABI   GraphicsSetCxRegIndirectPatchSetAddress(uint32_t*                      cmd,
                                                             const volatile ShaderRegister* regs);
 int KYTY_SYSV_ABI   GraphicsSetShRegIndirectPatchSetAddress(uint32_t*                      cmd,
@@ -183,6 +185,7 @@ uint32_t KYTY_SYSV_ABI  GraphicsAcbAcquireMemGetSize();
 uint32_t* KYTY_SYSV_ABI GraphicsAcbCondExec(CommandBuffer* buf, const volatile uint32_t* address,
                                             uint32_t num_dwords);
 uint32_t KYTY_SYSV_ABI  GraphicsAcbCondExecGetSize();
+uint32_t KYTY_SYSV_ABI  GraphicsAcbJumpGetSize();
 uint32_t* KYTY_SYSV_ABI GraphicsAcbWaitRegMem(CommandBuffer* buf, uint8_t size,
                                               uint8_t compare_function, uint8_t cache_policy,
                                               const volatile void* address, uint64_t reference,
@@ -241,6 +244,7 @@ uint32_t KYTY_SYSV_ABI  GraphicsGetPacketSize(uint32_t* packet);
 int KYTY_SYSV_ABI       GraphicsSetPacketPredication(uint32_t* packet, uint32_t predication);
 int KYTY_SYSV_ABI       GraphicsSetRangePredication(uint32_t* start, const volatile uint32_t* end,
                                                     uint32_t predication);
+int KYTY_SYSV_ABI       GraphicsRewindPatchSetRewindState(uint32_t* cmd, uint8_t state);
 int KYTY_SYSV_ABI       GraphicsCondExecPatchSetEnd(uint32_t* cmd, const volatile uint32_t* buffer);
 int KYTY_SYSV_ABI       GraphicsCondExecPatchSetCommandAddress(uint32_t*                cmd,
                                                                const volatile uint32_t* command);

src/libs/audio.cpp

@@ -364,7 +364,12 @@ bool Audio::QueueSdlAudio(PortOut* port, const void* data, bool blocking) {
 	}
 
 	if (blocking) {
-		const auto min_queued_size = queue_size * 2u;
+		constexpr uint64_t target_latency_us = 40000;
+		const auto buffer_us = port->freq != 0 ? (1000000ULL * port->samples_num) / port->freq : 0;
+		const auto buffers =
+		    buffer_us != 0 ? static_cast<uint32_t>((target_latency_us + buffer_us - 1) / buffer_us)
+		                   : 2u;
+		const auto min_queued_size = queue_size * std::clamp(buffers, 2u, 16u);
 		const auto wait_start      = LibKernel::KernelGetProcessTime();
 		while (SDL_GetQueuedAudioSize(port->audio_device) > min_queued_size) {
 			if (LibKernel::KernelGetProcessTime() - wait_start > 200000) {
@@ -511,7 +516,16 @@ uint32_t Audio::AudioOutOutputs(OutputParam* params, uint32_t num, bool blocking
 		max_wait_time      = (wait_time > max_wait_time ? wait_time : max_wait_time);
 	}
 
-	if (blocking && max_wait_time != 0) {
+	bool all_ports_have_device = true;
+	for (uint32_t i = 0; i < num; i++) {
+		if (m_out_ports[params[i].handle.GetId()].audio_device == 0) {
+			all_ports_have_device = false;
+			break;
+		}
+	}
+
+	// Device-backed ports are paced by the SDL queue above.
+	if (blocking && max_wait_time != 0 && !all_ports_have_device) {
 		Common::Thread::SleepMicro(max_wait_time);
 	}
 

src/libs/libAudio2.cpp

@@ -336,18 +336,6 @@ static AudioOut2PortStateEntry* audioout2_find_port_locked(AudioOut2PortHandle p
 	return nullptr;
 }
 
-static uint32_t audioout2_context_grains(AudioOut2ContextHandle ctx) {
-	uint32_t samples_num = 512;
-
-	g_audioout2_context_mutex.Lock();
-	if (auto* state = audioout2_find_context_locked(ctx); state != nullptr) {
-		samples_num = (state->num_grains == 0 ? 512u : state->num_grains);
-	}
-	g_audioout2_context_mutex.Unlock();
-
-	return samples_num;
-}
-
 static void audioout2_queue_context_audio(AudioOut2ContextHandle ctx, bool blocking) {
 	std::vector<AudioInternal::OutputParam> params;
 	params.reserve(AudioInternal::OUT_PORTS_MAX);
@@ -455,6 +443,12 @@ int KYTY_SYSV_ABI AudioOut2ContextDestroy(AudioOut2ContextHandle ctx) {
 	PRINT_NAME();
 	LOGF("\t ctx = 0x%016" PRIx64 "\n", ctx);
 
+	g_audioout2_context_mutex.Lock();
+	if (auto* state = audioout2_find_context_locked(ctx); state != nullptr) {
+		*state = AudioOut2ContextState {};
+	}
+	g_audioout2_context_mutex.Unlock();
+
 	std::array<int, 256> audio_handles {};
 	size_t               audio_handles_num = 0;
 
@@ -473,12 +467,6 @@ int KYTY_SYSV_ABI AudioOut2ContextDestroy(AudioOut2ContextHandle ctx) {
 		audioout2_close_audio_handle(audio_handles[i]);
 	}
 
-	g_audioout2_context_mutex.Lock();
-	if (auto* state = audioout2_find_context_locked(ctx); state != nullptr) {
-		*state = AudioOut2ContextState {};
-	}
-	g_audioout2_context_mutex.Unlock();
-
 	return OK;
 }
 
@@ -559,30 +547,45 @@ int KYTY_SYSV_ABI AudioOut2PortCreate(AudioOut2ContextHandle ctx, const AudioOut
 	EXIT_NOT_IMPLEMENTED(params == nullptr);
 	EXIT_NOT_IMPLEMENTED(port == nullptr);
 
-	const auto next_port = g_audioout2_next_port.fetch_add(1, std::memory_order_relaxed);
+	const auto next_port    = g_audioout2_next_port.fetch_add(1, std::memory_order_relaxed);
+	const auto audio_format = audioout2_data_format_to_audio_format(params->data_format);
+	const auto audio_type   = audioout2_port_type_to_audio_out_type(params->port_type);
+
+	g_audioout2_context_mutex.Lock();
+	const auto* context_state = audioout2_find_context_locked(ctx);
+	if (context_state == nullptr) {
+		g_audioout2_context_mutex.Unlock();
+		return AUDIO_OUT2_ERROR_INVALID_PARAM;
+	}
+	const auto samples_num = context_state->num_grains == 0 ? 512u : context_state->num_grains;
 
 	g_audioout2_port_mutex.Lock();
-	auto* port_state = audioout2_find_port_locked(0);
-	if (port_state == nullptr) {
-		for (auto& candidate: g_audioout2_ports) {
-			if (!candidate.used) {
-				port_state = &candidate;
-				break;
-			}
+	AudioOut2PortStateEntry* port_state = nullptr;
+	for (auto& candidate: g_audioout2_ports) {
+		if (!candidate.used) {
+			port_state = &candidate;
+			break;
 		}
 	}
+	if (port_state != nullptr) {
+		*port_state               = AudioOut2PortStateEntry {};
+		port_state->used          = true;
+		port_state->handle        = next_port;
+		port_state->context       = ctx;
+		port_state->port_type     = params->port_type;
+		port_state->data_format   = params->data_format;
+		port_state->sampling_freq = params->sampling_freq;
+		port_state->samples_num   = samples_num;
+		port_state->audio_format  = audio_format;
+	}
 	g_audioout2_port_mutex.Unlock();
+	g_audioout2_context_mutex.Unlock();
 
-	if (next_port > g_audioout2_ports.size() || port_state == nullptr) {
+	if (port_state == nullptr) {
 		return AUDIO_OUT2_ERROR_PORT_FULL;
 	}
 
-	*port = next_port;
-
-	const auto samples_num  = audioout2_context_grains(ctx);
-	const auto audio_format = audioout2_data_format_to_audio_format(params->data_format);
-	const auto audio_type   = audioout2_port_type_to_audio_out_type(params->port_type);
-	int        audio_handle = 0;
+	int audio_handle = 0;
 
 	if (audio_format != AudioInternal::Format::Unknown &&
 	    !audioout2_port_type_is_object(params->port_type)) {
@@ -591,17 +594,17 @@ int KYTY_SYSV_ABI AudioOut2PortCreate(AudioOut2ContextHandle ctx, const AudioOut
 	}
 
 	g_audioout2_port_mutex.Lock();
-	*port_state               = AudioOut2PortStateEntry {};
-	port_state->used          = true;
-	port_state->handle        = *port;
-	port_state->context       = ctx;
-	port_state->port_type     = params->port_type;
-	port_state->data_format   = params->data_format;
-	port_state->sampling_freq = params->sampling_freq;
-	port_state->samples_num   = samples_num;
-	port_state->audio_format  = audio_format;
-	port_state->audio_handle  = audio_handle;
+	const bool reserved = port_state->used && port_state->handle == next_port;
+	if (reserved) {
+		port_state->audio_handle = audio_handle;
+	}
 	g_audioout2_port_mutex.Unlock();
+	if (!reserved) {
+		audioout2_close_audio_handle(audio_handle);
+		return AUDIO_OUT2_ERROR_INVALID_PARAM;
+	}
+
+	*port = next_port;
 
 	if (next_port <= 16 || (next_port % 600) == 0) {
 		PRINT_NAME();

src/libs/libC.cpp

@@ -633,6 +633,15 @@ LIB_VERSION("LibcInternalExt", 1, "LibcInternal", 1, 1);
 static uint64_t g_mspace_atomic_id_mask = 0;
 static uint64_t g_mstate_table[64]      = {0};
 
+using thread_atexit_destructor_t = KYTY_SYSV_ABI void (*)(void*);
+
+struct ThreadAtexitDestructor {
+	thread_atexit_destructor_t destructor;
+	void*                      object;
+};
+
+static thread_local std::vector<ThreadAtexitDestructor> g_thread_atexit_destructors;
+
 struct Info {
 	uint64_t  size;
 	uint32_t  unknown1;
@@ -650,25 +659,29 @@ void KYTY_SYSV_ABI LibcHeapGetTraceInfo(Info* info) {
 	info->mstate_table          = g_mstate_table;
 }
 
-uint64_t KYTY_SYSV_ABI LibcInternalExtUnknownQBS714Jr3g(uint64_t arg0, uint64_t arg1, uint64_t arg2,
-                                                        uint64_t arg3, uint64_t arg4,
-                                                        uint64_t arg5) {
+int KYTY_SYSV_ABI LibcInternalExtCxaThreadAtexit(thread_atexit_destructor_t destructor, void* object,
+                                                 void* /*module_id*/) {
 	PRINT_NAME();
 
-	LOGF("\t arg0 = 0x%016" PRIx64 "\n"
-	     "\t arg1 = 0x%016" PRIx64 "\n"
-	     "\t arg2 = 0x%016" PRIx64 "\n"
-	     "\t arg3 = 0x%016" PRIx64 "\n"
-	     "\t arg4 = 0x%016" PRIx64 "\n"
-	     "\t arg5 = 0x%016" PRIx64 "\n",
-	     arg0, arg1, arg2, arg3, arg4, arg5);
+	g_thread_atexit_destructors.push_back({destructor, object});
 
 	return 0;
 }
 
+void RunThreadAtexitDestructors() {
+	while (!g_thread_atexit_destructors.empty()) {
+		auto destructor = g_thread_atexit_destructors.back();
+		g_thread_atexit_destructors.pop_back();
+
+		if (destructor.destructor != nullptr) {
+			destructor.destructor(destructor.object);
+		}
+	}
+}
+
 LIB_DEFINE(InitLibcInternalExt_1) {
 	LIB_FUNC("NWtTN10cJzE", LibcInternalExt::LibcHeapGetTraceInfo);
-	LIB_FUNC("qBS714-Jr3g", LibcInternalExt::LibcInternalExtUnknownQBS714Jr3g);
+	LIB_FUNC("qBS714-Jr3g", LibcInternalExt::LibcInternalExtCxaThreadAtexit);
 }
 
 } // namespace LibcInternalExt

src/libs/libGraphicsDriver.cpp

@@ -159,6 +159,7 @@ LIB_DEFINE(InitGraphicsDriver_1) {
 	LIB_FUNC("f3dg2CSgRKY", Gen5::GraphicsCreateShader);
 	LIB_FUNC("dolOmWH+huQ", Gen5::GraphicsUnknownGetFusedShaderSize);
 	LIB_FUNC("fd5Bp5tGTgo", Gen5::GraphicsUnknownFuseShaderHalves);
+	LIB_FUNC("nApJjpKNBl4", Gen5::GraphicsUnknownNApJjpKNBl4);
 	LIB_FUNC("vcmNN+AAXnY", Gen5::GraphicsSetCxRegIndirectPatchSetAddress);
 	LIB_FUNC("Qrj4c+61z4A", Gen5::GraphicsSetShRegIndirectPatchSetAddress);
 	LIB_FUNC("6lNcCp+fxi4", Gen5::GraphicsSetUcRegIndirectPatchSetAddress);
@@ -207,6 +208,7 @@ LIB_DEFINE(InitGraphicsDriver_1) {
 	LIB_FUNC("ewobAQeMo5k", Gen5::GraphicsAcbAcquireMemGetSize);
 	LIB_FUNC("qyM2bxYFPAk", Gen5::GraphicsAcbCondExec);
 	LIB_FUNC("ozKzBP4aki4", Gen5::GraphicsAcbCondExecGetSize);
+	LIB_FUNC("b-oySn+G2tE", Gen5::GraphicsAcbJumpGetSize);
 	LIB_FUNC("htn36gPnBk4", Gen5::GraphicsAcbWaitRegMem);
 	LIB_FUNC("-RnpfpxIhec", Gen5::GraphicsAcbDmaData);
 	LIB_FUNC("qzMN2XKGA4k", Gen5::GraphicsAcbCopyData);
@@ -282,6 +284,7 @@ LIB_DEFINE(InitGraphicsDriver_1) {
 	LIB_FUNC("YWTKOju587o", Gen5::GraphicsCondExecPatchSetCommandAddress);
 	LIB_FUNC("k-JpyR2dYAM", Gen5::GraphicsCondExecPatchSetEnd);
 	LIB_FUNC("3ZWa3AoyWZQ", Gen5::GraphicsCondExecPatchSetCommandAddress);
+	LIB_FUNC("ziVA3whp3p4", Gen5::GraphicsRewindPatchSetRewindState);
 	LIB_FUNC("YUeqkyT7mEQ", Gen5::GraphicsDcbSetFlip);
 }
 

src/libs/libRtc.cpp

@@ -4,6 +4,7 @@
 #include "libs/libs.h"
 #include "loader/symbolDatabase.h"
 
+#include <chrono>
 #include <cstdint>
 #include <cstdio>
 #include <limits>
@@ -247,20 +248,12 @@ static int KYTY_SYSV_ABI RtcGetCurrentTick(RtcTick* tick) {
 		return RTC_ERROR_DATETIME_UNINITIALIZED;
 	}
 
-	const auto now  = Common::DateTime::FromSystemUTC();
-	const auto date = now.GetDate();
-	const auto tod  = now.GetTime();
-
-	RtcDateTime time {};
-	time.year        = static_cast<uint16_t>(date.Year());
-	time.month       = static_cast<uint16_t>(date.Month());
-	time.day         = static_cast<uint16_t>(date.Day());
-	time.hour        = static_cast<uint16_t>(tod.Hour24());
-	time.minute      = static_cast<uint16_t>(tod.Minute());
-	time.second      = static_cast<uint16_t>(tod.Second());
-	time.microsecond = static_cast<uint32_t>(tod.Msec() * 1000);
-
-	return RtcGetTick(&time, tick);
+	const auto now_us =
+	    static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::microseconds>(
+	                              std::chrono::system_clock::now().time_since_epoch())
+	                              .count());
+	tick->tick = RTC_UNIX_EPOCH_TICKS + now_us;
+	return OK;
 }
 
 static int KYTY_SYSV_ABI RtcGetCurrentNetworkTick(RtcTick* tick) {

src/loader/runtimeLinker.cpp

@@ -148,7 +148,7 @@ static bool PatchGuestMemory64(uint64_t vaddr, uint64_t value) {
 }
 
 static uint64_t AllocateUnresolvedImportThunk(uint64_t record_id) {
-	constexpr uint64_t thunk_size = 162;
+	constexpr uint64_t thunk_size = 165;
 
 	if (g_unresolved_stub_thunk_pages.empty() ||
 	    g_unresolved_stub_thunk_offset + thunk_size > UNRESOLVED_STUB_PAGE_SIZE) {
@@ -254,6 +254,10 @@ static uint64_t AllocateUnresolvedImportThunk(uint64_t record_id) {
 	emit(0x41);
 	emit(0xff);
 	emit(0xe3); // jmp r11
+	// Match the integer fallback for floating-point return values.
+	emit(0x0f);
+	emit(0x57);
+	emit(0xc0); // xorps xmm0, xmm0
 	emit(0x31);
 	emit(0xc0); // xor eax, eax
 	emit(0xc3); // ret

tests/AudioOut2PortTests.cpp

@@ -0,0 +1,249 @@
+#include "libs/audio.h"
+#include "libs/audio_internal.h"
+#include "libs/errno.h"
+
+#include <algorithm>
+#include <atomic>
+#include <condition_variable>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <mutex>
+#include <thread>
+#include <vector>
+
+namespace {
+
+namespace AudioOut2 = Libs::Audio::AudioOut2;
+
+std::mutex              g_device_mutex;
+std::condition_variable g_device_cv;
+std::vector<int>        g_live_devices;
+int                     g_next_device  = 1;
+int                     g_open_waiters = 0;
+bool                    g_block_opens  = false;
+
+void Check(bool value, const char* text) {
+	if (!value) {
+		std::fprintf(stderr, "AudioOut2PortTests: failed: %s\n", text);
+		std::abort();
+	}
+}
+
+struct PortParam {
+	uint16_t port_type;
+	uint16_t pad;
+	uint32_t data_format;
+	uint32_t sampling_freq;
+	uint32_t flags;
+	uint64_t user_handle;
+	uint32_t reserved[10];
+};
+
+struct ContextParam {
+	uint32_t max_ports;
+	uint32_t max_object_ports;
+	uint32_t guarantee_object_ports;
+	uint32_t queue_depth;
+	uint32_t num_grains;
+	uint32_t flags;
+	uint32_t reserved[10];
+};
+
+struct PortState {
+	uint16_t output;
+	uint8_t  num_channels;
+	uint8_t  pad1;
+	int16_t  volume;
+	uint16_t reroute_counter;
+	uint32_t flags;
+	uint32_t pad2;
+	uint64_t reserved[6];
+};
+
+const auto* AsParam(const PortParam* param) {
+	return reinterpret_cast<const AudioOut2::AudioOut2PortParam*>(param);
+}
+
+const auto* AsParam(const ContextParam* param) {
+	return reinterpret_cast<const AudioOut2::AudioOut2ContextParam*>(param);
+}
+
+auto* AsState(PortState* state) {
+	return reinterpret_cast<AudioOut2::AudioOut2PortState*>(state);
+}
+
+PortParam MakeParam(uint32_t data_format = 0x200) {
+	PortParam param {};
+	param.data_format   = data_format;
+	param.sampling_freq = 48000;
+	return param;
+}
+
+AudioOut2::AudioOut2ContextHandle CreateContext() {
+	ContextParam param {};
+	param.queue_depth                         = 4;
+	param.num_grains                          = 512;
+	AudioOut2::AudioOut2ContextHandle context = 0;
+	Check(AudioOut2::AudioOut2ContextCreate(AsParam(&param), nullptr, 0, &context) == OK,
+	      "context create failed");
+	return context;
+}
+
+void BlockDeviceOpens() {
+	std::lock_guard lock(g_device_mutex);
+	g_open_waiters = 0;
+	g_block_opens  = true;
+}
+
+void WaitForDeviceOpens(int count) {
+	std::unique_lock lock(g_device_mutex);
+	g_device_cv.wait(lock, [count]() { return g_open_waiters >= count; });
+}
+
+void ReleaseDeviceOpens() {
+	std::lock_guard lock(g_device_mutex);
+	g_block_opens = false;
+	g_device_cv.notify_all();
+}
+
+int LiveDeviceCount() {
+	std::lock_guard lock(g_device_mutex);
+	return static_cast<int>(g_live_devices.size());
+}
+
+void TestSlotReuse() {
+	const auto context = CreateContext();
+	const auto param   = MakeParam();
+	for (int i = 0; i < 300; i++) {
+		AudioOut2::AudioOut2PortHandle port = 0;
+		Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &port) == OK,
+		      "port slot was not reusable");
+		Check(port != 0, "port handle is zero");
+		AudioOut2::AudioOut2PortDestroy(port);
+	}
+	AudioOut2::AudioOut2ContextDestroy(context);
+}
+
+void TestFullTableRecovers() {
+	const auto                                  context = CreateContext();
+	const auto                                  param   = MakeParam();
+	std::vector<AudioOut2::AudioOut2PortHandle> ports;
+	ports.reserve(256);
+
+	for (int i = 0; i < 256; i++) {
+		AudioOut2::AudioOut2PortHandle port = 0;
+		Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &port) == OK,
+		      "port table filled early");
+		ports.push_back(port);
+	}
+
+	AudioOut2::AudioOut2PortHandle overflow = 0;
+	Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &overflow) != OK,
+	      "full port table accepted another port");
+
+	for (auto port: ports) {
+		AudioOut2::AudioOut2PortDestroy(port);
+	}
+
+	AudioOut2::AudioOut2PortHandle port = 0;
+	Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &port) == OK,
+	      "port table did not recover");
+	AudioOut2::AudioOut2PortDestroy(port);
+	AudioOut2::AudioOut2ContextDestroy(context);
+}
+
+void TestConcurrentCreates() {
+	constexpr int                               thread_count = 8;
+	const auto                                  context      = CreateContext();
+	const auto                                  param        = MakeParam(0x800);
+	std::vector<AudioOut2::AudioOut2PortHandle> ports(thread_count);
+	std::vector<int>                            results(thread_count);
+	std::vector<std::thread>                    threads;
+
+	BlockDeviceOpens();
+	for (int i = 0; i < thread_count; i++) {
+		threads.emplace_back([&, i]() {
+			results[i] = AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &ports[i]);
+		});
+	}
+	WaitForDeviceOpens(thread_count);
+	ReleaseDeviceOpens();
+	for (auto& thread: threads) {
+		thread.join();
+	}
+
+	for (int i = 0; i < thread_count; i++) {
+		Check(results[i] == OK, "concurrent port create failed");
+		PortState state {};
+		AudioOut2::AudioOut2PortGetState(ports[i], AsState(&state));
+		Check(state.num_channels == 8, "concurrent create lost its reserved slot");
+		AudioOut2::AudioOut2PortDestroy(ports[i]);
+	}
+	Check(LiveDeviceCount() == 0, "concurrent create leaked a device");
+	AudioOut2::AudioOut2ContextDestroy(context);
+}
+
+void TestContextDestroyCancelsPendingCreate() {
+	const auto                     context = CreateContext();
+	const auto                     param   = MakeParam();
+	AudioOut2::AudioOut2PortHandle port    = 0;
+	int                            result  = OK;
+
+	BlockDeviceOpens();
+	std::thread creator(
+	    [&]() { result = AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &port); });
+	WaitForDeviceOpens(1);
+	AudioOut2::AudioOut2ContextDestroy(context);
+	ReleaseDeviceOpens();
+	creator.join();
+
+	Check(result != OK, "destroyed context retained a pending port create");
+	Check(LiveDeviceCount() == 0, "cancelled port create leaked a device");
+}
+
+} // namespace
+
+namespace Libs::Audio::AudioInternal {
+
+int AudioOutOpen(int /*type*/, uint32_t /*samples_num*/, uint32_t /*freq*/, Format /*format*/) {
+	std::unique_lock lock(g_device_mutex);
+	const int        handle = g_next_device++;
+	g_live_devices.push_back(handle);
+	g_open_waiters++;
+	g_device_cv.notify_all();
+	g_device_cv.wait(lock, []() { return !g_block_opens; });
+	return handle;
+}
+
+void AudioOutClose(int handle) {
+	std::lock_guard lock(g_device_mutex);
+	const auto      it = std::find(g_live_devices.begin(), g_live_devices.end(), handle);
+	if (it != g_live_devices.end()) {
+		g_live_devices.erase(it);
+	}
+}
+
+uint32_t AudioOutOutputs(const OutputParam* /*params*/, uint32_t /*num*/, bool /*blocking*/) {
+	return 0;
+}
+
+} // namespace Libs::Audio::AudioInternal
+
+namespace Libs::LibKernel {
+
+uint64_t KYTY_SYSV_ABI KernelGetProcessTime() {
+	static std::atomic_uint64_t now {0};
+	return now.fetch_add(1000);
+}
+
+} // namespace Libs::LibKernel
+
+int main() {
+	TestSlotReuse();
+	TestFullTableRecovers();
+	TestConcurrentCreates();
+	TestContextDestroyCancelsPendingCreate();
+	std::printf("AudioOut2PortTests: all cases passed\n");
+	return 0;
+}

tests/BitArrayTests.cpp

@@ -0,0 +1,310 @@
+#include "common/bitArray.h"
+
+#include <array>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <utility>
+
+namespace {
+
+using Bits = Common::BitArray<128>;
+
+static_assert(sizeof(Common::BitArray<1024>) == 128);
+
+void Check(bool value, const char *message) {
+  if (!value) {
+    std::fprintf(stderr, "BitArrayTests: failed: %s\n", message);
+    std::abort();
+  }
+}
+
+void TestPointAndRangeOperations() {
+  Bits bits;
+  Check(bits.None() && !bits.Any(), "default state is not empty");
+
+  for (const auto index : {size_t{0}, size_t{63}, size_t{64}, size_t{127}}) {
+    bits.Set(index);
+    Check(bits.Get(index), "Set did not set a boundary bit");
+    bits.Unset(index);
+    Check(!bits.Get(index), "Unset did not clear a boundary bit");
+  }
+
+  bits.SetRange(60, 68);
+  for (size_t index = 0; index < 128; index++) {
+    Check(bits.Get(index) == (index >= 60 && index < 68),
+          "cross-word SetRange changed the wrong bits");
+  }
+  bits.Fill();
+  bits.UnsetRange(60, 68);
+  for (size_t index = 0; index < 128; index++) {
+    Check(bits.Get(index) == !(index >= 60 && index < 68),
+          "cross-word UnsetRange changed the wrong bits");
+  }
+  bits.Clear();
+  bits.SetRange(0, 128);
+  Check(!bits.None(), "full SetRange left the array empty");
+  bits.UnsetRange(0, 128);
+  Check(bits.None(), "full UnsetRange left set bits");
+
+  bits.Set(7);
+  bits.SetRange(9, 9);
+  bits.SetRange(0, 129);
+  bits.UnsetRange(9, 9);
+  bits.UnsetRange(0, 129);
+  Check(bits.Get(7), "invalid or empty range modified the array");
+}
+
+void TestMaskedConstructionAndBitwiseOperations() {
+  Bits source;
+  source.Fill();
+  const Bits masked(source, 31, 97);
+  for (size_t index = 0; index < 128; index++) {
+    Check(masked.Get(index) == (index >= 31 && index < 97),
+          "masked constructor retained a bit outside its range");
+  }
+  Check(Bits(source, 12, 12).None(), "empty masked constructor produced bits");
+  Check(Bits(source, 0, 129).None(),
+        "invalid masked constructor produced bits");
+
+  Bits left;
+  left.SetRange(0, 80);
+  Bits right;
+  right.SetRange(40, 120);
+  const auto exclusive = left ^ right;
+  for (size_t index = 0; index < 128; index++) {
+    const bool expected = (index < 80) != (index >= 40 && index < 120);
+    Check(exclusive.Get(index) == expected, "XOR produced the wrong bit");
+  }
+  const auto inverted = ~left;
+  for (size_t index = 0; index < 128; index++) {
+    Check(inverted.Get(index) == (index >= 80), "NOT produced the wrong bit");
+  }
+}
+
+void TestRangeDiscoveryAndIteration() {
+  Bits bits;
+  Check(bits.FirstRange() == Bits::Range{128, 128},
+        "empty FirstRange is wrong");
+  Check(bits.LastRange() == Bits::Range{0, 0}, "empty LastRange is wrong");
+
+  bits.SetRange(3, 8);
+  bits.SetRange(63, 70);
+  bits.Set(127);
+  Check(bits.FirstRange() == Bits::Range{3, 8}, "FirstRange is wrong");
+  Check(bits.FirstRangeFrom(5) == Bits::Range{5, 8},
+        "FirstRangeFrom inside a run is wrong");
+  Check(bits.FirstRangeFrom(8) == Bits::Range{63, 70},
+        "FirstRangeFrom gap is wrong");
+  Check(bits.LastRange() == Bits::Range{127, 128}, "LastRange is wrong");
+  Check(bits.LastRangeFrom(69) == Bits::Range{63, 69},
+        "LastRangeFrom inside a run is wrong");
+  Check(bits.LastRangeFrom(63) == Bits::Range{3, 8},
+        "LastRangeFrom gap is wrong");
+
+  constexpr std::array expected{Bits::Range{3, 8}, Bits::Range{63, 70},
+                                Bits::Range{127, 128}};
+  size_t range_index = 0;
+  for (const auto range : bits) {
+    Check(range_index < expected.size() && range == expected[range_index],
+          "range iterator produced the wrong run");
+    range_index++;
+  }
+  Check(range_index == expected.size(), "range iterator omitted a run");
+}
+
+void TestRandomizedDifferential() {
+  Bits bits;
+  std::array<bool, 128> reference{};
+  uint64_t random = 0x53a9'7f11'ced4'29b5ull;
+  const auto next_random = [&random] {
+    random ^= random << 13;
+    random ^= random >> 7;
+    random ^= random << 17;
+    return random;
+  };
+
+  for (size_t operation = 0; operation < 10000; operation++) {
+    const auto first = static_cast<size_t>(next_random() % 128);
+    const auto last =
+        first + 1 + static_cast<size_t>(next_random() % (128 - first));
+    if ((next_random() & 1) != 0) {
+      bits.SetRange(first, last);
+      for (auto index = first; index < last; index++) {
+        reference[index] = true;
+      }
+    } else {
+      bits.UnsetRange(first, last);
+      for (auto index = first; index < last; index++) {
+        reference[index] = false;
+      }
+    }
+
+    bool any = false;
+    for (size_t index = 0; index < reference.size(); index++) {
+      Check(bits.Get(index) == reference[index],
+            "randomized bit state diverged");
+      any |= reference[index];
+    }
+    Check(bits.Any() == any && bits.None() == !any,
+          "randomized Any/None diverged");
+
+    const auto range_start = static_cast<size_t>(next_random() % 129);
+    auto expected_first_begin = range_start;
+    while (expected_first_begin < reference.size() &&
+           !reference[expected_first_begin]) {
+      expected_first_begin++;
+    }
+    if (expected_first_begin == reference.size()) {
+      Check(bits.FirstRangeFrom(range_start) == Bits::Range{128, 128},
+            "randomized FirstRangeFrom empty suffix diverged");
+    } else {
+      auto expected_first_end = expected_first_begin;
+      while (expected_first_end < reference.size() &&
+             reference[expected_first_end]) {
+        expected_first_end++;
+      }
+      Check(bits.FirstRangeFrom(range_start) ==
+                Bits::Range{expected_first_begin, expected_first_end},
+            "randomized FirstRangeFrom diverged");
+    }
+
+    const auto range_end = static_cast<size_t>(next_random() % 129);
+    auto expected_last_end = range_end;
+    while (expected_last_end != 0 && !reference[expected_last_end - 1]) {
+      expected_last_end--;
+    }
+    if (expected_last_end == 0) {
+      Check(bits.LastRangeFrom(range_end) == Bits::Range{0, 0},
+            "randomized LastRangeFrom empty prefix diverged");
+    } else {
+      auto expected_last_begin = expected_last_end;
+      while (expected_last_begin != 0 && reference[expected_last_begin - 1]) {
+        expected_last_begin--;
+      }
+      Check(bits.LastRangeFrom(range_end) ==
+                Bits::Range{expected_last_begin, expected_last_end},
+            "randomized LastRangeFrom diverged");
+    }
+
+    const auto masked_start = static_cast<size_t>(next_random() % 129);
+    const auto masked_end =
+        masked_start +
+        static_cast<size_t>(next_random() % (129 - masked_start));
+    const Bits masked(bits, masked_start, masked_end);
+    for (size_t index = 0; index < reference.size(); index++) {
+      Check(masked.Get(index) == (index >= masked_start && index < masked_end &&
+                                  reference[index]),
+            "randomized masked constructor diverged");
+    }
+
+    size_t first_begin = 0;
+    while (first_begin < reference.size() && !reference[first_begin]) {
+      first_begin++;
+    }
+    if (first_begin == reference.size()) {
+      Check(bits.FirstRange() == Bits::Range{128, 128},
+            "randomized empty FirstRange diverged");
+      Check(bits.LastRange() == Bits::Range{0, 0},
+            "randomized empty LastRange diverged");
+    } else {
+      auto first_end = first_begin;
+      while (first_end < reference.size() && reference[first_end]) {
+        first_end++;
+      }
+      Check(bits.FirstRange() == Bits::Range{first_begin, first_end},
+            "randomized FirstRange diverged");
+
+      auto last_end = reference.size();
+      while (!reference[last_end - 1]) {
+        last_end--;
+      }
+      auto last_begin = last_end;
+      while (last_begin != 0 && reference[last_begin - 1]) {
+        last_begin--;
+      }
+      Check(bits.LastRange() == Bits::Range{last_begin, last_end},
+            "randomized LastRange diverged");
+    }
+
+    size_t expected_begin = 0;
+    for (const auto [begin, end] : bits) {
+      while (expected_begin < reference.size() && !reference[expected_begin]) {
+        expected_begin++;
+      }
+      Check(begin == expected_begin, "randomized iterator run start diverged");
+      while (expected_begin < reference.size() && reference[expected_begin]) {
+        expected_begin++;
+      }
+      Check(end == expected_begin, "randomized iterator run end diverged");
+    }
+    while (expected_begin < reference.size() && !reference[expected_begin]) {
+      expected_begin++;
+    }
+    Check(expected_begin == reference.size(),
+          "randomized iterator omitted a run");
+  }
+}
+
+void TestTrackerSizedRandomizedDifferential() {
+  using TrackerBits = Common::BitArray<1024>;
+  TrackerBits bits;
+  std::array<bool, 1024> reference{};
+  uint64_t random = 0x9e37'79b9'7f4a'7c15ull;
+  const auto next_random = [&random] {
+    random ^= random << 13;
+    random ^= random >> 7;
+    random ^= random << 17;
+    return random;
+  };
+
+  for (size_t operation = 0; operation < 4096; operation++) {
+    const auto first = static_cast<size_t>(next_random() % reference.size());
+    const auto last =
+        first + 1 +
+        static_cast<size_t>(next_random() % (reference.size() - first));
+    const bool set = (next_random() & 1) != 0;
+    if (set) {
+      bits.SetRange(first, last);
+    } else {
+      bits.UnsetRange(first, last);
+    }
+    for (auto index = first; index < last; index++) {
+      reference[index] = set;
+    }
+
+    for (size_t index = 0; index < reference.size(); index++) {
+      Check(bits.Get(index) == reference[index],
+            "tracker-sized randomized bit state diverged");
+    }
+
+    size_t expected = 0;
+    for (const auto [begin, end] : bits) {
+      while (expected < reference.size() && !reference[expected]) {
+        expected++;
+      }
+      Check(begin == expected, "tracker-sized randomized range start diverged");
+      while (expected < reference.size() && reference[expected]) {
+        expected++;
+      }
+      Check(end == expected, "tracker-sized randomized range end diverged");
+    }
+    while (expected < reference.size() && !reference[expected]) {
+      expected++;
+    }
+    Check(expected == reference.size(),
+          "tracker-sized randomized iterator omitted a run");
+  }
+}
+
+} // namespace
+
+int main() {
+  TestPointAndRangeOperations();
+  TestMaskedConstructionAndBitwiseOperations();
+  TestRangeDiscoveryAndIteration();
+  TestRandomizedDifferential();
+  TestTrackerSizedRandomizedDifferential();
+  std::puts("BitArrayTests: all cases passed");
+  return 0;
+}

tests/ImagePageTableTests.cpp

@@ -117,12 +117,15 @@ void TestStrictByteFilteringAndPredicate() {
 	OwnerIndex index;
 	Check(index.Register(31, {{0x300100, 0x100}}), "first byte-disjoint owner registers");
 	Check(index.Register(32, {{0x300800, 0x100}}), "second byte-disjoint owner registers");
+	Check(index.Register(33, {{0x30f000, 0x100}}), "coarse-only owner registers");
 	Check(index.TrackingMembershipCount(0x300) == 2,
 	      "byte-disjoint owners share one tracking page");
+	Check(index.CoarseMembershipCount(3) == 3,
+	      "all owners share one coarse candidate bucket");
 	Check(index.Query(0x300400, 0x40).empty(), "page hit without byte overlap is filtered out");
 	const auto page_candidates = index.QueryCandidates(0x300400, 0x40);
 	Check(page_candidates.size() == 2,
-	      "fault candidate query retains byte-disjoint owners on the touched page");
+	      "fault candidate query retains touched-page owners and rejects coarse-only owners");
 	const auto first = index.Query(0x300180, 0x10);
 	Check(first.size() == 1 && first.front() == 31,
 	      "strict byte overlap selects only the matching owner");
@@ -132,6 +135,20 @@ void TestStrictByteFilteringAndPredicate() {
 	      "supplied predicate filters query owners");
 }
 
+void TestOwnerIndexAddressSpaceBoundary() {
+	OwnerIndex         index;
+	constexpr uint64_t last_byte = OwnerIndex::CoarseTable::kAddressSpaceSize - 1;
+	Check(index.Register(41, {{last_byte, 1}}), "final guest byte registers");
+	const auto exact = index.Query(last_byte, 1);
+	Check(exact.size() == 1 && exact.front() == 41,
+	      "strict query finds an exact overlap at the final guest byte");
+	Check(index.Query(last_byte - 1, 1).empty(),
+	      "strict query preserves half-open overlap boundaries");
+	const auto page_candidates = index.QueryCandidates(last_byte - 1, 1);
+	Check(page_candidates.size() == 1 && page_candidates.front() == 41,
+	      "page candidate query retains a byte-disjoint owner on the final tracking page");
+}
+
 } // namespace
 
 int main() {
@@ -142,6 +159,7 @@ int main() {
 	TestMultiRangeRegistrationDeduplicatesPages();
 	TestSharedPageUnregisterLifecycle();
 	TestStrictByteFilteringAndPredicate();
+	TestOwnerIndexAddressSpaceBoundary();
 	std::printf("ImagePageTableTests: all cases passed\n");
 	return 0;
 }

tests/MemoryTrackerTests.cpp

@@ -27,7 +27,6 @@ namespace {
 
 using Libs::Graphics::MemoryTracker;
 using Libs::Graphics::PageManager;
-using Libs::Graphics::PageWatchMode;
 using Libs::Graphics::RangeSet;
 
 void Check(bool value, const char *text) {
@@ -127,6 +126,21 @@ bool IsWritable(const void *address) {
   return Protection(address) == PAGE_READWRITE;
 }
 
+uint64_t g_protection_calls = 0;
+
+struct ProtectionCall {
+  uint64_t address;
+  uint64_t size;
+  Common::VirtualMemory::Mode mode;
+};
+
+std::vector<ProtectionCall> g_protection_log;
+
+void ResetProtectionLog() {
+  g_protection_calls = 0;
+  g_protection_log.clear();
+}
+
 bool ProtectAddressSpace(uint64_t vaddr, uint64_t size,
                          Common::VirtualMemory::Mode mode) {
   uint32_t protection = PAGE_NOACCESS;
@@ -136,14 +150,14 @@ bool ProtectAddressSpace(uint64_t vaddr, uint64_t size,
     protection = PAGE_READWRITE;
   }
   DWORD old_protection = 0;
+  g_protection_calls++;
+  g_protection_log.push_back({vaddr, size, mode});
   return VirtualProtect(reinterpret_cast<void *>(vaddr), size, protection,
                         &old_protection) != 0;
 }
 
 struct TrackerHarness {
-  explicit TrackerHarness(
-      PageWatchMode gpu_watch_mode = PageWatchMode::ReadWrite)
-      : tracker(page_manager, gpu_watch_mode) {}
+  TrackerHarness() : tracker(page_manager) {}
 
   PageManager page_manager;
   MemoryTracker tracker;
@@ -287,6 +301,124 @@ void TestGpuDirtyBits() {
   Release(page_manager, memory, page_size * 2);
 }
 
+void TestExactDirtyIntervalsSharingTrackerPage() {
+  TrackerHarness harness;
+  auto &tracker = harness.tracker;
+  auto &page_manager = harness.page_manager;
+  const auto page_size = page_manager.GetPageSize();
+  auto *memory = Allocate(page_manager, 1);
+  const auto address = reinterpret_cast<uint64_t>(memory);
+
+  tracker.ForEachUploadRange(
+      address, page_size, false, [](uint64_t, uint64_t) noexcept {},
+      []() noexcept {});
+  RangeSet exact_dirty;
+  exact_dirty.Add(address + 64, 16);
+  exact_dirty.Add(address + 192, 32);
+
+  ResetProtectionLog();
+  tracker.MarkRegionAsGpuModified(address + 64, 16);
+  tracker.MarkRegionAsGpuModified(address + 192, 32);
+  Check(g_protection_calls == 1 &&
+            tracker.IsRegionGpuModified(address, page_size) &&
+            Protection(memory) == PAGE_NOACCESS,
+        "disjoint byte dirtiness duplicated the page watcher");
+
+  exact_dirty.Subtract(address + 64, 16);
+  if (exact_dirty.Intersections(address, page_size).empty()) {
+    tracker.UnmarkRegionAsGpuModified(address, page_size);
+  }
+  Check(g_protection_calls == 1 &&
+            tracker.IsRegionGpuModified(address, page_size) &&
+            Protection(memory) == PAGE_NOACCESS,
+        "draining one exact interval prematurely released its shared page");
+
+  exact_dirty.Subtract(address + 192, 32);
+  if (exact_dirty.Intersections(address, page_size).empty()) {
+    tracker.UnmarkRegionAsGpuModified(address, page_size);
+  }
+  Check(g_protection_calls == 2 &&
+            !tracker.IsRegionGpuModified(address, page_size) &&
+            Protection(memory) == PAGE_READONLY,
+        "draining the final exact interval did not release its tracker page");
+
+  tracker.UntrackMemory(address, page_size);
+  Release(page_manager, memory, page_size);
+}
+
+void TestGpuDownloadProtectionMirrors() {
+  TrackerHarness harness;
+  auto &tracker = harness.tracker;
+  auto &page_manager = harness.page_manager;
+  const auto page_size = page_manager.GetPageSize();
+  auto *memory = Allocate(page_manager, 4);
+  const auto address = reinterpret_cast<uint64_t>(memory);
+
+  tracker.ForEachUploadRange(
+      address, page_size * 4, false, [](uint64_t, uint64_t) noexcept {},
+      []() noexcept {});
+  tracker.MarkRegionAsGpuModified(address + 16, 32);
+  tracker.MarkRegionAsGpuModified(address + page_size * 2 + 16, 32);
+
+  std::vector<RangeSet::Range> visited;
+  ResetProtectionLog();
+  tracker.ForEachDownloadRange<false>(
+      address, page_size * 3,
+      [&](uint64_t range_address, uint64_t range_size) noexcept {
+        visited.push_back({range_address, range_size});
+      });
+  Check(visited.size() == 2 && visited[0].address == address &&
+            visited[0].size == page_size &&
+            visited[1].address == address + page_size * 2 &&
+            visited[1].size == page_size && g_protection_calls == 0 &&
+            tracker.IsRegionGpuModified(address, page_size * 3),
+        "non-clearing download changed protection or lost sparse ranges");
+
+  visited.clear();
+  tracker.ForEachDownloadRange<true>(
+      address + 16, 32,
+      [&](uint64_t range_address, uint64_t range_size) noexcept {
+        visited.push_back({range_address, range_size});
+      });
+  Check(visited.size() == 1 && visited[0].address == address &&
+            visited[0].size == page_size && g_protection_log.size() == 1 &&
+            g_protection_log[0].address == address &&
+            g_protection_log[0].size == page_size &&
+            g_protection_log[0].mode == Common::VirtualMemory::Mode::Read &&
+            !tracker.IsRegionGpuModified(address, page_size) &&
+            tracker.IsRegionGpuModified(address + page_size * 2, page_size) &&
+            Protection(memory) == PAGE_READONLY &&
+            Protection(memory + page_size * 2) == PAGE_NOACCESS,
+        "partial download did not preserve the CPU/GPU protection mirrors");
+
+  visited.clear();
+  ResetProtectionLog();
+  tracker.ForEachDownloadRange<true>(
+      address + 16, 32,
+      [&](uint64_t range_address, uint64_t range_size) noexcept {
+        visited.push_back({range_address, range_size});
+      });
+  Check(visited.empty() && g_protection_calls == 0 &&
+            tracker.IsRegionGpuModified(address + page_size * 2, page_size),
+        "idempotent partial download disturbed another GPU-owned page");
+
+  tracker.UnmarkRegionAsGpuModified(address, page_size * 3);
+  Check(!tracker.IsRegionGpuModified(address, page_size * 3) &&
+            Protection(memory + page_size * 2) == PAGE_READONLY,
+        "broad final unmark did not restore write-only tracking");
+  ResetProtectionLog();
+  tracker.MarkRegionAsCpuModified(address + 16, 32);
+  Check(
+      g_protection_log.size() == 1 && g_protection_log[0].address == address &&
+          g_protection_log[0].size == page_size &&
+          g_protection_log[0].mode == Common::VirtualMemory::Mode::ReadWrite &&
+          IsWritable(memory) && !IsWritable(memory + page_size),
+      "CPU-dirty transition did not release only its write watcher");
+
+  tracker.UntrackMemory(address, page_size * 4);
+  Release(page_manager, memory, page_size * 4);
+}
+
 void TestCrossRegionUpload() {
   constexpr uintptr_t base = 0x0000000200010000ull;
   constexpr uint64_t region_size = 4ull * 1024ull * 1024ull;
@@ -315,35 +447,106 @@ void TestCrossRegionUpload() {
   Release(page_manager, memory, region_size * 2);
 }
 
-void TestBackingWritePublication() {
+void TestGpuUnmarkUsesRegionMask() {
+  constexpr auto region_size = Libs::Graphics::TRACKER_REGION_SIZE;
+  constexpr auto page_size = Libs::Graphics::TRACKER_PAGE_SIZE;
   TrackerHarness harness;
   auto &tracker = harness.tracker;
   auto &page_manager = harness.page_manager;
-  const auto page_size = page_manager.GetPageSize();
-  auto *memory = Allocate(page_manager, 1);
-  const auto address = reinterpret_cast<uint64_t>(memory);
+  auto *memory = Allocate(page_manager, region_size * 2 / page_size);
+  const auto allocation_base = reinterpret_cast<uint64_t>(memory);
+  const auto region_base =
+      (allocation_base + region_size - 1) & ~(region_size - 1);
+  Check(region_base + region_size + page_size <=
+            allocation_base + region_size * 2,
+        "test allocation does not span two complete tracker regions");
+
+  const auto sparse_begin = region_base + page_size;
+  tracker.ForEachUploadRange(
+      sparse_begin, page_size * 3, false, [](uint64_t, uint64_t) noexcept {},
+      []() noexcept {});
+  tracker.MarkRegionAsGpuModified(sparse_begin, page_size);
+  tracker.MarkRegionAsGpuModified(sparse_begin + page_size * 2, page_size);
+  ResetProtectionLog();
+  tracker.UnmarkRegionAsGpuModified(sparse_begin, page_size * 3);
+  Check(
+      g_protection_calls == 1 && g_protection_log.size() == 1 &&
+          g_protection_log[0].address == sparse_begin &&
+          g_protection_log[0].size == page_size * 3 &&
+          g_protection_log[0].mode == Common::VirtualMemory::Mode::Read &&
+          !tracker.IsRegionGpuModified(sparse_begin, page_size * 3) &&
+          Protection(reinterpret_cast<void *>(sparse_begin)) == PAGE_READONLY &&
+          Protection(reinterpret_cast<void *>(sparse_begin + page_size)) ==
+              PAGE_READONLY &&
+          Protection(reinterpret_cast<void *>(sparse_begin + page_size * 2)) ==
+              PAGE_READONLY,
+      "GPU unmark did not coalesce a sparse 4 MiB region mask");
+  ResetProtectionLog();
+  tracker.UnmarkRegionAsGpuModified(sparse_begin, page_size * 3);
+  Check(g_protection_calls == 0,
+        "idempotent GPU unmark performed a protection call");
+
+  const auto boundary = region_base + region_size;
+  const auto cross_begin = boundary - page_size;
+  tracker.ForEachUploadRange(
+      cross_begin, page_size * 2, false, [](uint64_t, uint64_t) noexcept {},
+      []() noexcept {});
+  tracker.MarkRegionAsGpuModified(cross_begin, page_size * 2);
+  ResetProtectionLog();
+  tracker.UnmarkRegionAsGpuModified(cross_begin, page_size * 2);
+  Check(g_protection_calls == 2 && g_protection_log.size() == 2 &&
+            g_protection_log[0].address == cross_begin &&
+            g_protection_log[0].size == page_size &&
+            g_protection_log[0].mode == Common::VirtualMemory::Mode::Read &&
+            g_protection_log[1].address == boundary &&
+            g_protection_log[1].size == page_size &&
+            g_protection_log[1].mode == Common::VirtualMemory::Mode::Read &&
+            !tracker.IsRegionGpuModified(cross_begin, page_size * 2),
+        "cross-region GPU unmark did not use one update per 4 MiB region");
+
+  tracker.UntrackMemory(allocation_base, region_size * 2);
+  Release(page_manager, memory, region_size * 2);
+}
+
+void TestFullRegionGpuUnmarkBatching() {
+  constexpr auto region_size = Libs::Graphics::TRACKER_REGION_SIZE;
+  constexpr auto page_size = Libs::Graphics::TRACKER_PAGE_SIZE;
+  TrackerHarness harness;
+  auto &tracker = harness.tracker;
+  auto &page_manager = harness.page_manager;
+  auto *memory = Allocate(page_manager, region_size * 2 / page_size);
+  const auto allocation_base = reinterpret_cast<uint64_t>(memory);
+  const auto region_base =
+      (allocation_base + region_size - 1) & ~(region_size - 1);
+  Check(region_base + region_size <= allocation_base + region_size * 2,
+        "test allocation does not contain a complete tracker region");
 
   tracker.ForEachUploadRange(
-      address, page_size, true, [](uint64_t, uint64_t) noexcept {},
+      region_base, region_size, false, [](uint64_t, uint64_t) noexcept {},
       []() noexcept {});
-  Check(tracker.IsRegionGpuModified(address, page_size) &&
-            Protection(memory) == PAGE_NOACCESS,
-        "backing publication setup did not establish GPU ownership");
-  std::vector<RangeSet::Range> dirty{{address, page_size}};
-  auto writes = page_manager.ReserveBackingWrites(dirty);
-  Check(writes.size() == 1 && Protection(memory) == PAGE_NOACCESS,
-        "backing reservation exposed protected guest memory");
-  uint32_t downloads = 0;
-  tracker.ForEachDownloadRange<true>(
-      address, page_size, [&](uint64_t, uint64_t) noexcept { downloads++; });
-  tracker.MarkRegionAsCpuModified(address, page_size);
-  writes.clear();
-  Check(downloads == 1 && !tracker.IsRegionGpuModified(address, page_size) &&
-            tracker.IsRegionCpuModified(address, page_size) &&
-            IsWritable(memory),
-        "backing publication did not restore CPU ownership");
-  tracker.UntrackMemory(address, page_size);
-  Release(page_manager, memory, page_size);
+  tracker.MarkRegionAsGpuModified(region_base, region_size);
+  Check(tracker.IsRegionGpuModified(region_base, region_size) &&
+            Protection(reinterpret_cast<void *>(region_base)) ==
+                PAGE_NOACCESS &&
+            Protection(reinterpret_cast<void *>(region_base + region_size -
+                                                page_size)) == PAGE_NOACCESS,
+        "full-region setup did not establish GPU read protection");
+
+  ResetProtectionLog();
+  tracker.UnmarkRegionAsGpuModified(region_base, region_size);
+  Check(
+      g_protection_log.size() == 1 &&
+          g_protection_log[0].address == region_base &&
+          g_protection_log[0].size == region_size &&
+          g_protection_log[0].mode == Common::VirtualMemory::Mode::Read &&
+          !tracker.IsRegionGpuModified(region_base, region_size) &&
+          Protection(reinterpret_cast<void *>(region_base)) == PAGE_READONLY &&
+          Protection(reinterpret_cast<void *>(region_base + region_size -
+                                              page_size)) == PAGE_READONLY,
+      "full-region GPU unmark did not use one exact 4 MiB protection request");
+
+  tracker.UntrackMemory(allocation_base, region_size * 2);
+  Release(page_manager, memory, region_size * 2);
 }
 
 [[noreturn]] void RunDeathCase(const char *name) {
@@ -433,8 +636,11 @@ int main(int argc, char **argv) {
   TestCpuDirtyUpload();
   TestRangeInvalidation();
   TestGpuDirtyBits();
+  TestExactDirtyIntervalsSharingTrackerPage();
+  TestGpuDownloadProtectionMirrors();
   TestCrossRegionUpload();
-  TestBackingWritePublication();
+  TestGpuUnmarkUsesRegionMask();
+  TestFullRegionGpuUnmarkBatching();
   TestFatalPaths();
   std::puts("MemoryTrackerTests: all cases passed");
   return 0;

tests/PageManagerTests.cpp

@@ -26,7 +26,9 @@
 namespace {
 
 using Libs::Graphics::PageManager;
-using Libs::Graphics::PageWatchMode;
+using Libs::Graphics::RegionBits;
+using Libs::Graphics::TRACKER_PAGE_SIZE;
+using Libs::Graphics::TRACKER_REGION_SIZE;
 
 void Check(bool value, const char *text) {
   if (!value) {
@@ -113,6 +115,11 @@ bool IsWritable(const void *address) {
 }
 
 uint64_t g_protection_calls = 0;
+struct ProtectionCall {
+  uint64_t address;
+  uint64_t size;
+};
+std::vector<ProtectionCall> g_protection_ranges;
 
 bool ProtectAddressSpace(uint64_t vaddr, uint64_t size,
                          Common::VirtualMemory::Mode mode) {
@@ -124,6 +131,7 @@ bool ProtectAddressSpace(uint64_t vaddr, uint64_t size,
   }
   DWORD old_protection = 0;
   g_protection_calls++;
+  g_protection_ranges.push_back({vaddr, size});
   return VirtualProtect(reinterpret_cast<void *>(vaddr), size, protection,
                         &old_protection) != 0;
 }
@@ -155,12 +163,12 @@ void TestWatchAndUnwatch() {
   const auto address = reinterpret_cast<uint64_t>(memory);
 
   manager.OnGpuMap(address, page_size * 2);
-  manager.UpdatePageWatchers(true, address, page_size);
+  manager.UpdatePageWatchers<true>(address, page_size);
   Check(Protection(memory) == PAGE_READONLY && IsWritable(memory + page_size),
         "write watch installed incorrect protections");
   Check(g_protection_calls != 0,
         "watch protection bypassed the address-space owner callback");
-  manager.UpdatePageWatchers(false, address, page_size);
+  manager.UpdatePageWatchers<false>(address, page_size);
   Check(IsWritable(memory), "write unwatch did not restore access");
   manager.OnGpuUnmap(address, page_size * 2);
   Check(VirtualFree(memory, 0, MEM_RELEASE) != 0, "VirtualFree failed");
@@ -173,39 +181,17 @@ void TestSharedWatcherCounts() {
   const auto address = reinterpret_cast<uint64_t>(memory);
 
   manager.OnGpuMap(address, page_size);
-  manager.UpdatePageWatchers(true, address + 8, 32);
-  manager.UpdatePageWatchers(true, address + 128, 64);
-  manager.UpdatePageWatchers(false, address + 8, 32);
+  manager.UpdatePageWatchers<true>(address + 8, 32);
+  manager.UpdatePageWatchers<true>(address + 128, 64);
+  manager.UpdatePageWatchers<false>(address + 8, 32);
   Check(Protection(memory) == PAGE_READONLY,
         "first unwatch released a shared watcher");
-  manager.UpdatePageWatchers(false, address + 128, 64);
+  manager.UpdatePageWatchers<false>(address + 128, 64);
   Check(IsWritable(memory), "last unwatch did not restore access");
   manager.OnGpuUnmap(address, page_size);
   Check(VirtualFree(memory, 0, MEM_RELEASE) != 0, "VirtualFree failed");
 }
 
-void TestMixedWatcherModes() {
-  PageManager manager;
-  const auto page_size = manager.GetPageSize();
-  auto *memory = Allocate(page_size);
-  const auto address = reinterpret_cast<uint64_t>(memory);
-
-  manager.OnGpuMap(address, page_size);
-  manager.UpdatePageWatchers(true, address, page_size, PageWatchMode::Write);
-  manager.UpdatePageWatchers(true, address, page_size,
-                             PageWatchMode::ReadWrite);
-  Check(Protection(memory) == PAGE_NOACCESS,
-        "read/write watcher did not deny access");
-  manager.UpdatePageWatchers(false, address, page_size, PageWatchMode::Write);
-  Check(Protection(memory) == PAGE_NOACCESS,
-        "write unwatch released a read/write watcher");
-  manager.UpdatePageWatchers(false, address, page_size,
-                             PageWatchMode::ReadWrite);
-  Check(IsWritable(memory), "read/write unwatch did not restore access");
-  manager.OnGpuUnmap(address, page_size);
-  Check(VirtualFree(memory, 0, MEM_RELEASE) != 0, "VirtualFree failed");
-}
-
 void TestCrossRegionRange() {
   PageManager manager;
   const auto page_size = manager.GetPageSize();
@@ -218,11 +204,27 @@ void TestCrossRegionRange() {
         "test allocation does not contain a region boundary");
 
   manager.OnGpuMap(base, region_size * 2);
-  manager.UpdatePageWatchers(true, boundary - page_size, page_size * 2);
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchers<true>(boundary - page_size, page_size * 2);
+  Check(g_protection_calls == 2 && g_protection_ranges.size() == 2 &&
+            g_protection_ranges[0].address == boundary - page_size &&
+            g_protection_ranges[0].size == page_size &&
+            g_protection_ranges[1].address == boundary &&
+            g_protection_ranges[1].size == page_size,
+        "cross-region watch was not split only at the region boundary");
   Check(!IsWritable(reinterpret_cast<void *>(boundary - page_size)) &&
             !IsWritable(reinterpret_cast<void *>(boundary)),
         "cross-region watch did not protect both pages");
-  manager.UpdatePageWatchers(false, boundary - page_size, page_size * 2);
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchers<false>(boundary - page_size, page_size * 2);
+  Check(g_protection_calls == 2 && g_protection_ranges.size() == 2 &&
+            g_protection_ranges[0].address == boundary - page_size &&
+            g_protection_ranges[0].size == page_size &&
+            g_protection_ranges[1].address == boundary &&
+            g_protection_ranges[1].size == page_size,
+        "cross-region unwatch was not split only at the region boundary");
   Check(IsWritable(reinterpret_cast<void *>(boundary - page_size)) &&
             IsWritable(reinterpret_cast<void *>(boundary)),
         "cross-region unwatch did not restore both pages");
@@ -239,26 +241,26 @@ void TestBatchedWatcherRanges() {
   const auto address = reinterpret_cast<uint64_t>(memory);
 
   manager.OnGpuMap(address, allocation_size);
-  manager.UpdatePageWatchers(true, address + page_size, page_size);
-  manager.UpdatePageWatchers(true, address + page_size * 3, page_size);
-  manager.UpdatePageWatchers(true, address, page_size * 5);
-  manager.UpdatePageWatchers(false, address, page_size * 5);
+  manager.UpdatePageWatchers<true>(address + page_size, page_size);
+  manager.UpdatePageWatchers<true>(address + page_size * 3, page_size);
+  manager.UpdatePageWatchers<true>(address, page_size * 5);
+  manager.UpdatePageWatchers<false>(address, page_size * 5);
   Check(IsWritable(memory) && Protection(memory + page_size) == PAGE_READONLY &&
             IsWritable(memory + page_size * 2) &&
             Protection(memory + page_size * 3) == PAGE_READONLY &&
             IsWritable(memory + page_size * 4),
         "fragmented unwatch lost overlapping watcher counts");
-  manager.UpdatePageWatchers(false, address + page_size, page_size);
-  manager.UpdatePageWatchers(false, address + page_size * 3, page_size);
+  manager.UpdatePageWatchers<false>(address + page_size, page_size);
+  manager.UpdatePageWatchers<false>(address + page_size * 3, page_size);
 
   g_protection_calls = 0;
-  manager.UpdatePageWatchers(true, address, allocation_size);
+  manager.UpdatePageWatchers<true>(address, allocation_size);
   Check(g_protection_calls == 4 && Protection(memory) == PAGE_READONLY &&
             Protection(memory + region_size) == PAGE_READONLY &&
             Protection(memory + region_size * 2) == PAGE_READONLY &&
             Protection(memory + allocation_size - page_size) == PAGE_READONLY,
         "large watch was not batched and protected by tracking region");
-  manager.UpdatePageWatchers(false, address, allocation_size);
+  manager.UpdatePageWatchers<false>(address, allocation_size);
   Check(IsWritable(memory) && IsWritable(memory + region_size) &&
             IsWritable(memory + region_size * 2) &&
             IsWritable(memory + allocation_size - page_size),
@@ -268,20 +270,266 @@ void TestBatchedWatcherRanges() {
   Check(VirtualFree(memory, 0, MEM_RELEASE) != 0, "VirtualFree failed");
 }
 
+void TestRegionMaskWatcherRanges() {
+  PageManager manager;
+  constexpr auto page_size = TRACKER_PAGE_SIZE;
+  constexpr auto region_size = TRACKER_REGION_SIZE;
+  auto *memory = Allocate(region_size * 2);
+  const auto allocation_base = reinterpret_cast<uint64_t>(memory);
+  const auto region_base =
+      (allocation_base + region_size - 1) & ~(region_size - 1);
+  Check(region_base + region_size <= allocation_base + region_size * 2,
+        "test allocation does not contain a complete tracking region");
+  manager.OnGpuMap(allocation_base, region_size * 2);
+
+  RegionBits full_mask;
+  full_mask.Fill();
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<true>(region_base, full_mask);
+  Check(g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+            g_protection_ranges[0].address == region_base &&
+            g_protection_ranges[0].size == region_size,
+        "full region mask did not use one protection span");
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<false>(region_base, full_mask);
+  Check(g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+            g_protection_ranges[0].address == region_base &&
+            g_protection_ranges[0].size == region_size,
+        "full region unmask did not use one protection span");
+
+  RegionBits sparse_mask;
+  sparse_mask.Set(1);
+  sparse_mask.Set(3);
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<true>(region_base, sparse_mask);
+  Check(g_protection_calls == 2 &&
+            Protection(reinterpret_cast<void *>(region_base + page_size)) ==
+                PAGE_READONLY &&
+            IsWritable(reinterpret_cast<void *>(region_base + page_size * 2)) &&
+            Protection(reinterpret_cast<void *>(region_base + page_size * 3)) ==
+                PAGE_READONLY,
+        "sparse region mask installed incorrect write watchers");
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<false>(region_base, sparse_mask);
+  Check(g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+            g_protection_ranges[0].address == region_base + page_size &&
+            g_protection_ranges[0].size == page_size * 3,
+        "sparse unmask did not bridge a compatible gap");
+
+  manager.UpdatePageWatchersForRegion<true>(region_base, sparse_mask);
+  g_protection_calls = 0;
+  manager.UpdatePageWatchersForRegion<true>(region_base, sparse_mask);
+  Check(g_protection_calls == 0,
+        "duplicate sparse watch changed an already protected range");
+  manager.UpdatePageWatchersForRegion<false>(region_base, sparse_mask);
+  Check(g_protection_calls == 0,
+        "first sparse unwatch released a duplicate watcher");
+  manager.UpdatePageWatchersForRegion<false>(region_base, sparse_mask);
+
+  manager.UpdatePageWatchers<true>(region_base + page_size * 2, page_size);
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<true>(region_base, sparse_mask);
+  Check(g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+            g_protection_ranges[0].address == region_base + page_size &&
+            g_protection_ranges[0].size == page_size * 3,
+        "sparse mask did not bridge a compatible protected gap");
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<false>(region_base, sparse_mask);
+  Check(g_protection_calls == 2 && g_protection_ranges.size() == 2 &&
+            g_protection_ranges[0].address == region_base + page_size &&
+            g_protection_ranges[0].size == page_size &&
+            g_protection_ranges[1].address == region_base + page_size * 3 &&
+            g_protection_ranges[1].size == page_size,
+        "sparse unmask crossed an incompatible protected gap");
+  manager.UpdatePageWatchers<false>(region_base + page_size * 2, page_size);
+
+  g_protection_calls = 0;
+  manager.UpdatePageWatchersForRegion<true, true>(region_base, sparse_mask);
+  Check(g_protection_calls == 2 &&
+            Protection(reinterpret_cast<void *>(region_base + page_size)) ==
+                PAGE_NOACCESS &&
+            Protection(reinterpret_cast<void *>(region_base + page_size * 3)) ==
+                PAGE_NOACCESS,
+        "sparse read mask did not deny access");
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<false, true>(region_base, sparse_mask);
+  Check(g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+            g_protection_ranges[0].size == page_size * 3,
+        "sparse read unmask did not bridge a compatible gap");
+
+  manager.OnGpuUnmap(allocation_base, region_size * 2);
+  Check(VirtualFree(memory, 0, MEM_RELEASE) != 0, "VirtualFree failed");
+}
+
+void TestRegionEndpointBatching() {
+  PageManager manager;
+  constexpr auto page_size = TRACKER_PAGE_SIZE;
+  constexpr auto region_size = TRACKER_REGION_SIZE;
+  constexpr auto last_page = region_size / page_size - 1;
+  auto *memory = Allocate(region_size * 2);
+  const auto allocation_base = reinterpret_cast<uint64_t>(memory);
+  const auto region_base =
+      (allocation_base + region_size - 1) & ~(region_size - 1);
+  Check(region_base + region_size <= allocation_base + region_size * 2,
+        "test allocation does not contain a complete tracking region");
+  manager.OnGpuMap(allocation_base, region_size * 2);
+
+  RegionBits endpoints;
+  endpoints.Set(0);
+  endpoints.Set(last_page);
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<true>(region_base, endpoints);
+  Check(g_protection_calls == 2 && g_protection_ranges.size() == 2 &&
+            g_protection_ranges[0].address == region_base &&
+            g_protection_ranges[0].size == page_size &&
+            g_protection_ranges[1].address ==
+                region_base + region_size - page_size &&
+            g_protection_ranges[1].size == page_size,
+        "endpoint watch did not protect only the selected pages");
+
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<false>(region_base, endpoints);
+  Check(g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+            g_protection_ranges[0].address == region_base &&
+            g_protection_ranges[0].size == region_size,
+        "endpoint unwatch did not coalesce the compatible 4 MiB span");
+
+  RegionBits full_mask;
+  full_mask.Fill();
+  manager.UpdatePageWatchersForRegion<true>(region_base, full_mask);
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<true>(region_base, full_mask);
+  Check(g_protection_calls == 0,
+        "duplicate full-region watch issued a redundant protection call");
+  manager.UpdatePageWatchersForRegion<false>(region_base, full_mask);
+  Check(g_protection_calls == 0,
+        "first full-region unwatch released overlapping watcher counts");
+  manager.UpdatePageWatchersForRegion<false>(region_base, full_mask);
+  Check(g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+            g_protection_ranges[0].address == region_base &&
+            g_protection_ranges[0].size == region_size,
+        "last full-region unwatch did not use one 4 MiB protection call");
+
+  manager.OnGpuUnmap(allocation_base, region_size * 2);
+  Check(VirtualFree(memory, 0, MEM_RELEASE) != 0, "VirtualFree failed");
+}
+
+void TestReadWriteWatcherInteractions() {
+  PageManager manager;
+  constexpr auto page_size = TRACKER_PAGE_SIZE;
+  constexpr auto region_size = TRACKER_REGION_SIZE;
+  auto *memory = Allocate(region_size * 2);
+  const auto allocation_base = reinterpret_cast<uint64_t>(memory);
+  const auto region_base =
+      (allocation_base + region_size - 1) & ~(region_size - 1);
+  Check(region_base + region_size <= allocation_base + region_size * 2,
+        "test allocation does not contain a complete tracking region");
+  manager.OnGpuMap(allocation_base, region_size * 2);
+
+  RegionBits write_mask;
+  write_mask.SetRange(10, 15);
+  RegionBits read_mask;
+  read_mask.Set(11);
+  read_mask.Set(13);
+
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<true>(region_base, write_mask);
+  Check(g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+            g_protection_ranges[0].address == region_base + page_size * 10 &&
+            g_protection_ranges[0].size == page_size * 5,
+        "contiguous write watch was not batched");
+
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<true, true>(region_base, read_mask);
+  Check(
+      g_protection_calls == 2 &&
+          Protection(reinterpret_cast<void *>(region_base + page_size * 11)) ==
+              PAGE_NOACCESS &&
+          Protection(reinterpret_cast<void *>(region_base + page_size * 12)) ==
+              PAGE_READONLY &&
+          Protection(reinterpret_cast<void *>(region_base + page_size * 13)) ==
+              PAGE_NOACCESS,
+      "read watchers did not compose with write-only watchers");
+
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<false>(region_base, write_mask);
+  Check(
+      g_protection_calls == 3 &&
+          IsWritable(reinterpret_cast<void *>(region_base + page_size * 10)) &&
+          Protection(reinterpret_cast<void *>(region_base + page_size * 11)) ==
+              PAGE_NOACCESS &&
+          IsWritable(reinterpret_cast<void *>(region_base + page_size * 12)) &&
+          Protection(reinterpret_cast<void *>(region_base + page_size * 13)) ==
+              PAGE_NOACCESS &&
+          IsWritable(reinterpret_cast<void *>(region_base + page_size * 14)),
+      "write unwatch changed pages still owned by read watchers");
+
+  g_protection_calls = 0;
+  g_protection_ranges.clear();
+  manager.UpdatePageWatchersForRegion<false, true>(region_base, read_mask);
+  Check(
+      g_protection_calls == 1 && g_protection_ranges.size() == 1 &&
+          g_protection_ranges[0].address == region_base + page_size * 11 &&
+          g_protection_ranges[0].size == page_size * 3 &&
+          IsWritable(reinterpret_cast<void *>(region_base + page_size * 11)) &&
+          IsWritable(reinterpret_cast<void *>(region_base + page_size * 13)),
+      "read unwatch did not coalesce through compatible writable state");
+
+  manager.OnGpuUnmap(allocation_base, region_size * 2);
+  Check(VirtualFree(memory, 0, MEM_RELEASE) != 0, "VirtualFree failed");
+}
+
 [[noreturn]] void RunDeathCase(const char *name) {
   PageManager manager;
   const auto page_size = manager.GetPageSize();
   if (std::strcmp(name, "invalid-range") == 0) {
-    manager.UpdatePageWatchers(true, (1ull << 40u) - 1, 2);
+    manager.UpdatePageWatchers<true>((1ull << 40u) - 1, 2);
   } else if (std::strcmp(name, "unknown-untrack") == 0) {
-    manager.UpdatePageWatchers(false, 0x1000, page_size);
+    manager.UpdatePageWatchers<false>(0x1000, page_size);
   } else if (std::strcmp(name, "destructor-watch") == 0) {
     auto doomed = std::make_unique<PageManager>();
     auto *memory = Allocate(page_size);
     const auto address = reinterpret_cast<uint64_t>(memory);
     doomed->OnGpuMap(address, page_size);
-    doomed->UpdatePageWatchers(true, address, page_size);
+    doomed->UpdatePageWatchers<true>(address, page_size);
     doomed.reset();
+  } else if (std::strcmp(name, "known-write-underflow") == 0) {
+    auto *memory = Allocate(page_size);
+    const auto address = reinterpret_cast<uint64_t>(memory);
+    manager.OnGpuMap(address, page_size);
+    manager.UpdatePageWatchers<true>(address, page_size);
+    manager.UpdatePageWatchers<false>(address, page_size);
+    manager.UpdatePageWatchers<false>(address, page_size);
+  } else if (std::strcmp(name, "read-overflow") == 0) {
+    auto *memory = Allocate(page_size);
+    const auto address = reinterpret_cast<uint64_t>(memory);
+    manager.OnGpuMap(address, page_size);
+    RegionBits mask;
+    const auto region_base = address & ~(TRACKER_REGION_SIZE - 1);
+    const auto page = static_cast<size_t>((address - region_base) / page_size);
+    mask.Set(page);
+    manager.UpdatePageWatchersForRegion<true, true>(region_base, mask);
+    manager.UpdatePageWatchersForRegion<true, true>(region_base, mask);
+  } else if (std::strcmp(name, "write-overflow") == 0) {
+    auto *memory = Allocate(page_size);
+    const auto address = reinterpret_cast<uint64_t>(memory);
+    manager.OnGpuMap(address, page_size);
+    for (uint32_t count = 0; count < 128; count++) {
+      manager.UpdatePageWatchers<true>(address, page_size);
+    }
   }
   std::_Exit(0x7f);
 }
@@ -328,7 +576,8 @@ void CheckDeathCase(const char *name) {
 
 void TestFatalPaths() {
   for (const char *name :
-       {"invalid-range", "unknown-untrack", "destructor-watch"}) {
+       {"invalid-range", "unknown-untrack", "destructor-watch",
+        "known-write-underflow", "read-overflow", "write-overflow"}) {
     CheckDeathCase(name);
   }
 }
@@ -350,9 +599,11 @@ int main(int argc, char **argv) {
   }
   TestWatchAndUnwatch();
   TestSharedWatcherCounts();
-  TestMixedWatcherModes();
   TestCrossRegionRange();
   TestBatchedWatcherRanges();
+  TestRegionMaskWatcherRanges();
+  TestRegionEndpointBatching();
+  TestReadWriteWatcherInteractions();
   TestFatalPaths();
   std::puts("PageManagerTests: all cases passed");
   return 0;

tests/ResourceTrackingTests.cpp

@@ -1468,6 +1468,32 @@ void TestNativeBindingLayoutOneDimensionalImages() {
 	      "binding allocator did not preserve first-class 1D image groups");
 }
 
+void TestNativeBindingLayoutExpandsDynamicStorageMips() {
+	Program program;
+	program.stage = ShaderType::Compute;
+	program.blocks.resize(1);
+	auto dynamic = ImageUse(8, Opcode::ImageStore, ResourceKind::StorageImageUint,
+	                        Decoder::ImageDimension::Dim2D, 4);
+	dynamic.memory.image_has_mip = true;
+	program.blocks[0].instructions = {
+	    dynamic, ImageUse(12, Opcode::ImageStore, ResourceKind::StorageImageUint,
+	                      Decoder::ImageDimension::Dim2D, 6)};
+	Prepare(program);
+
+	ShaderComputeInputInfo compute;
+	compute.thread_ids_num = 1;
+	std::string error;
+	Check(CollectShaderInfo(program, {.compute = &compute}, &error), error.c_str());
+	Check(program.info.images.size() == 2 &&
+	          program.info.images[0].mip_mode == ImageMipMode::DynamicStorage,
+	      "dynamic storage image was not tracked independently");
+	program.info.images[0].mip_levels = 3;
+	Check(AllocateBindings(program, {}, &error), error.c_str());
+	const auto* storage = FindBinding(program.bindings, DescriptorBindingKind::StorageUint2D);
+	Check(storage != nullptr && storage->resources == std::vector<uint32_t>({0, 0, 0, 1}),
+	      "dynamic storage mip span overlapped the following logical image");
+}
+
 void TestNativeBindingLayoutSrtAndUserDataOverflow() {
 	Program srt;
 	srt.stage = ShaderType::Compute;
@@ -1936,6 +1962,7 @@ int main() {
 		RUN(TestTrackedProgramIsImmutable);
 		RUN(TestNativeBindingLayout);
 		RUN(TestNativeBindingLayoutOneDimensionalImages);
+		RUN(TestNativeBindingLayoutExpandsDynamicStorageMips);
 		RUN(TestNativeBindingLayoutSrtAndUserDataOverflow);
 		RUN(TestNativeBindingLayoutGds);
 		RUN(TestNativeBindingLayoutIsTransactional);

tests/shaderCfgTests.cpp

@@ -5394,7 +5394,224 @@ void TestNewShaderRecompilerCfgSharedOuterAndLoopMerge() {
 	CheckSpirvBinaryValidates(result.spirv);
 }
 
-void TestNewShaderRecompilerCfgLoopSharedContinueSelectionMerges() {
+void TestNewShaderRecompilerCfgLoopEarlyBreakNoSelection() {
+	const uint32_t shader[] = {
+	    EncodeSopc(0x0a, 0, 129),    // loop: s_cmp_lt_u32 s0, 1
+	    EncodeSopp(0x04, 4),         // loop exit -> end
+	    EncodeSopc(0x06, 1, 1),      // s_cmp_eq_u32 s1, s1
+	    EncodeSopp(0x04, 2),         // early break -> same loop end
+	    EncodeSop2(0x00, 0, 0, 129), // s_add_u32 s0, s0, 1
+	    EncodeSopp(0x02, 0xfffau),   // backedge -> loop header
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage   = ShaderType::Compute;
+	options.dump_ir = true;
+
+	ShaderRecompiler::CompileResult result;
+	std::string                     error;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(Common::ContainsStr(result.ir_dump, "mode=structured"),
+	      "loop early-break CFG did not stay on structured path");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 246) != 0,
+	      "loop early-break SPIR-V lacks OpLoopMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 247) == 0,
+	      "loop early-break SPIR-V unexpectedly used OpSelectionMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) == 0,
+	      "loop early-break CFG unexpectedly used dispatcher OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestNewShaderRecompilerCfgNestedLoopNonlocalExitDispatcher() {
+	const uint32_t shader[] = {
+	    EncodeSopc(0x0a, 0, 129),    // outer loop: s_cmp_lt_u32 s0, 1
+	    EncodeSopp(0x04, 9),         // outer exit -> end
+	    EncodeSopc(0x0a, 1, 129),    // inner loop: s_cmp_lt_u32 s1, 1
+	    EncodeSopp(0x04, 5),         // inner exit -> outer continue
+	    EncodeSopc(0x06, 2, 2),      // s_cmp_eq_u32 s2, s2
+	    EncodeSopp(0x05, 5),         // nonlocal exit -> outer end
+	    EncodeSMovB32(3, 129),       // inner work
+	    EncodeSop2(0x00, 1, 1, 129), // s_add_u32 s1, s1, 1
+	    EncodeSopp(0x02, 0xfff9u),   // inner backedge
+	    EncodeSop2(0x00, 0, 0, 129), // outer continue: s_add_u32 s0, s0, 1
+	    EncodeSopp(0x02, 0xfff5u),   // outer backedge
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage   = ShaderType::Compute;
+	options.dump_ir = true;
+
+	ShaderRecompiler::CompileResult result;
+	std::string                     error;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(Common::ContainsStr(result.ir_dump, "mode=dispatcher"),
+	      "nested-loop nonlocal exit did not select dispatcher fallback");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) != 0,
+	      "nested-loop nonlocal exit dispatcher SPIR-V lacks OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestNewShaderRecompilerCfgNestedLoopLocalExitNoSelection() {
+	const uint32_t shader[] = {
+	    EncodeSopc(0x0a, 0, 129),    // outer loop: s_cmp_lt_u32 s0, 1
+	    EncodeSopp(0x04, 6),         // outer exit -> end
+	    EncodeSopc(0x0a, 1, 129),    // inner loop: s_cmp_lt_u32 s1, 1
+	    EncodeSopp(0x04, 2),         // inner exit -> outer continue
+	    EncodeSMovB32(2, 129),       // inner work
+	    EncodeSopp(0x02, 0xfffcu),   // inner backedge
+	    EncodeSop2(0x00, 0, 0, 129), // outer continue: s_add_u32 s0, s0, 1
+	    EncodeSopp(0x02, 0xfff8u),   // outer backedge
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage   = ShaderType::Compute;
+	options.dump_ir = true;
+
+	ShaderRecompiler::CompileResult result;
+	std::string                     error;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(Common::ContainsStr(result.ir_dump, "mode=structured"),
+	      "nested local loop exit did not stay on structured path");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 246) >= 2,
+	      "nested local loop exit SPIR-V lacks both OpLoopMerge instructions");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 247) == 0,
+	      "nested local loop exit SPIR-V unexpectedly used OpSelectionMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) == 0,
+	      "nested local loop exit unexpectedly used dispatcher OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestNewShaderRecompilerCfgNestedLoopExitTailMergeSplit() {
+	const uint32_t shader[] = {
+	    EncodeSopc(0x0a, 0, 129),    // outer loop: s_cmp_lt_u32 s0, 1
+	    EncodeSopp(0x04, 11),        // outer exit -> end
+	    EncodeSopc(0x06, 1, 1),      // inner loop first exit condition
+	    EncodeSopp(0x05, 3),         // first inner exit -> tail A
+	    EncodeSopc(0x06, 2, 2),      // inner loop second exit condition
+	    EncodeSopp(0x05, 3),         // second inner exit -> tail B
+	    EncodeSopp(0x02, 0xfffbu),   // inner backedge
+	    EncodeSMovB32(3, 129),       // tail A
+	    EncodeSopp(0x02, 2),         // tail A -> outer continue
+	    EncodeSMovB32(4, 129),       // tail B
+	    EncodeSopp(0x02, 0),         // tail B -> outer continue
+	    EncodeSop2(0x00, 0, 0, 129), // outer continue: s_add_u32 s0, s0, 1
+	    EncodeSopp(0x02, 0xfff3u),   // outer backedge
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::Decoder::Program program;
+	std::string                        error;
+	Check(ShaderRecompiler::Decoder::DecodeProgram(std::span {shader}, program, &error),
+	      error.c_str());
+
+	ShaderRecompiler::CFG::Graph graph;
+	Check(ShaderRecompiler::CFG::BuildGraph(program, graph, &error), error.c_str());
+	const auto original_block_count = graph.blocks.size();
+	Check(ShaderRecompiler::CFG::Structurize(graph, &error), error.c_str());
+	Check(graph.blocks.size() > original_block_count,
+	      "nested loop exit tails did not create a private inner merge");
+
+	const auto* outer_header = graph.FindBlockByPc(0);
+	const auto* inner_header = graph.FindBlockByPc(8);
+	Check(outer_header != nullptr && inner_header != nullptr &&
+	          outer_header->terminator.loop_header && inner_header->terminator.loop_header,
+	      "nested loop exit-tail fixture did not retain both loop headers");
+	Check(inner_header->terminator.merge_block != outer_header->terminator.continue_block,
+	      "inner loop merge still aliases the outer continue target");
+	const auto* inner_merge = graph.FindBlock(inner_header->terminator.merge_block);
+	Check(inner_merge != nullptr && inner_merge->inst_begin == inner_merge->inst_end &&
+	          inner_merge->terminator.kind == ShaderRecompiler::CFG::TerminatorKind::Branch &&
+	          inner_merge->terminator.true_block == outer_header->terminator.continue_block,
+	      "private inner merge does not forward to the outer continue target");
+}
+
+void TestNewShaderRecompilerCfgMixedContinueNonmergeExitDispatcher() {
+	const uint32_t shader[] = {
+	    EncodeSopc(0x06, 7, 7),      // entry branch bypasses loop -> exit X
+	    EncodeSopp(0x05, 5),         // entry -> X
+	    EncodeSopc(0x0a, 0, 129),    // loop: s_cmp_lt_u32 s0, 1
+	    EncodeSopp(0x04, 5),         // loop exit -> Y
+	    EncodeSopc(0x06, 1, 1),      // inner condition
+	    EncodeSopp(0x05, 1),         // nonmerge exit -> X, else continue
+	    EncodeSopp(0x02, 0xfffbu),   // loop backedge
+	    EncodeSMovB32(2, 129),       // X
+	    EncodeSopp(0x02, 2),         // X -> end
+	    EncodeSMovB32(3, 129),       // Y
+	    EncodeSopp(0x02, 0),         // Y -> end
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage   = ShaderType::Compute;
+	options.dump_ir = true;
+
+	ShaderRecompiler::CompileResult result;
+	std::string                     error;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(Common::ContainsStr(result.ir_dump, "mode=dispatcher"),
+	      "mixed continue/nonmerge exit did not select dispatcher fallback");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) != 0,
+	      "mixed continue/nonmerge exit dispatcher SPIR-V lacks OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestNewShaderRecompilerCfgConditionalLatchNoSelection() {
+	const uint32_t shader[] = {
+	    EncodeSopp(0x02, 0),       // loop header -> conditional block
+	    EncodeSopc(0x06, 0, 0),    // s_cmp_eq_u32 s0, s0
+	    EncodeSopp(0x05, 1),       // loop exit -> end
+	    EncodeSopp(0x02, 0xfffcu), // separate latch -> loop header
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage   = ShaderType::Compute;
+	options.dump_ir = true;
+
+	ShaderRecompiler::CompileResult result;
+	std::string                     error;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(Common::ContainsStr(result.ir_dump, "mode=structured"),
+	      "conditional latch did not stay on structured path");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 246) != 0,
+	      "conditional latch SPIR-V lacks OpLoopMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 247) == 0,
+	      "conditional latch SPIR-V unexpectedly used OpSelectionMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) == 0,
+	      "conditional latch unexpectedly used dispatcher OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestNewShaderRecompilerCfgDirectConditionalLatchNoSelection() {
+	const uint32_t shader[] = {
+	    EncodeSopp(0x02, 0),       // loop header -> conditional latch
+	    EncodeSopc(0x06, 0, 0),    // s_cmp_eq_u32 s0, s0
+	    EncodeSopp(0x05, 0xfffdu), // direct latch backedge -> loop header
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage   = ShaderType::Compute;
+	options.dump_ir = true;
+
+	ShaderRecompiler::CompileResult result;
+	std::string                     error;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(Common::ContainsStr(result.ir_dump, "mode=structured"),
+	      "direct conditional latch did not stay on structured path");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 246) != 0,
+	      "direct conditional latch SPIR-V lacks OpLoopMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 247) == 0,
+	      "direct conditional latch SPIR-V unexpectedly used OpSelectionMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) == 0,
+	      "direct conditional latch unexpectedly used dispatcher OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestNewShaderRecompilerCfgLoopEarlyContinuesNoSelection() {
 	const uint32_t shader[] = {
 	    EncodeSMovB32(0, 128),       // s0 = 0
 	    EncodeSopc(0x0a, 0, 130),    // loop: s_cmp_lt_u32 s0, 2
@@ -5419,15 +5636,111 @@ void TestNewShaderRecompilerCfgLoopSharedContinueSelectionMerges() {
 	std::string                     error;
 	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
 	Check(Common::ContainsStr(result.ir_dump, "mode=structured"),
-	      "shared loop continue selections should stay on structured path");
-	Check(!Common::ContainsStr(result.ir_dump, "duplicate structured merge block"),
-	      "shared loop continue selections were not split before structurization");
-	Check(SpirvContainsOpcode(result.spirv, 246),
-	      "shared loop continue selections SPIR-V lacks OpLoopMerge");
-	Check(SpirvContainsOpcode(result.spirv, 247),
-	      "shared loop continue selections SPIR-V lacks OpSelectionMerge");
-	Check(!SpirvContainsOpcode(result.spirv, 251),
-	      "shared loop continue selections unexpectedly used dispatcher OpSwitch");
+	      "loop early continues should stay on structured path");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 246) != 0,
+	      "loop early continues SPIR-V lacks OpLoopMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 247) == 0,
+	      "loop early continues SPIR-V unexpectedly used OpSelectionMerge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) == 0,
+	      "loop early continues unexpectedly used dispatcher OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestNewShaderRecompilerCfgConditionalLoopHeaderSelection() {
+	const uint32_t shader[] = {
+	    EncodeSopc(0x06, 0, 0),    // loop body selection condition
+	    EncodeSopp(0x05, 2),       // select path B
+	    EncodeSMovB32(1, 129),     // path A
+	    EncodeSopp(0x02, 1),       // path A -> join
+	    EncodeSMovB32(2, 129),     // path B
+	    EncodeSMovB32(3, 129),     // join
+	    EncodeSopc(0x06, 4, 4),    // repeat condition
+	    EncodeSopp(0x05, 0xfff8u), // repeat -> guest header
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::Decoder::Program decoded;
+	std::string                        error;
+	Check(ShaderRecompiler::Decoder::DecodeProgram(std::span {shader}, decoded, &error),
+	      error.c_str());
+	ShaderRecompiler::CFG::Graph graph;
+	Check(ShaderRecompiler::CFG::BuildGraph(decoded, graph, &error), error.c_str());
+	const auto original_block_count = graph.blocks.size();
+	Check(ShaderRecompiler::CFG::Structurize(graph, &error), error.c_str());
+	Check(graph.blocks.size() > original_block_count,
+	      "conditional guest loop header did not create a synthetic header");
+
+	uint32_t loop_headers      = 0;
+	uint32_t selection_headers = 0;
+	for (const auto& block: graph.blocks) {
+		if (block.terminator.loop_header) {
+			loop_headers++;
+			Check(block.inst_begin == block.inst_end &&
+			          block.terminator.kind == ShaderRecompiler::CFG::TerminatorKind::Branch,
+			      "canonical loop header is not an empty unconditional block");
+		} else if (block.terminator.kind ==
+		               ShaderRecompiler::CFG::TerminatorKind::ConditionalBranch &&
+		           block.terminator.merge_block != UINT32_MAX) {
+			selection_headers++;
+		}
+	}
+	Check(loop_headers == 1u && selection_headers == 1u,
+	      "guest conditional was not separated from the loop header");
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage = ShaderType::Compute;
+	ShaderRecompiler::CompileResult result;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(SpirvInstructionOpcodeCount(result.spirv, 246) == 1u,
+	      "conditional loop-header SPIR-V has the wrong loop-merge count");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 247) == 1u,
+	      "conditional loop-header SPIR-V has the wrong selection-merge count");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) == 0u,
+	      "conditional loop-header unexpectedly used dispatcher OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestNewShaderRecompilerCfgMultipleLoopLatches() {
+	const uint32_t shader[] = {
+	    EncodeSopc(0x0a, 0, 129),  // loop condition
+	    EncodeSopp(0x04, 5),       // loop exit -> end
+	    EncodeSopc(0x06, 1, 1),    // early repeat condition
+	    EncodeSopp(0x05, 0xfffcu), // early repeat -> header
+	    EncodeSMovB32(2, 129),     // body
+	    EncodeSMovB32(3, 129),     // body tail
+	    EncodeSopp(0x02, 0xfff9u), // ordinary latch -> header
+	    0xbf810000u,
+	};
+
+	ShaderRecompiler::Decoder::Program decoded;
+	std::string                        error;
+	Check(ShaderRecompiler::Decoder::DecodeProgram(std::span {shader}, decoded, &error),
+	      error.c_str());
+	ShaderRecompiler::CFG::Graph graph;
+	Check(ShaderRecompiler::CFG::BuildGraph(decoded, graph, &error), error.c_str());
+	const auto original_block_count = graph.blocks.size();
+	Check(graph.back_edges.size() == 2u, "multiple-latch fixture lacks two native backedges");
+	Check(ShaderRecompiler::CFG::Structurize(graph, &error), error.c_str());
+	Check(graph.blocks.size() == original_block_count + 1u,
+	      "multiple native latches did not create one synthetic continue");
+	Check(graph.back_edges.size() == 1u && graph.natural_loops.size() == 1u,
+	      "multiple native latches were not coalesced to one SPIR-V backedge");
+	const auto& loop           = graph.natural_loops.front();
+	const auto* continue_block = graph.FindBlock(loop.continue_block);
+	Check(continue_block != nullptr && continue_block->inst_begin == continue_block->inst_end &&
+	          continue_block->predecessors.size() == 2u,
+	      "canonical continue does not join both native latches");
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage = ShaderType::Compute;
+	ShaderRecompiler::CompileResult result;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(SpirvInstructionOpcodeCount(result.spirv, 246) == 1u,
+	      "multiple-latch SPIR-V has the wrong loop-merge count");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 247) == 0u,
+	      "multiple-latch SPIR-V unexpectedly used a selection merge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) == 0u,
+	      "multiple-latch SPIR-V unexpectedly used dispatcher OpSwitch");
 	CheckSpirvBinaryValidates(result.spirv);
 }
 
@@ -5457,6 +5770,85 @@ void TestNewShaderRecompilerCfgDuplicateMergeStructuredSplit() {
 	CheckSpirvBinaryValidates(result.spirv);
 }
 
+void TestNewShaderRecompilerCfgOverlappingEarlyExitLadder() {
+	const uint32_t shader[] = {
+	    EncodeSopc(0x06, 0, 0), // block 0
+	    EncodeSopp(0x04, 2),    // block 0 -> 2 or 1
+	    EncodeSopc(0x06, 1, 1), // block 1
+	    EncodeSopp(0x04, 6),    // block 1 -> 5 or 2
+	    EncodeSopc(0x06, 2, 2), // block 2
+	    EncodeSopp(0x04, 4),    // block 2 -> 5 or 3
+	    EncodeSopc(0x06, 3, 3), // block 3
+	    EncodeSopp(0x04, 2),    // block 3 -> 5 or 4
+	    EncodeSMovB32(4, 129),  // block 4
+	    0xbf810000u,             // block 4 -> 6
+	    EncodeSMovB32(5, 129),  // block 5
+	    0xbf810000u,             // block 5 -> 6
+	};
+
+	ShaderRecompiler::Decoder::Program decoded;
+	std::string                        error;
+	Check(ShaderRecompiler::Decoder::DecodeProgram(std::span {shader}, decoded, &error),
+	      error.c_str());
+	ShaderRecompiler::CFG::Graph graph;
+	Check(ShaderRecompiler::CFG::BuildGraph(decoded, graph, &error), error.c_str());
+	Check(graph.blocks.size() == 7u && graph.blocks[0].successors == std::vector<uint32_t>({1, 2}) &&
+	          graph.blocks[0].terminator.true_block == 2u &&
+	          graph.blocks[0].terminator.false_block == 1u &&
+	          graph.blocks[1].successors == std::vector<uint32_t>({2, 5}) &&
+	          graph.blocks[1].terminator.true_block == 5u &&
+	          graph.blocks[1].terminator.false_block == 2u &&
+	          graph.blocks[2].successors == std::vector<uint32_t>({3, 5}) &&
+	          graph.blocks[2].terminator.true_block == 5u &&
+	          graph.blocks[2].terminator.false_block == 3u &&
+	          graph.blocks[3].successors == std::vector<uint32_t>({4, 5}) &&
+	          graph.blocks[3].terminator.true_block == 5u &&
+	          graph.blocks[3].terminator.false_block == 4u &&
+	          graph.blocks[4].successors == std::vector<uint32_t>({6}) &&
+	          graph.blocks[5].successors == std::vector<uint32_t>({6}),
+	      "overlapping early-exit fixture does not match the observed shader CFG");
+	Check(ShaderRecompiler::CFG::Structurize(graph, &error), error.c_str());
+	std::vector<bool>     reachable(graph.blocks.size());
+	std::vector<uint32_t> pending = {graph.entry_block};
+	while (!pending.empty()) {
+		const auto block_id = pending.back();
+		pending.pop_back();
+		if (reachable[block_id]) {
+			continue;
+		}
+		reachable[block_id] = true;
+		pending.insert(pending.end(), graph.blocks[block_id].successors.begin(),
+		               graph.blocks[block_id].successors.end());
+	}
+	Check(std::all_of(reachable.begin(), reachable.end(), [](bool value) { return value; }),
+	      "overlapping early-exit structurization left unreachable blocks");
+	std::vector<uint32_t> merges;
+	for (const auto& block: graph.blocks) {
+		if (block.terminator.kind == ShaderRecompiler::CFG::TerminatorKind::ConditionalBranch) {
+			Check(block.terminator.merge_block != UINT32_MAX &&
+			          std::find(merges.begin(), merges.end(), block.terminator.merge_block) ==
+			              merges.end(),
+			      "overlapping early-exit structurization retained a shared merge");
+			merges.push_back(block.terminator.merge_block);
+		}
+	}
+
+	ShaderRecompiler::CompileOptions options;
+	options.stage   = ShaderType::Pixel;
+	options.dump_ir = true;
+	ShaderRecompiler::CompileResult result;
+	Check(ShaderRecompiler::TryRecompile(shader, options, result, &error), error.c_str());
+	Check(Common::ContainsStr(result.ir_dump, "mode=structured"),
+	      "overlapping early-exit ladder did not stay on the structured path");
+	Check(!Common::ContainsStr(result.ir_dump, "duplicate structured merge block"),
+	      "overlapping early-exit ladder retained a shared merge");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 247) >= 4u,
+	      "overlapping early-exit ladder lost its selections");
+	Check(SpirvInstructionOpcodeCount(result.spirv, 251) == 0u,
+	      "overlapping early-exit ladder used dispatcher OpSwitch");
+	CheckSpirvBinaryValidates(result.spirv);
+}
+
 void TestNewShaderRecompilerCfgIrreducibleDispatcher() {
 	const uint32_t shader[] = {
 	    EncodeSopp(0x05, 2),       // entry -> B, fallthrough A
@@ -7014,7 +7406,6 @@ int main() {
 	using namespace Libs::Graphics;
 
 	EnsureConfigInitialized();
-
 	TestResourceDescriptorClassification();
 	TestNativeShaderResourceDependencies();
 	TestNormalizedImageContracts();
@@ -7086,8 +7477,18 @@ int main() {
 	TestNewShaderRecompilerCfgLoopHeaderBufferLoadDispatcher();
 	TestNewShaderRecompilerCfgLoopHeaderDsAppendConsumeDispatcher();
 	TestNewShaderRecompilerCfgSharedOuterAndLoopMerge();
-	TestNewShaderRecompilerCfgLoopSharedContinueSelectionMerges();
+	TestNewShaderRecompilerCfgLoopEarlyBreakNoSelection();
+	TestNewShaderRecompilerCfgNestedLoopNonlocalExitDispatcher();
+	TestNewShaderRecompilerCfgNestedLoopLocalExitNoSelection();
+	TestNewShaderRecompilerCfgNestedLoopExitTailMergeSplit();
+	TestNewShaderRecompilerCfgMixedContinueNonmergeExitDispatcher();
+	TestNewShaderRecompilerCfgConditionalLatchNoSelection();
+	TestNewShaderRecompilerCfgDirectConditionalLatchNoSelection();
+	TestNewShaderRecompilerCfgLoopEarlyContinuesNoSelection();
+	TestNewShaderRecompilerCfgConditionalLoopHeaderSelection();
+	TestNewShaderRecompilerCfgMultipleLoopLatches();
 	TestNewShaderRecompilerCfgDuplicateMergeStructuredSplit();
+	TestNewShaderRecompilerCfgOverlappingEarlyExitLadder();
 	TestNewShaderRecompilerCfgIrreducibleDispatcher();
 	TestNewShaderRecompilerExecMaskHelpers();
 	TestComputeShaderInputWaveSize();