graphics: accept video-out buffers with 128-byte independent DCC blocks (Gran Turismo 7) (#1945)

graphics: accept video-out buffers with 128-byte independent DCC blocks

3rdparty/CMakeLists.txt

@@ -42,6 +42,28 @@ if(NOT KYTY_CLANG_CL)
 endif()
 install(FILES "${openxr_SOURCE_DIR}/LICENSE" DESTINATION "licenses/openxr")
 
+# Noto Sans v2.015 (SIL Open Font License 1.1) is libFont's fallback font.
+set(KYTY_NOTO_SANS_REPOSITORY "https://raw.githubusercontent.com/notofonts/notofonts.github.io")
+set(KYTY_NOTO_SANS_COMMIT 42af67ccd12bc0e31ed85384bf73e39248265374)
+set(KYTY_NOTO_SANS_BASE_URL "${KYTY_NOTO_SANS_REPOSITORY}/${KYTY_NOTO_SANS_COMMIT}/fonts")
+set(KYTY_NOTO_SANS_DIR "${FETCHCONTENT_BASE_DIR}/noto-sans")
+function(kyty_fetch_noto_sans path sha256)
+	get_filename_component(name "${path}" NAME)
+	file(DOWNLOAD "${KYTY_NOTO_SANS_BASE_URL}/${path}" "${KYTY_NOTO_SANS_DIR}/${name}"
+		EXPECTED_HASH SHA256=${sha256})
+endfunction()
+kyty_fetch_noto_sans(LICENSE f2095b08bed08b23a6fe26112fcd679a2bee3f002eef077eb05d215ed1051bd8)
+kyty_fetch_noto_sans(NotoSans/hinted/ttf/NotoSans-Regular.ttf
+	478c558ea716033cd60c03438f628dfa75694dcf6b5f6d505a2f05fd2b4f3823)
+kyty_fetch_noto_sans(NotoSans/hinted/ttf/NotoSans-Bold.ttf
+	1df075a380fc7cb898acf64c1f7b3b4dd780de3caa860178bf929de35817a913)
+kyty_fetch_noto_sans(NotoSans/hinted/ttf/NotoSans-Italic.ttf
+	467e3f89eeca4108bb8710a2b9e0cf2281ac56d5b0609211a83776d0505eecb5)
+kyty_fetch_noto_sans(NotoSans/hinted/ttf/NotoSans-BoldItalic.ttf
+	1b602a9d6353be42c91df097a4857b69fa2696f26703d7a33b54a15d87c2622c)
+install(FILES "${KYTY_NOTO_SANS_DIR}/LICENSE" DESTINATION "licenses/noto-sans")
+set(KYTY_NOTO_SANS_DIR "${KYTY_NOTO_SANS_DIR}" PARENT_SCOPE)
+
 # The guest instruction patcher needs a decoder/encoder and a small
 # trampoline assembler. Pin the same revisions used by shadPS4's implementation.
 set(XBYAK_BUILD_TESTS OFF CACHE BOOL "" FORCE)

CMakeLists.txt

@@ -407,6 +407,8 @@ add_executable(scalar_provenance_tests EXCLUDE_FROM_ALL
 target_link_libraries(scalar_provenance_tests fmt::fmt xbyak::xbyak)
 target_include_directories(scalar_provenance_tests PRIVATE ${inc_headers})
 
+add_executable(build_flags_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/BuildFlagsTests.cpp")
+
 add_executable(page_manager_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/PageManagerTests.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/host_gpu/pageManager.cpp"
@@ -650,6 +652,7 @@ if(BUILD_TESTING)
 	add_test(NAME image_page_table COMMAND $<TARGET_FILE:image_page_table_tests>)
 	add_test(NAME memory_tracker COMMAND $<TARGET_FILE:memory_tracker_tests>)
 	add_test(NAME page_manager COMMAND $<TARGET_FILE:page_manager_tests>)
+	add_test(NAME build_flags COMMAND $<TARGET_FILE:build_flags_tests>)
 	add_test(NAME archive_file COMMAND $<TARGET_FILE:archive_file_tests>)
 	add_test(NAME trophy_system COMMAND $<TARGET_FILE:trophy_system_tests>)
 	add_test(NAME avplayer_file COMMAND $<TARGET_FILE:avplayer_file_tests>)
@@ -697,8 +700,6 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --stream-buffer-only)
 	add_test(NAME gpu_command_lane
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --gpu-command-lane-only)
-	add_test(NAME wait_until_safe_for_rendering
-		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --wait-until-safe-for-rendering-only)
 	add_test(NAME pm4_flip
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --flip-only)
 	add_test(NAME depth_bits
@@ -743,6 +744,7 @@ if(BUILD_TESTING)
 		image_page_table_tests
 		memory_tracker_tests
 		page_manager_tests
+		build_flags_tests
 		archive_file_tests
 		trophy_system_tests
 		avplayer_file_tests
@@ -811,6 +813,19 @@ add_custom_target(kyty_touch_cursors
 add_dependencies(kyty_emulator kyty_touch_cursors)
 install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/assets/cursors" DESTINATION assets)
 
+set(KYTY_NOTO_SANS_FONTS
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Regular.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Bold.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Italic.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-BoldItalic.ttf")
+add_custom_target(kyty_noto_sans_fonts
+	COMMAND ${CMAKE_COMMAND} -E make_directory "$<TARGET_FILE_DIR:kyty_emulator>/assets/fonts"
+	COMMAND ${CMAKE_COMMAND} -E copy_if_different ${KYTY_NOTO_SANS_FONTS}
+		"$<TARGET_FILE_DIR:kyty_emulator>/assets/fonts"
+	VERBATIM)
+add_dependencies(kyty_emulator kyty_noto_sans_fonts)
+install(FILES ${KYTY_NOTO_SANS_FONTS} DESTINATION assets/fonts)
+
 target_link_libraries(kyty_emulator ${kyty_emulator_link_libraries})
 if (WIN32)
 	target_link_libraries(kyty_emulator opengl32)

src/common/emulatorConfig.cpp

@@ -105,6 +105,10 @@ bool HideCursorEnabled() {
 	return g_config->hide_cursor_enabled;
 }
 
+bool StretchEnabled() {
+	return g_config->stretch_enabled;
+}
+
 bool VrEnabled() {
 	return g_config->vr_enabled;
 }

src/common/emulatorConfig.h

@@ -64,6 +64,7 @@ struct ConfigOptions {
 	bool                   fullscreen_enabled          = false;
 	bool                   hdr_enabled                 = false;
 	bool                   hide_cursor_enabled         = false;
+	bool                   stretch_enabled             = false;
 	bool                   splash_enabled              = true;
 	bool                   vr_enabled                  = false;
 	bool                   amd_cpu_enabled             = false;
@@ -125,6 +126,7 @@ int32_t GetGpuIndex();
 bool     FullscreenEnabled();
 bool     HdrEnabled();
 bool     HideCursorEnabled();
+bool     StretchEnabled();
 bool     VrEnabled();
 bool     AmdCpuEnabled();
 bool     IsNetworkConnected();

src/emulator.cpp

@@ -99,10 +99,6 @@ static void Init(const Config::ConfigOptions& cfg, const std::filesystem::path&
 
 	if (Common::File::IsFileExisting(param_json)) {
 		Loader::SystemContentLoadParamSfo(param_json);
-		if (const auto flexible_memory_size = Loader::SystemContentGetFlexibleMemorySize();
-		    flexible_memory_size != 0) {
-			Libs::LibKernel::Memory::SetFlexibleMemorySize(flexible_memory_size);
-		}
 	}
 	if (cfg.splash_enabled) {
 		Libs::Graphics::InitializeSplashScreen(param_json.parent_path());

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -98,6 +98,7 @@ public:
 
 	template <typename T>
 	void WaitRegMem(uint32_t func, const T* addr, T ref, T mask, uint32_t poll, uint32_t wait_op);
+	void MemSemaphore(uint64_t* counter, bool signal, bool write);
 	void WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw_num, uint32_t write_control);
 	void WriteReferenceClock(uint64_t dst_address, uint32_t num_bytes);
 	void DmaData(uint8_t engine, uint8_t dst_sel, uint8_t dst_cache_policy,

src/graphics/guest_gpu/command_processor/pm4Dispatch.cpp

@@ -202,6 +202,7 @@ constexpr auto MakeOpcodeDispatchTable() {
 	g_cp_op_func[Pm4::IT_DMA_DATA]                  = CpOpDmaData;
 	g_cp_op_func[Pm4::IT_ACQUIRE_MEM]               = CpOpAcquireMem;
 	g_cp_op_func[Pm4::IT_RELEASE_MEM]               = CpOpReleaseMem;
+	g_cp_op_func[Pm4::IT_MEM_SEMAPHORE]             = CpOpMemSemaphore;
 	g_cp_op_func[Pm4::IT_REWIND]                    = CpOpRewind;
 	g_cp_op_func[Pm4::IT_SET_CONTEXT_REG]           = CpOpSetContextReg;
 	g_cp_op_func[Pm4::IT_SET_SH_REG]                = CpOpSetShaderReg;

src/graphics/guest_gpu/command_processor/pm4Dispatch.h

@@ -136,6 +136,7 @@ uint32_t CpOpEventWrite(CommandProcessor&, uint32_t, const uint32_t*, uint32_t,
 uint32_t CpOpDmaData(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpAcquireMem(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpReleaseMem(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
+uint32_t CpOpMemSemaphore(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetContextReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetShaderReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetUconfigReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -2060,6 +2060,21 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 	return 7;
 }
 
+KYTY_CP_OP_PARSER(CpOpMemSemaphore) {
+	KYTY_PROFILER_FUNCTION();
+
+	const auto control = buffer[2];
+	EXIT_NOT_IMPLEMENTED(cmd_id != 0xc0023900 || (control & ~0xe0110000u) != 0 ||
+	                     (control >> 29u) < 6u);
+
+	auto* counter = reinterpret_cast<uint64_t*>(
+	    (buffer[0] & ~0x7u) | (static_cast<uint64_t>(buffer[1] & 0x3ffffu) << 32u));
+	// A signal updates the counter before the next packet, so its mailbox stall never waits.
+	cp.MemSemaphore(counter, (control >> 29u) == 6u, ((control >> 20u) & 0x1u) != 0);
+
+	return 3;
+}
+
 KYTY_CP_OP_PARSER(CpOpSetContextReg) {
 	KYTY_PROFILER_FUNCTION();
 

src/graphics/guest_gpu/graphicsRun.cpp

@@ -30,7 +30,6 @@
 #include <mutex>
 #include <semaphore>
 #include <thread>
-#include <utility>
 #include <vector>
 
 namespace Libs::Graphics {
@@ -92,12 +91,6 @@ void GuestGpu::SendCommand(Common::UniqueFunction<void>&& command) {
 	m_work_available.Signal();
 }
 
-void GuestGpu::Wake() {
-	Common::LockGuard lock(m_queue_mutex);
-	m_wake_pending = true;
-	m_work_available.Signal();
-}
-
 void GuestGpu::ProcessCommands() {
 	EXIT_IF(!IsGpuThread());
 	while (m_pending_commands.load(std::memory_order_acquire) != 0) {
@@ -281,6 +274,16 @@ template void CommandProcessor::WaitRegMem<uint32_t>(uint32_t, const uint32_t*,
 template void CommandProcessor::WaitRegMem<uint64_t>(uint32_t, const uint64_t*, uint64_t, uint64_t,
                                                      uint32_t, uint32_t);
 
+void CommandProcessor::MemSemaphore(uint64_t* counter, bool signal, bool write) {
+	if (signal) {
+		*counter = write ? 1u : *counter + 1u;
+	} else if (*counter == 0) {
+		SuspendPm4();
+	} else {
+		(*counter)--;
+	}
+}
+
 void CommandProcessor::WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw_num,
                                  uint32_t write_control) {
 	const uint32_t dst_sel = ((write_control >> 30u) & 0x1u) | ((write_control >> 7u) & 0x1eu);
@@ -433,10 +436,7 @@ void GuestGpu::ThreadRun(void* data) {
 				}
 				if (selected_queue < 0) {
 					gpu->m_processing = false;
-					// A wake that came before this wait must not be lost.
-					if (!std::exchange(gpu->m_wake_pending, false)) {
-						gpu->m_work_available.WaitFor(&gpu->m_queue_mutex, 100);
-					}
+					gpu->m_work_available.WaitFor(&gpu->m_queue_mutex, 100);
 					for (auto& queue: gpu->m_queues) {
 						if (!queue.empty()) {
 							queue.front().blocked = false;
@@ -804,6 +804,27 @@ static void PatchIndirectDrawOffsets(CommandProcessor& cp, IndirectDrawRegisters
 	write(registers.draw_index, draw_index);
 }
 
+// Writes the GS user-data slots the packet patches; false when the shaders cannot read the
+// offsets from GPU records: a first-index or subdraw-index destination, or another stage's
+// user data.
+static bool IndirectDrawOffsets(IndirectDrawRegisters registers, ShaderIndirectOffsets& offsets) {
+	if (registers.index_offset != Pm4::SH_NOP || registers.draw_index != Pm4::SH_NOP) {
+		return false;
+	}
+	for (const auto [location, slot]: {std::pair {registers.vertex_offset, &offsets.vertex},
+	                                   std::pair {registers.instance_offset, &offsets.instance}}) {
+		if (location == Pm4::SH_NOP) {
+			continue;
+		}
+		const uint32_t index = location - Pm4::SPI_SHADER_USER_DATA_GS_0;
+		if (index >= HW::UserSgprInfo::SGPRS_MAX) {
+			return false;
+		}
+		*slot = static_cast<uint8_t>(index);
+	}
+	return true;
+}
+
 void CommandProcessor::DrawIndirect(uint32_t data_offset, IndirectDrawRegisters registers,
                                     uint32_t draw_initiator, bool indexed) {
 	const auto args_size =
@@ -835,13 +856,10 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 	// NUM_INSTANCES keeps the last draw's instance count.
 	m_num_instances_addr = args_addr + static_cast<uint64_t>(draw_count - 1) * stride_in_bytes +
 	                       offsetof(vk::DrawIndirectCommand, instanceCount);
-	// With no SGPRs to patch, the GPU reads the records if the shaders allow. Shaders apply
-	// the records' offsets only through those SGPRs; Vulkan applies them, which matches
-	// while they are zero.
-	const DrawIndirectArgs records {args_addr, draw_count, stride_in_bytes};
-	if (registers.vertex_offset == Pm4::SH_NOP && registers.instance_offset == Pm4::SH_NOP &&
-	    registers.index_offset == Pm4::SH_NOP && registers.draw_index == Pm4::SH_NOP &&
-	    (!indexed || m_index_buffer_size != 0) &&
+	// The GPU reads the records if the shaders allow; the VS prolog removes Vulkan's bases
+	// from the index VGPRs and copies them only to the user SGPRs the packet patches.
+	DrawIndirectArgs records {args_addr, draw_count, stride_in_bytes};
+	if (IndirectDrawOffsets(registers, records.offsets) && (!indexed || m_index_buffer_size != 0) &&
 	    (indexed ? executor.DrawIndex(m_submit_id, CurrentBuffer(),
 	                                  {.index_count         = m_index_buffer_size,
 	                                   .index_addr          = IndexAddress(0),
@@ -931,11 +949,10 @@ void CommandProcessor::DrawIndexAuto(DrawAutoArgs args) {
 }
 
 void CommandProcessor::WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index) {
-	// Stall only this command processor, not the host command lane or other queues.
-	if (!m_renderer.GetVideoOut().IsSafeForRendering(static_cast<int>(video_out_handle),
-	                                                 static_cast<int>(display_buffer_index))) {
-		SuspendPm4();
-	}
+	BufferFlush();
+
+	m_renderer.GetVideoOut().WaitFlipDone(static_cast<int>(video_out_handle),
+	                                      static_cast<int>(display_buffer_index));
 }
 
 void CommandProcessor::ReleaseMem(uint32_t data_sel, uint32_t interrupt_selector, void* address,
@@ -966,16 +983,28 @@ void CommandProcessor::ReleaseMem(uint32_t data_sel, uint32_t interrupt_selector
 	// Dumped occlusion counters are written before any end-of-pipe signal reaches the guest.
 	m_renderer.GetOcclusionQueries().Publish();
 
+	// A label beside GPU-written bytes goes to the backing store, so they are not downloaded.
+	// Other labels go through the guest mapping: its fault keeps a clean cached copy current and
+	// downloads GPU-written bytes under the label before the store.
+	const auto write_label = [this, address](const void* value, uint64_t size) {
+		const auto vaddr   = reinterpret_cast<uint64_t>(address);
+		auto&      buffers = m_renderer.GetBufferCache();
+		if (!GuestRange {vaddr, size}.Valid() || !buffers.IsRegionGpuModified(vaddr, size) ||
+		    buffers.HasGpuDirtyBytes(vaddr, size) ||
+		    !LibKernel::Memory::TryWriteBacking(vaddr, value, size)) {
+			std::memcpy(address, value, size);
+		}
+	};
 	// EndOfPipeData: kNone, kImmediate32b, kImmediate64b, kRefClk; 5 reads GDS.
 	switch (data_sel) {
 		case 0: break;
 		case 1: {
 			const auto value = static_cast<uint32_t>(data);
-			std::memcpy(address, &value, sizeof(value));
+			write_label(&value, sizeof(value));
 			break;
 		}
 		case 3: data = Sync::ReadReferenceClock(); [[fallthrough]];
-		case 2: std::memcpy(address, &data, sizeof(data)); break;
+		case 2: write_label(&data, sizeof(data)); break;
 		case 5: {
 			GetScheduler().Finish();
 			const auto gds = static_cast<uint32_t>(data);

src/graphics/guest_gpu/graphicsRun.h

@@ -31,8 +31,6 @@ public:
 	[[nodiscard]] bool IsStopping();
 	void               SendCommand(Common::UniqueFunction<void>&& command);
 	void               SendCommandSync(Common::UniqueFunction<void>&& command);
-	// Retries blocked submissions now instead of after the 100 ms stall timeout.
-	void               Wake();
 
 	// Submitted command memory is borrowed and must remain valid until GPU execution completes.
 	void Submit(std::span<const uint32_t> commands);
@@ -84,7 +82,6 @@ private:
 	bool                                           m_processing        = false;
 	bool                                           m_accepting         = true;
 	bool                                           m_stopping          = false;
-	bool                                           m_wake_pending      = false;
 	bool                                           m_shutdown_complete = false;
 	// Completion callbacks can outlive GuestGpu during renderer shutdown.
 	std::shared_ptr<std::binary_semaphore> m_suspend_point_ready =

src/graphics/host_gpu/rangeSet.h

@@ -5,6 +5,7 @@
 
 #include <algorithm>
 #include <cstdint>
+#include <iterator>
 #include <map>
 
 namespace Libs::Graphics {
@@ -17,14 +18,19 @@ public:
 		if (it != m_ranges.begin() && std::prev(it)->second >= address) {
 			it = std::prev(it);
 		}
-		uint64_t begin = address;
-		uint64_t last  = end;
-		while (it != m_ranges.end() && it->first <= last) {
-			begin = std::min(begin, it->first);
-			last  = std::max(last, it->second);
-			it    = m_ranges.erase(it);
+		if (it == m_ranges.end() || it->first > address) {
+			it = m_ranges.emplace_hint(it, address, end);
+		} else {
+			if (it->second >= end) {
+				return;
+			}
+			it->second = end;
+		}
+		auto next = std::next(it);
+		while (next != m_ranges.end() && next->first <= it->second) {
+			it->second = std::max(it->second, next->second);
+			next       = m_ranges.erase(next);
 		}
-		m_ranges.emplace(begin, last);
 	}
 
 	void Subtract(uint64_t address, uint64_t size) {

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

@@ -8,6 +8,7 @@
 #include <cstddef>
 #include <cstdint>
 #include <cstring>
+#include <iterator>
 #include <memory>
 #include <type_traits>
 #include <utility>
@@ -118,6 +119,13 @@ public:
 
 	class const_iterator final {
 	public:
+		// libstdc++'s algorithms need these. Input, not forward, because operator* returns a copy.
+		using iterator_category = std::input_iterator_tag;
+		using value_type        = OwnerT;
+		using difference_type   = std::ptrdiff_t;
+		using pointer           = void;
+		using reference         = OwnerT;
+
 		[[nodiscard]] OwnerT operator*() const noexcept { return m_owners->At(m_index); }
 		const_iterator&      operator++() noexcept {
 			++m_index;

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

@@ -488,16 +488,15 @@ void TextureCache::RefreshCopySource(ImageId id) {
 }
 
 bool TextureCache::CopyD16(Image& destination, Image& source) {
-	const bool source_depth      = source.info.IsDepth();
-	const bool destination_depth = destination.info.IsDepth();
-	if (source_depth == destination_depth) {
+	// Converts guest D16 between a four-byte host depth aspect and two-byte color or D16 depth.
+	const bool source_wide = DepthAspectTransferBytes(source.backing.format) == sizeof(uint32_t);
+	if (source_wide == (DepthAspectTransferBytes(destination.backing.format) == sizeof(uint32_t))) {
 		return false;
 	}
-	auto&      depth          = source_depth ? source : destination;
-	auto&      color          = source_depth ? destination : source;
-	const auto transfer_bytes = DepthAspectTransferBytes(depth.backing.format);
+	auto& depth  = source_wide ? source : destination;
+	auto& narrow = source_wide ? destination : source;
 	if (depth.info.bytes_per_block != sizeof(uint16_t) ||
-	    color.info.bytes_per_block != sizeof(uint16_t) || transfer_bytes != sizeof(uint32_t)) {
+	    narrow.info.bytes_per_block != sizeof(uint16_t)) {
 		return false;
 	}
 	EXIT_IF(source.backing.samples != 1 || destination.backing.samples != 1 ||
@@ -507,42 +506,42 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 
 	const auto     layers = depth.info.resources.layers;
 	const uint64_t depth_slice =
-	    static_cast<uint64_t>(depth.info.pitch) * depth.info.extent.height * transfer_bytes;
-	const uint64_t color_slice =
-	    static_cast<uint64_t>(color.info.pitch) * color.info.extent.height * sizeof(uint16_t);
-	EXIT_IF(layers == 0 || depth_slice > UINT64_MAX / layers || color_slice > UINT64_MAX / layers);
-	const auto                       depth_size = depth_slice * layers;
-	const auto                       color_size = color_slice * layers;
-	const std::array                 depth_copies {
+	    static_cast<uint64_t>(depth.info.pitch) * depth.info.extent.height * sizeof(uint32_t);
+	const uint64_t narrow_slice =
+	    static_cast<uint64_t>(narrow.info.pitch) * narrow.info.extent.height * sizeof(uint16_t);
+	EXIT_IF(layers == 0 || depth_slice > UINT64_MAX / layers || narrow_slice > UINT64_MAX / layers);
+	const std::array depth_copies {
 	    vk::BufferImageCopy {0,
 	                         depth.info.pitch,
 	                         depth.info.extent.height,
 	                         {vk::ImageAspectFlagBits::eDepth, 0, 0, layers},
 	                         {},
 	                         depth.info.extent}};
-	const std::array color_copies {
+	const std::array narrow_copies {
 	    vk::BufferImageCopy {0,
-	                         color.info.pitch,
-	                         color.info.extent.height,
-	                         {vk::ImageAspectFlagBits::eColor, 0, 0, layers},
+	                         narrow.info.pitch,
+	                         narrow.info.extent.height,
+	                         {narrow.info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
+	                                                : vk::ImageAspectFlagBits::eColor,
+	                          0, 0, layers},
 	                         {},
-	                         color.info.extent}};
+	                         narrow.info.extent}};
 
-	auto                         depth_buffer = m_tiler.GetScratchBuffer(depth_size);
-	auto                         color_buffer = m_tiler.GetScratchBuffer(color_size, depth_buffer.buffer);
+	auto depth_buffer  = m_tiler.GetScratchBuffer(depth_slice * layers);
+	auto narrow_buffer = m_tiler.GetScratchBuffer(narrow_slice * layers, depth_buffer.buffer);
 	const TileManager::D16Layout promote_layout {
 	    .width               = depth.info.extent.width,
 	    .height              = depth.info.extent.height,
 	    .layers              = layers,
-	    .source_row_stride   = static_cast<uint64_t>(color.info.pitch) * sizeof(uint16_t),
-	    .target_row_stride   = static_cast<uint64_t>(depth.info.pitch) * transfer_bytes,
-	    .source_slice_stride = color_slice,
+	    .source_row_stride   = static_cast<uint64_t>(narrow.info.pitch) * sizeof(uint16_t),
+	    .target_row_stride   = static_cast<uint64_t>(depth.info.pitch) * sizeof(uint32_t),
+	    .source_slice_stride = narrow_slice,
 	    .target_slice_stride = depth_slice,
 	};
 	const bool d32 = DepthAspectTransferFormat(depth.backing.format) == vk::Format::eD32Sfloat;
-	if (source_depth) {
+	if (source_wide) {
 		source.Download(depth_copies, depth_buffer.buffer, depth_buffer.offset, depth_buffer.size);
-		m_tiler.ConvertD16(depth_buffer, color_buffer, TileManager::D16Direction::Demote, d32,
+		m_tiler.ConvertD16(depth_buffer, narrow_buffer, TileManager::D16Direction::Demote, d32,
 		                   {.width               = promote_layout.width,
 		                    .height              = promote_layout.height,
 		                    .layers              = promote_layout.layers,
@@ -550,11 +549,12 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 		                    .target_row_stride   = promote_layout.source_row_stride,
 		                    .source_slice_stride = promote_layout.target_slice_stride,
 		                    .target_slice_stride = promote_layout.source_slice_stride});
-		destination.Upload(color_copies, color_buffer.buffer, color_buffer.offset,
-		                   color_buffer.size);
+		destination.Upload(narrow_copies, narrow_buffer.buffer, narrow_buffer.offset,
+		                   narrow_buffer.size);
 	} else {
-		source.Download(color_copies, color_buffer.buffer, color_buffer.offset, color_buffer.size);
-		m_tiler.ConvertD16(color_buffer, depth_buffer, TileManager::D16Direction::Promote, d32,
+		source.Download(narrow_copies, narrow_buffer.buffer, narrow_buffer.offset,
+		                narrow_buffer.size);
+		m_tiler.ConvertD16(narrow_buffer, depth_buffer, TileManager::D16Direction::Promote, d32,
 		                   promote_layout);
 		destination.Upload(depth_copies, depth_buffer.buffer, depth_buffer.offset,
 		                   depth_buffer.size);

src/graphics/host_gpu/renderer/dynamicState.cpp

@@ -125,6 +125,14 @@ void DynamicState::Commit(const GraphicContext& graphics, vk::CommandBuffer cmdb
 			}
 		}
 	}
+	if (dirty_state.cull_mode) {
+		dirty_state.cull_mode = false;
+		cmdbuf.setCullMode(cull_mode);
+	}
+	if (dirty_state.front_face) {
+		dirty_state.front_face = false;
+		cmdbuf.setFrontFace(front_face);
+	}
 	if (dirty_state.blend_constants) {
 		dirty_state.blend_constants = false;
 		cmdbuf.setBlendConstants(blend_constants.data());

src/graphics/host_gpu/renderer/dynamicState.h

@@ -52,6 +52,9 @@ struct DynamicState {
 		bool stencil_back_write_mask    : 1;
 		bool stencil_back_compare_mask  : 1;
 
+		bool cull_mode  : 1;
+		bool front_face : 1;
+
 		bool blend_constants       : 1;
 		bool color_write_enables   : 1;
 		bool line_width            : 1;
@@ -86,6 +89,9 @@ struct DynamicState {
 	uint32_t   stencil_back_write_mask {};
 	uint32_t   stencil_back_compare_mask {};
 
+	vk::CullModeFlags cull_mode {};
+	vk::FrontFace     front_face {};
+
 	std::array<float, 4>                               blend_constants {};
 	std::array<vk::Bool32, RENDER_COLOR_ATTACHMENTS_MAX> color_write_enables {};
 	uint32_t                                           num_color_write_enables {};
@@ -221,6 +227,20 @@ struct DynamicState {
 		}
 	}
 
+	void SetCullMode(const vk::CullModeFlags cull_mode_) {
+		if (cull_mode != cull_mode_) {
+			cull_mode             = cull_mode_;
+			dirty_state.cull_mode = true;
+		}
+	}
+
+	void SetFrontFace(const vk::FrontFace front_face_) {
+		if (front_face != front_face_) {
+			front_face             = front_face_;
+			dirty_state.front_face = true;
+		}
+	}
+
 	void SetBlendConstants(const std::array<float, 4> blend_constants_) {
 		if (blend_constants != blend_constants_) {
 			blend_constants             = blend_constants_;

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

@@ -19,7 +19,13 @@ namespace Libs::Graphics {
 
 enum class ColorTransform : uint32_t { None, SwapBgra16, Reverse10_11_11 };
 
-enum class VideoOutCompression : uint8_t { Uncompressed, Dcc256_256_0, Dcc256_64_64, Unsupported };
+enum class VideoOutCompression : uint8_t {
+	Uncompressed,
+	Dcc256_256_0,
+	Dcc256_64_64,
+	Dcc128_128_128,
+	Unsupported
+};
 
 enum class ImageMetadataKind : uint8_t { None, Htile, Dcc, Cmask };
 
@@ -287,8 +293,9 @@ inline bool ImageInfo::IsDepth() const noexcept {
 [[nodiscard]] inline VideoOutCompression
 ClassifyVideoOutCompression(bool compressed, uint64_t metadata_address, uint32_t dcc_control,
                             uint64_t dcc_clear_color) noexcept {
-	constexpr uint32_t DCC_256_256_0 = 0x00000048u;
-	constexpr uint32_t DCC_256_64_64 = 0x00000208u;
+	constexpr uint32_t DCC_256_256_0   = 0x00000048u;
+	constexpr uint32_t DCC_256_64_64   = 0x00000208u;
+	constexpr uint32_t DCC_128_128_128 = 0x00100024u;
 	if (!compressed) {
 		return metadata_address == 0 && dcc_control == 0 && dcc_clear_color == 0
 		           ? VideoOutCompression::Uncompressed
@@ -300,6 +307,7 @@ ClassifyVideoOutCompression(bool compressed, uint64_t metadata_address, uint32_t
 	switch (dcc_control) {
 		case DCC_256_256_0: return VideoOutCompression::Dcc256_256_0;
 		case DCC_256_64_64: return VideoOutCompression::Dcc256_64_64;
+		case DCC_128_128_128: return VideoOutCompression::Dcc128_128_128;
 		default: return VideoOutCompression::Unsupported;
 	}
 }

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

@@ -484,7 +484,7 @@ void TileManager::ConvertD16(Result source, Result target, D16Direction directio
 	                            vk::AccessFlagBits::eHostWrite |
 	                            vk::AccessFlagBits::eTransferWrite |
 	                            vk::AccessFlagBits::eShaderWrite;
-	barriers[1].dstAccessMask = vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite;
+	barriers[1].dstAccessMask = vk::AccessFlagBits::eShaderWrite;
 	command.pipelineBarrier(
 	    vk::PipelineStageFlagBits::eAllCommands | vk::PipelineStageFlagBits::eHost,
 	    vk::PipelineStageFlagBits::eComputeShader, {}, 0, nullptr, 2, barriers, 0, nullptr);

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

@@ -20,7 +20,7 @@ namespace {
 
 namespace IR = ShaderRecompiler::IR;
 
-// Smaller tables are cheap to materialize on each draw and are often transient.
+// Smaller tables are often written for one draw, so they enter on a second draw of a frame.
 constexpr uint64_t MinTableBytes = 64ull * 1024ull;
 constexpr uint32_t ElementMask   = (1u << IR::ImageHeapKindShift) - 1u;
 
@@ -133,12 +133,16 @@ ImageHeap::~ImageHeap() {
 bool ImageHeap::Acquire(const IR::ImageHeapRequest& request, IR::ImageHeapView& result) {
 	KYTY_PROFILER_FUNCTION();
 	const auto size = request.size & ~uint64_t {3};
-	if (size < MinTableBytes || request.count == 0u || request.step == 0u ||
+	if (request.count == 0u || request.step == 0u ||
 	    ((request.base | request.first | request.step) & 3u) != 0u) {
 		return false;
 	}
 	auto memory = std::ranges::find_if(
 	    m_memories, [&](const auto& candidate) { return candidate->base == request.base; });
+	if (memory == m_memories.end() && size < MinTableBytes &&
+	    m_sightings.try_emplace(request.base, m_draw).first->second == m_draw) {
+		return false;
+	}
 	// A table follows its descriptor's size, as the first program of a draw acquires it.
 	if (memory == m_memories.end() || ((*memory)->size != size && (*memory)->draw != m_draw)) {
 		if (!m_context.IsMapped(request.base, size)) return false;
@@ -593,26 +597,42 @@ void ImageHeap::UnmapMemory(uint64_t address, uint64_t size) {
 	} else {
 		m_unmapped.push_back({address, size});
 	}
-	auto&           scheduler = m_context.GetCommandScheduler();
 	std::scoped_lock lock {m_mutex};
 	std::erase_if(m_memories, [&](const std::unique_ptr<Memory>& memory) {
 		if (address >= memory->base + memory->size || memory->base >= address + size) {
 			return false;
 		}
-		Unwatch(*memory);
-		for (auto& view: memory->views) {
-			std::erase(m_active, view.get());
-			for (const auto& [value, index]: view->elements) {
-				Unreference(m_elements[index]);
-				view->released.push_back(index);
-			}
-			Retire(*view);
-			if (scheduler.Active()) {
-				scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
-			}
-		}
+		Drop(*memory);
 		return true;
 	});
 }
 
+void ImageHeap::RunGarbageCollector() {
+	// A smaller table rebuilt each frame moves to a new base. Larger ones keep their dead words.
+	std::scoped_lock lock {m_mutex};
+	std::erase_if(m_memories, [&](const std::unique_ptr<Memory>& memory) {
+		if (memory->size >= MinTableBytes || memory->draw > m_suspend_draw) return false;
+		Drop(*memory);
+		return true;
+	});
+	m_suspend_draw = m_draw;
+	m_sightings.clear();
+}
+
+void ImageHeap::Drop(Memory& memory) {
+	auto& scheduler = m_context.GetCommandScheduler();
+	Unwatch(memory);
+	for (auto& view: memory.views) {
+		std::erase(m_active, view.get());
+		for (const auto& [value, index]: view->elements) {
+			Unreference(m_elements[index]);
+			view->released.push_back(index);
+		}
+		Retire(*view);
+		if (scheduler.Active()) {
+			scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
+		}
+	}
+}
+
 } // namespace Libs::Graphics

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

@@ -26,7 +26,7 @@ struct ImageHeapTestAccess;
 // of a table owns device slot words, one per table dword, that select an array element. CPU
 // writes to a table fault in its snapshot, which each draw refreshes once; a view takes the
 // changes when a draw first acquires it. Changes to the images behind elements bump
-// Image::heap_epoch.
+// Image::heap_epoch. A smaller table leaves at a suspend point if the frame did not acquire it.
 class ImageHeap final: public ShaderRecompiler::IR::ImageHeapSource {
 public:
 	ImageHeap(RenderContext& context, PageManager& page_manager);
@@ -47,6 +47,8 @@ public:
 	// CPU write fault, on any thread.
 	void InvalidateMemory(uint64_t address, uint64_t size);
 	void UnmapMemory(uint64_t address, uint64_t size);
+	// At a suspend point.
+	void RunGarbageCollector();
 
 private:
 	using DescriptorValue = ShaderRecompiler::IR::DescriptorValue;
@@ -107,6 +109,7 @@ private:
 	template <bool track>
 	void     Watch(const Memory& memory, size_t begin, size_t end);
 	void     Unwatch(const Memory& memory);
+	void     Drop(Memory& memory);
 	void     UpdateProbes(View& view, uint64_t begin, uint64_t end);
 	void     UpdateProbe(View& view, uint64_t probe);
 	void     Assign(View& view, size_t slot, uint32_t entry);
@@ -125,9 +128,11 @@ private:
 	std::mutex                           m_mutex; // Fault-visible Memory state and m_memories.
 	std::vector<std::unique_ptr<Memory>> m_memories;
 	std::vector<View*>                   m_active; // Acquired since the last commit.
-	uint64_t                             m_draw = 1;
+	uint64_t                             m_draw         = 1;
+	uint64_t                             m_suspend_draw = 0; // m_draw at the last suspend point.
 	std::vector<Element>                 m_elements;
 	std::vector<uint32_t>                m_free;
+	std::unordered_map<uint64_t, uint64_t>    m_sightings; // Smaller tables: base, first draw.
 	std::deque<std::pair<uint64_t, uint32_t>> m_retired; // Tick, element.
 	std::vector<std::pair<size_t, size_t>>    m_runs;
 	std::vector<uint32_t>                     m_scratch;

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

@@ -687,6 +687,10 @@ void PipelineCache::GetGraphicsPrograms(const CommandBuffer& command, DrawRender
 		vertex_params[0] = PrepareProgram(vertex_regs, context, user_config, vertex_info[0]);
 	}
 	const bool mesh_active = vertex_info[0].logical_stage == ShaderType::Mesh;
+	if (!tess_active && !mesh_active) {
+		vertex_info[0].indirect_offsets = state.indirect_offsets;
+		vertex_info[0].indirect_draw    = state.indirect_draw;
+	}
 	if (object_primitive_id) {
 		// CxObjectIdMode: register source plus a 32-bit pipelined primitive-ID addition.
 		EXIT_NOT_IMPLEMENTED(sh.m_paClObjprimIdCntl != 2u || mesh_active || tess_active ||
@@ -1003,15 +1007,11 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(const CommandBuffer&
 	}
 	static_params.alpha_to_coverage_enable = ps_active && ctx.GetShaderRegisters().AlphaToMask();
 	const bool rect_list = Prospero::IsRectList(command.GetUserConfig().GetPrimType());
-	static_params.cull_back  = !rect_list && mc.cull_back;
-	static_params.cull_front = !rect_list && mc.cull_front;
-	static_params.face       = mc.face;
 	static_params.provoking_vtx_last = mc.provoking_vtx_last;
 	static_params.polygon_mode =
-	    ResolvePolygonMode(mc, static_params.cull_front, static_params.cull_back);
+	    ResolvePolygonMode(mc, !rect_list && mc.cull_front, !rect_list && mc.cull_back);
 	if (tessellation) {
-		static_params.patch_control_points =
-		    static_cast<uint8_t>(vs_input_info.tess.input_control_points);
+		static_params.patch_control_points = vs_input_info.tess.input_control_points;
 	} else if (!mesh && static_params.topology == vk::PrimitiveTopology::ePatchList) {
 		// Rectangle lists expand through generated tessellation stages.
 		static_params.patch_control_points = 3;

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

@@ -54,15 +54,12 @@ struct PipelineStaticParameters {
 	bool                       depth_clip_enable        = true;
 	vk::PrimitiveTopology      topology                 = vk::PrimitiveTopology::ePointList;
 	// 0 without tessellation, 3 for a rectangle list, else the input control points.
-	uint8_t                    patch_control_points     = 0;
+	uint32_t                   patch_control_points     = 0;
 	bool                       primitive_restart_enable = false;
 	uint32_t                   samples                  = 1;
 	bool                       sample_shading_enable    = false;
 	bool                       alpha_to_coverage_enable = false;
 	uint32_t                   color_mask[RENDER_COLOR_ATTACHMENTS_MAX]           = {};
-	bool                       cull_front                                         = false;
-	bool                       cull_back                                          = false;
-	bool                       face                                               = false;
 	bool                       provoking_vtx_last                                 = false;
 	vk::PolygonMode            polygon_mode                                       = vk::PolygonMode::eFill;
 	uint8_t                    color_srcblend[RENDER_COLOR_ATTACHMENTS_MAX]       = {};
@@ -140,6 +137,9 @@ struct DrawRenderState {
 	bool                                 ps_active                                = false;
 	std::array<ShaderVertexInputInfo, 3> vertex_info;
 	ShaderPixelInputInfo                 ps_input_info;
+	// Input: the user-data slots a GPU-read indirect draw patches.
+	ShaderIndirectOffsets                indirect_offsets;
+	bool                                 indirect_draw = false;
 
 	[[nodiscard]] std::span<const ShaderVertexInputInfo> VertexStages() const {
 		return {vertex_info.data(), vertex_info[1].stage ? 3u : 1u};

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

@@ -219,17 +219,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	vk::PipelineViewportStateCreateInfo viewport_state {};
 	viewport_state.pNext = &depth_clip_control;
 
-	vk::CullModeFlags cull_mode = vk::CullModeFlagBits::eNone;
-	if (static_params.cull_back) {
-		cull_mode |= vk::CullModeFlagBits::eBack;
-	}
-	if (static_params.cull_front) {
-		cull_mode |= vk::CullModeFlagBits::eFront;
-	}
-
-	vk::FrontFace front_face =
-	    (static_params.face ? vk::FrontFace::eClockwise : vk::FrontFace::eCounterClockwise);
-
 	vk::PipelineRasterizationDepthClipStateCreateInfoEXT clip_ext {};
 	clip_ext.depthClipEnable = static_params.depth_clip_enable ? VK_TRUE : VK_FALSE;
 
@@ -250,8 +239,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		provoking_vertex.pNext = rasterizer.pNext;
 		rasterizer.pNext = &provoking_vertex;
 	}
-	rasterizer.cullMode  = cull_mode;
-	rasterizer.frontFace = front_face;
 	rasterizer.polygonMode = static_params.polygon_mode;
 	rasterizer.lineWidth = 1.0f;
 
@@ -347,6 +334,8 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	    vk::DynamicState::eViewportWithCount,
 	    vk::DynamicState::eScissorWithCount,
 	    vk::DynamicState::eLineWidth,
+	    vk::DynamicState::eCullMode,
+	    vk::DynamicState::eFrontFace,
 	    vk::DynamicState::eDepthTestEnable,
 	    vk::DynamicState::eDepthWriteEnable,
 	    vk::DynamicState::eDepthCompareOp,

src/graphics/host_gpu/renderer/render.h

@@ -45,9 +45,10 @@ enum class DrawOffsetSource : uint8_t {
 
 // Nonzero address: the GPU reads count Vulkan indirect draw records, stride bytes apart.
 struct DrawIndirectArgs {
-	uint64_t address = 0;
-	uint32_t count   = 0;
-	uint32_t stride  = 0;
+	uint64_t              address = 0;
+	uint32_t              count   = 0;
+	uint32_t              stride  = 0;
+	ShaderIndirectOffsets offsets;
 };
 
 struct DrawIndexArgs {

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -147,6 +147,7 @@ void RenderContext::RunGarbageCollector() {
 		m_fault_process_pending = false;
 		m_buffer_cache.ProcessFaultBuffer();
 	}
+	m_image_heap.RunGarbageCollector();
 	m_texture_cache.ProcessDownloadImages();
 	m_texture_cache.RunGarbageCollector();
 	m_buffer_cache.RunGarbageCollector();

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -52,7 +52,7 @@ std::pair<int32_t, uint32_t> ResolveDrawOffsets(uint32_t index_offset,
 	                                           const ShaderVertexInputInfo& vs_input_info) {
 	auto     vertex_offset   = static_cast<int32_t>(index_offset);
 	uint32_t instance_offset = 0;
-	if (!vs_input_info.fetch_embedded) {
+	if (!vs_input_info.has_dynamic_vertex_fetch) {
 		return {vertex_offset, instance_offset};
 	}
 
@@ -403,6 +403,20 @@ static void UpdateGraphicsDynamicState(const CommandBuffer& buffer, DynamicState
 		dynamic_state.SetDepthBias(constant_factor, poly_offset.clamp, slope_factor);
 	}
 
+	// Rect lists are drawn without culling.
+	vk::CullModeFlags cull_mode {};
+	if (!Prospero::IsRectList(buffer.GetUserConfig().GetPrimType())) {
+		if (mode.cull_front) {
+			cull_mode |= vk::CullModeFlagBits::eFront;
+		}
+		if (mode.cull_back) {
+			cull_mode |= vk::CullModeFlagBits::eBack;
+		}
+	}
+	dynamic_state.SetCullMode(cull_mode);
+	dynamic_state.SetFrontFace(mode.face ? vk::FrontFace::eClockwise
+	                                     : vk::FrontFace::eCounterClockwise);
+
 	dynamic_state.SetStencilTestEnabled(depth.stencil_test_enable);
 	if (depth.stencil_test_enable) {
 		const auto& front = depth.stencil_front;
@@ -1236,24 +1250,36 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	}
 }
 
-// Vulkan offsets for a prepared draw, or nullopt when its indirect records need CPU counts, as
-// mesh and quad-list draws do. Indirect records carry their own offsets.
+// Vulkan offsets for a prepared draw, or nullopt when its indirect records need the CPU: mesh
+// and quad-list draws for counts, tessellation draws for the index/SGPR prolog. Indirect records
+// carry their own offsets.
 static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&        ucfg,
                                                    const ShaderVertexInputInfo& vs,
                                                    DrawOffsetSource             source,
                                                    int32_t                      first_vertex,
                                                    uint32_t                     first_instance,
                                                    const DrawIndirectArgs&      indirect) {
+	if ((indirect.address != 0 || source == DrawOffsetSource::IndirectArgs) &&
+	    ucfg.GetIndexOffset() != 0) {
+		static std::once_flag warning_once;
+		std::call_once(warning_once, [&] {
+			Log::WriteToConsoleAndLog(fmt::format(
+			    "GPU: indirect draw ignores GE_INDX_OFFSET=0x{:08x}; geometry may be incorrect.\n",
+			    ucfg.GetIndexOffset()));
+		});
+	}
 	if (indirect.address != 0) {
-		if (vs.stage.module->info.stage == ShaderType::Mesh ||
-		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy) {
+		if (!vs.indirect_draw ||
+		    vs.stage.module->info.stage == ShaderType::Mesh ||
+		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy ||
+		    vs.indirect_offsets != indirect.offsets) {
 			return std::nullopt;
 		}
 		return DrawEmitInfo {.indirect = indirect};
 	}
 	if (source == DrawOffsetSource::IndirectArgs) {
 		// Native fetches consume the patched SGPRs; rewritten fetches need Vulkan offsets.
-		return vs.fetch_embedded ? DrawEmitInfo {first_vertex, first_instance} : DrawEmitInfo {};
+		return vs.has_dynamic_vertex_fetch ? DrawEmitInfo {first_vertex, first_instance} : DrawEmitInfo {};
 	}
 	const auto [vertex_offset, instance_offset] = ResolveDrawOffsets(ucfg.GetIndexOffset(), vs);
 	return DrawEmitInfo {vertex_offset + first_vertex, instance_offset};
@@ -1310,7 +1336,8 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	hw_check(buffer);
 
-	DrawRenderState state {};
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets,
+	                       .indirect_draw    = args.indirect.address != 0};
 	if (!GetDrawTopology(ucfg, state.topology)) {
 		return true;
 	}
@@ -1415,7 +1442,8 @@ bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndexAuto,
 	                         args.vertex_count, args.instance_count, args.first_instance};
 
-	DrawRenderState state {};
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets,
+	                       .indirect_draw    = args.indirect.address != 0};
 	if (!GetDrawTopology(ucfg, state.topology)) {
 		ResetBindings();
 		return true;

src/graphics/host_gpu/shaders/gpu_tiler_demote_d16.comp

@@ -1,10 +1,11 @@
 #version 450
+#extension GL_EXT_shader_16bit_storage : require
 
 layout(constant_id = 0) const uint D32 = 0u;
 layout(local_size_x = 64) in;
 
 layout(set = 0, binding = 0, std430) readonly buffer Input { uint data[]; } input_buffer;
-layout(set = 0, binding = 1, std430) buffer Output { uint data[]; } output_buffer;
+layout(set = 0, binding = 1, std430) buffer Output { uint16_t data[]; } output_buffer;
 
 layout(push_constant, std430) uniform Conversion {
 	uint src_base;
@@ -32,9 +33,5 @@ void main() {
 	}
 	uint src = params.src_base + y * params.source_pitch + x * 4u;
 	uint dst = params.dst_base + y * params.target_pitch + x * 2u;
-	uint shift = (dst & 2u) * 8u;
-	uint mask = 0xffffu << shift;
-	uint word = dst >> 2u;
-	atomicAnd(output_buffer.data[word], ~mask);
-	atomicOr(output_buffer.data[word], decode(input_buffer.data[src >> 2u]) << shift);
+	output_buffer.data[dst >> 1u] = uint16_t(decode(input_buffer.data[src >> 2u]));
 }

src/graphics/presentation/performanceOverlay.cpp

@@ -533,16 +533,16 @@ void DrawMemoryAndIo(const Snapshot& shown) {
 		    Bytes(guest.pooled_allocated) + " (" + Bytes(guest.pool_committed) + " committed)");
 		Row("Flexible", Bytes(guest.flexible_used) + " / " + Bytes(guest.flexible_total));
 		Row("Extended", Bytes(guest.automatic_allocated));
-		// Past the pool size the guest sees no page-table entries available.
-		const auto entries = std::max(guest.cpu_page_entries, guest.gpu_page_entries);
+		// Past its total a side sees no page-table entries available.
+		const auto percent = std::max(guest.cpu_page_entries * 100 / guest.cpu_page_entries_total,
+		                              guest.gpu_page_entries * 100 / guest.gpu_page_entries_total);
 		Row("Page table (2 MB)",
-		    Fmt("CPU %llu  GPU %llu  / %llu",
+		    Fmt("CPU %llu / %llu  GPU %llu / %llu",
 		        static_cast<unsigned long long>(guest.cpu_page_entries),
+		        static_cast<unsigned long long>(guest.cpu_page_entries_total),
 		        static_cast<unsigned long long>(guest.gpu_page_entries),
-		        static_cast<unsigned long long>(guest.page_entries_total)),
-		    entries > guest.page_entries_total            ? BAD
-		    : entries * 10 > guest.page_entries_total * 9 ? MEDIUM
-		                                                  : GOOD);
+		        static_cast<unsigned long long>(guest.gpu_page_entries_total)),
+		    LevelColor(static_cast<double>(percent), 90, 100));
 		ImGui::EndTable();
 	}
 	if (BeginSection("Emulator process")) {

src/graphics/presentation/videoOut.cpp

@@ -221,7 +221,7 @@ public:
 	void WaitForSubmitSlot(VideoOutConfig& cfg);
 	bool Flip(uint32_t micros);
 	void GetFlipStatus(VideoOutConfig& cfg, VideoOutFlipStatus& out);
-	bool IsSafeForRendering(VideoOutConfig& cfg, int index);
+	void Wait(VideoOutConfig& cfg, int index);
 
 private:
 	enum class RequestState { Reserved, Recording, Ready, Presenting };
@@ -240,7 +240,6 @@ private:
 		Graphics::Presenter::Frame* frame;
 		bool                        premultiplied_alpha;
 	};
-	void SignalDoneLocked();
 
 	Graphics::Presenter& m_presenter;
 	Common::Mutex        m_mutex;
@@ -251,7 +250,6 @@ private:
 	std::list<Request>   m_cpu_requests;
 	std::list<Request>   m_cancelled_requests;
 	bool                 m_processing      = false;
-	bool                 m_gpu_waiting     = false;
 	uint64_t             m_next_request_id = 1;
 };
 
@@ -955,7 +953,7 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe
 				}
 			}
 		}
-		SignalDoneLocked();
+		m_done_cond_var.SignalAll();
 	}
 	if (cfg.flip_status.flipPendingNum >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
 		Common::DebugCounters::Add(Common::DebugCounters::Counter::FlipsRejected);
@@ -1035,7 +1033,7 @@ void FlipQueue::Cancel(VideoOutConfig& cfg) {
 			it = queue->erase(it);
 		}
 	}
-	SignalDoneLocked();
+	m_done_cond_var.SignalAll();
 	m_submit_slot_cond_var.SignalAll();
 	m_submit_cond_var.SignalAll();
 	m_mutex.Unlock();
@@ -1207,24 +1205,17 @@ void FlipQueue::WaitForSubmitSlot(VideoOutConfig& cfg) {
 	}
 }
 
-void FlipQueue::SignalDoneLocked() {
-	m_done_cond_var.SignalAll();
-	// A stalled GPU cannot drain and shut down without re-checking under m_mutex, so it is alive here.
-	if (m_gpu_waiting) {
-		m_gpu_waiting = false;
-		m_presenter.Renderer().GetGpu().Wake();
-	}
-}
-
-// A display buffer is safe for rendering once it is neither displayed nor waiting to be.
-bool FlipQueue::IsSafeForRendering(VideoOutConfig& cfg, int index) {
+void FlipQueue::Wait(VideoOutConfig& cfg, int index) {
 	Common::LockGuard lock(m_mutex);
-	auto matches = [&cfg, index](const auto& r) { return r.cfg == &cfg && r.index == index; };
-	const bool safe = cfg.flip_status.currentBuffer != index &&
-	                  std::none_of(m_requests.begin(), m_requests.end(), matches) &&
-	                  std::none_of(m_cpu_requests.begin(), m_cpu_requests.end(), matches);
-	m_gpu_waiting = !safe;
-	return safe;
+
+	auto has_request = [this, &cfg, index] {
+		auto matches = [&cfg, index](const auto& r) { return r.cfg == &cfg && r.index == index; };
+		return std::any_of(m_requests.begin(), m_requests.end(), matches) ||
+		       std::any_of(m_cpu_requests.begin(), m_cpu_requests.end(), matches);
+	};
+	while (has_request()) {
+		m_done_cond_var.Wait(&m_mutex);
+	}
 }
 
 bool FlipQueue::Flip(uint32_t micros) {
@@ -1320,7 +1311,7 @@ bool FlipQueue::Flip(uint32_t micros) {
 		it = m_requests.erase(it);
 	}
 	m_processing = false;
-	SignalDoneLocked();
+	m_done_cond_var.SignalAll();
 	m_submit_slot_cond_var.SignalAll();
 	m_mutex.Unlock();
 	for (size_t i = count; i != 0; i--) {
@@ -1703,12 +1694,12 @@ void VideoOutDriver::WaitForSubmitSlot(int handle) {
 	m_impl->GetFlipQueue().WaitForSubmitSlot(*cfg);
 }
 
-bool VideoOutDriver::IsSafeForRendering(int handle, int index) {
+void VideoOutDriver::WaitFlipDone(int handle, int index) {
 	auto* ctx = m_impl->Get(handle);
 	EXIT_IF(ctx == nullptr);
 
 	EXIT_NOT_IMPLEMENTED(!IsValidBufferIndex(index));
-	return m_impl->GetFlipQueue().IsSafeForRendering(*ctx, index);
+	m_impl->GetFlipQueue().Wait(*ctx, index);
 }
 
 KYTY_SYSV_ABI int VideoOutGetFlipStatus(int handle, VideoOutFlipStatus* status) {

src/graphics/presentation/videoOut.h

@@ -38,7 +38,7 @@ public:
 	void CompleteFlip(uint64_t request_id);
 	void SubmitFlipPreparation(uint64_t request_id);
 	void WaitForSubmitSlot(int handle);
-	[[nodiscard]] bool IsSafeForRendering(int handle, int index);
+	void WaitFlipDone(int handle, int index);
 
 	[[nodiscard]] Impl& State() noexcept;
 

src/graphics/presentation/window/swapchain.cpp

@@ -290,7 +290,8 @@ public:
 private:
 	vk::SurfaceFormatKHR SelectFormat();
 	void Destroy();
-	void DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer);
+	void DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer,
+	                 const vk::Rect2D& rect);
 
 	WindowContext&   m_window;
 	vk::SwapchainKHR m_handle = nullptr;
@@ -637,7 +638,23 @@ bool Swapchain::PrepareSystemOverlay() {
 	return m_system_overlay->PrepareFrame(m_extent, m_format, ImageCount());
 }
 
-void Swapchain::DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer) {
+static vk::Rect2D FitAspect(const vk::Extent3D& source, const vk::Extent2D& target) {
+	const uint64_t src_w = std::max(source.width, 1u);
+	const uint64_t src_h = std::max(source.height, 1u);
+	uint32_t       w     = target.width;
+	uint32_t       h     = target.height;
+	if (src_w * h > src_h * w) {
+		h = std::max<uint32_t>(1, static_cast<uint32_t>(src_h * w / src_w));
+	} else {
+		w = std::max<uint32_t>(1, static_cast<uint32_t>(src_w * h / src_h));
+	}
+	return {{static_cast<int32_t>((target.width - w) / 2),
+	         static_cast<int32_t>((target.height - h) / 2)},
+	        {w, h}};
+}
+
+void Swapchain::DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer,
+                            const vk::Rect2D& rect) {
 	auto device = m_window.graphic_ctx.device;
 	if (m_overlay_pipeline == nullptr) {
 		vk::SamplerCreateInfo sampler {};
@@ -756,12 +773,14 @@ void Swapchain::DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& l
 	rendering.colorAttachmentCount = 1;
 	rendering.pColorAttachments    = &attachment;
 	command.beginRendering(&rendering);
-	const vk::Viewport viewport {
-	    0.0f, 0.0f, static_cast<float>(m_extent.width), static_cast<float>(m_extent.height),
-	    0.0f, 1.0f};
-	const vk::Rect2D scissor {{0, 0}, m_extent};
+	const vk::Viewport viewport {static_cast<float>(rect.offset.x),
+	                             static_cast<float>(rect.offset.y),
+	                             static_cast<float>(rect.extent.width),
+	                             static_cast<float>(rect.extent.height),
+	                             0.0f,
+	                             1.0f};
 	command.setViewport(0, 1, &viewport);
-	command.setScissor(0, 1, &scissor);
+	command.setScissor(0, 1, &rect);
 	command.draw(3, 1, 0, 0);
 	command.endRendering();
 }
@@ -799,6 +818,24 @@ void Swapchain::RecordPresentCommands(CommandBuffer& command, Presenter::Frame*
 	                           vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlags {}, 0,
 	                           nullptr, 0, nullptr, 1, &to_transfer);
 
+	const auto* aspect_source = source != nullptr ? source : overlay.frame;
+	const auto  game_rect     = !Config::StretchEnabled() && aspect_source != nullptr
+	                               ? FitAspect(aspect_source->image.extent, m_extent)
+	                               : vk::Rect2D {{0, 0}, m_extent};
+	if (source == nullptr || game_rect.extent != m_extent) {
+		const vk::ClearColorValue black {};
+		vk_command.clearColorImage(m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal,
+		                           &black, 1, &to_transfer.subresourceRange);
+		if (source != nullptr) {
+			vk::ImageMemoryBarrier after_clear = to_transfer;
+			after_clear.srcAccessMask          = vk::AccessFlagBits::eTransferWrite;
+			after_clear.oldLayout              = vk::ImageLayout::eTransferDstOptimal;
+			vk_command.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
+			                           vk::PipelineStageFlagBits::eTransfer,
+			                           vk::DependencyFlags {}, 0, nullptr, 0, nullptr, 1,
+			                           &after_clear);
+		}
+	}
 	if (source != nullptr) {
 		vk::ImageBlit region {};
 		region.srcSubresource.aspectMask     = vk::ImageAspectFlagBits::eColor;
@@ -812,16 +849,14 @@ void Swapchain::RecordPresentCommands(CommandBuffer& command, Presenter::Frame*
 		region.dstSubresource.mipLevel       = 0;
 		region.dstSubresource.baseArrayLayer = 0;
 		region.dstSubresource.layerCount     = 1;
-		region.dstOffsets[1].x               = static_cast<int>(m_extent.width);
-		region.dstOffsets[1].y               = static_cast<int>(m_extent.height);
+		region.dstOffsets[0].x = game_rect.offset.x;
+		region.dstOffsets[0].y = game_rect.offset.y;
+		region.dstOffsets[1].x = game_rect.offset.x + static_cast<int>(game_rect.extent.width);
+		region.dstOffsets[1].y = game_rect.offset.y + static_cast<int>(game_rect.extent.height);
 		region.dstOffsets[1].z               = 1;
 		vk_command.blitImage(source->image.image, vk::ImageLayout::eTransferSrcOptimal,
 		                     m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal, 1,
 		                     &region, vk::Filter::eLinear);
-	} else {
-		const vk::ClearColorValue black {};
-		vk_command.clearColorImage(m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal,
-		                           &black, 1, &to_transfer.subresourceRange);
 	}
 
 	vk::ImageMemoryBarrier to_present {};
@@ -848,7 +883,7 @@ void Swapchain::RecordPresentCommands(CommandBuffer& command, Presenter::Frame*
 	                           &to_present);
 	if (draw_attachment) {
 		if (draw_overlay) {
-			DrawOverlay(vk_command, overlay);
+			DrawOverlay(vk_command, overlay, game_rect);
 			if (draw_system_overlay) {
 				to_present.srcAccessMask = vk::AccessFlagBits::eColorAttachmentWrite;
 				to_present.oldLayout     = vk::ImageLayout::eColorAttachmentOptimal;

src/graphics/presentation/window/vulkanWindow.cpp

@@ -51,6 +51,7 @@ struct VulkanExtensions {
 vk::PhysicalDeviceVulkan11Features WindowContext::RequiredVulkan11Features() noexcept {
 	vk::PhysicalDeviceVulkan11Features features {};
 	features.storageBuffer16BitAccess = VK_TRUE;
+	features.shaderDrawParameters     = VK_TRUE;
 	return features;
 }
 
@@ -278,6 +279,8 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		              required_features12.samplerMirrorClampToEdge);
 		check_feature(features11.storageBuffer16BitAccess, "storageBuffer16BitAccess",
 		              WindowContext::RequiredVulkan11Features().storageBuffer16BitAccess);
+		check_feature(features11.shaderDrawParameters, "shaderDrawParameters",
+		              WindowContext::RequiredVulkan11Features().shaderDrawParameters);
 		check_feature(features12.storageBuffer8BitAccess, "storageBuffer8BitAccess",
 		              required_features12.storageBuffer8BitAccess);
 		check_feature(features12.timelineSemaphore, "timelineSemaphore",

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -814,7 +814,7 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	const auto& cfg = *selected_cfg;
 	Frontend::EmbeddedFetchPlan embedded_fetch;
 	if ((options.stage == ShaderType::Vertex || options.stage == ShaderType::Local) &&
-	    options.input_info.vertex->fetch_embedded) {
+	    options.input_info.vertex->has_dynamic_vertex_fetch) {
 		embedded_fetch = DetectEmbeddedVertexFetch(
 		    decoded, *options.input_info.vertex, options.user_data_base,
 		    static_cast<uint32_t>(options.user_data.size()), options.wave_size);

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

@@ -63,7 +63,8 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 	}
 	if (kind == IR::StageInputKind::VertexIndex || kind == IR::StageInputKind::InstanceIndex ||
 	    kind == IR::StageInputKind::InvocationId || kind == IR::StageInputKind::PrimitiveId ||
-	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId) {
+	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId ||
+	    kind == IR::StageInputKind::BaseVertex || kind == IR::StageInputKind::BaseInstance) {
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpLoad, TypeI32(state), value, variable);

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

@@ -38,7 +38,8 @@ bool HasFlag(const IR::MemoryInfo& mem, uint32_t flag) {
 
 uint32_t AddressU32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::Inst& address,
                     uint32_t component) {
-	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags, component);
+	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags,
+	                                                         mem.image_dimension, component);
 	const auto packed = layout.bit_offset / 32u;
 	if (packed >= address.NumArgs()) return ConstantU32(ctx.state, 0);
 	auto value = ctx.Def(address.Arg(packed));
@@ -55,10 +56,11 @@ uint32_t AddressU32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::
 
 uint32_t AddressF32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::Inst& address,
                     uint32_t component) {
-	const auto value = AddressU32(ctx, mem, address, component);
-	return Decoder::ImageAddressComponentLayout(mem.image_sample_flags, component).bit_width == 16u
-	           ? EmitF16BitsToF32(ctx.state, value)
-	           : Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), value);
+	const auto value  = AddressU32(ctx, mem, address, component);
+	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags,
+	                                                         mem.image_dimension, component);
+	return layout.bit_width == 16u ? EmitF16BitsToF32(ctx.state, value)
+	                               : Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), value);
 }
 
 ImageSampleLayout Layout(const IR::MemoryInfo& mem) {
@@ -811,6 +813,25 @@ uint32_t EmitImageResult(ValueEmitContext& ctx, const IR::Inst& inst, const Emit
 	return mixed ? value : result_bits(mem.resource, value);
 }
 
+// Ordinals outside the table select candidates[0]. Past ordinal 0, a consecutive table is an
+// offset of ordinal - 1, which wraps at 0 so that one unsigned compare checks both ends.
+uint32_t SamplerArrayIndex(EmitterState& state, uint32_t ordinal,
+                           const std::vector<uint32_t>& candidates) {
+	const auto count = static_cast<uint32_t>(candidates.size());
+	const auto gap = std::adjacent_find(candidates.begin() + 1, candidates.end(),
+	    [](uint32_t lhs, uint32_t rhs) { return rhs != lhs + 1u; });
+	if (count == 1u || gap != candidates.end()) {
+		return EmitIndexSwitch(state, ordinal, count, TypeU32(state),
+		                       [&](uint32_t item) { return ConstantU32(state, candidates[item]); });
+	}
+	const auto offset = Binary(state, spv::OpISub, TypeU32(state), ordinal, ConstantU32(state, 1u));
+	const auto in_table = Binary(state, spv::OpULessThan, TypeBool(state), offset,
+	                             ConstantU32(state, count - 1u));
+	const auto index = Binary(state, spv::OpIAdd, TypeU32(state), offset,
+	                          ConstantU32(state, candidates[1]));
+	return Select(state, TypeU32(state), in_table, index, ConstantU32(state, candidates[0]));
+}
+
 struct ImageSampler {
 	uint32_t id = 0;
 	bool dynamic = false;
@@ -846,9 +867,7 @@ std::array<ImageSampler, 2> LoadImageSamplers(ValueEmitContext& ctx, const IR::I
 				selected = EmitIndirectResourceIndex(state, selected,
 				    samplers[index].indirect_mapping_offset,
 				    samplers[index].indirect_search_iterations, 0u);
-			array_index = EmitIndexSwitch(
-			    state, selected, static_cast<uint32_t>(candidates.size()),
-			    TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, candidates[ordinal]); });
+			array_index = SamplerArrayIndex(state, selected, candidates);
 		}
 		sampler = {LoadSamplerDescriptor(state, index, array_index), array_index != 0u};
 	};

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

@@ -291,6 +291,7 @@ inline constexpr auto EmitBufferAtomicIAdd32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicISub32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicSMin32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMin32    = EmitAtomic32;
+inline constexpr auto EmitBufferAtomicUMin64    = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd32     = EmitAtomic32;

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

@@ -585,6 +585,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::ImageAtomicUMin32:
 		case IR::ValueOpcode::ImageAtomicUMin64:
 		case IR::ValueOpcode::BufferAtomicUMin32:
+		case IR::ValueOpcode::BufferAtomicUMin64:
 		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
 		case IR::ValueOpcode::ImageAtomicSMax32:
 		case IR::ValueOpcode::BufferAtomicSMax32:

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

@@ -476,6 +476,8 @@ uint32_t BuiltInForInput(IR::StageInputKind kind) {
 		case IR::StageInputKind::LocalInvocationId: return spv::BuiltInLocalInvocationId;
 		case IR::StageInputKind::LocalInvocationIndex: return spv::BuiltInLocalInvocationIndex;
 		case IR::StageInputKind::GlobalInvocationId: return spv::BuiltInGlobalInvocationId;
+		case IR::StageInputKind::BaseVertex: return spv::BuiltInBaseVertex;
+		case IR::StageInputKind::BaseInstance: return spv::BuiltInBaseInstance;
 		default: return UINT32_MAX;
 	}
 }
@@ -526,7 +528,9 @@ void DefineInputs(EmitterState& state) {
 			case IR::StageInputKind::PrimitiveId:
 			case IR::StageInputKind::InstanceIndex:
 			case IR::StageInputKind::Layer:
-			case IR::StageInputKind::SampleId: type = TypeI32(state); break;
+			case IR::StageInputKind::SampleId:
+			case IR::StageInputKind::BaseVertex:
+			case IR::StageInputKind::BaseInstance: type = TypeI32(state); break;
 			case IR::StageInputKind::WorkgroupId:
 			case IR::StageInputKind::NumWorkgroups:
 			case IR::StageInputKind::LocalInvocationId:
@@ -802,6 +806,10 @@ void DefineModule(EmitterState& state) {
 	if (InputVariableForKind(state, IR::StageInputKind::SampleId) != 0) {
 		state.builder.RequireCapability(spv::CapabilitySampleRateShading);
 	}
+	if (InputVariableForKind(state, IR::StageInputKind::BaseVertex) != 0 ||
+	    InputVariableForKind(state, IR::StageInputKind::BaseInstance) != 0) {
+		state.builder.RequireCapability(spv::CapabilityDrawParameters);
+	}
 	if (state.program.info.image_gather_extended) {
 		state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
 	}

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

@@ -14,12 +14,6 @@ struct MimgSampleInfo {
 	uint32_t    flags    = 0;
 };
 
-struct MimgGatherInfo {
-	uint32_t encoding = 0;
-	Opcode   decoded  = Opcode::UNSUPPORTED;
-	uint32_t flags    = 0;
-};
-
 struct MimgAtomicInfo {
 	uint32_t encoding    = 0;
 	Opcode   decoded     = Opcode::UNSUPPORTED;
@@ -181,23 +175,72 @@ constexpr MimgSampleInfo MIMG_SAMPLE_OPCODE_LIST[] = {
          ImageSampleFlagOffset | ImageSampleFlagAdjust},
 };
 
-constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
-    {0x44u, Opcode::IMAGE_GATHER4_L, ImageSampleFlagLod},
-    {0x47u, Opcode::IMAGE_GATHER4_LZ, ImageSampleFlagLevelZero},
-    {0x48u, Opcode::IMAGE_GATHER4_C, ImageSampleFlagCompare},
-    {0x4cu, Opcode::IMAGE_GATHER4_C_L, ImageSampleFlagCompare | ImageSampleFlagLod},
-    {0x4fu, Opcode::IMAGE_GATHER4_C_LZ,
-     ImageSampleFlagCompare | ImageSampleFlagLevelZero},
-    {0x57u, Opcode::IMAGE_GATHER4_LZ_O,
-     ImageSampleFlagLevelZero | ImageSampleFlagOffset},
-    {0x58u, Opcode::IMAGE_GATHER4_C_O,
-     ImageSampleFlagCompare | ImageSampleFlagOffset},
-    {0x5fu, Opcode::IMAGE_GATHER4_C_LZ_O,
+constexpr MimgSampleInfo MIMG_GATHER_OPCODE_LIST[] = {
+    {0x40u, "image_gather4", 0},
+    {0x41u, "image_gather4_cl", ImageSampleFlagLodClamp},
+    {0x44u, "image_gather4_l", ImageSampleFlagLod},
+    {0x45u, "image_gather4_b", ImageSampleFlagBias},
+    {0x46u, "image_gather4_b_cl", ImageSampleFlagBias | ImageSampleFlagLodClamp},
+    {0x47u, "image_gather4_lz", ImageSampleFlagLevelZero},
+    {0x48u, "image_gather4_c", ImageSampleFlagCompare},
+    {0x49u, "image_gather4_c_cl", ImageSampleFlagCompare | ImageSampleFlagLodClamp},
+    {0x4cu, "image_gather4_c_l", ImageSampleFlagCompare | ImageSampleFlagLod},
+    {0x4du, "image_gather4_c_b", ImageSampleFlagCompare | ImageSampleFlagBias},
+    {0x4eu, "image_gather4_c_b_cl",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp},
+    {0x4fu, "image_gather4_c_lz", ImageSampleFlagCompare | ImageSampleFlagLevelZero},
+    {0x50u, "image_gather4_o", ImageSampleFlagOffset},
+    {0x51u, "image_gather4_cl_o", ImageSampleFlagLodClamp | ImageSampleFlagOffset},
+    {0x54u, "image_gather4_l_o", ImageSampleFlagLod | ImageSampleFlagOffset},
+    {0x55u, "image_gather4_b_o", ImageSampleFlagBias | ImageSampleFlagOffset},
+    {0x56u, "image_gather4_b_cl_o",
+     ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagOffset},
+    {0x57u, "image_gather4_lz_o", ImageSampleFlagLevelZero | ImageSampleFlagOffset},
+    {0x58u, "image_gather4_c_o", ImageSampleFlagCompare | ImageSampleFlagOffset},
+    {0x59u, "image_gather4_c_cl_o",
+     ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagOffset},
+    {0x5cu, "image_gather4_c_l_o",
+     ImageSampleFlagCompare | ImageSampleFlagLod | ImageSampleFlagOffset},
+    {0x5du, "image_gather4_c_b_o",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagOffset},
+    {0x5eu, "image_gather4_c_b_cl_o",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
+         ImageSampleFlagOffset},
+    {0x5fu, "image_gather4_c_lz_o",
      ImageSampleFlagCompare | ImageSampleFlagLevelZero | ImageSampleFlagOffset},
-    {0x61u, Opcode::IMAGE_GATHER4H, ImageSampleFlagGatherHorizontal},
-    {0xc8u, Opcode::IMAGE_GATHER4_C, ImageSampleFlagCompare | ImageSampleFlagAdjust},
-    {0xd8u, Opcode::IMAGE_GATHER4_C_O,
+    {0x61u, "image_gather4h", ImageSampleFlagGatherHorizontal},
+    {0xc0u, "image_gather4_a", ImageSampleFlagAdjust},
+    {0xc1u, "image_gather4_cl_a", ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
+    {0xc5u, "image_gather4_b_a", ImageSampleFlagBias | ImageSampleFlagAdjust},
+    {0xc6u, "image_gather4_b_cl_a",
+     ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
+    {0xc8u, "image_gather4_c_a", ImageSampleFlagCompare | ImageSampleFlagAdjust},
+    {0xc9u, "image_gather4_c_cl_a",
+     ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
+    {0xcdu, "image_gather4_c_b_a",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagAdjust},
+    {0xceu, "image_gather4_c_b_cl_a",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
+         ImageSampleFlagAdjust},
+    {0xd0u, "image_gather4_o_a", ImageSampleFlagOffset | ImageSampleFlagAdjust},
+    {0xd1u, "image_gather4_cl_o_a",
+     ImageSampleFlagLodClamp | ImageSampleFlagOffset | ImageSampleFlagAdjust},
+    {0xd5u, "image_gather4_b_o_a",
+     ImageSampleFlagBias | ImageSampleFlagOffset | ImageSampleFlagAdjust},
+    {0xd6u, "image_gather4_b_cl_o_a",
+     ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagOffset |
+         ImageSampleFlagAdjust},
+    {0xd8u, "image_gather4_c_o_a",
      ImageSampleFlagCompare | ImageSampleFlagOffset | ImageSampleFlagAdjust},
+    {0xd9u, "image_gather4_c_cl_o_a",
+     ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagOffset |
+         ImageSampleFlagAdjust},
+    {0xddu, "image_gather4_c_b_o_a",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagOffset |
+         ImageSampleFlagAdjust},
+    {0xdeu, "image_gather4_c_b_cl_o_a",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
+         ImageSampleFlagOffset | ImageSampleFlagAdjust},
 };
 
 constexpr MimgAtomicInfo MIMG_ATOMIC_OPCODE_LIST[] = {
@@ -219,13 +262,13 @@ constexpr auto MIMG_SAMPLE_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_SAMPLE_OPCO
 constexpr auto MIMG_GATHER_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_GATHER_OPCODE_LIST);
 constexpr auto MIMG_ATOMIC_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_ATOMIC_OPCODE_LIST);
 
-Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgGatherInfo* gather,
+Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgSampleInfo* gather,
                         const MimgAtomicInfo* atomic) {
 	if (sample != nullptr) {
 		return Opcode::IMAGE_SAMPLE;
 	}
 	if (gather != nullptr) {
-		return gather->decoded;
+		return Opcode::IMAGE_GATHER4;
 	}
 	if (atomic != nullptr) {
 		return atomic->decoded;
@@ -244,24 +287,10 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
 	}
 }
 
-uint32_t DecodeMimgSampleFlags(const MimgSampleInfo* sample, const MimgGatherInfo* gather) {
-	if (sample != nullptr) {
-		return sample->flags;
-	}
-	if (gather != nullptr) {
-		return gather->flags;
-	}
-	return 0;
-}
-
 uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
-                                     const MimgSampleInfo* sample, const MimgGatherInfo* gather,
-                                     const MimgAtomicInfo* atomic) {
-	if (sample != nullptr) {
-		return ImageSampleAddressComponents(sample->flags, dimension);
-	}
-	if (gather != nullptr) {
-		return ImageSampleAddressComponents(gather->flags, dimension);
+                                     const MimgSampleInfo* sampled, const MimgAtomicInfo* atomic) {
+	if (sampled != nullptr) {
+		return ImageSampleAddressComponents(sampled->flags, dimension);
 	}
 	if (atomic != nullptr) {
 		return ImageCoordComponents(dimension);
@@ -290,33 +319,31 @@ uint32_t CountDmaskComponents(uint32_t dmask) {
 
 } // namespace
 
-ImageAddressComponent ImageAddressComponentLayout(uint32_t flags, uint32_t component) {
-	const auto width = [flags](uint32_t index) {
-		if ((flags & ImageSampleFlagA16) == 0u) return 32u;
-		uint32_t cursor = 0;
-		if ((flags & ImageSampleFlagOffset) != 0u) {
-			if (index == cursor++) return 32u;
-		}
-		if ((flags & ImageSampleFlagBias) != 0u) {
-			if (index == cursor++) return 16u;
-		}
-		if ((flags & ImageSampleFlagCompare) != 0u && index == cursor) return 32u;
-		return 16u;
-	};
-	uint32_t offset = 0;
-	for (uint32_t index = 0; index < component; index++) {
-		const auto bits = width(index);
-		if (bits == 32u) offset = (offset + 31u) & ~31u;
-		offset += bits;
+ImageAddressComponent ImageAddressComponentLayout(uint32_t flags, ImageDimension dimension,
+                                                  uint32_t component) {
+	if ((flags & ImageSampleFlagA16) == 0u) return {component * 32u, 32u};
+	// A16 offsets, bias and PCF reference each take a whole VGPR, the bias as a half in its low
+	// 16 bits. The other components pack two per VGPR, except that each 1D derivative and the last
+	// two 3D ones take the low half of their own VGPR.
+	const bool     has_bias    = (flags & ImageSampleFlagBias) != 0u;
+	const uint32_t bias_index  = (flags & ImageSampleFlagOffset) != 0u ? 1u : 0u;
+	const uint32_t whole_vgprs = std::popcount(
+	    flags & (ImageSampleFlagOffset | ImageSampleFlagBias | ImageSampleFlagCompare));
+	if (component < whole_vgprs) {
+		return {component * 32u, has_bias && component == bias_index ? 16u : 32u};
 	}
-	const auto bits = width(component);
-	if (bits == 32u) offset = (offset + 31u) & ~31u;
-	return {offset, bits};
+	const uint32_t axes =
+	    (flags & ImageSampleFlagDerivative) != 0u ? ImageGradientComponents(dimension) : 0u;
+	const uint32_t unpaired = axes % 2u * 2u;
+	const uint32_t paired   = axes * 2u - unpaired;
+	const uint32_t index    = component - whole_vgprs;
+	const uint32_t half     = index + std::clamp(index, paired, paired + unpaired) - paired;
+	return {whole_vgprs * 32u + half * 16u, 16u};
 }
 
-uint32_t ImageAddressDwordCount(uint32_t flags, uint32_t components) {
+uint32_t ImageAddressDwordCount(uint32_t flags, ImageDimension dimension, uint32_t components) {
 	if (components == 0u) return 0u;
-	const auto last = ImageAddressComponentLayout(flags, components - 1u);
+	const auto last = ImageAddressComponentLayout(flags, dimension, components - 1u);
 	return (last.bit_offset + last.bit_width + 31u) / 32u;
 }
 
@@ -329,16 +356,17 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	const auto     dimension  = DecodeImageDimension((word0 >> 3u) & 0x7u);
 	const uint32_t word_count = 2u + nsa_dwords;
 
-	const uint32_t vdata  = (word1 >> 8u) & 0xffu;
-	const uint32_t vaddr  = word1 & 0xffu;
-	const uint32_t srsrc  = (word1 >> 16u) & 0x1fu;
-	const uint32_t ssamp  = (word1 >> 21u) & 0x1fu;
-	const bool     r128   = ((word0 >> 15u) & 0x1u) != 0u;
-	const bool     a16    = ((word1 >> 30u) & 0x1u) != 0u;
-	const bool     d16    = ((word1 >> 31u) & 0x1u) != 0u;
-	const auto*    sample = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
-	const auto*    gather = Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
-	const auto*    atomic = Detail::FindOpcode(MIMG_ATOMIC_OPS, opcode);
+	const uint32_t vdata   = (word1 >> 8u) & 0xffu;
+	const uint32_t vaddr   = word1 & 0xffu;
+	const uint32_t srsrc   = (word1 >> 16u) & 0x1fu;
+	const uint32_t ssamp   = (word1 >> 21u) & 0x1fu;
+	const bool     r128    = ((word0 >> 15u) & 0x1u) != 0u;
+	const bool     a16     = ((word1 >> 30u) & 0x1u) != 0u;
+	const bool     d16     = ((word1 >> 31u) & 0x1u) != 0u;
+	const auto*    sample  = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
+	const auto*    gather  = Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
+	const auto*    atomic  = Detail::FindOpcode(MIMG_ATOMIC_OPS, opcode);
+	const auto*    sampled = sample != nullptr ? sample : gather;
 
 	inst.pc                 = pc;
 	inst.word_count         = word_count;
@@ -351,11 +379,11 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	inst.data_dwords        = d16 ? (inst.data_components + 1u) / 2u : inst.data_components;
 	inst.glc                = ((word0 >> 13u) & 1u) != 0;
 	inst.slc                = ((word0 >> 25u) & 1u) != 0;
-	inst.image_sample_flags = DecodeMimgSampleFlags(sample, gather);
+	inst.image_sample_flags = sampled != nullptr ? sampled->flags : 0u;
 	if (a16) {
 		inst.image_sample_flags |= ImageSampleFlagA16;
 	}
-	if ((sample != nullptr || gather != nullptr) && (word0 & (1u << 12u)) != 0u) {
+	if (sampled != nullptr && (word0 & (1u << 12u)) != 0u) {
 		inst.image_sample_flags |= ImageSampleFlagUnnormalized;
 	}
 	inst.image_dimension  = dimension;
@@ -364,8 +392,7 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	for (uint32_t i = 0; i < nsa_dwords * 4u; i++) {
 		inst.image_nsa_addr[i] = (code[word_index + 2u + i / 4u] >> ((i % 4u) * 8u)) & 0xffu;
 	}
-	inst.image_address_components =
-	    DecodeMimgAddressComponents(opcode, dimension, sample, gather, atomic);
+	inst.image_address_components = DecodeMimgAddressComponents(opcode, dimension, sampled, atomic);
 	SetRawWords(inst, code, word_index, word_count);
 
 	if (inst.opcode == Opcode::UNSUPPORTED) {
@@ -389,8 +416,8 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 			}
 		}
 	}
-	const bool supports_d16 = sample != nullptr || gather != nullptr || opcode == 0x00u ||
-	                          opcode == 0x01u || opcode == 0x08u || opcode == 0x09u;
+	const bool supports_d16 = sampled != nullptr || opcode == 0x00u || opcode == 0x01u ||
+	                          opcode == 0x08u || opcode == 0x09u;
 	if (d16 && !supports_d16) {
 		SetUnsupported(inst, Family::MIMG, opcode, "MIMG opcode does not support D16 data");
 	}
@@ -407,9 +434,10 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	inst.src_count = 3;
 }
 
-const char* MimgSampleOpcodeName(uint32_t opcode) {
-	const auto* sample = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
-	return sample != nullptr ? sample->name : nullptr;
+const char* MimgSampledOpcodeName(uint32_t opcode) {
+	const auto* sample  = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
+	const auto* sampled = sample != nullptr ? sample : Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
+	return sampled != nullptr ? sampled->name : nullptr;
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::Decoder

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

@@ -8,15 +8,16 @@ namespace Libs::Graphics::ShaderRecompiler::Decoder {
 void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
                 Instruction& inst);
 
-const char* MimgSampleOpcodeName(uint32_t opcode);
+const char* MimgSampledOpcodeName(uint32_t opcode);
 
 struct ImageAddressComponent {
 	uint32_t bit_offset;
 	uint32_t bit_width;
 };
 
-ImageAddressComponent ImageAddressComponentLayout(uint32_t flags, uint32_t component);
-uint32_t ImageAddressDwordCount(uint32_t flags, uint32_t components);
+ImageAddressComponent ImageAddressComponentLayout(uint32_t flags, ImageDimension dimension,
+                                                  uint32_t component);
+uint32_t ImageAddressDwordCount(uint32_t flags, ImageDimension dimension, uint32_t components);
 
 } // namespace Libs::Graphics::ShaderRecompiler::Decoder
 

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

@@ -22,7 +22,7 @@ constexpr MemoryOpcodeInfo SMEM_OPCODE_LIST[] = {
     {0x04u, Opcode::S_LOAD_DWORDX16, 16, 32},      {0x08u, Opcode::S_BUFFER_LOAD_DWORD, 1, 32},
     {0x09u, Opcode::S_BUFFER_LOAD_DWORDX2, 2, 32}, {0x0au, Opcode::S_BUFFER_LOAD_DWORDX4, 4, 32},
     {0x0bu, Opcode::S_BUFFER_LOAD_DWORDX8, 8, 32}, {0x0cu, Opcode::S_BUFFER_LOAD_DWORDX16, 16, 32},
-    {0x25u, Opcode::S_MEMREALTIME, 2, 32},
+    {0x20u, Opcode::S_DCACHE_INV, 0, 0},           {0x25u, Opcode::S_MEMREALTIME, 2, 32},
 };
 
 constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
@@ -69,6 +69,7 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {0x40u, Opcode::BUFFER_ATOMIC_FMAX, 1, 32},
     {0x50u, Opcode::BUFFER_ATOMIC_SWAP_X2, 2, 32},
     {0x52u, Opcode::BUFFER_ATOMIC_ADD_X2, 2, 32},
+    {0x56u, Opcode::BUFFER_ATOMIC_UMIN_X2, 2, 32},
     {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
     {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
     {0x71u, Opcode::BUFFER_GL0_INV, 0, 0},
@@ -95,14 +96,14 @@ constexpr MemoryOpcodeInfo MTBUF_OPCODE_LIST[] = {
 };
 
 constexpr MemoryOpcodeInfo FLAT_OPCODE_LIST[] = {
-    {0x08u, Opcode::FLAT_LOAD_UBYTE, 1, 8},     {0x09u, Opcode::FLAT_LOAD_SBYTE, 1, 8, true},
-    {0x0au, Opcode::FLAT_LOAD_USHORT, 1, 16},   {0x0bu, Opcode::FLAT_LOAD_SSHORT, 1, 16, true},
-    {0x0cu, Opcode::FLAT_LOAD_DWORD, 1, 32},    {0x0du, Opcode::FLAT_LOAD_DWORDX2, 2, 32},
-    {0x0eu, Opcode::FLAT_LOAD_DWORDX4, 4, 32},  {0x0fu, Opcode::FLAT_LOAD_DWORDX3, 3, 32},
-    {0x18u, Opcode::FLAT_STORE_BYTE, 1, 8},     {0x1au, Opcode::FLAT_STORE_SHORT, 1, 16},
-    {0x1cu, Opcode::FLAT_STORE_DWORD, 1, 32},   {0x1du, Opcode::FLAT_STORE_DWORDX2, 2, 32},
-    {0x1eu, Opcode::FLAT_STORE_DWORDX4, 4, 32}, {0x1fu, Opcode::FLAT_STORE_DWORDX3, 3, 32},
-    {0x24u, Opcode::FLAT_LOAD_SHORT_D16, 1, 16},
+    {0x08u, Opcode::FLAT_LOAD_UBYTE, 1, 8},      {0x09u, Opcode::FLAT_LOAD_SBYTE, 1, 8, true},
+    {0x0au, Opcode::FLAT_LOAD_USHORT, 1, 16},    {0x0bu, Opcode::FLAT_LOAD_SSHORT, 1, 16, true},
+    {0x0cu, Opcode::FLAT_LOAD_DWORD, 1, 32},     {0x0du, Opcode::FLAT_LOAD_DWORDX2, 2, 32},
+    {0x0eu, Opcode::FLAT_LOAD_DWORDX4, 4, 32},   {0x0fu, Opcode::FLAT_LOAD_DWORDX3, 3, 32},
+    {0x18u, Opcode::FLAT_STORE_BYTE, 1, 8},      {0x1au, Opcode::FLAT_STORE_SHORT, 1, 16},
+    {0x1cu, Opcode::FLAT_STORE_DWORD, 1, 32},    {0x1du, Opcode::FLAT_STORE_DWORDX2, 2, 32},
+    {0x1eu, Opcode::FLAT_STORE_DWORDX4, 4, 32},  {0x1fu, Opcode::FLAT_STORE_DWORDX3, 3, 32},
+    {0x24u, Opcode::FLAT_LOAD_SHORT_D16, 1, 16}, {0x25u, Opcode::FLAT_LOAD_SHORT_D16_HI, 1, 16},
 };
 
 constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
@@ -285,6 +286,9 @@ void DecodeSmem(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	if (inst.opcode == Opcode::UNSUPPORTED) {
 		SetUnsupported(inst, Family::SMEM, opcode, "SMEM opcode is not implemented");
 	}
+	if (inst.opcode == Opcode::S_DCACHE_INV) {
+		return;
+	}
 
 	DecodeScalarDestination(sdst, pc, inst.dst);
 	if (inst.opcode == Opcode::S_MEMREALTIME) {
@@ -419,9 +423,10 @@ void DecodeFlat(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	}
 
 	DecodeVectorGpr(IsFlatStoreOpcode(inst.opcode) ? data : vdst, inst.dst);
-	if (inst.opcode == Opcode::FLAT_LOAD_SHORT_D16) {
-		// D16 loads reuse partial destinations to preserve the untouched high half.
-		inst.dst.sdwa_sel = 4u;
+	if (inst.opcode == Opcode::FLAT_LOAD_SHORT_D16 ||
+	    inst.opcode == Opcode::FLAT_LOAD_SHORT_D16_HI) {
+		// D16 loads reuse partial destinations to preserve the untouched half.
+		inst.dst.sdwa_sel = inst.opcode == Opcode::FLAT_LOAD_SHORT_D16_HI ? 5u : 4u;
 	}
 	DecodeVectorGpr(addr, inst.src0);
 	inst.src_count = 1;

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

@@ -123,10 +123,9 @@ std::string ImageSampleFlagsToString(uint32_t flags) {
 }
 
 std::string FormatMimg(const Instruction& inst) {
-	const char* sample_name =
-	    inst.opcode == Opcode::IMAGE_SAMPLE ? MimgSampleOpcodeName(inst.opcode_id) : nullptr;
+	const char*            sampled_name = MimgSampledOpcodeName(inst.opcode_id);
 	const std::string_view name =
-	    sample_name != nullptr ? sample_name : magic_enum::enum_name(inst.opcode);
+	    sampled_name != nullptr ? sampled_name : magic_enum::enum_name(inst.opcode);
 	std::string text =
 	    fmt::format("0x{:08x}: {} {}, {}, {}, {} ; dmask=0x{:x} image_dim={}", inst.pc, name,
 	                OperandToString(inst.dst).c_str(), OperandToString(inst.src0).c_str(),
@@ -138,22 +137,10 @@ std::string FormatMimg(const Instruction& inst) {
 	if (inst.image_r128) {
 		text += " r128=1";
 	}
-	switch (inst.opcode) {
-		case Opcode::IMAGE_SAMPLE:
-		case Opcode::IMAGE_GATHER4_L:
-		case Opcode::IMAGE_GATHER4_LZ:
-		case Opcode::IMAGE_GATHER4_C:
-		case Opcode::IMAGE_GATHER4_C_L:
-		case Opcode::IMAGE_GATHER4_C_LZ:
-		case Opcode::IMAGE_GATHER4_LZ_O:
-		case Opcode::IMAGE_GATHER4_C_O:
-		case Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Opcode::IMAGE_GATHER4H:
-			text += fmt::format(" sample_flags={} addr_components={}",
-			                    ImageSampleFlagsToString(inst.image_sample_flags).c_str(),
-			                    inst.image_address_components);
-			break;
-		default: break;
+	if (sampled_name != nullptr) {
+		text += fmt::format(" sample_flags={} addr_components={}",
+		                    ImageSampleFlagsToString(inst.image_sample_flags).c_str(),
+		                    inst.image_address_components);
 	}
 	if (inst.image_nsa_dwords != 0) {
 		text += fmt::format(" nsa_dwords={} nsa_addr=", inst.image_nsa_dwords);
@@ -570,6 +557,7 @@ std::string InstructionToString(const Instruction& inst) {
 			                                               OperandToString(inst.src0).c_str()));
 		case Opcode::S_BARRIER:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: s_barrier", inst.pc));
+		case Opcode::S_DCACHE_INV:
 		case Opcode::BUFFER_GL0_INV:
 		case Opcode::BUFFER_GL1_INV:
 			return fmt::format("0x{:08x}: {}", inst.pc, magic_enum::enum_name(inst.opcode));
@@ -616,15 +604,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::IMAGE_LOAD_MIP:
 		case Opcode::IMAGE_GET_RESINFO:
 		case Opcode::IMAGE_GET_LOD:
-		case Opcode::IMAGE_GATHER4_L:
-		case Opcode::IMAGE_GATHER4_LZ:
-		case Opcode::IMAGE_GATHER4_C:
-		case Opcode::IMAGE_GATHER4_C_L:
-		case Opcode::IMAGE_GATHER4_C_LZ:
-		case Opcode::IMAGE_GATHER4_LZ_O:
-		case Opcode::IMAGE_GATHER4_C_O:
-		case Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Opcode::IMAGE_GATHER4H: return WithUnsupportedReason(inst, FormatMimg(inst));
+		case Opcode::IMAGE_GATHER4: return WithUnsupportedReason(inst, FormatMimg(inst));
 		case Opcode::S_LOAD_DWORD:
 		case Opcode::S_LOAD_DWORDX2:
 		case Opcode::S_LOAD_DWORDX4:
@@ -685,6 +665,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_ATOMIC_SUB:
 		case Opcode::BUFFER_ATOMIC_SMIN:
 		case Opcode::BUFFER_ATOMIC_UMIN:
+		case Opcode::BUFFER_ATOMIC_UMIN_X2:
 		case Opcode::BUFFER_ATOMIC_SMAX:
 		case Opcode::BUFFER_ATOMIC_UMAX:
 		case Opcode::BUFFER_ATOMIC_AND:
@@ -701,6 +682,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::FLAT_LOAD_USHORT:
 		case Opcode::FLAT_LOAD_SSHORT:
 		case Opcode::FLAT_LOAD_SHORT_D16:
+		case Opcode::FLAT_LOAD_SHORT_D16_HI:
 		case Opcode::FLAT_LOAD_DWORD:
 		case Opcode::FLAT_LOAD_DWORDX2:
 		case Opcode::FLAT_LOAD_DWORDX3:

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

@@ -544,6 +544,7 @@ enum class Opcode {
 	S_BUFFER_LOAD_DWORDX8,
 	S_BUFFER_LOAD_DWORDX16,
 	S_MEMREALTIME,
+	S_DCACHE_INV,
 	BUFFER_LOAD_FORMAT_X,
 	BUFFER_LOAD_FORMAT_XY,
 	BUFFER_LOAD_FORMAT_XYZ,
@@ -596,6 +597,7 @@ enum class Opcode {
 	BUFFER_ATOMIC_SUB,
 	BUFFER_ATOMIC_SMIN,
 	BUFFER_ATOMIC_UMIN,
+	BUFFER_ATOMIC_UMIN_X2,
 	BUFFER_ATOMIC_SMAX,
 	BUFFER_ATOMIC_UMAX,
 	BUFFER_ATOMIC_AND,
@@ -612,6 +614,7 @@ enum class Opcode {
 	FLAT_LOAD_USHORT,
 	FLAT_LOAD_SSHORT,
 	FLAT_LOAD_SHORT_D16,
+	FLAT_LOAD_SHORT_D16_HI,
 	FLAT_LOAD_DWORD,
 	FLAT_LOAD_DWORDX2,
 	FLAT_LOAD_DWORDX3,
@@ -707,15 +710,7 @@ enum class Opcode {
 	IMAGE_ATOMIC_FMIN,
 	IMAGE_ATOMIC_FMAX,
 	IMAGE_SAMPLE,
-	IMAGE_GATHER4_L,
-	IMAGE_GATHER4_LZ,
-	IMAGE_GATHER4_C,
-	IMAGE_GATHER4_C_L,
-	IMAGE_GATHER4_C_LZ,
-	IMAGE_GATHER4_LZ_O,
-	IMAGE_GATHER4_C_O,
-	IMAGE_GATHER4_C_LZ_O,
-	IMAGE_GATHER4H,
+	IMAGE_GATHER4,
 	V_INTERP_P1_F32,
 	V_INTERP_P2_F32,
 	V_INTERP_MOV_F32,

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

@@ -291,7 +291,7 @@ void Translator::V_MAD_U64_U32(const Decoder::Instruction& inst) {
 	    inst.dst2.kind != Decoder::OperandKind::Unknown) {
 		// The 32x32 product fits in 64 bits, so only the addition can carry.
 		const auto carry = IR::U1(ir.Emit(IR::ValueOpcode::ULessThan64, {result, product}));
-		WriteMask(inst.dst2, carry, VectorMask(carry));
+		WriteVectorMask(inst.dst2, carry);
 	}
 }
 

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

@@ -322,8 +322,8 @@ IR::Value Translator::MakeImageAddress(const Decoder::Instruction& inst,
                                        const Decoder::Operand&     base) {
 	std::array<IR::Value, 13> components {};
 	components.fill(IR::Value(0u));
-	const auto count =
-	    Decoder::ImageAddressDwordCount(inst.image_sample_flags, inst.image_address_components);
+	const auto count = Decoder::ImageAddressDwordCount(
+	    inst.image_sample_flags, inst.image_dimension, inst.image_address_components);
 	EXIT_IF(count > components.size());
 	const auto nsa_components =
 	    std::min(inst.image_nsa_dwords * 4u, Decoder::MaxImageNsaAddressComponents);
@@ -941,6 +941,8 @@ void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
+		// Translated scalar loads have no cache of their own to invalidate.
+		case Decoder::Opcode::S_DCACHE_INV: return;
 		case Decoder::Opcode::BUFFER_GL0_INV:
 		case Decoder::Opcode::BUFFER_GL1_INV:
 			ir.Emit(IR::ValueOpcode::DataCacheInvalidate);
@@ -1029,6 +1031,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicSMin32);
 		case Decoder::Opcode::BUFFER_ATOMIC_UMIN:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicUMin32);
+		case Decoder::Opcode::BUFFER_ATOMIC_UMIN_X2:
+			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicUMin64);
 		case Decoder::Opcode::BUFFER_ATOMIC_SMAX:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicSMax32);
 		case Decoder::Opcode::BUFFER_ATOMIC_UMAX:
@@ -1146,6 +1150,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::FLAT_LOAD_USHORT:
 		case Decoder::Opcode::FLAT_LOAD_SSHORT:
 		case Decoder::Opcode::FLAT_LOAD_SHORT_D16:
+		case Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI:
 		case Decoder::Opcode::FLAT_LOAD_DWORD:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX2:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX3:
@@ -1165,15 +1170,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::IMAGE_STORE:
 		case Decoder::Opcode::IMAGE_STORE_MIP: return IMAGE_STORE(inst);
 		case Decoder::Opcode::IMAGE_SAMPLE: return IMAGE_SAMPLE(inst);
-		case Decoder::Opcode::IMAGE_GATHER4_L:
-		case Decoder::Opcode::IMAGE_GATHER4_LZ:
-		case Decoder::Opcode::IMAGE_GATHER4_C:
-		case Decoder::Opcode::IMAGE_GATHER4_C_L:
-		case Decoder::Opcode::IMAGE_GATHER4_C_LZ:
-		case Decoder::Opcode::IMAGE_GATHER4_LZ_O:
-		case Decoder::Opcode::IMAGE_GATHER4_C_O:
-		case Decoder::Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Decoder::Opcode::IMAGE_GATHER4H: return IMAGE_GATHER(inst);
+		case Decoder::Opcode::IMAGE_GATHER4: return IMAGE_GATHER(inst);
 
 		case Decoder::Opcode::DS_MIN_F32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicFMin32, false);

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

@@ -1389,10 +1389,28 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			                      IR::U32(IR::Value(options.wave_size << 12u)));
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3),
 			                      entry_ir.BitwiseOr(vertex_count, IR::U32(IR::Value(1u << 28u))));
+			// Vulkan adds an indirect record's bases to the indices; the hardware passes them
+			// only in the user SGPRs the packet patches. SH_NOP leaves those SGPRs unchanged;
+			// it does not make the record's bases zero.
+			const auto& offsets = options.input_info.vertex->indirect_offsets;
+			const auto  index   = [&](uint8_t slot, IR::StageInputKind kind,
+			                          IR::StageInputKind base_kind) {
+				if (!options.input_info.vertex->indirect_draw) {
+					return builtin(kind);
+				}
+				const auto base = builtin(base_kind);
+				if (slot < options.user_data_count) {
+					entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(options.user_data_base + slot),
+					                      base);
+				}
+				return entry_ir.ISub(builtin(kind), base);
+			};
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
-			                      builtin(IR::StageInputKind::VertexIndex));
+			                      index(offsets.vertex, IR::StageInputKind::VertexIndex,
+			                            IR::StageInputKind::BaseVertex));
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
-			                      builtin(IR::StageInputKind::InstanceIndex));
+			                      index(offsets.instance, IR::StageInputKind::InstanceIndex,
+			                            IR::StageInputKind::BaseInstance));
 		}
 	}
 	const bool flush_f32_inputs = options.stage == ShaderType::Compute &&

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

@@ -118,15 +118,14 @@ ResourcePlan& ResourcePlan::operator=(ResourcePlan&& other) noexcept {
 }
 
 Program::~Program() {
-	// Planning expressions can refer to block values but outlive block storage in the base class.
+	// Planning expressions can refer to block values but outlive block storage in the base class,
+	// and values cross block boundaries: detach everything before any storage is destroyed.
 	for (auto& inst: value_storage) {
-		inst.Invalidate();
+		inst.DropArgs();
 	}
-	// Values may cross block boundaries. Detach all arguments before any block starts destroying
-	// its instruction storage so reverse-use links always point to live definitions.
 	for (auto* block: blocks) {
 		for (auto& inst: *block) {
-			inst.Invalidate();
+			inst.DropArgs();
 		}
 	}
 }
@@ -169,6 +168,8 @@ bool EquivalentSelectorDomain(const ResourcePlan& program,
                               const DescriptorSource::IndirectDescriptor& right) {
 	return left.selector == right.selector && left.signed_count == right.signed_count &&
 	       left.workgroup_axis == right.workgroup_axis &&
+	       left.workgroup_multiplier == right.workgroup_multiplier &&
+	       left.workgroup_shift == right.workgroup_shift &&
 	       EquivalentValue(program, left.key_count, right.key_count) &&
 	       EquivalentValue(program, left.selector_first, right.selector_first) &&
 	       EquivalentValue(program, left.selector_mask, right.selector_mask);

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

@@ -157,12 +157,15 @@ KYTY_IO_FIELDS(StageOutput, a, b, c, d)
 KYTY_IO_FIELDS(ShaderInfo, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w, x, y, z,
                A, B, C, D, E, F, G, H, I, J, K)
 KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::CountedRows, a, b)
-KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead, a, b, c, d, e, f, g, h, i, j)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::MaskRanks, a, b, c)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead, a, b, c, d, e, f, g, h, i, j, k,
+               l)
 KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::PointerTable, a, b, c, d)
-KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p,
+               q, r)
 KYTY_IO_FIELDS(DescriptorSource, a, b, c)
 KYTY_IO_FIELDS(SrtRead, a, b)
-KYTY_IO_FIELDS(ResourceBlock, a, b, c, d)
+KYTY_IO_FIELDS(ResourceBlock, a, b, c, d, e)
 KYTY_IO_FIELDS(UniformFill, a, b, c, d, e, f)
 KYTY_IO_FIELDS(UniformFillPlan, a, b, c, d)
 KYTY_IO_FIELDS(ResourceSpecialization::Buffer, a, b, c, d, e, f, g, h)

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

@@ -251,6 +251,8 @@ enum class StageInputKind {
 	LocalInvocationId,
 	LocalInvocationIndex,
 	GlobalInvocationId,
+	BaseVertex,
+	BaseInstance,
 	Parameter,
 };
 
@@ -572,6 +574,13 @@ struct DescriptorSource {
 
 				bool operator==(const CountedRows& other) const = default;
 			};
+			struct MaskRanks {
+				uint32_t source = UINT32_MAX;
+				uint32_t stride = 0; // Byte spacing of the mask words its load can address.
+				uint32_t offset = 0;
+
+				bool operator==(const MaskRanks& other) const = default;
+			};
 			uint32_t source = UINT32_MAX;
 			uint32_t stride = 0;
 			uint32_t offset = 0;
@@ -579,10 +588,12 @@ struct DescriptorSource {
 			uint32_t bit_offset = 0;
 			uint32_t bit_mask = UINT32_MAX;
 			bool global = false;
-			// Four DWORD bitmap words per record; set-bit ranks subtract from the last key.
+			// Four or eight DWORD bitmap words per record; set ranks subtract from the last key.
 			bool bitset = false;
 			std::optional<CountedRows> rows;
 			std::vector<std::pair<uint32_t, uint32_t>> byte_ranges;
+			std::optional<uint32_t> excluded; // A guard keeps this key from the read.
+			std::optional<MaskRanks> ranks;   // Records are the bits some mask word sets.
 
 			bool operator==(const SelectorRead& other) const = default;
 		};
@@ -606,8 +617,12 @@ struct DescriptorSource {
 		bool     indexed = false;
 		bool     signed_count = false;
 		uint32_t workgroup_axis  = UINT32_MAX;
+		// Nonzero shift: workgroup keys are (selector_first + ID) * multiplier >> shift.
+		uint32_t workgroup_multiplier = 0;
+		uint32_t workgroup_shift      = 0;
 		Value    key_count;
-		// First selector index, or first descriptor key for a direct or bitmap range.
+		// First selector index, workgroup key base, or first descriptor key for a direct or bitmap
+		// range.
 		Value                 selector_first;
 		Value    selector_mask;
 		std::vector<uint32_t> sources;
@@ -634,6 +649,7 @@ struct ResourceBlock {
 	Value                 condition;
 	std::vector<uint32_t> successors;
 	std::vector<uint32_t> sources;
+	std::vector<uint32_t> images;
 	std::vector<uint32_t> srt_reads;
 };
 

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

@@ -25,22 +25,6 @@ Value Value::F32(float value) {
 	return Value(Type::F32, std::bit_cast<uint32_t>(value));
 }
 
-bool Value::IsEmpty() const {
-	return type == Type::Void;
-}
-
-bool Value::IsImmediate() const {
-	return type != Type::Opaque;
-}
-
-bool Value::IsIdentity() const {
-	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Identity;
-}
-
-bool Value::IsPhi() const {
-	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Phi;
-}
-
 Type Value::GetType() const {
 	if (IsPhi()) {
 		return inst->Flags<Type>();
@@ -56,19 +40,6 @@ Inst* Value::Instruction() const {
 	return inst;
 }
 
-Inst* Value::TryInstruction() const {
-	return type == Type::Opaque ? inst : nullptr;
-}
-
-Inst* Value::ResolveInstruction() const {
-	EXIT_IF(type != Type::Opaque);
-	return IsIdentity() ? inst->Arg(0).ResolveInstruction() : inst;
-}
-
-Value Value::Resolve() const {
-	return IsIdentity() ? inst->Arg(0).Resolve() : *this;
-}
-
 ScalarReg Value::ScalarRegister() const {
 	EXIT_IF(type != Type::ScalarReg);
 	return scalar_reg;
@@ -94,11 +65,6 @@ uint16_t Value::U16() const {
 	return imm_u16;
 }
 
-uint32_t Value::U32() const {
-	EXIT_IF(type != Type::U32);
-	return imm_u32;
-}
-
 uint64_t Value::U64() const {
 	EXIT_IF(type != Type::U64);
 	return imm_u64;
@@ -152,10 +118,6 @@ Inst::~Inst() {
 	ClearArgs();
 }
 
-ValueOpcode Inst::GetOpcode() const {
-	return opcode;
-}
-
 Type Inst::GetType() const {
 	if (opcode == ValueOpcode::Phi) {
 		return static_cast<Type>(flags);
@@ -178,51 +140,25 @@ size_t Inst::UseCount() const {
 	return uses.size();
 }
 
-size_t Inst::NumArgs() const {
-	return num_args == PhiArity ? phi_args.size() : num_args;
-}
-
 size_t Inst::NumPhiBlocks() const {
 	return num_args == PhiArity ? phi_args.size() : 0;
 }
 
-Value Inst::Arg(size_t index) const {
-	EXIT_IF(index >= NumArgs());
-	if (num_args <= InlineArity) {
-		return fixed_args[index];
-	}
-	return num_args == PhiArity ? phi_args[index].second : large_args[index];
-}
-
 Block* Inst::PhiBlock(size_t index) const {
 	EXIT_IF(opcode != ValueOpcode::Phi || index >= phi_args.size());
 	return phi_args[index].first;
 }
 
-Block* Inst::Parent() const {
-	return parent;
-}
-
-const std::vector<Use>& Inst::Uses() const {
-	return uses;
-}
-
 void Inst::SetParent(Block* block) {
 	parent = block;
 }
 
 void Inst::SetArg(size_t index, Value value) {
-	const auto old = Arg(index);
-	if (auto* old_inst = old.TryInstruction(); old_inst != nullptr) {
+	auto& arg = ArgSlot(*this, index);
+	if (auto* old_inst = arg.TryInstruction(); old_inst != nullptr) {
 		RemoveUse(old_inst, index);
 	}
-	if (num_args <= InlineArity) {
-		fixed_args[index] = value;
-	} else if (num_args == PhiArity) {
-		phi_args[index].second = value;
-	} else {
-		large_args[index] = value;
-	}
+	arg = value;
 	if (auto* new_inst = value.TryInstruction(); new_inst != nullptr) {
 		AddUse(new_inst, index);
 	}
@@ -238,11 +174,15 @@ void Inst::AddPhiOperand(Block* predecessor, Value value) {
 }
 
 void Inst::ReplaceUsesWith(Value replacement, bool preserve) {
-	const auto old_uses = uses;
-	for (const auto& use: old_uses) {
-		use.user->SetArg(use.operand, replacement);
-	}
+	// Detach the arguments first (a phi can use itself), then move every use over in order.
 	Invalidate();
+	const auto old_uses = std::exchange(uses, {});
+	for (const auto& use: old_uses) {
+		ArgSlot(*use.user, use.operand) = replacement;
+	}
+	if (auto* target = replacement.TryInstruction(); target != nullptr) {
+		target->uses.insert(target->uses.end(), old_uses.begin(), old_uses.end());
+	}
 	if (preserve) {
 		opcode = ValueOpcode::Identity;
 		num_args = 1;
@@ -256,9 +196,6 @@ void Inst::Invalidate() {
 }
 
 void Inst::AddUse(Inst* used, size_t operand) {
-	const auto found = std::ranges::find_if(
-	    used->uses, [&](const Use& use) { return use.user == this && use.operand == operand; });
-	EXIT_IF(found != used->uses.end());
 	used->uses.push_back({this, operand});
 }
 
@@ -275,6 +212,10 @@ void Inst::ClearArgs() {
 			RemoveUse(value_inst, index);
 		}
 	}
+	DropArgs();
+}
+
+void Inst::DropArgs() {
 	if (num_args == PhiArity) {
 		std::destroy_at(&phi_args);
 		std::construct_at(&fixed_args);

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

@@ -33,14 +33,14 @@ public:
 	static Value F16(uint16_t bits);
 	static Value F32(float value);
 
-	[[nodiscard]] bool IsEmpty() const;
-	[[nodiscard]] bool IsImmediate() const;
+	[[nodiscard]] bool IsEmpty() const { return type == Type::Void; }
+	[[nodiscard]] bool IsImmediate() const { return type != Type::Opaque; }
 	[[nodiscard]] bool IsIdentity() const;
 	[[nodiscard]] bool IsPhi() const;
 	[[nodiscard]] Type GetType() const;
 
 	[[nodiscard]] Inst*     Instruction() const;
-	[[nodiscard]] Inst*     TryInstruction() const;
+	[[nodiscard]] Inst*     TryInstruction() const { return type == Type::Opaque ? inst : nullptr; }
 	[[nodiscard]] Inst*     ResolveInstruction() const;
 	[[nodiscard]] Value     Resolve() const;
 	[[nodiscard]] ScalarReg ScalarRegister() const;
@@ -112,7 +112,7 @@ public:
 	Inst(Inst&&)                 = delete;
 	Inst& operator=(Inst&&)      = delete;
 
-	[[nodiscard]] ValueOpcode             GetOpcode() const;
+	[[nodiscard]] ValueOpcode             GetOpcode() const { return opcode; }
 	[[nodiscard]] Type                    GetType() const;
 	[[nodiscard]] bool                    MayHaveSideEffects() const;
 	[[nodiscard]] bool                    HasUses() const;
@@ -121,8 +121,8 @@ public:
 	[[nodiscard]] size_t                  NumPhiBlocks() const;
 	[[nodiscard]] Value                   Arg(size_t index) const;
 	[[nodiscard]] Block*                  PhiBlock(size_t index) const;
-	[[nodiscard]] Block*                  Parent() const;
-	[[nodiscard]] const std::vector<Use>& Uses() const;
+	[[nodiscard]] Block*                  Parent() const { return parent; }
+	[[nodiscard]] const std::vector<Use>& Uses() const { return uses; }
 	// SRT walker output index, assigned by the resource plan that owns this instruction.
 	[[nodiscard]] uint32_t WalkOutput() const { return walk_output; }
 	void SetWalkOutput(uint32_t index) const { walk_output = index; }
@@ -138,6 +138,8 @@ public:
 	void AddPhiOperand(Block* predecessor, Value value);
 	void ReplaceUsesWith(Value replacement, bool preserve = true);
 	void Invalidate();
+	// Drops the arguments without updating their use lists, for teardown that destroys both sides.
+	void DropArgs();
 
 	template <typename T>
 	requires(sizeof(T) <= sizeof(uint64_t) && std::is_trivially_copyable_v<T>)
@@ -157,6 +159,8 @@ public:
 private:
 	friend void EliminateDeadCode(const std::vector<Block*>& blocks);
 
+	template <typename Self>
+	static auto& ArgSlot(Self& self, size_t index);
 	void AddUse(Inst* used, size_t operand);
 	void RemoveUse(Inst* used, size_t operand);
 	void ClearArgs();
@@ -181,4 +185,44 @@ private:
 
 static_assert(sizeof(Inst) <= 120, "Inst operand and backend storage unintentionally increased");
 
+// Called per operand by the IR passes and the resource walker; inline, as the build has no LTO.
+inline bool Value::IsIdentity() const {
+	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Identity;
+}
+
+inline bool Value::IsPhi() const {
+	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Phi;
+}
+
+inline Inst* Value::ResolveInstruction() const {
+	EXIT_IF(type != Type::Opaque);
+	return IsIdentity() ? inst->Arg(0).ResolveInstruction() : inst;
+}
+
+inline Value Value::Resolve() const {
+	return IsIdentity() ? inst->Arg(0).Resolve() : *this;
+}
+
+inline uint32_t Value::U32() const {
+	EXIT_IF(type != Type::U32);
+	return imm_u32;
+}
+
+inline size_t Inst::NumArgs() const {
+	return num_args == PhiArity ? phi_args.size() : num_args;
+}
+
+template <typename Self>
+auto& Inst::ArgSlot(Self& self, size_t index) {
+	EXIT_IF(index >= self.NumArgs());
+	if (self.num_args <= InlineArity) {
+		return self.fixed_args[index];
+	}
+	return self.num_args == PhiArity ? self.phi_args[index].second : self.large_args[index];
+}
+
+inline Value Inst::Arg(size_t index) const {
+	return ArgSlot(*this, index);
+}
+
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -123,6 +123,7 @@ BufferAccess BufferAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::BufferAtomicISub32:
 		case ValueOpcode::BufferAtomicSMin32:
 		case ValueOpcode::BufferAtomicUMin32:
+		case ValueOpcode::BufferAtomicUMin64:
 		case ValueOpcode::BufferAtomicSMax32:
 		case ValueOpcode::BufferAtomicUMax32:
 		case ValueOpcode::BufferAtomicAnd32:
@@ -140,6 +141,7 @@ uint32_t BufferComponentCount(ValueOpcode opcode) {
 	switch (opcode) {
 		case ValueOpcode::BufferAtomicSwap64:
 		case ValueOpcode::BufferAtomicIAdd64:
+		case ValueOpcode::BufferAtomicUMin64:
 		case ValueOpcode::BufferAtomicAnd64:
 		case ValueOpcode::BufferAtomicOr64:
 		case ValueOpcode::LoadBufferU32x2:

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

@@ -269,6 +269,7 @@ VALUE_OPCODE(BufferAtomicIAdd64, U64, BufferResource, U32, U32, U32, U64, U1)
 VALUE_OPCODE(BufferAtomicISub32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicSMin32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicUMin32, U32, BufferResource, U32, U32, U32, U32, U1)
+VALUE_OPCODE(BufferAtomicUMin64, U64, BufferResource, U32, U32, U32, U64, U1)
 VALUE_OPCODE(BufferAtomicSMax32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicUMax32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicAnd32, U32, BufferResource, U32, U32, U32, U32, U1)

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

@@ -331,6 +331,18 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
 		keys.resize(count);
 		return true;
 	};
+	// Reads the key of each record whose rank is set.
+	const auto read_ranks = [&](uint64_t base, uint64_t size, uint64_t first, uint64_t step,
+	                            uint32_t ranks) {
+		keys.reserve(std::popcount(ranks));
+		for (; ranks != 0u; ranks &= ranks - 1u) {
+			const auto offset = first + std::countr_zero(ranks) * step;
+			uint32_t key = 0;
+			if (offset > UINT32_MAX || !read_selector(base, size, offset, {&key, 1})) return false;
+			keys.push_back(key);
+		}
+		return true;
+	};
 	if (selector->rows) {
 		const auto& rows = *selector->rows;
 		DescriptorValue count_value;
@@ -377,19 +389,21 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
 		}
 	} else if (selector->bitset) {
 		ShaderBufferResource material;
-		std::array<uint32_t, 4> masks {};
+		std::array<uint32_t, 8> storage {};
+		const auto record_words = std::min<size_t>(selector->stride / 4u, storage.size());
+		const auto masks = std::span(storage).first(record_words);
 		uint32_t first = 0, count = 0;
 		if (!DecodeBufferDescriptor(material_value, material) ||
 		    !WalkedValue(program, indirect.selector_first, first) ||
 		    !EvaluateIndirectCount(program, indirect, runtime, count) || count != masks.size() * 32u ||
-		    selector->stride != sizeof(masks) || selector->offset != 0u ||
-		    material.Type() != 0u || material.Stride() != sizeof(masks) ||
+		    selector->stride != masks.size_bytes() || selector->offset != 0u ||
+		    material.Type() != 0u || material.Stride() != selector->stride ||
 		    material.SwizzleEnabled() || material.AddTid() || material.OutOfBounds() != 0u ||
 		    (material.Base48() & 3u) != 0u) return false;
 		if (material.Format() != Prospero::BufferFormat::kInvalid) {
 			const auto words = material.GetSize() / sizeof(uint32_t);
-			// Bitmap storage can grow with resolution while its candidate domain stays
-			// at 128 keys. Bound scratch memory, preserving each coherent read's capture.
+			// Bitmap storage can grow with resolution while its candidate domain stays at 32 keys
+			// per record word. Bound scratch memory, preserving each coherent read's capture.
 			for (uint64_t first_word = 0; first_word < words; first_word += keys.size()) {
 				keys.resize(std::min<uint64_t>(words - first_word, MaxIndirectDescriptorProbes));
 				if (!ReadScalarTable(material.Base48(), material.GetSize(), first_word * 4u,
@@ -445,12 +459,35 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
 		     uint64_t {first} + count > material.NumRecords())) return false;
 		const auto last = count != 0u ? uint64_t {first} + count - 1u : 0u;
 		if (last > UINT32_MAX) return false;
+		// Out-of-bounds scalar reads return zero.
 		const auto end = count != 0u ? last * step + selector->offset + 4u : 0u;
-		if (end > uint64_t {UINT32_MAX} + 1u ||
-		    (indirect.workgroup_axis == UINT32_MAX && end > size))
-			return false;
-		if (!read_keys(base, size, uint64_t {first} * step + selector->offset, step, count))
+		if (end > uint64_t {UINT32_MAX} + 1u) return false;
+		const auto start = uint64_t {first} * step + selector->offset;
+		if (selector->ranks) {
+			// A scan selects the ranks some word of its mask's address class sets.
+			const auto& mask = *selector->ranks;
+			DescriptorValue mask_value;
+			ShaderBufferResource masks;
+			if (count > 32u || !WalkedDescriptor(program, mask.source, mask_value) ||
+			    !DecodeBufferDescriptor(mask_value, masks)) return false;
+			// Each chunk holds whole strides of the power-of-two class, or one of its words.
+			const auto words = mask.stride / 4u;
+			const auto chunk =
+			    words <= MaxIndirectDescriptorProbes ? MaxIndirectDescriptorProbes : 1u;
+			uint32_t ranks = 0;
+			for (auto at = uint64_t {mask.offset}; at + 4u <= masks.GetSize();
+			     at += uint64_t {std::max(chunk, words)} * 4u) {
+				keys.resize(std::min<uint64_t>((masks.GetSize() - at) / 4u, chunk));
+				if (!ReadScalarTable(masks.Base48(), masks.GetSize(), at, runtime, keys))
+					return false;
+				for (size_t word = 0; word < keys.size(); word += words) ranks |= keys[word];
+			}
+			keys.clear();
+			if (count < 32u) ranks &= (1u << count) - 1u;
+			if (!read_ranks(base, size, start, step, ranks)) return false;
+		} else if (!read_keys(base, size, start, step, count)) {
 			return false;
+		}
 	} else if (!indirect.selector_mask.IsEmpty()) {
 		uint32_t mask = 0;
 		uint32_t count = 0;
@@ -462,18 +499,8 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
 		if (count < 32u) mask &= (1u << count) - 1u;
 		const auto material_base =
 		    (static_cast<uint64_t>(material_value.dwords[1]) << 32u) | material_value.dwords[0];
-		keys.reserve(std::popcount(mask));
-		while (mask != 0u) {
-			const auto index = std::countr_zero(mask);
-			const auto offset = static_cast<uint64_t>(selector->offset) +
-			                    static_cast<uint64_t>(index) * selector->stride;
-			if (offset > UINT32_MAX) return false;
-			uint32_t key = 0;
-			if (!read_selector(material_base, UINT64_MAX, static_cast<uint32_t>(offset),
-			                   {&key, 1})) return false;
-			keys.push_back(key);
-			mask &= mask - 1u;
-		}
+		if (!read_ranks(material_base, UINT64_MAX, selector->offset, selector->stride, mask))
+			return false;
 	} else if (selector->stride == 0u) {
 		ShaderBufferResource material;
 		if (!DecodeBufferDescriptor(material_value, material) ||
@@ -503,6 +530,7 @@ bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::In
 		if (selector->bit_offset != 0u || selector->bit_mask != UINT32_MAX || selector->maximum != UINT32_MAX)
 			for (auto& key: keys)
 				key = std::min((key >> selector->bit_offset) & selector->bit_mask, selector->maximum);
+		if (selector->excluded) std::erase(keys, *selector->excluded);
 		std::ranges::sort(keys);
 		keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
 	}
@@ -756,6 +784,12 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			std::iota(keys.begin(), keys.end(), first_key);
 			if (!keys.empty() && keys.back() < keys.front())
 				std::rotate(keys.begin(), keys.begin() + (UINT32_MAX - first_key + 1ull), keys.end());
+			if (indirect.workgroup_shift != 0u) {
+				for (auto& key: keys)
+					key = static_cast<uint32_t>(uint64_t {key} * indirect.workgroup_multiplier >>
+					                            indirect.workgroup_shift);
+				keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
+			}
 		} else {
 			if (!indirect.selector_mask.IsEmpty() && table_value.dword_count != 2u) return false;
 			if (!selector_keys->resolved) {
@@ -1041,6 +1075,26 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			return dimension == Decoder::ImageDimension::Dim1D ||
 			       dimension == Decoder::ImageDimension::Dim2D;
 		};
+		// The descriptor's type selects the address layout. A 2D or 3D candidate reads the leading
+		// coordinates of an equal or larger dimension, but an array with more than one slice reads
+		// the next as its slice. Derivative, LOD, clamp and mip operands shift with the dimension.
+		const auto leading = [&](Decoder::ImageDimension dimension, const DescriptorValue& value) {
+			const bool sliced =
+			    static_cast<Prospero::ImageType>((value.dwords[3] >> 28u) & 0xfu) ==
+			        Prospero::ImageType::kColor2DArray &&
+			    (value.dwords[4] & 0x1fffu) != ((value.dwords[4] >> 16u) & 0x1fffu);
+			return !sliced &&
+			       (dimension == Decoder::ImageDimension::Dim2D || dimension == base.dimension);
+		};
+		const bool leading_coordinates = (base.dimension == Decoder::ImageDimension::Dim2D ||
+		                                  base.dimension == Decoder::ImageDimension::Dim3D) &&
+		    leading(image_class.dimension, snapshot.images[exemplar]) &&
+		    std::ranges::none_of(program.memory_info, [&](const MemoryInfo& memory) {
+			    return memory.kind == ResourceKind::Image && memory.resource == root_index &&
+			           (memory.image_has_mip || (memory.image_sample_flags &
+			               (Decoder::ImageSampleFlagLod | Decoder::ImageSampleFlagDerivative |
+			                Decoder::ImageSampleFlagLodClamp)) != 0u);
+		    });
 		for (uint32_t candidate = 0; candidate < specialization.images.size(); candidate++) {
 			auto& image = specialization.images[candidate];
 			if (image.indirect_root != root_index) {
@@ -1060,8 +1114,10 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			    base.dimension == Decoder::ImageDimension::Dim2D &&
 			    !image.cube && !image_class.cube && is_1d_or_2d(image.dimension) &&
 			    is_1d_or_2d(image_class.dimension);
+			const bool prefix_coordinates =
+			    leading_coordinates && leading(image.dimension, snapshot.images[candidate]);
 			if ((!same_coordinates && !(is_2d(image.dimension) && is_2d(image_class.dimension)) &&
-			     !unfiltered_coordinates) ||
+			     !unfiltered_coordinates && !prefix_coordinates) ||
 			    image.conversion_format != image_class.conversion_format ||
 			    (!storage_swizzle && image.shader_swizzle != image_class.shader_swizzle)) {
 				return SpecializationFail(
@@ -1164,14 +1220,16 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 				block.sources.push_back(program.info.buffers.at(memory.resource).source);
 			} else {
 				block.sources.push_back(program.info.images.at(memory.resource).source);
+				block.images.push_back(memory.resource);
 				if (image.needs_sampler) {
 					block.sources.push_back(program.info.samplers.at(memory.sampler).source);
 				}
 			}
 		}
-		std::ranges::sort(block.sources);
-		block.sources.erase(std::unique(block.sources.begin(), block.sources.end()),
-		                    block.sources.end());
+		for (auto* list: {&block.sources, &block.images}) {
+			std::ranges::sort(*list);
+			list->erase(std::unique(list->begin(), list->end()), list->end());
+		}
 	}
 	for (auto& block: blocks) block.condition = predicate(block.condition);
 	if (std::ranges::none_of(
@@ -1645,7 +1703,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		}
 		if (source->indirect_descriptor.has_value()) {
 			snapshot.images[i] = {.dword_count = 8u};
-			if (!WalkedActive(program, image.source)) {
+			if (!WalkedActive(program, image.source) || !WalkedImageActive(program, i)) {
 				continue;
 			}
 			const auto& indirect = *source->indirect_descriptor;
@@ -1665,7 +1723,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			if (!evaluate(image.source, snapshot.images[i])) {
 				return false;
 			}
-			if (!ValidImageDescriptor(snapshot.images[i], image.r128)) {
+			// An image listed only by unreached blocks reads zero, not its source's descriptor.
+			if (!WalkedImageActive(program, i) ||
+			    !ValidImageDescriptor(snapshot.images[i], image.r128)) {
 				snapshot.images[i].dwords.fill(0);
 			}
 		}
@@ -1711,8 +1771,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 }
 
 uint32_t ImageHeapKind(const ImageResource& root, DescriptorValue& value) {
-	// Mirrors BuildResourceSpecialization for sampled roots: entries it would reject, and
-	// entries that need shader-side conversion, read as null instead of failing the draw.
+	// Mirrors BuildResourceSpecialization for sampled roots: entries it would reject, entries of
+	// another dimension and entries that need shader-side conversion read as null instead of
+	// failing the draw.
 	const auto format = static_cast<Prospero::BufferFormat>((value.dwords[1] >> 20u) & 0x1ffu);
 	const auto tile = static_cast<Prospero::TileMode>((value.dwords[3] >> 20u) & 0x1fu);
 	const auto numeric_class =

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

@@ -7,6 +7,8 @@
 
 #include <algorithm>
 #include <bit>
+#include <cmath>
+#include <cstdio>
 #include <fmt/format.h>
 #include <map>
 #include <numeric>
@@ -1014,7 +1016,9 @@ private:
 				    (!shared_image_heap && (a.table_offset != b.table_offset ||
 				                           a.table_immediate != b.table_immediate ||
 				                           a.table_stride != b.table_stride)) || a.indexed != b.indexed ||
-				    a.workgroup_axis != b.workgroup_axis || a.sources != b.sources ||
+				    a.workgroup_axis != b.workgroup_axis ||
+				    a.workgroup_multiplier != b.workgroup_multiplier ||
+				    a.workgroup_shift != b.workgroup_shift || a.sources != b.sources ||
 				    a.signed_count != b.signed_count ||
 				    !EquivalentValue(m_program, a.key_count, b.key_count) ||
 				    a.selector_first.IsEmpty() != b.selector_first.IsEmpty() ||
@@ -1136,31 +1140,63 @@ private:
 		LaneQuantifier lanes = LaneQuantifier::All;
 	};
 
+	// The immediate that edge proves value differs from on every lane, if any.
+	std::optional<uint32_t> ExcludedBy(const EdgePredicate& edge, Value value) const {
+		if (edge.lanes != LaneQuantifier::All) return {};
+		const auto* test = edge.condition.TryInstruction();
+		if (test != nullptr && test->NumArgs() == 2u &&
+		    ((test->GetOpcode() == ValueOpcode::INotEqual32 && edge.positive) ||
+		     (test->GetOpcode() == ValueOpcode::IEqual32 && !edge.positive))) {
+			for (uint32_t arg = 0; arg < 2u; ++arg) {
+				uint32_t immediate;
+				if (ImmediateU32(test->Arg(arg), immediate) &&
+				    EquivalentValue(m_program, test->Arg(arg ^ 1u), value)) return immediate;
+			}
+		}
+		return {};
+	}
+
 	bool NonzeroOnEntry(Value value, const Block* block) const {
 		if (block == nullptr) return false;
 		bool witnessed = false;
 		const bool guarded = GuardedOnEntry(block, [](const Block* current) {
 			return current->ImmPredecessors().size() != 1u;
 		}, [&](const EdgePredicate& edge) {
-			if (edge.lanes != LaneQuantifier::All) return false;
-			const auto* test = edge.condition.TryInstruction();
-			if (test != nullptr && test->NumArgs() == 2u &&
-			    ((test->GetOpcode() == ValueOpcode::INotEqual32 && edge.positive) ||
-			     (test->GetOpcode() == ValueOpcode::IEqual32 && !edge.positive))) {
-				for (uint32_t arg = 0; arg < 2u; ++arg) {
-					uint32_t immediate;
-					if (ImmediateU32(test->Arg(arg), immediate) && immediate == 0u &&
-					    EquivalentValue(m_program, test->Arg(arg ^ 1u), value)) return witnessed = true;
-				}
-			}
-			return false;
+			return ExcludedBy(edge, value) == 0u && (witnessed = true);
 		});
 		return guarded && witnessed;
 	}
 
-	uint32_t WorkgroupAxis(Value key) const {
+	// The key that every path from its definition to block rules out, if any.
+	std::optional<uint32_t> ExcludedKeyOnEntry(Value key, const Block* block) const {
+		const auto* definition = key.Resolve().TryInstruction();
+		if (definition == nullptr) return {};
+		std::optional<uint32_t> excluded;
+		const auto defined = [&](const Block* current) { return current == definition->Parent(); };
+		const bool guarded = GuardedOnEntry(block, defined, [&](const EdgePredicate& edge) {
+			const auto immediate = ExcludedBy(edge, key);
+			if (!immediate || (excluded && excluded != immediate)) return false;
+			excluded = immediate;
+			return true;
+		});
+		return guarded ? excluded : std::nullopt;
+	}
+
+	// The workgroup ID axis of key; with first, also of a runtime first + workgroup ID sum.
+	uint32_t WorkgroupAxis(Value key, Value* first = nullptr) const {
 		const auto* builtin = key.Resolve().TryInstruction();
 		uint32_t axis = UINT32_MAX;
+		if (first != nullptr && builtin != nullptr && builtin->GetOpcode() == ValueOpcode::IAdd32) {
+			for (uint32_t arm = 0; arm < 2u; ++arm) {
+				axis = WorkgroupAxis(builtin->Arg(arm));
+				const auto base = builtin->Arg(arm ^ 1u);
+				if (axis == UINT32_MAX ||
+				    !ValidateRuntimeValue(m_program, base, RuntimeValueType::Integer)) continue;
+				*first = base;
+				return axis;
+			}
+			return UINT32_MAX;
+		}
 		if (m_program.stage != ShaderType::Compute || builtin == nullptr ||
 		    builtin->GetOpcode() != ValueOpcode::GetBuiltin || builtin->NumArgs() != 2u ||
 		    builtin->Arg(0) != Value(static_cast<uint32_t>(StageInputKind::WorkgroupId)) ||
@@ -1168,6 +1204,27 @@ private:
 		return axis;
 	}
 
+	// Matches key = (first + workgroup ID) * multiplier >> shift with shift >= 32; first and the
+	// division are optional.
+	bool MatchWorkgroupKey(Value key, DescriptorSource::IndirectDescriptor& indirect) const {
+		uint32_t bits = 0, multiplier = 0;
+		const auto* shift = key.Resolve().TryInstruction();
+		if (shift != nullptr && shift->GetOpcode() == ValueOpcode::ShiftRightLogical32) {
+			const auto* high = shift->Arg(0).Resolve().TryInstruction();
+			if (high != nullptr && high->GetOpcode() == ValueOpcode::UMulHi &&
+			    ImmediateU32(shift->Arg(1), bits) && bits < 32u &&
+			    ImmediateU32(high->Arg(1), multiplier)) {
+				key = high->Arg(0);
+				bits += 32u;
+			}
+		}
+		indirect.workgroup_axis = WorkgroupAxis(key, &indirect.selector_first);
+		if (indirect.workgroup_axis == UINT32_MAX) return false;
+		indirect.workgroup_multiplier = multiplier;
+		indirect.workgroup_shift = bits;
+		return true;
+	}
+
 	bool MatchTableOffset(Value value, Value& key, uint32_t& offset, uint32_t& stride) const {
 		offset = 0;
 		for (;;) {
@@ -1672,6 +1729,8 @@ private:
 		uint32_t initial_arm;
 	};
 
+	static bool ZeroBased(uint32_t initial, uint32_t step) { return initial == 0u && step == 1u; }
+
 	std::optional<Induction> MatchInduction(Value key, const auto& accepts_step) const {
 		key = key.Resolve();
 		const auto* phi = key.TryInstruction();
@@ -1733,8 +1792,7 @@ private:
 	}
 
 	uint32_t PostTestLoopMaximum(Value key) const {
-		const auto induction = MatchInduction(key,
-		    [](uint32_t initial, uint32_t step) { return initial == 0u && step == 1u; });
+		const auto induction = MatchInduction(key, ZeroBased);
 		if (!induction) return UINT32_MAX;
 		const auto [phi, step, initial_arm] = *induction;
 		const auto* incoming = phi->PhiBlock(initial_arm ^ 1u);
@@ -1767,8 +1825,7 @@ private:
 
 	const Inst* BoundedLoop(Value key, const Block* use, const auto& accepts_bound,
 	                       LaneQuantifier lanes = LaneQuantifier::Any) const {
-		return BoundedInduction(key, use,
-		    [](uint32_t initial, uint32_t step) { return initial == 0u && step == 1u; },
+		return BoundedInduction(key, use, ZeroBased,
 		    [&](const Inst& compare, uint32_t initial) {
 			    return (compare.GetOpcode() == ValueOpcode::SLessThan32 ||
 			            compare.GetOpcode() == ValueOpcode::ULessThan32) &&
@@ -1793,6 +1850,29 @@ private:
 		});
 	}
 
+	// A zero-based key that only per-lane data bounds takes the count all uniform zero-based
+	// loops in the program share, assuming each lane's count stays within it.
+	bool ProgramLoopCount(Value key, DescriptorSource::IndirectDescriptor& indirect) const {
+		if (!MatchInduction(key, ZeroBased) || BoundedLoop(key, nullptr) != nullptr) return false;
+		const Inst* count = nullptr;
+		for (auto* block: m_program.blocks) {
+			for (auto& phi: *block) {
+				if (phi.GetOpcode() != ValueOpcode::Phi) break;
+				const auto* compare = BoundedLoop(Value(&phi), nullptr);
+				if (compare == nullptr) continue;
+				if (count != nullptr && !EquivalentValue(m_program, compare->Arg(1), count->Arg(1)))
+					return false;
+				count = compare;
+			}
+		}
+		if (count == nullptr) return false;
+		std::printf("Warning: shader 0x%016llx bounds a per-lane table loop by its uniform loops' "
+		            "count\n", static_cast<unsigned long long>(m_program.shader_hash));
+		indirect.key_count    = count->Arg(1);
+		indirect.signed_count = count->GetOpcode() == ValueOpcode::SLessThan32;
+		return true;
+	}
+
 	Value ActiveValue(Value value, Value guard) const {
 		value = value.Resolve();
 		while (const auto* select = value.TryInstruction()) {
@@ -1829,6 +1909,10 @@ private:
 	}
 
 	bool Dominates(const Block* initial, const Block* use, Value witness = {}) const {
+		// The proven local-zero invocation fails every all-lane bypass condition that excludes it,
+		// such as an all-inactive or an all-nonzero bypass.
+		const auto* negated = SimplifyGuard(witness).TryInstruction();
+		const bool nonzero = negated != nullptr && negated->GetOpcode() == ValueOpcode::LogicalNot;
 		std::vector<const Block*> pending {use};
 		for (size_t index = 0; index < pending.size(); ++index) {
 			const auto* current = pending[index];
@@ -1837,9 +1921,9 @@ private:
 			for (const auto* previous: current->ImmPredecessors()) {
 				if (!witness.IsEmpty()) {
 					const auto edge = ConditionalEdge(previous, current);
-					// The proven local-zero invocation makes an all-inactive bypass impossible.
-					if (edge && edge->lanes == LaneQuantifier::All && !edge->positive &&
-					    Implies(witness, edge->condition)) continue;
+					if (edge && edge->lanes == LaneQuantifier::All &&
+					    (edge->positive ? nonzero && Implies(edge->condition, negated->Arg(0))
+					                    : Implies(witness, edge->condition))) continue;
 				}
 				if (std::ranges::find(pending, previous) == pending.end()) pending.push_back(previous);
 			}
@@ -1948,7 +2032,10 @@ private:
 
 	bool LocalZeroWitness(Value guard) const {
 		const auto* equal = SimplifyGuard(guard).TryInstruction();
-		if (equal == nullptr || equal->GetOpcode() != ValueOpcode::IEqual32) return false;
+		const bool negated = equal != nullptr && equal->GetOpcode() == ValueOpcode::LogicalNot;
+		if (negated) equal = SimplifyGuard(equal->Arg(0)).TryInstruction();
+		const auto expected = negated ? ValueOpcode::INotEqual32 : ValueOpcode::IEqual32;
+		if (equal == nullptr || equal->GetOpcode() != expected) return false;
 		for (uint32_t arg = 0; arg < 2u; ++arg) {
 			if (equal->Arg(arg).Resolve() != Value(0u)) continue;
 			Value induction;
@@ -1982,6 +2069,10 @@ private:
 		    EquivalentValue(m_program, test->Arg(0), value)) return test->Arg(1);
 		if ((positive ? op == ValueOpcode::UGreaterThan32 : op == ValueOpcode::ULessThanEqual32) &&
 		    EquivalentValue(m_program, test->Arg(1), value)) return test->Arg(0);
+		uint32_t limit = 0;
+		if ((positive ? op == ValueOpcode::ULessThanEqual32 : op == ValueOpcode::UGreaterThan32) &&
+		    ImmediateU32(test->Arg(1), limit) && limit != UINT32_MAX &&
+		    EquivalentValue(m_program, test->Arg(0), value)) return Value(limit + 1u);
 		const auto* scan = value.Resolve().TryInstruction();
 		if (finite_bitscan && scan != nullptr && scan->GetOpcode() == ValueOpcode::FindILsb32 &&
 		    (positive ? op == ValueOpcode::INotEqual32 : op == ValueOpcode::IEqual32)) {
@@ -2056,7 +2147,7 @@ private:
 		const auto* read = ActiveValue(value, active).TryInstruction();
 		if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadSharedU32 ||
 		    m_program.stage != ShaderType::Compute ||
-		    read->Arg(1).Resolve() != Value(true)) return {};
+		    !Implies(active, read->Arg(1))) return {};
 		uint64_t group_size = 1u;
 		for (const auto dimension: m_input.compute->threads_num) {
 			if (dimension == 0u || dimension > m_input.compute->wave_size) return {};
@@ -2087,10 +2178,16 @@ private:
 		const auto* limit = BoundedLoop(read_offset.index, read->Parent(),
 		    [](const Inst& compare, uint32_t) { return compare.GetOpcode() == ValueOpcode::ULessThan32; });
 		if (limit == nullptr) return {};
+		auto counter_active = active;
 		const auto* counter_read = ActiveValue(limit->Arg(1), active).TryInstruction();
+		// Every lane reads the same counter word, so a selected lane carries it.
+		if (counter_read != nullptr && counter_read->GetOpcode() == ValueOpcode::ReadFirstLane) {
+			counter_active = counter_read->Arg(1);
+			counter_read = ActiveValue(counter_read->Arg(0), counter_active).TryInstruction();
+		}
 		uint32_t consumed_counter = 0;
 		if (counter_read == nullptr || counter_read->GetOpcode() != ValueOpcode::LoadSharedU32 ||
-		    counter_read->Arg(1).Resolve() != Value(true) ||
+		    !Implies(counter_active, counter_read->Arg(1)) ||
 		    !constant_address(*counter_read, consumed_counter)) return {};
 
 		struct Initializer { const Inst* inst; uint32_t address; uint32_t words; };
@@ -2304,25 +2401,23 @@ private:
 		    ((offset | stride | memory->offset) & 3u) == 0u &&
 		    uint64_t {offset} + memory->offset <= UINT32_MAX;
 		if (affine) {
-			workgroup_axis = WorkgroupAxis(index);
-			if (workgroup_axis == UINT32_MAX) {
-				const auto* sum = index.Resolve().TryInstruction();
-				if (sum != nullptr && sum->GetOpcode() == ValueOpcode::IAdd32) {
-					for (uint32_t arm = 0; arm < 2u; ++arm) {
-						const auto axis = WorkgroupAxis(sum->Arg(arm));
-						const auto base = sum->Arg(arm ^ 1u);
-						if (axis == UINT32_MAX ||
-						    !ValidateRuntimeValue(m_program, base, RuntimeValueType::Integer)) continue;
-						workgroup_axis = axis;
-						first = base;
-						break;
-					}
-				}
-			}
+			workgroup_axis = WorkgroupAxis(index, &first);
 			const auto* bound = BoundedLoop(index, read->Parent());
 			if (bound != nullptr && bound->GetOpcode() == ValueOpcode::ULessThan32) {
 				count = bound->Arg(1);
+			} else if (workgroup_axis == UINT32_MAX && !require_bound) {
+				// Other bounded indices, such as bit scans, read the records up to their maximum;
+				// a bit scan of a nonzero mask load reads only the ranks its words set.
+				const auto maximum = SelectorMaximum(index);
+				if (maximum == 31u) selector.ranks = BitScanRanks(index);
+				if (maximum < MaxIndirectDescriptorProbes) count = Value(maximum + 1u);
 			}
+			// A record selected from a shared append list is below the list's bound.
+			const auto* lane = index.Resolve().TryInstruction();
+			if (count.IsEmpty() && lane != nullptr &&
+			    lane->GetOpcode() == ValueOpcode::ReadFirstLane &&
+			    HasActiveLane(lane->Arg(1), lane->Parent()))
+				count = SharedLoadBound(lane->Arg(0), lane->Arg(1));
 		}
 		// Raw pointers have no V# bounds for enumerating an unbounded selector domain.
 		if (memory->kind == ResourceKind::ScalarAddress && count.IsEmpty() &&
@@ -2498,7 +2593,8 @@ private:
 	// Each equation is an inductive subset obligation. Phis require every incoming
 	// value; AND needs either operand. Grounding rejects cycles without an initial
 	// value, then the greatest fixed point checks every initial and carried edge.
-	bool BitsetMaskSubset(Value mask, Value active, Value word, const Inst& read) const {
+	bool BitsetMaskSubset(Value mask, Value active, Value word,
+	                      std::span<const Inst* const> reads) const {
 		struct State {
 			Value value;
 			Value guard;
@@ -2510,7 +2606,10 @@ private:
 			bool all = true;
 			std::vector<size_t> terms;
 		};
-		const auto load_guard = read.Arg(4).Resolve();
+		// A scalar load reads one word without EXEC; four-word loads take their guard.
+		const bool scalar = reads[0]->GetOpcode() == ValueOpcode::ReadConstBuffer;
+		const auto load_guard = scalar ? Value(true) : reads[0]->Arg(4).Resolve();
+		const auto words = static_cast<uint32_t>(reads.size()) * (scalar ? 1u : 4u);
 		std::vector<State> states(2);
 		std::vector<Equation> equations(2);
 		const auto add = [&](State state) {
@@ -2521,9 +2620,9 @@ private:
 			equations.emplace_back();
 			return states.size() - 1u;
 		};
-		std::array<size_t, 5> roots {add({.value = active})};
-		for (uint32_t component = 0; component < 4u; ++component)
-			roots[component + 1u] = add({mask, active, component, component});
+		std::vector<size_t> roots {add({.value = active})};
+		for (uint32_t component = 0; component < words; ++component)
+			roots.push_back(add({mask, active, component, component}));
 		for (size_t index = 2; index < states.size(); ++index) {
 			const auto state = states[index];
 			const auto* inst = state.value.TryInstruction();
@@ -2573,11 +2672,14 @@ private:
 				switch (inst->GetOpcode()) {
 					case ValueOpcode::CompositeExtractU32x4: {
 						uint32_t component;
-						if (inst->Arg(0).Resolve().TryInstruction() == &read &&
-						    ImmediateU32(inst->Arg(1), component) && component == state.component)
+						if (inst->Arg(0).Resolve().TryInstruction() == reads[state.component / 4u] &&
+						    ImmediateU32(inst->Arg(1), component) && component == state.component % 4u)
 							equation.terms = {1u};
 						break;
 					}
+					case ValueOpcode::ReadConstBuffer:
+						if (scalar && inst == reads[state.component]) equation.terms = {1u};
+						break;
 					case ValueOpcode::Phi:
 						// The predecessor may be an earlier outer iteration. Establish
 						// this register's invariant without the current word or EXEC.
@@ -2598,17 +2700,17 @@ private:
 								const auto* masked = key.TryInstruction();
 								uint32_t bound = 0;
 								if (masked != nullptr && masked->GetOpcode() == ValueOpcode::BitwiseAnd32 &&
-								    DescendingLoop(word, masked->Parent(), bound) && bound == 3u) {
+								    DescendingLoop(word, masked->Parent(), bound) && bound == words - 1u) {
 									for (uint32_t side = 0; side < 2u; ++side) {
 										uint32_t bits;
-										if (ImmediateU32(masked->Arg(side), bits) && (bits & 3u) == 3u)
+										if (ImmediateU32(masked->Arg(side), bits) && (bits & bound) == bound)
 											key = masked->Arg(side ^ 1u).Resolve();
 									}
 								}
 								return EquivalentValue(m_program, key, word);
 							});
 							uint32_t literal;
-							if (ImmediateU32(equality, literal) && literal < 4u) selected_word = literal;
+							if (ImmediateU32(equality, literal) && literal < words) selected_word = literal;
 							equation.terms.push_back(term(inst->Arg(1), inst->Arg(0), selected_word));
 						}
 						if (!selected || !*selected)
@@ -2646,7 +2748,53 @@ private:
 		return std::ranges::all_of(roots, [&](size_t root) { return valid[root]; });
 	}
 
-	bool MatchBitsetSelector(Value rank, Value active, SelectorBounds bounds,
+	// The mask words whose set bits are every record a bit scan of index selects: the scanned
+	// value keeps only bits of one scalar load, which reads a word its address can reach.
+	std::optional<DescriptorSource::IndirectDescriptor::SelectorRead::MaskRanks>
+	BitScanRanks(Value index) {
+		const auto* scan = index.Resolve().TryInstruction();
+		if (scan == nullptr || scan->GetOpcode() != ValueOpcode::FindILsb32) return {};
+		const auto* masked = scan->Arg(0).Resolve().TryInstruction();
+		if (masked == nullptr) return {};
+		std::vector<const Inst*> values {masked};
+		const Inst* load = nullptr;
+		for (size_t i = 0; i < values.size() && load == nullptr; ++i) {
+			const auto* value = values[i];
+			if (value->GetOpcode() == ValueOpcode::ReadConstBuffer) load = value;
+			if (value->GetOpcode() != ValueOpcode::Phi &&
+			    value->GetOpcode() != ValueOpcode::BitwiseAnd32) continue;
+			for (size_t arg = 0; arg < value->NumArgs(); ++arg) {
+				const auto* next = value->Arg(arg).Resolve().TryInstruction();
+				if (next != nullptr && std::ranges::find(values, next) == values.end())
+					values.push_back(next);
+			}
+		}
+		uint32_t memory_index = 0;
+		const auto* memory = load != nullptr ? ScalarReadMemory(*load, memory_index) : nullptr;
+		if (memory == nullptr || memory->offset > INT32_MAX ||
+		    !MemoryIndexBelongsTo(memory_index, *load) ||
+		    !BitsetMaskSubset(scan->Arg(0), Value(true), Value(0u), std::array {load})) return {};
+		auto address = load->Arg(1);
+		auto active = Value(true);
+		const auto* lane = address.Resolve().TryInstruction();
+		if (lane != nullptr && lane->GetOpcode() == ValueOpcode::ReadFirstLane) {
+			if (!HasActiveLane(lane->Arg(1), lane->Parent())) return {};
+			active = lane->Arg(1);
+			address = lane->Arg(0);
+		}
+		// A left shift keeps the address a multiple of its scale; other addresses read any word.
+		const auto* shift = ActiveValue(address, active).TryInstruction();
+		uint32_t bits = 0;
+		if (shift == nullptr || shift->GetOpcode() != ValueOpcode::ShiftLeftLogical32 ||
+		    !ImmediateU32(shift->Arg(1), bits) || bits >= 32u) bits = 0;
+		DescriptorSource source;
+		if (!MakeRuntimeTableSource(*load, source)) return {};
+		return DescriptorSource::IndirectDescriptor::SelectorRead::MaskRanks {
+		    .source = InternSource(source), .stride = std::max(1u << bits, 4u),
+		    .offset = memory->offset & ~3u};
+	}
+
+	bool MatchBitsetSelector(Value rank, Value active, SelectorBounds bounds, uint32_t words,
 	                         DescriptorSource::IndirectDescriptor& indirect) {
 		const auto selected_value = [&](Value value) {
 			for (;;) {
@@ -2670,7 +2818,7 @@ private:
 		}
 		const auto* word_inst = word.TryInstruction();
 		if (word_inst == nullptr || std::ranges::none_of(bounds, [&](const auto& bound) {
-			return bound.first == word_inst && bound.second == 3u;
+			return bound.first == word_inst && bound.second == words - 1u;
 		})) return false;
 		// FFBH's leading-zero count is inverted again to recover the set-bit rank.
 		for (uint32_t inverse = 0; inverse < 2u; ++inverse) {
@@ -2682,21 +2830,34 @@ private:
 		const auto* scan = bit.TryInstruction();
 		if (scan == nullptr || scan->GetOpcode() != ValueOpcode::FindUMsb32 ||
 		    !ImpliesNonzero(active, scan->Arg(0))) return false;
+		const auto bitmap_load = [&](const Inst& read, uint32_t offset) {
+			const auto index = read.Flags<MemoryFlags>().index;
+			if (read.GetOpcode() != ValueOpcode::LoadBufferU32x4 || InCycle(read.Parent()) ||
+			    index >= m_program.memory_info.size()) return false;
+			const auto& memory = m_program.memory_info[index];
+			return memory.kind == ResourceKind::Buffer && !memory.typed && !memory.formatted &&
+			       memory.idxen && !memory.offen && memory.offset == offset &&
+			       memory.data_bits == 32u && memory.data_dwords == 4u &&
+			       read.Arg(2).Resolve() == Value(0u) && read.Arg(3).Resolve() == Value(0u);
+		};
 		for (const auto* block: m_program.blocks) {
 			for (const auto& read: *block) {
-				if (read.GetOpcode() != ValueOpcode::LoadBufferU32x4 || InCycle(block)) continue;
-				const auto index = read.Flags<MemoryFlags>().index;
-				if (index >= m_program.memory_info.size()) continue;
-				const auto& memory = m_program.memory_info[index];
-				if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
-				    !memory.idxen || memory.offen || memory.offset != 0u || memory.data_bits != 32u ||
-				    memory.data_dwords != 4u || read.Arg(2).Resolve() != Value(0u) ||
-				    read.Arg(3).Resolve() != Value(0u) ||
-				    !BitsetMaskSubset(scan->Arg(0), active, word, read)) continue;
+				if (!bitmap_load(read, 0u)) continue;
+				// An eight-word record continues in a second four-word load of the same buffer and EXEC.
+				std::array<const Inst*, 2> reads {&read};
+				if (words == 8u)
+					for (const auto* other: m_program.blocks)
+						for (const auto& load: *other)
+							if (reads[1] == nullptr && bitmap_load(load, 16u) &&
+							    EquivalentValue(m_program, load.Arg(0), read.Arg(0)) &&
+							    EquivalentValue(m_program, load.Arg(4), read.Arg(4))) reads[1] = &load;
+				const auto parts = std::span(reads).first(words / 4u);
 				DescriptorSource source;
-				if (!MakeRuntimeTableSource(read, source)) continue;
+				if (std::ranges::find(parts, nullptr) != parts.end() ||
+				    !BitsetMaskSubset(scan->Arg(0), active, word, parts) ||
+				    !MakeRuntimeTableSource(read, source)) continue;
 				indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead {
-				    .source = InternSource(source), .stride = 16u, .offset = 0u, .bitset = true});
+				    .source = InternSource(source), .stride = words * 4u, .offset = 0u, .bitset = true});
 				return true;
 			}
 		}
@@ -2726,7 +2887,8 @@ private:
 		first.SetArg(1, Value(maximum));
 		indirect.selector_first = Value(&first);
 		indirect.key_count = Value(maximum + 1u);
-		if (maximum == 127u) MatchBitsetSelector(lane->Arg(0), lane->Arg(1), bounds, indirect);
+		if (maximum == 127u || maximum == 255u)
+			MatchBitsetSelector(lane->Arg(0), lane->Arg(1), bounds, (maximum + 1u) / 32u, indirect);
 		return true;
 	}
 
@@ -2993,6 +3155,47 @@ private:
 		return true;
 	}
 
+	// The EXEC predicate of the image access through use; empty for any other use.
+	Value AccessPredicate(const Use& use, bool sampler) const {
+		const auto& access = *use.user;
+		if (use.operand != (sampler ? 1u : 0u) ||
+		    ImageOpcodeInfoOf(access.GetOpcode()).access == ImageAccess::None) return {};
+		if (sampler && (access.GetOpcode() != ValueOpcode::ImageSampleRaw ||
+		    (m_program.memory_info[access.Flags<MemoryFlags>().index].image_sample_flags &
+		     Decoder::ImageSampleFlagAdjust) != 0u)) return {};
+		const auto mask = access.Arg(access.NumArgs() - 1u);
+		return mask.GetType() == Type::U1 ? mask : Value {};
+	}
+
+	// A key no bound covers takes its maximum from a signed clamp to a constant, on the key or on
+	// the key plus a constant in an earlier-tracked sibling read of the same table. Negative keys
+	// and keys past that maximum read the first row.
+	uint32_t ClampedTableMaximum(Value key, const DescriptorSource& table,
+	                             const DescriptorSource::IndirectDescriptor& indirect) const {
+		uint32_t row = 0, offset = 0;
+		const auto clamped = [&](Value value) {
+			const auto* min = value.Resolve().TryInstruction();
+			if (min != nullptr && min->GetOpcode() == ValueOpcode::SMin32)
+				for (uint32_t arg = 0; arg < 2u; ++arg)
+					if (ImmediateU32(min->Arg(arg), row)) return min->Arg(arg ^ 1u);
+			return Value {};
+		};
+		if (!clamped(key).IsEmpty()) return row;
+		for (const auto& sibling: m_indirect_descriptors) {
+			const auto& read = *m_sources[sibling.source].indirect_descriptor;
+			const auto* add = clamped(sibling.key).Resolve().TryInstruction();
+			if (add == nullptr || add->GetOpcode() != ValueOpcode::IAdd32 ||
+			    m_sources[read.table_source] != table ||
+			    read.table_offset != indirect.table_offset ||
+			    read.table_stride != indirect.table_stride ||
+			    read.table_immediate != indirect.table_immediate) continue;
+			for (uint32_t arg = 0; arg < 2u; ++arg)
+				if (ImmediateU32(add->Arg(arg), offset) &&
+				    EquivalentValue(m_program, add->Arg(arg ^ 1u), key)) return row;
+		}
+		return UINT32_MAX;
+	}
+
 	bool TryMakeTableDescriptor(Inst& handle, const DescriptorSource& descriptor,
 	                            IndirectDescriptorPlan& plan) {
 		const bool sampler = handle.GetOpcode() == ValueOpcode::GetSamplerResource;
@@ -3019,6 +3222,7 @@ private:
 		} else if (!plan.table_indexed &&
 		           MatchScalarMaterialKey(key, indirect, material_source,
 		                                  !sampler && table_source.dword_count == 4u, sampler)) {
+			indirect.selector->excluded = ExcludedKeyOnEntry(key, handle.Parent());
 			// Keep live descriptor payloads, but project descriptor-only SMEM onto
 			// the GPU key using the same plan as other indirect image reads.
 			plan.retain_reads |= table_source.dword_count == 4u;
@@ -3080,36 +3284,42 @@ private:
 			if (indirect.key_count.IsEmpty() &&
 			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source, false) &&
 			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source, true) &&
-			    !MatchReversedImageRange(key, indirect)) {
-				const auto maximum = SelectorMaximum(key);
-				if (maximum >= MaxIndirectDescriptorProbes) return false;
-				indirect.key_count = Value(maximum + 1u);
+			    !MatchReversedImageRange(key, indirect) && !MatchWorkgroupKey(key, indirect)) {
+				auto maximum = SelectorMaximum(key);
+				if (maximum >= MaxIndirectDescriptorProbes) {
+					// Each image access bounds the key on the lanes it executes.
+					maximum = 0u;
+					for (const auto& use: handle.Uses()) {
+						const auto active = AccessPredicate(use, false);
+						if (active.IsEmpty()) {
+							maximum = UINT32_MAX;
+							break;
+						}
+						maximum = std::max(maximum, SelectorMaximum(key, {}, 0, active));
+					}
+				}
+				if (maximum >= MaxIndirectDescriptorProbes)
+					maximum = ClampedTableMaximum(key, table_source, indirect);
+				if (maximum < MaxIndirectDescriptorProbes)
+					indirect.key_count = Value(maximum + 1u);
+				else if (!ProgramLoopCount(key, indirect))
+					return false;
 			}
 			if ((table_offset & 3u) != 0u) return false;
 		} else {
 			// Keep the GPU selector and descriptor reads, and preserve the proven scalar
 			// offset domain within the bounded image table.
 			if (sampler || plan.split_offsets) return false;
-			const auto active = [&](const Use& use, bool sampler) -> Value {
-				const auto& access = *use.user;
-				if (use.operand != (sampler ? 1u : 0u) ||
-				    ImageOpcodeInfoOf(access.GetOpcode()).access == ImageAccess::None) return {};
-				if (sampler && (access.GetOpcode() != ValueOpcode::ImageSampleRaw ||
-				    (m_program.memory_info[access.Flags<MemoryFlags>().index].image_sample_flags &
-				     Decoder::ImageSampleFlagAdjust) != 0u)) return {};
-				const auto mask = access.Arg(access.NumArgs() - 1u);
-				return mask.GetType() == Type::U1 ? mask : Value {};
-			};
 			bool shared_selector = false;
 			if (!plan.table_indexed) {
 				for (const auto& owner: m_indirect_descriptors) {
 					const auto& selected = *m_sources[owner.source].indirect_descriptor;
 					if (!selected.selector || !EquivalentValue(m_program, key, owner.key) ||
 					    !std::ranges::all_of(handle.Uses(), [&](const Use& target) {
-						    const auto target_active = active(target, false);
+						    const auto target_active = AccessPredicate(target, false);
 						    return !target_active.IsEmpty() &&
 						        std::ranges::any_of(owner.handle->Uses(), [&](const Use& witness) {
-							        const auto owner_active = active(witness,
+							        const auto owner_active = AccessPredicate(witness,
 							            owner.handle->GetOpcode() == ValueOpcode::GetSamplerResource);
 							        return !owner_active.IsEmpty() && AvailableBefore(*witness.user, *target.user) &&
 							            Implies(target_active, owner_active);
@@ -3430,6 +3640,42 @@ private:
 		return result;
 	}
 
+	// Bounds of an F32 value on the lanes where guard holds; NaN when unknown.
+	std::pair<float, float> FloatRange(Value value, Value guard, uint32_t depth = 0) const {
+		constexpr std::pair unknown {NAN, NAN};
+		value = ActiveValue(value, guard);
+		if (value.IsImmediate() && value.GetType() == Type::F32)
+			return {value.F32Value(), value.F32Value()};
+		const auto* inst = value.TryInstruction();
+		// Each multiply-add evaluates three operands; keep the search shallow.
+		if (depth > 12u || inst == nullptr) return unknown;
+		const auto range = [&](uint32_t arg) {
+			return FloatRange(inst->Arg(arg), guard, depth + 1u);
+		};
+		switch (inst->GetOpcode()) {
+			case ValueOpcode::BitCastF32U32:
+			case ValueOpcode::BitCastU32F32: return range(0);
+			// Clamping maps NaN to zero.
+			case ValueOpcode::FPSaturate32: return {0.0f, 1.0f};
+			case ValueOpcode::FPFloor32: {
+				const auto [low, high] = range(0);
+				return {std::floor(low), std::floor(high)};
+			}
+			case ValueOpcode::FPMad32: {
+				// The rounded multiply and add are monotonic, so the bounds lie at the corners.
+				const auto a = range(0), b = range(1), c = range(2);
+				if (!std::isfinite(a.first) || !std::isfinite(a.second) ||
+				    !std::isfinite(b.first) || !std::isfinite(b.second)) return unknown;
+				const auto [low, high] = std::minmax({a.first * b.first, a.first * b.second,
+				                                      a.second * b.first, a.second * b.second});
+				const std::pair bounds {low + c.first, high + c.second};
+				if (!std::isfinite(bounds.first) || !std::isfinite(bounds.second)) return unknown;
+				return bounds;
+			}
+			default: return unknown;
+		}
+	}
+
 	uint32_t SelectorMaximum(Value value, SelectorBounds bounds = {}, uint32_t depth = 0,
 	                         Value guard = {}) const {
 		if (depth > 32u) return UINT32_MAX;
@@ -3445,9 +3691,13 @@ private:
 		switch (inst->GetOpcode()) {
 			case ValueOpcode::ReadFirstLane:
 				return SelectedSelectorDomain(*inst, guard).maximum;
-			case ValueOpcode::ReadLane:
-				// ReadLane ignores EXEC; only context-free bounds on every source lane apply.
-				return PossibleU32Bits(inst->Arg(0));
+			case ValueOpcode::ReadLane: {
+				// ReadLane ignores EXEC; only context-free bounds on every source lane apply,
+				// unless guard proves a waterfall key equal to the executing lanes' local value.
+				const auto local = EqualLocalKey(guard, value);
+				return local.IsEmpty() ? PossibleU32Bits(inst->Arg(0))
+				                       : SelectorMaximum(local, bounds, depth + 1u, guard);
+			}
 			case ValueOpcode::Phi: {
 				const auto initial = InitialCandidateMask(value);
 				if (!initial.IsEmpty()) return SelectorMaximum(initial, bounds, depth + 1u, guard);
@@ -3474,6 +3724,18 @@ private:
 				               NonzeroOnEntry(inst->Arg(0), inst->Parent()) ? 31u : UINT32_MAX;
 			case ValueOpcode::UMin32:
 			case ValueOpcode::BitwiseAnd32: return std::min(maximum(0), maximum(1));
+			case ValueOpcode::SMin32: {
+				// Signed and unsigned minimums agree on nonnegative operands.
+				const auto left = maximum(0), right = maximum(1);
+				if (std::max(left, right) <= INT32_MAX) return std::min(left, right);
+				break;
+			}
+			case ValueOpcode::ConvertS32F32: {
+				// The conversion truncates toward zero.
+				const auto [low, high] = FloatRange(inst->Arg(0), guard);
+				if (low >= 0.0f && high < 0x1p31f) return static_cast<uint32_t>(high);
+				break;
+			}
 			case ValueOpcode::SelectU32:
 				if (!guard.IsEmpty() && Implies(guard, inst->Arg(0))) return maximum(1);
 				return SelectorPredicateFalse(inst->Arg(0), bounds, depth + 1u, guard)

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

@@ -175,6 +175,8 @@ void AddBuiltinInput(ShaderInfo& info, StageInputKind kind) {
 		case StageInputKind::GlobalInvocationId:
 			AddInput(info, kind, 0, 3, "gl_GlobalInvocationID");
 			break;
+		case StageInputKind::BaseVertex: AddInput(info, kind, 0, 1, "gl_BaseVertex"); break;
+		case StageInputKind::BaseInstance: AddInput(info, kind, 0, 1, "gl_BaseInstance"); break;
 		case StageInputKind::PackedAncillary:
 		case StageInputKind::Parameter: break;
 	}
@@ -444,6 +446,8 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 				case StageInputKind::InstanceIndex:
 				case StageInputKind::FrontFacing:
 				case StageInputKind::LocalInvocationIndex:
+				case StageInputKind::BaseVertex:
+				case StageInputKind::BaseInstance:
 					if (component != 0u) {
 						return Fail("typed scalar builtin component is out of range");
 					}

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

@@ -82,6 +82,7 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::BitCastU32F32:
 		case ValueOpcode::BitCastF32U32:
 		case ValueOpcode::ConvertU32F32:
+		case ValueOpcode::ConvertS32F32:
 		case ValueOpcode::ConvertF32U32:
 		case ValueOpcode::CompositeConstructU64:
 		case ValueOpcode::CompositeExtractU64:
@@ -147,8 +148,8 @@ bool ImmediateBits(Value value, uint64_t& bits) {
 }
 
 // Scratch layout in bytes: per descriptor source its eight DWORDs and a known DWORD, one
-// activity byte per source, then one qword per walked value (all ones when unknown). The
-// walker's value slots, captured addresses, reachability and states follow.
+// activity byte per source, then per image, then one qword per walked value (all ones when
+// unknown). The walker's value slots, captured addresses, reachability and states follow.
 constexpr uint32_t DescriptorStride = 9u * sizeof(uint32_t);
 
 uint32_t ActiveOffset(const ResourcePlan& plan) {
@@ -156,7 +157,8 @@ uint32_t ActiveOffset(const ResourcePlan& plan) {
 }
 
 uint32_t OutputOffset(const ResourcePlan& plan) {
-	return (ActiveOffset(plan) + static_cast<uint32_t>(plan.descriptor_sources.size()) + 7u) & ~7u;
+	const auto active = plan.descriptor_sources.size() + plan.info.images.size();
+	return (ActiveOffset(plan) + static_cast<uint32_t>(active) + 7u) & ~7u;
 }
 
 const uint8_t* Scratch(const ResourcePlan& plan) {
@@ -1038,6 +1040,23 @@ void WalkerCompiler::EmitStrict(const Node& node) {
 			c.L(done);
 			break;
 		}
+		case ValueOpcode::ConvertS32F32: {
+			// The conversion saturates and maps NaN to zero. cvttss2si's out-of-range result,
+			// INT32_MIN, is already right for negative overflow.
+			Xbyak::Label done;
+			c.movd(c.xmm0, c.eax);
+			c.xor_(c.eax, c.eax);
+			c.ucomiss(c.xmm0, c.xmm0);
+			c.jp(done, Near);
+			c.mov(c.ecx, 0x4f000000u);
+			c.movd(c.xmm1, c.ecx);
+			c.mov(c.eax, INT32_MAX);
+			c.ucomiss(c.xmm0, c.xmm1);
+			c.jae(done, Near);
+			c.cvttss2si(c.eax, c.xmm0);
+			c.L(done);
+			break;
+		}
 		case ValueOpcode::FPMul32:
 			c.movd(c.xmm0, c.eax);
 			c.movd(c.xmm1, c.ecx);
@@ -1272,11 +1291,14 @@ void WalkerCompiler::Compile(std::span<const Value> values) {
 			for (const auto slot: blocks[block].srt_reads) flat(slot);
 			c.L(skip);
 		}
-		// A source no block lists is always active.
-		for (uint32_t source = 0; source < sources.size(); ++source) {
+		// A source or image no block lists is always active.
+		for (uint32_t index = 0; index < sources.size() + m_plan.info.images.size(); ++index) {
+			const bool     image = index >= sources.size();
+			const uint32_t id    = image ? index - static_cast<uint32_t>(sources.size()) : index;
 			bool listed = false;
 			for (uint32_t block = 0; block < blocks.size(); ++block) {
-				if (std::ranges::find(blocks[block].sources, source) == blocks[block].sources.end()) continue;
+				const auto& list = image ? blocks[block].images : blocks[block].sources;
+				if (std::ranges::find(list, id) == list.end()) continue;
 				if (listed) {
 					c.or_(c.al, c.byte[c.rdx + m_reach + block]);
 				} else {
@@ -1284,7 +1306,7 @@ void WalkerCompiler::Compile(std::span<const Value> values) {
 				}
 				listed = true;
 			}
-			const auto active = c.byte[c.rdx + ActiveOffset(m_plan) + source];
+			const auto active = c.byte[c.rdx + ActiveOffset(m_plan) + index];
 			if (listed) {
 				c.mov(active, c.al);
 			} else {
@@ -1391,6 +1413,10 @@ bool WalkedActive(const ResourcePlan& plan, uint32_t source) {
 	return Scratch(plan)[ActiveOffset(plan) + source] != 0u;
 }
 
+bool WalkedImageActive(const ResourcePlan& plan, uint32_t image) {
+	return Scratch(plan)[ActiveOffset(plan) + plan.descriptor_sources.size() + image] != 0u;
+}
+
 bool ValidateRuntimeValue(const ResourcePlan& program, Value value, RuntimeValueType type) {
 	return RuntimeValidator(program, type).Run(value);
 }

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

@@ -78,6 +78,7 @@ bool WalkSrt(const ResourcePlan& plan, const SrtRuntime& runtime, std::vector<ui
 bool WalkedValue(const ResourcePlan& plan, Value value, uint32_t& result);
 bool WalkedDescriptor(const ResourcePlan& plan, uint32_t source, DescriptorValue& result);
 bool WalkedActive(const ResourcePlan& plan, uint32_t source);
+bool WalkedImageActive(const ResourcePlan& plan, uint32_t image);
 // Length of the guest load at rip if it belongs to the walker running on this thread, else 0.
 // The exception handler completes such a load as a zero read.
 uint32_t SrtWalkerLoadLength(uint64_t rip);

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

@@ -348,14 +348,20 @@ void VisitInstruction(Pass& pass, Block* block, Inst& inst) {
 } // namespace
 
 void RewriteToSsa(const BlockList& blocks) {
-	Pass pass;
+	Pass                               pass;
+	std::unordered_map<Block*, size_t> filled_predecessors;
 	for (auto* block: blocks) {
+		if (block->ImmPredecessors().empty()) {
+			pass.Seal(block);
+		}
 		for (auto& inst: *block) {
 			VisitInstruction(pass, block, inst);
 		}
-	}
-	for (auto* block: blocks) {
-		pass.Seal(block);
+		for (auto* successor: block->ImmSuccessors()) {
+			if (++filled_predecessors[successor] == successor->ImmPredecessors().size()) {
+				pass.Seal(successor);
+			}
+		}
 	}
 }
 

src/graphics/shader/shader.cpp

@@ -291,10 +291,9 @@ static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserS
 	const auto& metadata = data.vertex;
 
 	if (metadata.vertex_attrib_reg >= 0) {
-		info.fetch_external   = false;
-		info.fetch_embedded   = true;
-		info.fetch_attrib_reg = metadata.vertex_attrib_reg;
-		info.fetch_buffer_reg = metadata.vertex_buffer_reg;
+		info.has_dynamic_vertex_fetch = true;
+		info.fetch_attrib_reg         = metadata.vertex_attrib_reg;
+		info.fetch_buffer_reg         = metadata.vertex_buffer_reg;
 
 		const auto* attrib = reinterpret_cast<const uint32_t*>(
 		    static_cast<uint64_t>(user_sgpr.value[metadata.vertex_attrib_reg]) |
@@ -413,7 +412,7 @@ static void ShaderGetStaticInputInfoCS(const HW::ComputeShaderInfo& regs,
 void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t>& key) {
 	EXIT_IF(info.resources_num < 0 || info.resources_num > ShaderVertexInputInfo::RES_MAX);
 	key.clear();
-	key.push_back(static_cast<uint32_t>(info.fetch_embedded));
+	key.push_back(static_cast<uint32_t>(info.has_dynamic_vertex_fetch));
 	key.push_back(static_cast<uint32_t>(info.fetch_attrib_reg));
 	key.push_back(static_cast<uint32_t>(info.fetch_buffer_reg));
 	key.push_back(info.resources_num);
@@ -422,6 +421,8 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 	key.push_back(info.lds_size_dwords);
 	key.push_back(info.pa_cl_vs_out_cntl);
 	key.push_back(info.user_clip_planes);
+	key.push_back(info.indirect_offsets.vertex | (uint32_t {info.indirect_offsets.instance} << 8u) |
+	              (uint32_t {info.indirect_draw} << 16u));
 	key.push_back(static_cast<uint32_t>(info.clip_space.enabled));
 	if (info.clip_space.enabled) {
 		for (const float value: info.clip_space.scale) {
@@ -692,9 +693,7 @@ void ShaderDbgDumpInputInfo(const ShaderVertexInputInfo& info) {
 
 	LOGF("ShaderDbgDumpInputInfo()\n");
 
-	LOGF("\t fetch_external = %s\n"
-	     "\t fetch_embedded = %s\n",
-	     info.fetch_external ? "true" : "false", info.fetch_embedded ? "true" : "false");
+	LOGF("\t has_dynamic_vertex_fetch = %s\n", info.has_dynamic_vertex_fetch ? "true" : "false");
 
 	for (int i = 0; i < info.resources_num; i++) {
 		LOGF("\t input %d\n", i);

src/graphics/shader/shader.h

@@ -124,6 +124,17 @@ struct ShaderTessellationInputInfo {
 	uint32_t output_topology       = 0;
 };
 
+// GS user-data slots an indirect draw packet writes its record's base vertex and first instance
+// to. A packet can leave both slots unused even when its record contains nonzero offsets.
+struct ShaderIndirectOffsets {
+	static constexpr uint8_t None = UINT8_MAX;
+
+	uint8_t vertex   = None;
+	uint8_t instance = None;
+
+	bool operator==(const ShaderIndirectOffsets&) const = default;
+};
+
 struct ShaderVertexInputInfo {
 	static constexpr int RES_MAX = 32;
 
@@ -146,8 +157,11 @@ struct ShaderVertexInputInfo {
 	ShaderClipSpaceTransform clip_space;
 	ShaderMeshInputInfo      mesh;
 	ShaderTessellationInputInfo tess;
-	bool                    fetch_external      = false;
-	bool                    fetch_embedded      = false;
+	ShaderIndirectOffsets   indirect_offsets;
+	// GPU-read records add Vulkan bases even when the packet patches no user SGPRs.
+	bool                    indirect_draw       = false;
+	// Attribute-table metadata enables dynamic fetch detection; loads are matched later.
+	bool                    has_dynamic_vertex_fetch = false;
 };
 
 struct ShaderComputeInputInfo: ShaderWorkgroupInputInfo {

src/kernel/memory.h

@@ -26,9 +26,10 @@ struct DebugStats {
 	uint64_t flexible_total      = 0;
 	uint64_t flexible_used       = 0;
 	uint64_t pool_committed      = 0;
-	uint64_t cpu_page_entries    = 0; // 2 MiB page-table entries in use.
-	uint64_t gpu_page_entries    = 0;
-	uint64_t page_entries_total  = 0; // Per side.
+	uint64_t cpu_page_entries       = 0; // 2 MiB page-table entries in use.
+	uint64_t gpu_page_entries       = 0;
+	uint64_t cpu_page_entries_total = 0;
+	uint64_t gpu_page_entries_total = 0;
 };
 [[nodiscard]] DebugStats GetDebugStats();
 
@@ -125,7 +126,6 @@ static_assert(sizeof(KernelMemoryPoolBlockStats) == 16,
               "KernelMemoryPoolBlockStats struct size is incorrect");
 
 void                   RegisterCallbacks(callback_func_t alloc_func, callback_func_t free_func);
-void                   SetFlexibleMemorySize(uint64_t size);
 int AllocateDirectMemory(int64_t search_start, int64_t search_end, size_t size, size_t alignment,
                          int memory_type, int64_t* phys_addr_out, bool automatic = false);
 int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot);
@@ -197,6 +197,9 @@ int KYTY_SYSV_ABI KernelMemoryPoolGetBlockStats(KernelMemoryPoolBlockStats* outp
 
 uint64_t AllocateProgramMemory(uint64_t search_addr, uint64_t size,
                                Common::VirtualMemory::Mode mode, const char* name);
+// Loaded modules' segments use flexible memory without taking its backing.
+void ChargeFlexibleMemory(uint64_t size);
+void ReleaseFlexibleMemory(uint64_t size);
 void SetProgramMemoryProtection(uint64_t vaddr, uint64_t size, Common::VirtualMemory::Mode mode);
 uint64_t AllocateRuntimeMemory(uint64_t search_addr, uint64_t size,
                                Common::VirtualMemory::Mode mode, const char* name,
@@ -222,6 +225,11 @@ bool     TestPlaceholderRangeIsFree(uint64_t vaddr, uint64_t size);
 bool     TestGuestAddressRangeIsOwned(uint64_t vaddr, uint64_t size);
 bool     TestGuestBackingOutsideAddressSpace();
 uint64_t TestGuestBackingSize();
+uint64_t TestFlexibleBackingHighWater();
+uint64_t TestGuestFreeRangeVisits();
+uint64_t TestDirectMappingCount();
+uint64_t TestFlexibleMappingCount();
+uint64_t TestBackingBytesZeroed();
 bool     TestGuestFreeRangeBounds();
 #endif
 

src/kernel/memoryAddressSpace.inc

@@ -203,6 +203,9 @@ public:
 		}
 		std::lock_guard lock(m_mutex);
 		std::memset(m_backing_base + backing_offset, 0, size);
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+		g_test_backing_bytes_zeroed += size;
+#endif
 		return true;
 	}
 
@@ -287,6 +290,18 @@ public:
 		return true;
 	}
 
+	bool ProtectViews(uint64_t vaddr, uint64_t size, VirtualMemory::Mode mode) {
+		std::shared_lock lock(m_mutex);
+		bool             ok = true;
+		ForEachSparseRange(
+		    m_maps, vaddr, size,
+		    [&](const auto&, uint64_t start, uint64_t bytes) {
+			    ok = ok && VirtualMemory::Protect(start, bytes, mode);
+		    },
+		    [&](uint64_t, uint64_t) { ok = false; });
+		return ok;
+	}
+
 	bool MapFixed(uint64_t vaddr, uint64_t size, uint64_t backing_offset,
 	              VirtualMemory::Mode mode, FailureReason* failure_reason = nullptr) {
 		if (!IsAvailable() || !RangeOk(backing_offset, size)) {
@@ -324,15 +339,11 @@ public:
 #endif
 	}
 
-	bool Unmap(uint64_t vaddr, uint64_t size, bool* placeholder_preserved = nullptr) {
+	bool Unmap(uint64_t vaddr, uint64_t size) {
 		if (size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
 
-		if (placeholder_preserved != nullptr) {
-			*placeholder_preserved = false;
-		}
-
 		const auto            end = vaddr + size;
 		std::vector<MapEntry> targets;
 		{
@@ -350,42 +361,36 @@ public:
 			}
 		}
 
-		std::vector<MapEntry> removed;
-		removed.reserve(targets.size());
+		size_t removed  = 0;
+		auto   rollback = [&] {
+			while (removed != 0) {
+				const auto& previous = targets[--removed];
+				EXIT_IF(!MapFixed(previous.vaddr, previous.size, previous.backing_offset,
+				                  previous.mode));
+			}
+		};
 		for (const auto& entry: targets) {
-			bool part_placeholder_preserved = false;
-			if (!UnmapOne(entry.vaddr, entry.size, &part_placeholder_preserved)) {
-				for (auto it = removed.rbegin(); it != removed.rend(); ++it) {
-					EXIT_IF(!MapFixed(it->vaddr, it->size, it->backing_offset, it->mode));
-				}
-				if (placeholder_preserved != nullptr) {
-					*placeholder_preserved = false;
-				}
+			if (!UnmapOne(entry.vaddr, entry.size)) {
+				rollback();
 				return false;
 			}
-			EXIT_IF(!part_placeholder_preserved);
-			removed.push_back(entry);
+			++removed;
 		}
 
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		// Each unmapped file view becomes a separate Windows placeholder. The address-space owner
 		// records this contiguous guest span as one free range, so normalize fragmented backing
 		// views into one placeholder before publishing it there.
-		if (removed.size() > 1 && !CoalescePlaceholders(vaddr, size)) {
-			for (auto it = removed.rbegin(); it != removed.rend(); ++it) {
-				EXIT_IF(!MapFixed(it->vaddr, it->size, it->backing_offset, it->mode));
-			}
+		if (removed > 1 && !CoalescePlaceholders(vaddr, size)) {
+			rollback();
 			return false;
 		}
 #endif
 
-		if (placeholder_preserved != nullptr) {
-			*placeholder_preserved = true;
-		}
 		return true;
 	}
 
-	bool UnmapOne(uint64_t vaddr, uint64_t size, bool* placeholder_preserved = nullptr) {
+	bool UnmapOne(uint64_t vaddr, uint64_t size) {
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
 		if (g_test_backing_store_unmaps_before_failure == 0) {
 			g_test_backing_store_unmaps_before_failure = UINT32_MAX;
@@ -428,9 +433,7 @@ public:
 				m_maps.emplace(old.vaddr, old);
 				return false;
 			}
-			if (placeholder_preserved != nullptr) {
-				*placeholder_preserved = true;
-			}
+
 			return true;
 		}
 
@@ -451,9 +454,6 @@ public:
 				     ok;
 			}
 			if (ok) {
-				if (placeholder_preserved != nullptr) {
-					*placeholder_preserved = true;
-				}
 				return true;
 			}
 
@@ -481,9 +481,6 @@ public:
 		std::lock_guard lock(m_mutex);
 		RemoveMappingsUnlocked(vaddr, size);
 
-		if (placeholder_preserved != nullptr) {
-			*placeholder_preserved = true;
-		}
 		return true;
 #endif
 	}
@@ -634,8 +631,6 @@ private:
 
 	HANDLE m_handle = nullptr;
 #elif KYTY_PLATFORM == KYTY_PLATFORM_LINUX
-	[[nodiscard]] static uint64_t GetMappableTestSize() { return 0x4000; }
-
 	static int GetProtect(VirtualMemory::Mode mode) {
 		switch (mode) {
 			case VirtualMemory::Mode::Read: return PROT_READ;
@@ -760,7 +755,7 @@ public:
 		if (size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return std::any_of(m_owned.begin(), m_owned.end(), [vaddr, size](const auto& range) {
 			return VirtualRangesOverlap(vaddr, size, range.first, range.second);
 		});
@@ -864,11 +859,7 @@ public:
 		if (!ContainsMappedUnlocked(vaddr, size, MappingType::Backing)) {
 			return false;
 		}
-		bool placeholder_preserved = false;
-		if (!m_backing->Unmap(vaddr, size, &placeholder_preserved)) {
-			return false;
-		}
-		if (!placeholder_preserved) {
+		if (!m_backing->Unmap(vaddr, size)) {
 			return false;
 		}
 		EXIT_IF(!RemoveMappedUnlocked(vaddr, size, MappingType::Backing));
@@ -876,12 +867,13 @@ public:
 		return true;
 	}
 
+	// Reservations live in VirtualRanges; these free-range checks only read the host placeholders.
 	uint64_t ReserveAligned(uint64_t search_addr, uint64_t size, uint64_t alignment) {
 		if (size == 0) {
 			return 0;
 		}
 		alignment = std::max<uint64_t>(alignment != 0 ? alignment : PageSize(), PageSize());
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		if (search_addr == 0) {
 			auto addr = FindFreeAlignedUnlocked(HOST_SYSTEM_MANAGED_MIN,
 			                                    HOST_SYSTEM_MANAGED_MAX + 1u, size, alignment);
@@ -898,7 +890,7 @@ public:
 			return 0;
 		}
 		alignment = std::max<uint64_t>(alignment, PageSize());
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return FindFreeAlignedUnlocked(search_start, search_end, size, alignment);
 	}
 
@@ -906,19 +898,19 @@ public:
 		if (vaddr == 0 || size == 0 || !IsPageAligned(vaddr) || !IsPageAligned(size)) {
 			return false;
 		}
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return FindContaining(vaddr, size) != m_free.end();
 	}
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
 	bool TestContainsFree(uint64_t vaddr, uint64_t size) {
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return FindContaining(vaddr, size) != m_free.end();
 	}
 #endif
 
 	bool ReleaseFree(uint64_t vaddr, uint64_t size) {
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return FindContaining(vaddr, size) != m_free.end();
 	}
 
@@ -1041,10 +1033,6 @@ private:
 		return (value & (PageSize() - 1u)) == 0;
 	}
 
-	[[nodiscard]] bool IsGranularityAligned(uint64_t value) const {
-		return (value & (m_granularity - 1u)) == 0;
-	}
-
 	bool ProtectMappedUnlocked(uint64_t vaddr, uint64_t size, VirtualMemory::Mode mode) {
 		const auto end = vaddr + size;
 		auto       it  = m_mapped.upper_bound(vaddr);
@@ -1055,8 +1043,20 @@ private:
 			const auto& region      = it->second;
 			const auto  range_start = std::max(vaddr, region.vaddr);
 			const auto  range_end   = std::min(end, region.vaddr + region.size);
-			if (range_start < range_end &&
-			    !VirtualMemory::Protect(range_start, range_end - range_start, mode)) {
+			if (range_start >= range_end) {
+				continue;
+			}
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+			// Rollback can split a backing view without changing the owner interval.
+			// Windows protection calls must stay within each actual view.
+			if (region.type == MappingType::Backing) {
+				if (!m_backing->ProtectViews(range_start, range_end - range_start, mode)) {
+					return false;
+				}
+				continue;
+			}
+#endif
+			if (!VirtualMemory::Protect(range_start, range_end - range_start, mode)) {
 				return false;
 			}
 		}
@@ -1312,12 +1312,21 @@ private:
 	}
 
 	uint64_t FindFreeAlignedUnlocked(uint64_t search_start, uint64_t search_end, uint64_t size,
-	                                 uint64_t alignment) {
+	                                 uint64_t alignment) const {
 		if (search_start >= search_end || size > search_end - search_start) {
 			return 0;
 		}
-		for (const auto& [base, range_size]: m_free) {
-			const auto candidate = AlignUp(std::max(base, search_start), alignment);
+		// Include the predecessor because search_start can lie inside a free range.
+		auto it = m_free.upper_bound(search_start);
+		if (it != m_free.begin()) {
+			--it;
+		}
+		for (; it != m_free.end() && it->first < search_end; ++it) {
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+			g_test_free_range_visits.fetch_add(1, std::memory_order_relaxed);
+#endif
+			const auto& [base, range_size] = *it;
+			const auto candidate           = AlignUp(std::max(base, search_start), alignment);
 			if (candidate < search_end && size <= search_end - candidate &&
 			    GuestFreeRangeContains(base, range_size, candidate, size)) {
 				return candidate;
@@ -1331,9 +1340,8 @@ private:
 		if (next == m_free.begin()) {
 			return m_free.end();
 		}
-		auto it  = std::prev(next);
-		auto end = it->first + it->second;
-		return (vaddr >= it->first && vaddr + size <= end ? it : m_free.end());
+		auto it = std::prev(next);
+		return GuestFreeRangeContains(it->first, it->second, vaddr, size) ? it : m_free.end();
 	}
 
 	bool SplitUnlocked(std::map<uint64_t, uint64_t>::iterator it, uint64_t vaddr, uint64_t size) {

src/launcher/CMakeLists.txt

@@ -125,6 +125,7 @@ if(APPLE)
 	install(TARGETS kyty_emulator DESTINATION "KytyPS5.app/Contents/MacOS")
 	install(TARGETS kyty_updater DESTINATION "KytyPS5.app/Contents/MacOS")
 	install(FILES "${KYTY_TROPHY_UNLOCK_SOUND}" DESTINATION "KytyPS5.app/Contents/Resources/assets/sounds")
+	install(FILES ${KYTY_NOTO_SANS_FONTS} DESTINATION "KytyPS5.app/Contents/Resources/assets/fonts")
 	install(DIRECTORY "${CMAKE_SOURCE_DIR}/assets/cursors" DESTINATION "KytyPS5.app/Contents/Resources/assets")
 	# Deploy and ad-hoc sign the Qt frameworks and plugins in the bundle.
 	find_program(KYTY_MACDEPLOYQT NAMES macdeployqt PATHS "${Qt6_DIR}/../../../bin" NO_DEFAULT_PATH)

src/launcher/forms/configuration_edit_dialog.ui

@@ -115,7 +115,17 @@
             </item>
             <item row="0" column="1">
              <widget class="QComboBox" name="comboBox_screen_resolution">
-              <property name="toolTip"><string>Window resolution</string></property>
+              <property name="toolTip"><string>Window resolution as WIDTHxHEIGHT; pick a preset or type a custom size. Non-16:9 sizes such as 21:9 show the game with black bars unless Stretch to window is on.</string></property>
+              <property name="editable"><bool>true</bool></property>
+              <property name="insertPolicy"><enum>QComboBox::NoInsert</enum></property>
+              <item><property name="text"><string notr="true">1280x720</string></property></item>
+              <item><property name="text"><string notr="true">1920x1080</string></property></item>
+              <item><property name="text"><string notr="true">2560x1440</string></property></item>
+              <item><property name="text"><string notr="true">3840x2160</string></property></item>
+              <item><property name="text"><string notr="true">2560x1080</string></property></item>
+              <item><property name="text"><string notr="true">3440x1440</string></property></item>
+              <item><property name="text"><string notr="true">3840x1600</string></property></item>
+              <item><property name="text"><string notr="true">5120x2160</string></property></item>
              </widget>
             </item>
             <item row="1" column="0">
@@ -256,6 +266,12 @@
               <property name="text"><string>Auto-hide cursor</string></property>
              </widget>
             </item>
+            <item>
+             <widget class="QCheckBox" name="checkBox_stretch">
+              <property name="toolTip"><string>Stretch the game image to fill the window instead of keeping its aspect ratio with black bars</string></property>
+              <property name="text"><string>Stretch to window</string></property>
+             </widget>
+            </item>
             <item>
              <widget class="QCheckBox" name="checkBox_perf_overlay">
               <property name="toolTip"><string>Show FPS, frame times, CPU/GPU usage, RAM and memory pressure in the game window. Press F2 (Cmd+P on macOS) to toggle it, and add Shift for the detailed view.</string></property>

src/launcher/include/configuration.h

@@ -21,10 +21,7 @@ inline QStringList EnumToList() {
 	auto        me    = QMetaEnum::fromType<T>();
 	int         count = me.keyCount();
 	for (int i = 0; i < count; i++) {
-		auto key = QString(me.key(i));
-		ret << (key.startsWith('R') && key.size() > 2 && key.at(1).isDigit()
-		            ? key.remove('R').toLower()
-		            : key);
+		ret << QString(me.key(i));
 	}
 	return ret;
 }
@@ -32,16 +29,12 @@ inline QStringList EnumToList() {
 template <class T>
 T TextToEnum(const QString& text) {
 	auto me = QMetaEnum::fromType<T>();
-	return static_cast<T>(me.keyToValue(
-	    ((text.size() > 1 && text.at(0).isDigit()) ? 'R' + text.toUpper() : text).toUtf8().data()));
+	return static_cast<T>(me.keyToValue(text.toUtf8().constData()));
 }
 
 template <class T>
 QString EnumToText(T value) {
-	auto me  = QMetaEnum::fromType<T>();
-	auto key = QString(me.valueToKey(static_cast<int>(value)));
-	return (key.startsWith('R') && key.size() > 2 && key.at(1).isDigit() ? key.remove('R').toLower()
-	                                                                     : key);
+	return QString(QMetaEnum::fromType<T>().valueToKey(static_cast<int>(value)));
 }
 
 struct ControllerSettings {
@@ -80,13 +73,10 @@ public:
 	static constexpr int DEFAULT_CONSOLE_LANGUAGE = 1;
 	static constexpr int MAX_CONSOLE_LANGUAGE     = 29;
 
-	enum class Resolution {
-		R1280X720,
-		R1920X1080,
-		R2560X1440,
-		R3840X2160,
-	};
-	Q_ENUM(Resolution)
+	static constexpr int MIN_SCREEN_SIZE = 320;
+	static constexpr int MAX_SCREEN_SIZE = 16384;
+
+	static QString NormalizeResolution(const QString& text);
 
 	enum class ShaderOptimizationType { None, Size, Performance };
 	Q_ENUM(ShaderOptimizationType)
@@ -117,7 +107,7 @@ public:
 	// Controller preferences always come from the global configuration.
 	ControllerSettings controller;
 
-	Resolution             screen_resolution           = Resolution::R1280X720;
+	QString                screen_resolution           = QStringLiteral("1280x720");
 	QString                user_name                   = "Kyty";
 	int                    user_id                     = Config::DEFAULT_USER_ID;
 	QString                audio_input_device;
@@ -126,6 +116,7 @@ public:
 	bool                   fullscreen_enabled          = false;
 	bool                   hdr_enabled                 = false;
 	bool                   hide_cursor_enabled         = false;
+	bool                   stretch_enabled             = false;
 	bool                   readback_linear_images      = false;
 	bool                   readback_tiled_images       = false;
 	bool                   sync_raw_image_buffers      = false;
@@ -168,6 +159,7 @@ public:
 		fullscreen_enabled          = other.fullscreen_enabled;
 		hdr_enabled                 = other.hdr_enabled;
 		hide_cursor_enabled         = other.hide_cursor_enabled;
+		stretch_enabled             = other.stretch_enabled;
 		readback_linear_images      = other.readback_linear_images;
 		readback_tiled_images       = other.readback_tiled_images;
 		sync_raw_image_buffers      = other.sync_raw_image_buffers;

src/launcher/include/configurationEditDialog.h

@@ -46,11 +46,15 @@ private:
 	QLineEdit*                   m_addons_dir_edit        = nullptr;
 	QToolButton*                 m_remove_game_dir_button = nullptr;
 	bool                         m_global_settings        = false;
+	bool                         m_geometry_restored      = false;
 
 protected:
 	void Init(const Configuration& info);
 	void InitGameDirectories();
 	void AddGameDirectoryItem(const QString& dir);
+	void MakeTabsScrollable();
+	void GrowToFitContents();
+	void ShowTab(QWidget* page);
 
 	void moveEvent(QMoveEvent* event) override;
 	void resizeEvent(QResizeEvent* event) override;

src/launcher/src/configuration.cpp

@@ -1,5 +1,23 @@
 #include "configuration.h"
 
+#include <QRegularExpression>
+
+QString Configuration::NormalizeResolution(const QString& text) {
+	// Older settings stored the enum key, e.g. "R1920X1080".
+	static const QRegularExpression re(QStringLiteral("^\\s*R?(\\d+)\\s*[xX]\\s*(\\d+)\\s*$"));
+	const auto match = re.match(text);
+	if (!match.hasMatch()) {
+		return {};
+	}
+	const int width  = match.captured(1).toInt();
+	const int height = match.captured(2).toInt();
+	if (width < MIN_SCREEN_SIZE || height < MIN_SCREEN_SIZE || width > MAX_SCREEN_SIZE ||
+	    height > MAX_SCREEN_SIZE) {
+		return {};
+	}
+	return QStringLiteral("%1x%2").arg(width).arg(height);
+}
+
 QVariantMap Configuration::GameSettings() const {
 	QVariantMap values;
 #define KYTY_CFG_SET(n) values.insert(QStringLiteral(#n), QVariant::fromValue(n).toString())
@@ -12,6 +30,7 @@ QVariantMap Configuration::GameSettings() const {
 	KYTY_CFG_SET(fullscreen_enabled);
 	KYTY_CFG_SET(hdr_enabled);
 	KYTY_CFG_SET(hide_cursor_enabled);
+	KYTY_CFG_SET(stretch_enabled);
 	KYTY_CFG_SET(readback_linear_images);
 	KYTY_CFG_SET(readback_tiled_images);
 	KYTY_CFG_SET(sync_raw_image_buffers);
@@ -48,7 +67,8 @@ QVariantMap Configuration::GameSettings() const {
 template <class Settings>
 void Configuration::ReadGameSettingsValues(const Settings& s) {
 #define KYTY_CFG_GET(n) n = s.value(QStringLiteral(#n)).template value<decltype(n)>()
-	KYTY_CFG_GET(screen_resolution);
+	const auto resolution = NormalizeResolution(s.value("screen_resolution").toString());
+	screen_resolution     = resolution.isEmpty() ? QStringLiteral("1280x720") : resolution;
 	user_name          = s.value("user_name", user_name).toString();
 	bool user_id_ok    = false;
 	auto saved_user_id = s.value("user_id", user_id).toInt(&user_id_ok);
@@ -64,6 +84,7 @@ void Configuration::ReadGameSettingsValues(const Settings& s) {
 	KYTY_CFG_GET(fullscreen_enabled);
 	KYTY_CFG_GET(hdr_enabled);
 	KYTY_CFG_GET(hide_cursor_enabled);
+	KYTY_CFG_GET(stretch_enabled);
 	KYTY_CFG_GET(readback_linear_images);
 	KYTY_CFG_GET(readback_tiled_images);
 	KYTY_CFG_GET(sync_raw_image_buffers);
@@ -134,7 +155,9 @@ bool Configuration::SetGameSettings(const QJsonObject& settings, QString& error)
 		if (!it.value().isString()) {
 			return fail(tr("Game setting %1 must be a string.").arg(it.key()));
 		}
-		values.insert(it.key(), it.value().toString());
+		const auto value = it.value().toString();
+		values.insert(it.key(), it.key() == "screen_resolution" ? NormalizeResolution(value)
+		                                                       : value);
 	}
 
 	ReadGameSettingsValues(values);

src/launcher/src/configurationEditDialog.cpp

@@ -24,6 +24,8 @@
 #include <QMessageBox>
 #include <QPushButton>
 #include <QResizeEvent>
+#include <QScreen>
+#include <QScrollArea>
 #include <QSettings>
 #include <QSlider>
 #include <QSpinBox>
@@ -122,10 +124,10 @@ static QString GameDirectoryKey(const QString& dir) {
 }
 
 ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* parent)
-    : QDialog(parent, Qt::WindowCloseButtonHint), m_ui(new Ui::ConfigurationEditDialog),
-      m_info(info) {
+    : QDialog(parent, Qt::WindowMaximizeButtonHint | Qt::WindowCloseButtonHint),
+      m_ui(new Ui::ConfigurationEditDialog), m_info(info) {
 	m_ui->setupUi(this);
-	setMinimumWidth(width());
+	MakeTabsScrollable();
 	InitGameDirectories();
 	m_ui->controller_group->setVisible(false);
 	m_ui->settings_button->setVisible(false);
@@ -170,9 +172,11 @@ ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* p
 
 	setSizeGripEnabled(true);
 
-	restoreGeometry(g_last_geometry);
-
 	Init(info);
+	m_geometry_restored = restoreGeometry(g_last_geometry);
+	if (!m_geometry_restored) {
+		GrowToFitContents();
+	}
 }
 
 QByteArray ConfigurationEditDialog::g_last_geometry;
@@ -248,7 +252,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 		                    info.audio_input_device);
 	}
 	microphone->setCurrentIndex(microphone->findData(info.audio_input_device));
-	ListInit(m_ui->comboBox_screen_resolution, info.screen_resolution);
+	m_ui->comboBox_screen_resolution->setCurrentText(info.screen_resolution);
 	ListInit(m_ui->comboBox_present_mode, info.present_mode);
 	m_ui->comboBox_gpu->clear();
 	m_ui->comboBox_gpu->addItem(tr("Auto"));
@@ -285,6 +289,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_fullscreen->setChecked(info.fullscreen_enabled);
 	m_ui->checkBox_hdr->setChecked(info.hdr_enabled);
 	m_ui->checkBox_hide_cursor->setChecked(info.hide_cursor_enabled);
+	m_ui->checkBox_stretch->setChecked(info.stretch_enabled);
 	m_ui->checkBox_readback->setChecked(info.readback_linear_images);
 	m_ui->checkBox_readback_tiled->setChecked(info.readback_tiled_images);
 	m_ui->checkBox_sync_raw_image_buffers->setChecked(info.sync_raw_image_buffers);
@@ -413,8 +418,46 @@ void ConfigurationEditDialog::SetGlobalSettings(const QStringList& dirs, const Q
 	m_game_dirs_group->setVisible(true);
 	m_ui->controller_group->setVisible(true);
 	update_game_directory_buttons();
+	if (!m_geometry_restored) {
+		GrowToFitContents();
+	}
+}
+
+void ConfigurationEditDialog::MakeTabsScrollable() {
+	auto* tabs = m_ui->settings_tabs;
+	for (int i = 0; i < tabs->count(); i++) {
+		auto* page   = tabs->widget(i);
+		auto  title  = tabs->tabText(i);
+		auto* scroll = new QScrollArea;
+		scroll->setWidgetResizable(true);
+		scroll->setFrameShape(QFrame::NoFrame);
+		tabs->removeTab(i);
+		scroll->setWidget(page);
+		tabs->insertTab(i, scroll, title);
+	}
+	tabs->setCurrentIndex(0);
+}
+
+void ConfigurationEditDialog::GrowToFitContents() {
 	layout()->activate();
-	resize(size().expandedTo(minimumSizeHint()));
+	auto* tabs = m_ui->settings_tabs;
+	QSize content;
+	for (int i = 0; i < tabs->count(); i++) {
+		content = content.expandedTo(
+		    static_cast<QScrollArea*>(tabs->widget(i))->widget()->minimumSizeHint());
+	}
+	const QSize chrome = size() - tabs->currentWidget()->size();
+	resize(size().expandedTo(content + chrome).boundedTo(screen()->availableGeometry().size()));
+}
+
+void ConfigurationEditDialog::ShowTab(QWidget* page) {
+	auto* tabs = m_ui->settings_tabs;
+	for (int i = 0; i < tabs->count(); i++) {
+		if (tabs->widget(i)->isAncestorOf(page)) {
+			tabs->setCurrentIndex(i);
+			return;
+		}
+	}
 }
 
 QString ConfigurationEditDialog::GetAddonsDirectory() const {
@@ -474,13 +517,14 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 		info.controller.keyboard_player     = ui.comboBox_keyboard_player->currentData().toInt();
 	}
 	info.screen_resolution =
-	    TextToEnum<Configuration::Resolution>(ui.comboBox_screen_resolution->currentText());
+	    Configuration::NormalizeResolution(ui.comboBox_screen_resolution->currentText());
 	info.present_mode =
 	    TextToEnum<Configuration::PresentMode>(ui.comboBox_present_mode->currentText());
 	info.gpu_index                 = ui.comboBox_gpu->currentIndex() - 1;
 	info.fullscreen_enabled        = ui.checkBox_fullscreen->isChecked();
 	info.hdr_enabled               = ui.checkBox_hdr->isChecked();
 	info.hide_cursor_enabled       = ui.checkBox_hide_cursor->isChecked();
+	info.stretch_enabled           = ui.checkBox_stretch->isChecked();
 	info.readback_linear_images    = ui.checkBox_readback->isChecked();
 	info.readback_tiled_images     = ui.checkBox_readback_tiled->isChecked();
 	info.sync_raw_image_buffers    = ui.checkBox_sync_raw_image_buffers->isChecked();
@@ -518,7 +562,7 @@ void ConfigurationEditDialog::save() {
 	for (auto* field: {m_ui->lineEdit_shader_log_folder, m_ui->lineEdit_printf_file,
 	                   m_ui->lineEdit_cmd_dump_folder}) {
 		if (field->isEnabled() && field->text().isEmpty()) {
-			m_ui->settings_tabs->setCurrentWidget(m_ui->debug_tab);
+			ShowTab(m_ui->debug_tab);
 			field->setFocus();
 			break;
 		}
@@ -540,6 +584,16 @@ void ConfigurationEditDialog::save() {
 		                      tr("User ID cannot be 254 (everyone) or 255 (system)"));
 		return;
 	}
+	if (Configuration::NormalizeResolution(m_ui->comboBox_screen_resolution->currentText())
+	        .isEmpty()) {
+		ShowTab(m_ui->settings_tab);
+		m_ui->comboBox_screen_resolution->setFocus();
+		QMessageBox::critical(this, tr("Save failed"),
+		                      tr("Screen resolution must be WIDTHxHEIGHT, each between %1 and %2")
+		                          .arg(Configuration::MIN_SCREEN_SIZE)
+		                          .arg(Configuration::MAX_SCREEN_SIZE));
+		return;
+	}
 
 	UpdateInfo(m_info, *m_ui, m_global_settings);
 

src/launcher/src/mainDialog.cpp

@@ -134,7 +134,8 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 	m_ui->close_terminal->setVisible(false); // The emulator runs without a terminal.
 #endif
 
-	main_dialog->setWindowFlags(Qt::Dialog /*| Qt::MSWindowsFixedSizeDialogHint*/);
+	main_dialog->setWindowFlags(Qt::Dialog | Qt::WindowMinMaxButtonsHint |
+	                            Qt::WindowCloseButtonHint);
 
 	connect(main_dialog, &MainDialog::Start, this, &MainDialogPrivate::FindInterpreter,
 	        Qt::QueuedConnection);
@@ -240,7 +241,7 @@ static void AddFlag(QStringList& args, bool enabled, const char* name) {
 
 static QStringList CreateEmulatorArgs(const Configuration& info) {
 	QStringList args;
-	auto        r = EnumToText(info.screen_resolution).split('x');
+	auto        r = Configuration::NormalizeResolution(info.screen_resolution).split('x');
 
 	if (r.size() != 2) {
 		return {};
@@ -266,6 +267,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	AddFlag(args, info.fullscreen_enabled, "--fullscreen");
 	AddFlag(args, info.hdr_enabled, "--hdr");
 	AddFlag(args, info.hide_cursor_enabled, "--hide-cursor");
+	AddFlag(args, info.stretch_enabled, "--stretch");
 	AddFlag(args, info.readback_linear_images, "--readback-linear-images");
 	AddFlag(args, info.readback_tiled_images, "--readback-tiled-images");
 	AddFlag(args, info.sync_raw_image_buffers, "--sync-raw-image-buffers");

src/libs/agc.cpp

@@ -2480,6 +2480,9 @@ static constexpr uint32_t AGC_DRAW_INDIRECT_MULTI_END_TAG       = 0xc6000000u;
 static constexpr uint32_t AGC_WAIT_USER_DATA_BEGIN_32_TAG       = 0xc8010000u;
 static constexpr uint32_t AGC_WAIT_USER_DATA_BEGIN_64_TAG       = 0xc8020000u;
 static constexpr uint32_t AGC_WAIT_USER_DATA_END_TAG            = 0xc8000000u;
+static constexpr uint32_t AGC_MEM_SEMAPHORE_SIGNAL_BEGIN_TAG    = 0xc9010000u;
+static constexpr uint32_t AGC_MEM_SEMAPHORE_WAIT_BEGIN_TAG      = 0xc9020000u;
+static constexpr uint32_t AGC_MEM_SEMAPHORE_END_TAG             = 0xc9000000u;
 static constexpr uint32_t AGC_WAIT_USER_DATA_BEGIN_DW           = 4u;
 static constexpr uint32_t AGC_INTERNAL_DATA_PACKET_SIZE_DW      = 3u;
 
@@ -3077,6 +3080,42 @@ uint32_t* KYTY_SYSV_ABI AgcAcbWaitRegMem(CommandBuffer* buf, uint8_t size, uint8
 	                        poll_cycles);
 }
 
+uint32_t* KYTY_SYSV_ABI AgcAcbMemSemaphore(CommandBuffer* buf, const volatile void* address,
+                                           uint8_t mailbox, uint8_t signal_type, uint8_t op) {
+	EXIT_NOT_IMPLEMENTED(op != 6u && op != 7u);
+
+	auto* cmd =
+	    buf->AllocateDW(AGC_WAIT_USER_DATA_BEGIN_DW + AGC_INTERNAL_DATA_PACKET_SIZE_DW + 4u);
+	if (cmd == nullptr) {
+		return nullptr;
+	}
+
+	const auto address_value = reinterpret_cast<uint64_t>(address);
+	const bool signal        = (op == 6u);
+	cmd[0] = KYTY_PM4(AGC_WAIT_USER_DATA_BEGIN_DW, Pm4::IT_SET_UCONFIG_REG, 1u);
+	cmd[1] = AGC_INTERNAL_DATA_REGISTER;
+	cmd[2] = (signal ? AGC_MEM_SEMAPHORE_SIGNAL_BEGIN_TAG : AGC_MEM_SEMAPHORE_WAIT_BEGIN_TAG) |
+	         static_cast<uint32_t>((address_value >> 32u) & 0xffffu);
+	cmd[3] = static_cast<uint32_t>(address_value);
+
+	// A signal pads with a NOP header where a wait ends with its end marker.
+	auto* semaphore = cmd + AGC_WAIT_USER_DATA_BEGIN_DW;
+	if (signal) {
+		semaphore[0] = KYTY_PM4(AGC_INTERNAL_DATA_PACKET_SIZE_DW, Pm4::IT_NOP, 0u);
+		semaphore += AGC_INTERNAL_DATA_PACKET_SIZE_DW;
+	}
+	semaphore[0] = KYTY_PM4(4, Pm4::IT_MEM_SEMAPHORE, 0u);
+	semaphore[1] = static_cast<uint32_t>(address_value) & ~0x7u;
+	semaphore[2] = static_cast<uint32_t>(address_value >> 32u);
+	semaphore[3] = (static_cast<uint32_t>(op) << 29u) | ((signal_type & 0x1u) << 20u) |
+	               ((mailbox & 0x1u) << 16u);
+	if (!signal) {
+		write_agc_internal_data_packet(semaphore + 4, AGC_MEM_SEMAPHORE_END_TAG);
+	}
+
+	return cmd;
+}
+
 uint64_t KYTY_SYSV_ABI AgcAcbWaitOnAddressGetSize(uint8_t size) {
 	PRINT_NAME();
 
@@ -3840,11 +3879,12 @@ bool AgcIsInternalDataPacket(uint32_t cmd_id, const uint32_t* payload) {
 	const auto size_dw = KYTY_PM4_LEN(cmd_id);
 	if (size_dw == AGC_WAIT_USER_DATA_BEGIN_DW) {
 		const auto tag = payload[1] & 0xffff0000u;
-		return tag == AGC_WAIT_USER_DATA_BEGIN_32_TAG || tag == AGC_WAIT_USER_DATA_BEGIN_64_TAG;
+		return tag == AGC_WAIT_USER_DATA_BEGIN_32_TAG || tag == AGC_WAIT_USER_DATA_BEGIN_64_TAG ||
+		       tag == AGC_MEM_SEMAPHORE_SIGNAL_BEGIN_TAG || tag == AGC_MEM_SEMAPHORE_WAIT_BEGIN_TAG;
 	}
 
 	return size_dw == AGC_INTERNAL_DATA_PACKET_SIZE_DW &&
-	       (payload[1] == AGC_WAIT_USER_DATA_END_TAG ||
+	       (payload[1] == AGC_WAIT_USER_DATA_END_TAG || payload[1] == AGC_MEM_SEMAPHORE_END_TAG ||
 	        payload[1] == AGC_DRAW_INDIRECT_MULTI_BEGIN_TAG ||
 	        payload[1] == AGC_DRAW_INDIRECT_MULTI_END_TAG);
 }

src/libs/agc.h

@@ -216,6 +216,8 @@ uint32_t* KYTY_SYSV_ABI AgcAcbWaitRegMem(CommandBuffer* buf, uint8_t size, uint8
                                          uint8_t cache_policy, const volatile void* address,
                                          uint64_t reference, uint64_t mask, uint32_t poll_cycles);
 uint64_t KYTY_SYSV_ABI  AgcAcbWaitOnAddressGetSize(uint8_t size);
+uint32_t* KYTY_SYSV_ABI AgcAcbMemSemaphore(CommandBuffer* buf, const volatile void* address,
+                                           uint8_t mailbox, uint8_t signal_type, uint8_t op);
 uint32_t* KYTY_SYSV_ABI AgcAcbDmaData(CommandBuffer* buf, uint8_t dst, uint8_t dst_cache_policy,
                                       uint64_t dst_address_or_offset, uint8_t src,
                                       uint8_t  src_cache_policy,

src/libs/controller.cpp

@@ -519,7 +519,20 @@ void GameController::SetDevices(int active_id) {
 	m_trigger_state      = {};
 }
 
+// Input that changes no button or touch contact updates the newest unread sample, so a burst of
+// stick or sensor events between two reads cannot push unread button or touch changes out.
 void GameController::AddState() {
+	if (m_states_num != 0) {
+		const auto newest = (m_first_state + m_states_num - 1) % STATES_MAX;
+		auto&      last   = m_states[newest];
+		if (!m_obtained[newest] && last.buttons == m_state.buttons &&
+		    last.touch[0].down == m_state.touch[0].down &&
+		    last.touch[1].down == m_state.touch[1].down) {
+			last = m_state;
+			return;
+		}
+	}
+
 	if (m_states_num >= STATES_MAX) {
 		m_states_num  = STATES_MAX - 1;
 		m_first_state = (m_first_state + 1) % STATES_MAX;

src/libs/libAgcDriver.cpp

@@ -78,6 +78,7 @@ LIB_DEFINE(InitAgcDriver_1) {
 	LIB_FUNC("b-oySn+G2tE", Gen5::AgcAcbJumpGetSize);
 	LIB_FUNC("DwICrVxerkY", Gen5::AgcAcbRewind);
 	LIB_FUNC("htn36gPnBk4", Gen5::AgcAcbWaitRegMem);
+	LIB_FUNC("q4VuU-QsLOE", Gen5::AgcAcbMemSemaphore);
 	LIB_FUNC("idlaArvdXEs", Gen5::AgcAcbWaitOnAddressGetSize);
 	LIB_FUNC("-RnpfpxIhec", Gen5::AgcAcbDmaData);
 	LIB_FUNC("qzMN2XKGA4k", Gen5::AgcAcbCopyData);

src/libs/libAudio2.cpp

@@ -13,6 +13,7 @@
 #include <atomic>
 #include <cmath>
 #include <cstring>
+#include <span>
 #include <vector>
 
 namespace Libs::Audio {
@@ -168,17 +169,12 @@ struct AudioOut2MasteringStatesV2 {
 
 static std::atomic_uint64_t g_audioout2_next_context {1};
 static std::atomic_uint64_t g_audioout2_next_port {1};
-static AudioOut2UserHandle g_audioout2_next_user = 1;
+static AudioOut2UserHandle  g_audioout2_next_user = 1;
 
-struct AudioOut2ContextState {
-	bool                   used                = false;
-	AudioOut2ContextHandle handle              = 0;
-	uint32_t               queue_depth         = 4;
-	uint32_t               queued              = 0;
-	uint32_t               num_grains          = 512;
-	uint64_t               last_update         = 0;
-	int                    object_audio_handle = 0;
-};
+static constexpr uint32_t AUDIOOUT2_CHANNEL_PORTS_MAX = 32;
+static constexpr uint32_t AUDIOOUT2_OBJECT_PORTS_MAX  = 512;
+static constexpr uint32_t AUDIOOUT2_PORT_SLOT_BITS    = 10;
+static constexpr uint32_t AUDIOOUT2_CONTEXT_SLOT_BITS = 4;
 
 struct AudioOut2PortStateEntry {
 	bool                   used          = false;
@@ -196,6 +192,19 @@ struct AudioOut2PortStateEntry {
 	bool                   attributes_locked = false;
 	std::vector<uint8_t>   pcm_data;
 };
+
+struct AudioOut2ContextState {
+	bool                   used                = false;
+	AudioOut2ContextHandle handle              = 0;
+	uint32_t               queue_depth         = 4;
+	uint32_t               queued              = 0;
+	uint32_t               num_grains          = 512;
+	uint64_t               last_update         = 0;
+	int                    object_audio_handle = 0;
+	uint32_t               max_ports           = 0;
+	// Channel slots precede object slots; each partition has its own creation limit.
+	std::vector<AudioOut2PortStateEntry> ports;
+};
 struct AudioOut2SpeakerArrayState {
 	bool                        used          = false;
 	uint32_t                    num_speakers  = 0;
@@ -223,13 +232,16 @@ struct AudioOut2Submission {
 static Common::Mutex                              g_audioout2_state_mutex;
 static std::array<AudioOut2ContextState, 16>      g_audioout2_contexts;
 static std::array<AudioOut2Submission, 16>        g_audioout2_submissions;
-static std::array<AudioOut2PortStateEntry, 256>   g_audioout2_ports;
-static std::vector<AudioOut2UserHandle>          g_audioout2_users;
+static std::vector<AudioOut2UserHandle>           g_audioout2_users;
 static Common::Mutex                              g_audioout2_speaker_array_mutex;
 static std::array<AudioOut2SpeakerArrayState, 32> g_audioout2_speaker_arrays;
 static Common::Mutex                              g_audioout2_latency_mutex;
 static std::array<AudioOut2LatencyState, 16>      g_audioout2_latencies;
 
+static_assert(AUDIOOUT2_CHANNEL_PORTS_MAX + AUDIOOUT2_OBJECT_PORTS_MAX <=
+              (1u << AUDIOOUT2_PORT_SLOT_BITS));
+static_assert(g_audioout2_contexts.size() == (1u << AUDIOOUT2_CONTEXT_SLOT_BITS));
+
 static constexpr int AUDIO_OUT2_ERROR_NOT_READY                   = -2144960504; /* 0x80268008 */
 static constexpr int AUDIO_OUT2_ERROR_PORT_FULL                   = -2144960494; /* 0x80268012 */
 static constexpr int AUDIO_OUT2_ERROR_INVALID_PARAM               = -2144960511; /* 0x80268001 */
@@ -262,6 +274,11 @@ static AudioOut2Submission& audioout2_submission(const AudioOut2ContextState& co
 	return g_audioout2_submissions[&context - g_audioout2_contexts.data()];
 }
 
+// Ticks of a sixth of a microsecond: one 48 kHz sample lasts exactly 125 of them.
+static uint64_t audioout2_clock() {
+	return LibKernel::KernelGetProcessTime() * 6;
+}
+
 static uint32_t audioout2_grain_micros(uint32_t grains) {
 	const auto sample_count = (grains == 0 ? 512u : grains);
 	return std::max<uint32_t>((sample_count * 1000000u) / 48000u, 1000u);
@@ -331,49 +348,51 @@ static void audioout2_update_context_locked(AudioOut2ContextState* state) {
 		return;
 	}
 
-	const auto now = LibKernel::KernelGetProcessTime();
+	const auto now = audioout2_clock();
 	if (state->last_update == 0 || state->queued == 0) {
 		state->last_update = now;
 		return;
 	}
-
-	const auto grain_micros = static_cast<uint64_t>(audioout2_grain_micros(state->num_grains));
-	if (now <= state->last_update || grain_micros == 0) {
+	if (now <= state->last_update) {
 		return;
 	}
 
-	const auto elapsed = now - state->last_update;
-	const auto drained = std::min<uint64_t>(state->queued, elapsed / grain_micros);
+	const auto grain   = uint64_t {state->num_grains} * 125;
+	const auto drained = std::min<uint64_t>(state->queued, (now - state->last_update) / grain);
 	if (drained == 0) {
 		return;
 	}
 
 	state->queued -= static_cast<uint32_t>(drained);
-	state->last_update += drained * grain_micros;
+	state->last_update += drained * grain;
 	if (state->queued == 0) {
 		state->last_update = now;
 	}
 }
 
 static AudioOut2PortStateEntry* audioout2_find_port_locked(AudioOut2PortHandle port) {
-	for (auto& state: g_audioout2_ports) {
-		if (state.used && state.handle == port) {
-			return &state;
-		}
+	const auto context_slot =
+	    (port >> AUDIOOUT2_PORT_SLOT_BITS) & ((1u << AUDIOOUT2_CONTEXT_SLOT_BITS) - 1u);
+	const auto port_slot = port & ((1u << AUDIOOUT2_PORT_SLOT_BITS) - 1u);
+	auto&      context   = g_audioout2_contexts[context_slot];
+	if (context.used && port_slot < context.ports.size()) {
+		auto& state = context.ports[port_slot];
+		if (state.used && state.handle == port) return &state;
 	}
 
 	return nullptr;
 }
 
 static size_t audioout2_pcm_size(const AudioOut2PortStateEntry& state) {
-	const auto bytes_per_sample = (state.data_format & 0x7fu) == 1 ? sizeof(int16_t) : sizeof(float);
+	const auto bytes_per_sample =
+	    (state.data_format & 0x7fu) == 1 ? sizeof(int16_t) : sizeof(float);
 	return static_cast<size_t>(state.samples_num) *
 	       audioout2_data_format_channels(state.data_format) * bytes_per_sample;
 }
 
 static bool audioout2_context_has_queueable_device(const AudioOut2ContextState& context) {
-	for (const auto& state: g_audioout2_ports) {
-		if (state.used && state.context == context.handle && !state.pcm_data.empty()) {
+	for (const auto& state: context.ports) {
+		if (state.used && !state.pcm_data.empty()) {
 			const int handle = audioout2_port_type_is_object(state.port_type)
 			                       ? context.object_audio_handle
 			                       : state.audio_handle;
@@ -427,18 +446,21 @@ static void audioout2_queue_context_audio(AudioOut2ContextHandle ctx, AudioOut2S
 	out.object_mix.clear();
 	std::vector<AudioInternal::OutputParam> params;
 	params.reserve(AudioInternal::OUT_PORTS_MAX);
-	for (const auto& state: g_audioout2_ports) {
-		if (!state.used || state.context != ctx || state.pcm_data.empty()) {
+	// A port's PCM covers one grain: the push consumes it, so a port the game stops feeding
+	// falls silent.
+	for (auto& state: context->ports) {
+		if (!state.used || state.pcm_data.empty()) {
 			continue;
 		}
 		if (audioout2_port_type_is_object(state.port_type)) {
 			audioout2_mix_object(out.object_mix, context->num_grains, state);
 		} else if (state.audio_handle > 0 && params.size() < AudioInternal::OUT_PORTS_MAX) {
 			const auto i = params.size();
-			out.pcm[i]   = state.pcm_data;
+			out.pcm[i].swap(state.pcm_data);
 			out.gains[i] = state.gains;
 			params.push_back({state.audio_handle, out.pcm[i].data(), out.gains[i].data()});
 		}
+		state.pcm_data.clear();
 	}
 	if (!out.object_mix.empty() && params.size() < AudioInternal::OUT_PORTS_MAX) {
 		params.push_back({context->object_audio_handle, out.object_mix.data()});
@@ -467,22 +489,28 @@ int KYTY_SYSV_ABI AudioOut2ContextResetParam(AudioOut2ContextParam* params) {
 	EXIT_NOT_IMPLEMENTED(params == nullptr);
 
 	std::memset(params, 0, sizeof(AudioOut2ContextParam));
-	params->max_ports              = 256;
-	params->max_object_ports       = 256;
-	params->guarantee_object_ports = 0;
-	params->queue_depth            = 4;
-	params->num_grains             = 512;
-	params->flags                  = 1;
+	params->max_ports   = 8;
+	params->queue_depth = 1;
+	params->num_grains  = 256;
 
 	return OK;
 }
 
+static bool audioout2_valid_context_capacity(const AudioOut2ContextParam& params) {
+	return params.max_ports <= AUDIOOUT2_CHANNEL_PORTS_MAX &&
+	       params.max_object_ports <= AUDIOOUT2_OBJECT_PORTS_MAX &&
+	       params.guarantee_object_ports <= params.max_object_ports;
+}
+
 int KYTY_SYSV_ABI AudioOut2ContextQueryMemory(const AudioOut2ContextParam* params,
                                               size_t*                      memory_size) {
 	PRINT_NAME();
 
 	EXIT_NOT_IMPLEMENTED(params == nullptr);
 	EXIT_NOT_IMPLEMENTED(memory_size == nullptr);
+	if (!audioout2_valid_context_capacity(*params)) {
+		return AUDIO_OUT2_ERROR_INVALID_PARAM;
+	}
 
 	const auto queue_depth = (params->queue_depth == 0 ? 4u : params->queue_depth);
 	*memory_size           = 0x10000u + static_cast<size_t>(queue_depth) * 0x590u;
@@ -498,6 +526,9 @@ int KYTY_SYSV_ABI AudioOut2ContextCreate(const AudioOut2ContextParam* params, vo
 
 	EXIT_NOT_IMPLEMENTED(params == nullptr);
 	EXIT_NOT_IMPLEMENTED(ctx == nullptr);
+	if (!audioout2_valid_context_capacity(*params)) {
+		return AUDIO_OUT2_ERROR_INVALID_PARAM;
+	}
 
 	const auto num_grains = params->num_grains == 0 ? 512u : params->num_grains;
 	const int  object_audio_handle =
@@ -524,8 +555,10 @@ int KYTY_SYSV_ABI AudioOut2ContextCreate(const AudioOut2ContextParam* params, vo
 	state->handle              = *ctx;
 	state->queue_depth         = params->queue_depth == 0 ? 4u : params->queue_depth;
 	state->num_grains          = num_grains;
-	state->last_update         = LibKernel::KernelGetProcessTime();
+	state->last_update         = audioout2_clock();
 	state->object_audio_handle = object_audio_handle;
+	state->max_ports           = params->max_ports;
+	state->ports.resize(params->max_ports + params->max_object_ports);
 	g_audioout2_state_mutex.Unlock();
 
 	LOGF("\t buffer      = 0x%016" PRIx64 "\n"
@@ -541,8 +574,8 @@ int KYTY_SYSV_ABI AudioOut2ContextDestroy(AudioOut2ContextHandle ctx) {
 	PRINT_NAME();
 	LOGF("\t ctx = 0x%016" PRIx64 "\n", ctx);
 
-	std::array<int, 257> audio_handles {};
-	size_t audio_handles_num = 0;
+	std::array<int, AUDIOOUT2_CHANNEL_PORTS_MAX + 1> audio_handles {};
+	size_t                                           audio_handles_num = 0;
 
 	g_audioout2_state_mutex.Lock();
 	auto* state = audioout2_find_context_locked(ctx);
@@ -550,15 +583,12 @@ int KYTY_SYSV_ABI AudioOut2ContextDestroy(AudioOut2ContextHandle ctx) {
 		if (state->object_audio_handle > 0) {
 			audio_handles[audio_handles_num++] = state->object_audio_handle;
 		}
-		*state = AudioOut2ContextState {};
-	}
-	for (auto& port_state: g_audioout2_ports) {
-		if (port_state.used && port_state.context == ctx) {
-			if (port_state.audio_handle > 0 && audio_handles_num < audio_handles.size()) {
+		for (const auto& port_state: state->ports) {
+			if (port_state.used && port_state.audio_handle > 0) {
 				audio_handles[audio_handles_num++] = port_state.audio_handle;
 			}
-			port_state = AudioOut2PortStateEntry {};
 		}
+		*state = AudioOut2ContextState {};
 	}
 	g_audioout2_state_mutex.Unlock();
 
@@ -586,9 +616,7 @@ int KYTY_SYSV_ABI AudioOut2ContextAdvance(AudioOut2ContextHandle ctx) {
 	g_audioout2_state_mutex.Lock();
 	if (auto* state = audioout2_find_context_locked(ctx); state != nullptr) {
 		audioout2_update_context_locked(state);
-	}
-	for (auto& port: g_audioout2_ports) {
-		if (port.used && port.context == ctx) {
+		for (auto& port: state->ports) {
 			port.attributes_locked = false;
 		}
 	}
@@ -612,7 +640,7 @@ int KYTY_SYSV_ABI AudioOut2ContextPush(AudioOut2ContextHandle ctx, uint32_t bloc
 			sleep_micros = audioout2_grain_micros(state->num_grains);
 			if (state->queued < state->queue_depth || use_device_clock) {
 				if (state->queued == 0) {
-					state->last_update = LibKernel::KernelGetProcessTime();
+					state->last_update = audioout2_clock();
 				}
 				if (state->queued < state->queue_depth) {
 					state->queued++;
@@ -663,21 +691,19 @@ 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 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);
+	const bool is_object    = audioout2_port_type_is_object(params->port_type);
 
 	g_audioout2_state_mutex.Lock();
-	const auto* context_state = audioout2_find_context_locked(ctx);
+	auto* context_state = audioout2_find_context_locked(ctx);
 	if (context_state == nullptr) {
 		g_audioout2_state_mutex.Unlock();
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}
-	if (audioout2_port_type_is_object(params->port_type) &&
-	    (params->port_type != 0x100 || context_state->object_audio_handle == 0 ||
-	     params->sampling_freq != 48000 ||
-	     (audio_format != AudioInternal::Format::FloatMono &&
-	      audio_format != AudioInternal::Format::Signed16bitMono))) {
+	if (is_object && (params->port_type != 0x100 || params->sampling_freq != 48000 ||
+	                  (audio_format != AudioInternal::Format::FloatMono &&
+	                   audio_format != AudioInternal::Format::Signed16bitMono))) {
 		g_audioout2_state_mutex.Unlock();
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}
@@ -688,42 +714,45 @@ int KYTY_SYSV_ABI AudioOut2PortCreate(AudioOut2ContextHandle ctx, const AudioOut
 		g_audioout2_state_mutex.Unlock();
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}
-	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->user          = params->user_handle;
-		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->gains.fill(1.0f);
-	}
-	g_audioout2_state_mutex.Unlock();
-
-	if (port_state == nullptr) {
+	auto slots = std::span(context_state->ports);
+	slots =
+	    is_object ? slots.subspan(context_state->max_ports) : slots.first(context_state->max_ports);
+	const auto available = std::find_if(slots.begin(), slots.end(),
+	                                    [](const auto& candidate) { return !candidate.used; });
+	if (available == slots.end()) {
+		g_audioout2_state_mutex.Unlock();
 		return AUDIO_OUT2_ERROR_PORT_FULL;
 	}
+	const auto serial       = g_audioout2_next_port.fetch_add(1, std::memory_order_relaxed);
+	const auto context_slot = static_cast<uint64_t>(context_state - g_audioout2_contexts.data());
+	const auto port_slot    = static_cast<uint64_t>(&*available - context_state->ports.data());
+	const auto next_port    = (serial << (AUDIOOUT2_PORT_SLOT_BITS + AUDIOOUT2_CONTEXT_SLOT_BITS)) |
+	                          (context_slot << AUDIOOUT2_PORT_SLOT_BITS) | port_slot;
+	auto*      port_state   = &*available;
+	*port_state             = AudioOut2PortStateEntry {};
+	port_state->used        = true;
+	port_state->handle      = next_port;
+	port_state->context     = ctx;
+	port_state->user        = params->user_handle;
+	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->gains.fill(1.0f);
+	g_audioout2_state_mutex.Unlock();
 
 	int audio_handle = 0;
 
-	if (audio_format != AudioInternal::Format::Unknown &&
-	    !audioout2_port_type_is_object(params->port_type)) {
+	if (audio_format != AudioInternal::Format::Unknown && !is_object) {
 		audio_handle = AudioInternal::AudioOutOpen(audio_type, samples_num, params->sampling_freq,
 		                                           audio_format);
 	}
 
 	g_audioout2_state_mutex.Lock();
-	const bool reserved = port_state->used && port_state->handle == next_port;
+	// Context destruction may have freed or replaced the slots while the backend opened.
+	port_state          = audioout2_find_port_locked(next_port);
+	const bool reserved = port_state != nullptr;
 	if (reserved) {
 		port_state->audio_handle = audio_handle;
 	}
@@ -735,7 +764,7 @@ int KYTY_SYSV_ABI AudioOut2PortCreate(AudioOut2ContextHandle ctx, const AudioOut
 
 	*port = next_port;
 
-	if (next_port <= 16 || (next_port % 600) == 0) {
+	if (serial <= 16 || (serial % 600) == 0) {
 		PRINT_NAME();
 		LOGF("\t ctx           = 0x%016" PRIx64 "\n"
 		     "\t port          = 0x%016" PRIx64 "\n"
@@ -901,21 +930,23 @@ int KYTY_SYSV_ABI AudioOut2UserDestroy(AudioOut2UserHandle handle) {
 	PRINT_NAME();
 	LOGF("\t handle = 0x%016" PRIx64 "\n", static_cast<uint64_t>(handle));
 	Common::LockGuard lock(g_audioout2_state_mutex);
-	const auto it = std::find(g_audioout2_users.begin(), g_audioout2_users.end(), handle);
+	const auto        it = std::find(g_audioout2_users.begin(), g_audioout2_users.end(), handle);
 	if (it == g_audioout2_users.end()) {
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}
-	if (std::any_of(g_audioout2_ports.begin(), g_audioout2_ports.end(),
-	                [handle](const auto& port) { return port.used && port.user == handle; })) {
-		return AUDIO_OUT2_ERROR_BUSY;
+	for (const auto& context: g_audioout2_contexts) {
+		if (std::any_of(context.ports.begin(), context.ports.end(),
+		                [handle](const auto& port) { return port.used && port.user == handle; })) {
+			return AUDIO_OUT2_ERROR_BUSY;
+		}
 	}
 	g_audioout2_users.erase(it);
 	return OK;
 }
 
 int KYTY_SYSV_ABI AudioOut2UserGetSupportedAttributes(AudioOut2UserHandle handle,
-                                                       uint32_t* context_attributes,
-                                                       uint32_t* port_attributes) {
+                                                      uint32_t*           context_attributes,
+                                                      uint32_t*           port_attributes) {
 	if (context_attributes == nullptr || port_attributes == nullptr) {
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}

src/libs/libDbgAsan.cpp

@@ -38,8 +38,8 @@ struct Global {
 };
 
 struct Shadow {
-	uintptr_t start = 0;
-	int       count = 0;
+	uintptr_t start         = 0;
+	uint64_t  covered_bytes = 0;
 };
 
 struct AsanContext {
@@ -50,6 +50,14 @@ struct AsanContext {
 static uint32_t     g_fake_stack = 0;
 static AsanContext* g_ctx        = nullptr;
 
+// Allocation and release callbacks can partition the same guest range differently.
+static uint64_t ShadowCoverage(uintptr_t page, uintptr_t min_addr, uintptr_t max_addr) {
+	const auto guest_start = (page - 0x10000000000ull) << 3u;
+	const auto guest_end   = guest_start + 0x80000ull - 1;
+	return std::min<uintptr_t>(max_addr, guest_end) - std::max<uintptr_t>(min_addr, guest_start) +
+	       1;
+}
+
 static void AllocShadow(uintptr_t min_addr, uintptr_t max_addr) {
 	EXIT_IF(g_ctx == nullptr);
 
@@ -67,10 +75,11 @@ static void AllocShadow(uintptr_t min_addr, uintptr_t max_addr) {
 	     static_cast<uint64_t>(min_shadow), static_cast<uint64_t>(size));
 
 	for (uintptr_t page = min_shadow; page <= max_shadow; page += page_size) {
-		const auto it = std::find_if(g_ctx->shadows.begin(), g_ctx->shadows.end(),
-		                             [page](const Shadow& s) { return s.start == page; });
+		const auto bytes = ShadowCoverage(page, min_addr, max_addr);
+		const auto it    = std::find_if(g_ctx->shadows.begin(), g_ctx->shadows.end(),
+		                                [page](const Shadow& s) { return s.start == page; });
 		if (it != g_ctx->shadows.end()) {
-			it->count++;
+			it->covered_bytes += bytes;
 		} else {
 			if (!Common::VirtualMemory::AllocFixed(page, page_size,
 			                                       Common::VirtualMemory::Mode::ReadWrite)) {
@@ -78,8 +87,8 @@ static void AllocShadow(uintptr_t min_addr, uintptr_t max_addr) {
 			}
 
 			Shadow s;
-			s.start = page;
-			s.count = 1;
+			s.start         = page;
+			s.covered_bytes = bytes;
 			g_ctx->shadows.push_back(s);
 		}
 	}
@@ -105,14 +114,15 @@ static void FreeShadow(uintptr_t min_addr, uintptr_t max_addr) {
 		const auto it = std::find_if(g_ctx->shadows.begin(), g_ctx->shadows.end(),
 		                             [page](const Shadow& s) { return s.start == page; });
 		if (it != g_ctx->shadows.end()) {
-			it->count--;
-
-			if (it->count <= 0) {
+			const auto bytes = ShadowCoverage(page, min_addr, max_addr);
+			if (bytes >= it->covered_bytes) {
 				if (!Common::VirtualMemory::Free(page)) {
 					EXIT("can't free shadow memory\n");
 				}
 
 				g_ctx->shadows.erase(it);
+			} else {
+				it->covered_bytes -= bytes;
 			}
 		}
 	}
@@ -126,6 +136,14 @@ static void KernelFree(uintptr_t addr, size_t size) {
 	FreeShadow(addr, addr + size - 1);
 }
 
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+void TestRegisterMemoryCallbacks() {
+	EXIT_IF(g_ctx != nullptr);
+	g_ctx = new AsanContext;
+	LibKernel::Memory::RegisterCallbacks(KernelAlloc, KernelFree);
+}
+#endif
+
 static void KYTY_SYSV_ABI asan_init() {
 	PRINT_NAME();
 

src/libs/libFont.cpp

@@ -2,10 +2,12 @@
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/logging/log.h"
+#include "common/stringUtils.h"
 #include "libs/errno.h"
 #include "libs/libs.h"
 #include "loader/symbolDatabase.h"
 
+#include <SDL3/SDL.h>
 #include <algorithm>
 #include <array>
 #include <cinttypes>
@@ -675,19 +677,6 @@ static bool font_is_italic(const FontState* font) {
 	       font->render_effect_slant != 0.0f;
 }
 
-static const char* fallback_font_file_name(bool bold, bool italic) {
-	if (bold && italic) {
-		return "Roboto-BoldItalic.ttf";
-	}
-	if (bold) {
-		return "Roboto-Bold.ttf";
-	}
-	if (italic) {
-		return "Roboto-Italic.ttf";
-	}
-	return "Roboto-Regular.ttf";
-}
-
 static std::vector<uint8_t> read_file_bytes(const std::filesystem::path& path) {
 	std::ifstream file(path, std::ios::binary | std::ios::ate);
 	if (!file) {
@@ -709,49 +698,34 @@ static std::vector<uint8_t> read_file_bytes(const std::filesystem::path& path) {
 	return data;
 }
 
-static std::vector<std::filesystem::path> fallback_font_candidates(const char* file_name) {
-	const auto rel =
-	    std::filesystem::path("3rdparty") / "tracy" / "profiler" / "src" / "font" / file_name;
-	std::vector<std::filesystem::path> paths;
-
-	auto source_root = std::filesystem::path(__FILE__).parent_path().parent_path().parent_path();
-	paths.push_back(source_root / rel);
-
-	std::error_code ec;
-	auto            cwd = std::filesystem::current_path(ec);
-	if (!ec) {
-		paths.push_back(cwd / rel);
-		paths.push_back(cwd / ".." / rel);
-		paths.push_back(cwd / ".." / ".." / rel);
+static std::vector<uint8_t> load_fallback_font_data(const char* file_name) {
+	const char* base_path = SDL_GetBasePath();
+	if (base_path == nullptr) {
+		LOGF("Font: SDL_GetBasePath failed: %s\n", SDL_GetError());
+		return {};
 	}
-
-	return paths;
-}
-
-static std::vector<uint8_t> load_fallback_font_data(bool bold, bool italic) {
-	for (const auto& path: fallback_font_candidates(fallback_font_file_name(bold, italic))) {
-		auto data = read_file_bytes(path);
-		if (!data.empty()) {
-			return data;
-		}
+	// SDL returns UTF-8; path(const char*) would read it in the Windows ANSI code page.
+	const auto path = Common::PathFromUtf8(base_path) / "assets" / "fonts" / file_name;
+	auto       data = read_file_bytes(path);
+	if (data.empty()) {
+		LOGF("Font: fallback font not found: %s\n", Common::PathToString(path).c_str());
 	}
-
-	return {};
+	return data;
 }
 
 static const std::vector<uint8_t>& fallback_font_data(bool bold, bool italic) {
-	static const auto regular     = load_fallback_font_data(false, false);
-	static const auto bold_data   = load_fallback_font_data(true, false);
-	static const auto italic_data = load_fallback_font_data(false, true);
-	static const auto bold_italic = load_fallback_font_data(true, true);
+	static const auto regular     = load_fallback_font_data("NotoSans-Regular.ttf");
+	static const auto bold_data   = load_fallback_font_data("NotoSans-Bold.ttf");
+	static const auto italic_data = load_fallback_font_data("NotoSans-Italic.ttf");
+	static const auto bold_italic = load_fallback_font_data("NotoSans-BoldItalic.ttf");
 
-	if (bold && italic && !bold_italic.empty()) {
+	if (bold && italic) {
 		return bold_italic;
 	}
-	if (bold && !bold_data.empty()) {
+	if (bold) {
 		return bold_data;
 	}
-	if (italic && !italic_data.empty()) {
+	if (italic) {
 		return italic_data;
 	}
 	return regular;

src/libs/libSaveData.cpp

@@ -615,7 +615,12 @@ int KYTY_SYSV_ABI SaveDataDirNameSearch(const SaveDataDirNameSearchCond* cond,
 	Common::LockGuard lock(g_mount_mutex);
 	const std::string title_id = cond->title_id != nullptr ? cond->title_id->data : get_title_id();
 	const auto        root     = save_user_directory(title_id, cond->user_id);
-	std::vector<std::string> dir_list;
+	struct Found {
+		uint64_t    key = 0;
+		std::string name;
+		auto        operator<=>(const Found&) const = default;
+	};
+	std::vector<Found> dir_list;
 
 	if (Common::File::IsDirectoryExisting(root)) {
 		for (const auto& entry: Common::File::GetDirEntries(root)) {
@@ -625,22 +630,34 @@ int KYTY_SYSV_ABI SaveDataDirNameSearch(const SaveDataDirNameSearchCond* cond,
 				if (cond->dir_name == nullptr || cond->dir_name->data[0] == '\0' ||
 				    dir_name_match(Common::ToLower(entry.name).c_str(),
 				                   Common::ToLower(std::string(cond->dir_name->data)).c_str())) {
-					dir_list.push_back(entry.name);
+					dir_list.push_back({.name = entry.name});
 				}
 			}
 		}
 	}
 
-	std::sort(dir_list.begin(), dir_list.end(), [](const std::string& a, const std::string& b) {
-		return std::strcmp(a.c_str(), b.c_str()) < 0;
-	});
-
-	if (cond->order == SaveDataSortOrder::Descent) {
-		std::vector<std::string> reversed;
-		for (size_t i = dir_list.size(); i > 0; i--) {
-			reversed.push_back(dir_list[i - 1]);
+	// A save whose metadata cannot be read sorts as 0.
+	for (auto& found: dir_list) {
+		const auto directory = root / found.name;
+		switch (cond->key) {
+			case SaveDataSortKey::UserParam: {
+				SaveDataParam param {};
+				read_save_param(save_param_path(directory), &param);
+				found.key = param.user_param;
+				break;
+			}
+			case SaveDataSortKey::Blocks:
+			case SaveDataSortKey::FreeBlocks: read_save_blocks(directory, &found.key); break;
+			case SaveDataSortKey::Mtime:
+				found.key = static_cast<uint64_t>(newest_save_time(directory));
+				break;
+			default: break;
 		}
-		dir_list = reversed;
+	}
+	// Directory names break ties; descending order reverses both.
+	std::sort(dir_list.begin(), dir_list.end());
+	if (cond->order == SaveDataSortOrder::Descent) {
+		std::reverse(dir_list.begin(), dir_list.end());
 	}
 
 	auto max_count =
@@ -651,9 +668,9 @@ int KYTY_SYSV_ABI SaveDataDirNameSearch(const SaveDataDirNameSearchCond* cond,
 
 	for (size_t i = 0; i < max_count; i++) {
 		std::snprintf(result->dir_names[i].data, sizeof(result->dir_names[i].data), "%s",
-		              dir_list[i].c_str());
+		              dir_list[i].name.c_str());
 		if (result->params != nullptr) {
-			const int status = load_save_param(root / dir_list[i], &result->params[i]);
+			const int status = load_save_param(root / dir_list[i].name, &result->params[i]);
 			if (status != OK) {
 				return status;
 			}
@@ -661,7 +678,7 @@ int KYTY_SYSV_ABI SaveDataDirNameSearch(const SaveDataDirNameSearchCond* cond,
 		if (result->infos != nullptr) {
 			auto& info = result->infos[i];
 			info = {};
-			const int status = read_save_blocks(root / dir_list[i], &info.blocks);
+			const int status = read_save_blocks(root / dir_list[i].name, &info.blocks);
 			if (status != OK) {
 				return status;
 			}

src/loader/runtimeLinker.cpp

@@ -1636,6 +1636,8 @@ void RuntimeLinker::DeleteTls(Program* program, int thread_id) {
 	}
 }
 
+constexpr uint64_t GUEST_PAGE_SIZE = 0x4000;
+
 static uint64_t CalcBaseSize(const Elf64_Ehdr* ehdr, const Elf64_Phdr* phdr) {
 	uint64_t base_size = 0;
 	for (Elf64_Half i = 0; i < ehdr->e_phnum; i++) {
@@ -1649,6 +1651,20 @@ static uint64_t CalcBaseSize(const Elf64_Ehdr* ehdr, const Elf64_Phdr* phdr) {
 	return base_size;
 }
 
+// A module's E, R and RW LOAD segments use flexible memory in 16 KiB pages; GNU_RELRO adds none
+// since it lies inside an RW LOAD.
+uint64_t CalcFlexibleMemorySize(const Elf64_Phdr* phdr, Elf64_Half phnum) {
+	uint64_t size = 0;
+	for (Elf64_Half i = 0; i < phnum; i++) {
+		const auto flags = phdr[i].p_flags;
+		if (phdr[i].p_type == PT_LOAD &&
+		    (flags == PF_X || flags == PF_R || flags == (PF_R | PF_W))) {
+			size += AlignUp(phdr[i].p_memsz, GUEST_PAGE_SIZE);
+		}
+	}
+	return size;
+}
+
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
 void RuntimeLinker::LoadProgramToMemory(Program* program) {
 	KYTY_PROFILER_FUNCTION();
@@ -1668,8 +1684,7 @@ void RuntimeLinker::LoadProgramToMemory(Program* program) {
 
 	EXIT_IF(phdr == nullptr || ehdr == nullptr);
 
-	program->base_size                 = CalcBaseSize(ehdr, phdr);
-	constexpr uint64_t GUEST_PAGE_SIZE = 0x4000;
+	program->base_size = CalcBaseSize(ehdr, phdr);
 	EXIT_IF(program->base_size > UINT64_MAX - (GUEST_PAGE_SIZE - 1));
 	program->base_size_aligned = AlignUp(program->base_size, GUEST_PAGE_SIZE);
 
@@ -1694,6 +1709,8 @@ void RuntimeLinker::LoadProgramToMemory(Program* program) {
 	    g_desired_base_addr, program->mapped_size, Common::VirtualMemory::Mode::ExecuteReadWrite,
 	    Common::PathToString(program->file_name.filename()).c_str());
 	EXIT_IF(program->base_vaddr == 0);
+	program->flexible_size = CalcFlexibleMemorySize(phdr, ehdr->e_phnum);
+	Libs::LibKernel::Memory::ChargeFlexibleMemory(program->flexible_size);
 
 	if (patch_guest_instructions) {
 		const auto trampoline_addr           = program->base_vaddr + program->base_size_aligned;
@@ -1879,6 +1896,7 @@ void RuntimeLinker::DeleteProgram(Program* p) {
 		}
 		EXIT_IF(
 		    !Libs::LibKernel::Memory::FreeGuestMemory(program->base_vaddr, program->mapped_size));
+		Libs::LibKernel::Memory::ReleaseFlexibleMemory(program->flexible_size);
 	}
 }
 

src/loader/runtimeLinker.h

@@ -21,6 +21,7 @@ struct EntryParams {
 };
 
 class Elf64;
+struct Elf64_Phdr;
 struct Elf64_Sym;
 struct Elf64_Rela;
 class RuntimeLinker;
@@ -128,6 +129,7 @@ struct Program {
 	uint64_t                     base_size         = 0;
 	uint64_t                     base_size_aligned = 0;
 	uint64_t                     mapped_size       = 0;
+	uint64_t                     flexible_size     = 0;
 	uint64_t instruction_trampoline_size = 0;
 	std::unique_ptr<SymbolDatabase> export_symbols;
 	std::unique_ptr<SymbolDatabase> import_symbols;
@@ -203,6 +205,8 @@ private:
 	application_heap_posix_memalign_func_t m_application_heap_posix_memalign = nullptr;
 };
 
+uint64_t CalcFlexibleMemorySize(const Elf64_Phdr* phdr, uint16_t phnum);
+
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
 bool TestMainEntryUsesGuestStack();
 bool TestModuleRelocationUsesWritableHostMapping();

src/loader/systemContent.cpp

@@ -35,7 +35,7 @@ public:
 	bool     GetParamInt(const char* name, int32_t* value);
 	bool     GetParamString(const char* name, std::string* value);
 	bool     GetParamString(const char* name, char* value, size_t value_size);
-	uint64_t GetFlexibleMemorySize();
+	MemoryParams GetMemoryParams();
 
 	void DbgPrint();
 
@@ -61,7 +61,7 @@ private:
 	Common::Mutex      m_mutex;
 	bool               m_opened = false;
 	std::vector<Param> m_params;
-	uint64_t           m_flexible_memory_size = 0;
+	MemoryParams       m_memory_params;
 };
 
 class PlayGo {
@@ -172,11 +172,21 @@ void Psf::OpenJson(const std::filesystem::path& file_name) {
 	add_int_param("userDefinedParam3", "USER_DEFINED_PARAM_3");
 	add_int_param("userDefinedParam4", "USER_DEFINED_PARAM_4");
 
+	auto read_size = [](const nlohmann::ordered_json& object, const char* key, uint64_t* size,
+	                    uint64_t unit = 1) {
+		if (auto it = object.find(key); it != object.end()) {
+			*size = it->get<uint64_t>() * unit;
+		}
+	};
 	if (auto kernel = param_json.find("kernel");
 	    kernel != param_json.end() && kernel->is_object()) {
-		if (auto it = kernel->find("flexibleMemorySize"); it != kernel->end()) {
-			m_flexible_memory_size = it->get<uint64_t>();
-		}
+		read_size(*kernel, "flexibleMemorySize", &m_memory_params.flexible_size);
+		read_size(*kernel, "cpuPageTableSize", &m_memory_params.cpu_table_memory);
+		read_size(*kernel, "gpuPageTableSize", &m_memory_params.gpu_table_memory);
+	}
+	if (auto amm = param_json.find("amm"); amm != param_json.end() && amm->is_object()) {
+		read_size(*amm, "pagetableMemorySizeInMib", &m_memory_params.amm_table_memory,
+		          1024ull * 1024ull);
 	}
 
 	auto localized = param_json.find("localizedParameters");
@@ -227,8 +237,8 @@ void Psf::Close() {
 	Common::LockGuard lock(m_mutex);
 
 	m_params.clear();
-	m_flexible_memory_size = 0;
-	m_opened               = false;
+	m_memory_params = {};
+	m_opened        = false;
 }
 
 bool Psf::IsValid() {
@@ -339,10 +349,10 @@ bool Psf::GetParamString(const char* name, char* value, size_t value_size) {
 	return false;
 }
 
-uint64_t Psf::GetFlexibleMemorySize() {
+MemoryParams Psf::GetMemoryParams() {
 	Common::LockGuard lock(m_mutex);
 
-	return m_flexible_memory_size;
+	return m_memory_params;
 }
 
 PlayGo::~PlayGo() {
@@ -646,10 +656,10 @@ bool SystemContentParamSfoGetString(const char* name, char* value, size_t value_
 	return sc->psf.GetParamString(name, value, value_size);
 }
 
-uint64_t SystemContentGetFlexibleMemorySize() {
+MemoryParams SystemContentGetMemoryParams() {
 	auto* sc = Common::Singleton<SystemContent>::Instance();
 
-	return sc->psf.GetFlexibleMemorySize();
+	return sc->psf.GetMemoryParams();
 }
 
 bool SystemContentGetIconPath(std::filesystem::path* path) {

src/loader/systemContent.h

@@ -14,7 +14,16 @@ void     SystemContentLoadParamSfo(const std::filesystem::path& file_name);
 bool     SystemContentParamSfoGetInt(const char* name, int32_t* value);
 bool     SystemContentParamSfoGetString(const char* name, std::string* value);
 bool     SystemContentParamSfoGetString(const char* name, char* value, size_t value_size);
-uint64_t SystemContentGetFlexibleMemorySize();
+// Memory sizes a game's param.json sets, in bytes; 0 where it sets none.
+struct MemoryParams {
+	uint64_t flexible_size    = 0;
+	uint64_t cpu_table_memory = 0;
+	uint64_t gpu_table_memory = 0;
+	uint64_t amm_table_memory = 0;
+};
+
+MemoryParams SystemContentGetMemoryParams();
+
 bool     SystemContentGetIconPath(std::filesystem::path* path);
 bool     SystemContentGetChunksNum(uint32_t* num);
 

src/main.cpp

@@ -70,6 +70,8 @@ static void PrintUsage() {
 	::printf("  --hdr                                Enable HDR10 output when supported.\n");
 	::printf(
 	    "  --hide-cursor                        Hide the cursor after 2 s idle. Default: off.\n");
+	::printf(
+	    "  --stretch                            Stretch the image to the window. Default: off.\n");
 	::printf("  --no-splash                          Disable the startup splash screen.\n");
 	::printf("  --vr                                 Enable the virtual VR headset.\n");
 	::printf("  --amd-cpu                            Apply AMD CPU instruction patches.\n");
@@ -199,6 +201,7 @@ static constexpr Flag FLAGS[] = {
     {"--fullscreen", &ConfigOptions::fullscreen_enabled},
     {"--hdr", &ConfigOptions::hdr_enabled},
     {"--hide-cursor", &ConfigOptions::hide_cursor_enabled},
+    {"--stretch", &ConfigOptions::stretch_enabled},
     {"--no-splash", &ConfigOptions::splash_enabled, false},
     {"--vr", &ConfigOptions::vr_enabled},
     {"--amd-cpu", &ConfigOptions::amd_cpu_enabled},

src/utils.cmake

@@ -65,6 +65,8 @@ else()
 	SET(CMAKE_NINJA_FORCE_RESPONSE_FILE 1 CACHE INTERNAL "")
 endif()
 
+# Guest code runs natively and needs x86-64-v3 (AVX2, BMI2, FMA), so every host has it.
+# No FMA contraction: a * b + c keeps both roundings.
 if(KYTY_CLANG_CL)
 	if(CMAKE_CXX_FLAGS MATCHES "/W[0-4]")
 		string(REGEX REPLACE "/W[0-4]" "/W3" CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
@@ -79,7 +81,7 @@ if(KYTY_CLANG_CL)
 	string(REGEX REPLACE "/MD" "/MT" CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE}")
 	string(REGEX REPLACE "/MD" "/MT" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
 			
-	set(KYTY_CPP_FLAGS "${KYTY_CPP_FLAGS} /utf-8 /Oy- /wd4244 /wd4305 /wd4800 /wd4345")
+	set(KYTY_CPP_FLAGS "${KYTY_CPP_FLAGS} /utf-8 /Oy- /clang:-march=x86-64-v3 /clang:-ffp-contract=off /wd4244 /wd4305 /wd4800 /wd4345")
   
 	if(KYTY_WARNINGS_ARE_ERRORS)
 		#set(KYTY_CPP_FLAGS "${KYTY_CPP_FLAGS} /WX")
@@ -103,6 +105,7 @@ elseif(CLANG OR GCC)
 	else()
 		set(KYTY_CPP_FLAGS "${KYTY_CPP_FLAGS} -fno-exceptions -fdiagnostics-color=always -finput-charset=UTF-8 -fexec-charset=UTF-8 -static-libgcc -static-libstdc++ -g -fno-strict-aliasing -fno-omit-frame-pointer -Wall -Wno-unused-value -fmessage-length=0")
 	endif()
+	set(KYTY_CPP_FLAGS "${KYTY_CPP_FLAGS} -march=x86-64-v3 -ffp-contract=off")
 	
     if(KYTY_WARNINGS_ARE_ERRORS)
         #set(KYTY_CPP_FLAGS "${KYTY_CPP_FLAGS} -Werror")

tests/AudioOut2PortTests.cpp

@@ -35,6 +35,8 @@ int                               g_open_waiters = 0;
 bool                              g_block_opens  = false;
 bool                              g_block_outputs = false;
 std::vector<int>                  g_blocked_output;
+std::atomic_uint64_t              g_process_time {0};
+std::atomic_bool                  g_process_time_frozen {false};
 
 void Check(bool value, const char* text) {
 	if (!value) {
@@ -109,8 +111,10 @@ PortParam MakeParam(uint32_t data_format = 0x200) {
 	return param;
 }
 
-AudioOut2::AudioOut2ContextHandle CreateContext(uint32_t queue_depth = 4, uint32_t max_object_ports = 0) {
+AudioOut2::AudioOut2ContextHandle
+CreateContext(uint32_t queue_depth = 4, uint32_t max_object_ports = 0, uint32_t max_ports = 8) {
 	ContextParam param {};
+	param.max_ports                           = max_ports;
 	param.max_object_ports                    = max_object_ports;
 	param.queue_depth                         = queue_depth;
 	param.num_grains                          = 512;
@@ -279,13 +283,139 @@ void TestSlotReuse() {
 	AudioOut2::AudioOut2ContextDestroy(context);
 }
 
+void TestObjectPortCapacityPerContext() {
+	constexpr int port_full    = static_cast<int32_t>(0x80268012u);
+	constexpr int invalid_port = static_cast<int32_t>(0x80268009u);
+	std::array<AudioOut2::AudioOut2ContextHandle, 3> contexts {};
+	std::array<AudioOut2::AudioOut2PortHandle, 3>    last_ports {};
+	auto                                             object = MakeParam(0x100);
+	object.port_type                                        = 0x100;
+	for (size_t i = 0; i < contexts.size(); i++) {
+		contexts[i] = CreateContext(4, 512, 32);
+		CreatePort(contexts[i], MakeParam());
+		for (size_t j = 0; j < 512; j++) {
+			AudioOut2::AudioOut2PortHandle port = 0;
+			Check(AudioOut2::AudioOut2PortCreate(contexts[i], AsParam(&object), &port) == OK,
+			      "context could not create all 512 object ports");
+			last_ports[i] = port;
+		}
+		AudioOut2::AudioOut2PortHandle overflow = UINT64_MAX;
+		Check(AudioOut2::AudioOut2PortCreate(contexts[i], AsParam(&object), &overflow) ==
+		              port_full &&
+		          overflow == UINT64_MAX,
+		      "object capacity did not reject the 513th port without changing its output");
+		CreatePort(contexts[i], MakeParam());
+	}
+
+	for (size_t i = 0; i < contexts.size(); i++) {
+		const auto stale = last_ports[i];
+		AudioOut2::AudioOut2PortDestroy(stale);
+		last_ports[i] = CreatePort(contexts[i], object);
+		Check(last_ports[i] != stale &&
+		          AudioOut2::AudioOut2PortSetAttributes(stale, nullptr, 0) == invalid_port,
+		      "reused object slot accepted its old handle");
+		std::array<float, 512> pcm;
+		pcm.fill(0.25f * static_cast<float>(i + 1));
+		SetPcm(last_ports[i], pcm.data());
+		const auto  output   = PushPcm(contexts[i], 512 * 2 * sizeof(float));
+		const float expected = pcm[0] * 0.70710678f;
+		Check(output.size() == 1, "last object port did not reach its context's mixer");
+		CheckSamples(output[0], {expected, expected}, "object audio crossed context boundaries");
+	}
+
+	AudioOut2::AudioOut2ContextDestroy(contexts[0]);
+	Check(AudioOut2::AudioOut2PortSetAttributes(last_ports[0], nullptr, 0) == invalid_port,
+	      "destroyed context retained its object handles");
+	const auto replacement = CreateContext(4, 512);
+	for (size_t i = 0; i < 512; i++) {
+		CreatePort(replacement, object);
+	}
+	Check(AudioOut2::AudioOut2PortSetAttributes(last_ports[0], nullptr, 0) == invalid_port,
+	      "reused context accepted an old object handle");
+	for (size_t i = 1; i < contexts.size(); i++) {
+		Check(AudioOut2::AudioOut2PortSetAttributes(last_ports[i], nullptr, 0) == OK,
+		      "destroying another context invalidated a live object port");
+		AudioOut2::AudioOut2ContextDestroy(contexts[i]);
+	}
+	AudioOut2::AudioOut2ContextDestroy(replacement);
+	Check(LiveDeviceCount() == 0, "object capacity test leaked audio devices");
+}
+
+void TestContextPortLimits() {
+	constexpr int port_full = static_cast<int32_t>(0x80268012u);
+	ContextParam  param {};
+	param.max_ports                           = 1;
+	param.max_object_ports                    = 2;
+	param.num_grains                          = 512;
+	AudioOut2::AudioOut2ContextHandle context = 0;
+	Check(AudioOut2::AudioOut2ContextCreate(AsParam(&param), nullptr, 0, &context) == OK,
+	      "small context create failed");
+	auto object             = MakeParam(0x100);
+	object.port_type        = 0x100;
+	const auto channel      = CreatePort(context, MakeParam());
+	const auto first_object = CreatePort(context, object);
+	CreatePort(context, object);
+	AudioOut2::AudioOut2PortHandle overflow      = 0;
+	const auto                     channel_param = MakeParam();
+	Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&channel_param), &overflow) ==
+	              port_full &&
+	          AudioOut2::AudioOut2PortCreate(context, AsParam(&object), &overflow) == port_full,
+	      "declared channel/object capacities were ignored");
+	AudioOut2::AudioOut2PortDestroy(channel);
+	Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&object), &overflow) == port_full,
+	      "object port borrowed a channel slot");
+	CreatePort(context, channel_param);
+	AudioOut2::AudioOut2PortDestroy(first_object);
+	Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&channel_param), &overflow) == port_full,
+	      "channel port borrowed an object slot");
+	CreatePort(context, object);
+	AudioOut2::AudioOut2ContextDestroy(context);
+	param.max_ports = param.max_object_ports = 0;
+	Check(AudioOut2::AudioOut2ContextCreate(AsParam(&param), nullptr, 0, &context) == OK,
+	      "empty context create failed");
+	Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&channel_param), &overflow) ==
+	              port_full &&
+	          AudioOut2::AudioOut2PortCreate(context, AsParam(&object), &overflow) == port_full,
+	      "zero port capacity was ignored");
+	AudioOut2::AudioOut2ContextDestroy(context);
+}
+
+void TestContextCapacityParameters() {
+	constexpr int invalid_param = static_cast<int32_t>(0x80268001u);
+	ContextParam  param;
+	std::memset(&param, 0xff, sizeof(param));
+	Check(AudioOut2::AudioOut2ContextResetParam(
+	          reinterpret_cast<AudioOut2::AudioOut2ContextParam*>(&param)) == OK &&
+	          param.max_ports == 8 && param.max_object_ports == 0 &&
+	          param.guarantee_object_ports == 0 && param.num_grains == 256 &&
+	          param.queue_depth == 1 && param.flags == 0 &&
+	          std::all_of(std::begin(param.reserved), std::end(param.reserved),
+	                      [](uint32_t word) { return word == 0; }),
+	      "context reset did not produce valid default parameters");
+	const auto defaults = param;
+	for (size_t i = 0; i < 3; i++) {
+		param = defaults;
+		if (i == 0) param.max_ports = 33;
+		if (i == 1) param.max_object_ports = 513;
+		if (i == 2) param.guarantee_object_ports = 1;
+		size_t                            memory  = SIZE_MAX;
+		AudioOut2::AudioOut2ContextHandle context = UINT64_MAX;
+		Check(AudioOut2::AudioOut2ContextQueryMemory(AsParam(&param), &memory) == invalid_param &&
+		          memory == SIZE_MAX &&
+		          AudioOut2::AudioOut2ContextCreate(AsParam(&param), nullptr, 0, &context) ==
+		              invalid_param &&
+		          context == UINT64_MAX && LiveDeviceCount() == 0,
+		      "invalid capacities allocated resources or modified outputs");
+	}
+}
+
 void TestFullTableRecovers() {
 	const auto                                  context = CreateContext();
 	const auto                                  param   = MakeParam();
 	std::vector<AudioOut2::AudioOut2PortHandle> ports;
-	ports.reserve(256);
+	ports.reserve(8);
 
-	for (int i = 0; i < 256; i++) {
+	for (int i = 0; i < 8; i++) {
 		AudioOut2::AudioOut2PortHandle port = 0;
 		Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &port) == OK,
 		      "port table filled early");
@@ -322,6 +452,11 @@ void TestConcurrentCreates() {
 		});
 	}
 	WaitForDeviceOpens(thread_count);
+	AudioOut2::AudioOut2PortHandle overflow = 0;
+	Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &overflow) ==
+	              static_cast<int32_t>(0x80268012u) &&
+	          overflow == 0,
+	      "pending creates did not reserve their channel capacity");
 	ReleaseDeviceOpens();
 	for (auto& thread: threads) {
 		thread.join();
@@ -349,17 +484,31 @@ void TestContextDestroyCancelsPendingCreate() {
 	    [&]() { result = AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &port); });
 	WaitForDeviceOpens(1);
 	AudioOut2::AudioOut2ContextDestroy(context);
+	const auto                     replacement        = CreateContext();
+	AudioOut2::AudioOut2PortHandle replacement_port   = 0;
+	int                            replacement_result = OK;
+	std::thread                    replacement_creator([&]() {
+		replacement_result =
+		    AudioOut2::AudioOut2PortCreate(replacement, AsParam(&param), &replacement_port);
+	});
+	WaitForDeviceOpens(2);
 	ReleaseDeviceOpens();
 	creator.join();
+	replacement_creator.join();
 
 	Check(result != OK, "destroyed context retained a pending port create");
+	Check(replacement_result == OK &&
+	          AudioOut2::AudioOut2PortSetAttributes(replacement_port, nullptr, 0) == OK &&
+	          LiveDeviceCount() == 1,
+	      "cancelled create corrupted the replacement context or leaked a device");
+	AudioOut2::AudioOut2ContextDestroy(replacement);
 	Check(LiveDeviceCount() == 0, "cancelled port create leaked a device");
 }
 
 void TestSynchronousDevicePushBypassesModelledQueue() {
-	const auto context = CreateContext(1);
-	const auto param   = MakeParam();
-	AudioOut2::AudioOut2PortHandle port = 0;
+	const auto                     context = CreateContext(1);
+	const auto                     param   = MakeParam();
+	AudioOut2::AudioOut2PortHandle port    = 0;
 	Check(AudioOut2::AudioOut2PortCreate(context, AsParam(&param), &port) == OK,
 	      "device port create failed");
 
@@ -368,6 +517,7 @@ void TestSynchronousDevicePushBypassesModelledQueue() {
 	ResetOutputCalls();
 
 	Check(AudioOut2::AudioOut2ContextPush(context, 1) == OK, "first sync push failed");
+	SetPcm(port, pcm);
 	Check(AudioOut2::AudioOut2ContextPush(context, 1) == OK,
 	      "device-paced sync push was blocked by modelled queue");
 	const auto calls = OutputCalls();
@@ -447,6 +597,44 @@ void TestHandleWithoutPcmDoesNotBypassQueue() {
 	AudioOut2::AudioOut2ContextDestroy(context);
 }
 
+void TestPushConsumesPcm() {
+	const auto context = CreateContext();
+	const auto port    = CreatePort(context, MakeParam());
+	uint32_t   pcm[512] {};
+	SetPcm(port, pcm);
+	ResetOutputCalls();
+	Check(AudioOut2::AudioOut2ContextPush(context, 1) == OK, "push with PCM failed");
+	Check(AudioOut2::AudioOut2ContextPush(context, 1) == OK, "push without new PCM failed");
+	SetPcm(port, pcm);
+	SetPcm(port, nullptr);
+	Check(AudioOut2::AudioOut2ContextPush(context, 1) == OK, "push after clearing PCM failed");
+	Check(OutputCalls().size() == 1, "a push replayed PCM the game did not set again");
+
+	AudioOut2::AudioOut2PortDestroy(port);
+	AudioOut2::AudioOut2ContextDestroy(context);
+}
+
+void TestQueueDrainsAtExactGrainRate() {
+	// Three 512-sample grains play in exactly 32000 us.
+	const auto context = CreateContext();
+	g_process_time_frozen = true;
+	const auto start      = g_process_time.load();
+	for (int push = 0; push < 3; push++) {
+		Check(AudioOut2::AudioOut2ContextPush(context, 0) == OK, "async push failed");
+	}
+	const auto queued_after = [&](uint64_t micros) {
+		g_process_time  = start + micros;
+		uint32_t queued = 0;
+		Check(AudioOut2::AudioOut2ContextGetQueueLevel(context, &queued, nullptr) == OK,
+		      "queue-level query failed");
+		return queued;
+	};
+	Check(queued_after(31999) == 1, "the third grain drained before it finished playing");
+	Check(queued_after(32000) == 0, "three grains did not drain in their playing time");
+	g_process_time_frozen = false;
+	AudioOut2::AudioOut2ContextDestroy(context);
+}
+
 void TestPcmCopiedBeforeScratchBufferReuse() {
 	const auto context = CreateContext();
 	const auto param   = MakeParam();
@@ -486,7 +674,6 @@ void TestPortGainAndValidation() {
 	const auto context = CreateContext();
 	const auto port = CreatePort(context, MakeParam());
 	std::vector<float> pcm(512 * 2, 0.25f);
-	SetPcm(port, pcm.data());
 	auto check_output = [&](const std::array<float, 2>& gain, float expected_pcm) {
 		CheckSamples(PushPcm(context, pcm.size() * sizeof(float)).at(0), {expected_pcm},
 		             "gain modified stored PCM");
@@ -494,10 +681,11 @@ void TestPortGainAndValidation() {
 		Check(g_output_gains.size() == 1 && g_output_gains[0] == gain,
 		      "incorrect per-channel gain forwarded to backend");
 	};
+	SetPcm(port, pcm.data());
 	check_output({1.0f, 1.0f}, 0.25f);
 	for (const std::array<float, 2> gain: {std::array {2.0f, 0.5f}, {0.0f, 0.0f}, {1.0f, 1.0f}}) {
 		SetAttribute(port, 1, gain);
-		check_output(gain, 0.25f);
+		SetPcm(port, pcm.data());
 		check_output(gain, 0.25f);
 	}
 
@@ -507,10 +695,11 @@ void TestPortGainAndValidation() {
 	SetPcm(port, pcm.data());
 	check_output(gain, 0.75f);
 
-	std::fill(pcm.begin(), pcm.end(), -0.25f);
-	const Pcm replacement {pcm.data()};
+	const std::vector<float> replacement_pcm(pcm.size(), -0.25f);
+	const Pcm                replacement {replacement_pcm.data()};
 	for (const float invalid: {-1.0f, std::numeric_limits<float>::infinity(),
 	                           std::numeric_limits<float>::quiet_NaN()}) {
+		SetPcm(port, pcm.data());
 		const std::array invalid_gain {1.0f, invalid};
 		const Attribute attributes[] {{0, 0, &replacement, sizeof(replacement)},
 		                              {1, 0, &invalid_gain, sizeof(invalid_gain)}};
@@ -523,6 +712,7 @@ void TestPortGainAndValidation() {
 		Check(AudioOut2::AudioOut2PortSetAttributes(port, AsAttribute(&attribute), 1) ==
 		          static_cast<int32_t>(0x80268001u), "malformed gain was accepted");
 	}
+	SetPcm(port, pcm.data());
 	const Attribute duplicates[] {{0, 0, &replacement, sizeof(replacement)},
 	                              {1, 0, &gain, sizeof(gain)}, {1, 0, &gain, sizeof(gain)}};
 	Check(AudioOut2::AudioOut2PortSetAttributes(port, AsAttribute(duplicates), 3) ==
@@ -553,12 +743,10 @@ void TestObjectGainAndAmbisonics() {
 		std::fill(integer.begin(), integer.end(), 0);
 		SetAttribute(object, 1, 0.0f);
 		SetAttribute(object, 8, 65u);
-		for (int push = 0; push < 2; push++) {
-			const auto output = PushPcm(context, output_bytes);
-			Check(output.size() == 1, "object-only push did not output exactly once");
-			CheckSamples(output[0], {0.5f * half_power, 0.5f * half_power},
-			             "object PCM ownership, default gain, or attribute freeze failed");
-		}
+		const auto output = PushPcm(context, output_bytes);
+		Check(output.size() == 1, "object-only push did not output exactly once");
+		CheckSamples(output[0], {0.5f * half_power, 0.5f * half_power},
+		             "object PCM ownership, default gain, or attribute freeze failed");
 		std::fill(floating.begin(), floating.end(), 0.5f);
 		std::fill(integer.begin(), integer.end(), 16384);
 		for (const auto& test: cases) {
@@ -569,10 +757,8 @@ void TestObjectGainAndAmbisonics() {
 			                              {8, 0, &test.channel, sizeof(test.channel)}};
 			Check(AudioOut2::AudioOut2PortSetAttributes(object, AsAttribute(attributes), 3) == OK,
 			      "object attributes could not be set atomically with PCM first");
-			for (int push = 0; push < 2; push++) {
-				CheckSamples(PushPcm(context, output_bytes).at(0), {0.25f * test.left, 0.25f * test.right},
-				             "object gain or ambisonics decoded incorrectly or accumulated across pushes");
-			}
+			CheckSamples(PushPcm(context, output_bytes).at(0), {0.25f * test.left, 0.25f * test.right},
+			             "object gain or ambisonics decoded incorrectly or accumulated across pushes");
 		}
 		AudioOut2::AudioOut2ContextAdvance(context);
 		SetAttribute(object, 8, UINT32_MAX);
@@ -599,22 +785,21 @@ void TestObjectBusRouting() {
 	SetPcm(second, bed.data());
 	auto param = MakeParam(0x100);
 	param.port_type = 0x100;
-	std::vector<float> object_pcm(512, 0.25f);
-	SetPcm(CreatePort(context, param), object_pcm.data());
+	const auto object = CreatePort(context, param);
 	param.user_handle = other_user;
-	SetPcm(CreatePort(context, param), object_pcm.data());
-	std::fill(object_pcm.begin(), object_pcm.end(), 1.0f);
-	SetPcm(CreatePort(other_context, param), object_pcm.data());
+	const auto other_user_object = CreatePort(context, param);
+	std::vector<float> object_pcm(512, 0.25f);
+	SetPcm(object, object_pcm.data());
+	SetPcm(other_user_object, object_pcm.data());
+	SetPcm(CreatePort(other_context, param), std::vector<float>(512, 1.0f).data());
 	constexpr auto bytes = 512 * 2 * sizeof(float);
-	for (int push = 0; push < 2; push++) {
-		const auto output = PushPcm(context, bytes);
-		Check(output.size() == 3, "object bus was missing or queued more than once");
-		CheckSamples(output[0], {0.125f}, "objects modified first bed");
-		CheckSamples(output[1], {0.125f}, "objects modified second bed");
-		CheckSamples(output[2], {0.35355339f}, "MAIN objects failed to mix across users or leaked contexts");
-	}
-	SetPcm(first, nullptr);
-	SetPcm(second, nullptr);
+	const auto output = PushPcm(context, bytes);
+	Check(output.size() == 3, "object bus was missing or queued more than once");
+	CheckSamples(output[0], {0.125f}, "objects modified first bed");
+	CheckSamples(output[1], {0.125f}, "objects modified second bed");
+	CheckSamples(output[2], {0.35355339f}, "MAIN objects failed to mix across users or leaked contexts");
+	SetPcm(object, object_pcm.data());
+	SetPcm(other_user_object, object_pcm.data());
 	CheckSamples(PushPcm(context, bytes).at(0), {0.35355339f}, "object output depended on bed PCM");
 	AudioOut2::AudioOut2ContextDestroy(context);
 	Check(LiveDeviceCount() == 1, "context destroy closed another context's object device");
@@ -731,8 +916,7 @@ uint32_t AudioOutOutputs(const OutputParam* params, uint32_t num, bool blocking)
 namespace Libs::LibKernel {
 
 uint64_t KYTY_SYSV_ABI KernelGetProcessTime() {
-	static std::atomic_uint64_t now {0};
-	return now.fetch_add(1000);
+	return g_process_time_frozen ? g_process_time.load() : g_process_time.fetch_add(1000);
 }
 
 } // namespace Libs::LibKernel
@@ -741,6 +925,9 @@ int main() {
 	Check(AudioOut2::AudioOut2UserCreate(1000, &g_user_handle) == OK, "test user create failed");
 	TestUserSupportedAttributes();
 	TestSlotReuse();
+	TestObjectPortCapacityPerContext();
+	TestContextPortLimits();
+	TestContextCapacityParameters();
 	TestFullTableRecovers();
 	TestConcurrentCreates();
 	TestContextDestroyCancelsPendingCreate();
@@ -748,6 +935,8 @@ int main() {
 	TestFloat12ChannelPortOutputsPcm();
 	TestAsynchronousDevicePushKeepsQueueBounded();
 	TestHandleWithoutPcmDoesNotBypassQueue();
+	TestPushConsumesPcm();
+	TestQueueDrainsAtExactGrainRate();
 	TestPcmCopiedBeforeScratchBufferReuse();
 	TestPortGainAndValidation();
 	TestObjectGainAndAmbisonics();

tests/BuildFlagsTests.cpp

@@ -0,0 +1,17 @@
+#include <cstdio>
+
+#if !defined(__AVX2__) || !defined(__BMI2__) || !defined(__FMA__)
+#error "x86-64 builds target x86-64-v3"
+#endif
+
+int main() {
+  // (1 + 2^-23)^2 = 1 + 2^-22 + 2^-46; only a fused multiply-add keeps the 2^-46.
+  volatile float a = 1.0f + 0x1p-23f;
+  const float result = a * a - (1.0f + 0x1p-22f);
+  if (result != 0.0f) {
+    std::fprintf(stderr, "BuildFlagsTests: failed: a * a - b was contracted (%a)\n", result);
+    return 1;
+  }
+  std::puts("BuildFlagsTests: all cases passed");
+  return 0;
+}

tests/ControllerSettingsTests.cpp

@@ -689,6 +689,27 @@ void TestTriggerEffectState() {
 	check_state(0, 0, "switching controllers retained effect states");
 }
 
+void TestUnreadPressSurvivesStickBurst() {
+	Controller              controller;
+	std::array<PadData, 64> data {};
+	SetButton(1, PAD_BUTTON_CROSS, true);
+	for (int value = 0; value < 100; value++) {
+		SetAxis(1, Axis::LeftX, value);
+	}
+	SetButton(1, PAD_BUTTON_CROSS, false);
+	for (int value = 100; value < 200; value++) {
+		SetAxis(1, Axis::LeftX, value);
+	}
+	Check(PadRead(1, data.data(), 64) == 2 && data[0].buttons == PAD_BUTTON_CROSS &&
+	          data[0].left_stick_x == 99 && data[1].buttons == 0 && data[1].left_stick_x == 199,
+	      "a stick burst pushed an unread press out of the pad history");
+	SetAxis(1, Axis::LeftX, 50);
+	SetButton(1, PAD_BUTTON_CROSS, true);
+	Check(PadRead(1, data.data(), 64) == 2 && data[0].buttons == 0 &&
+	          data[0].left_stick_x == 50 && data[1].buttons == PAD_BUTTON_CROSS,
+	      "a stick move after a read did not add a sample");
+}
+
 } // namespace
 
 int main() {
@@ -702,6 +723,7 @@ int main() {
 	TestMaskedTriggersAndValidation();
 	TestIndependentOutputsAndPadSwitch();
 	TestTriggerEffectState();
+	TestUnreadPressSurvivesStickBurst();
 	TestMultiplayerRouting();
 	TestDedicatedKeyboardOutputs();
 	Config::Shutdown();

tests/ImagePageTableTests.cpp

@@ -3,6 +3,8 @@
 #include <cstdint>
 #include <cstdio>
 #include <cstdlib>
+#include <iterator>
+#include <type_traits>
 #include <utility>
 #include <vector>
 
@@ -13,6 +15,11 @@ using Table      = Libs::Graphics::MultiLevelPageTable<Owners, 20, 44, 14>;
 using PageOwners = Libs::Graphics::InlinePageOwnerList<uint32_t, 16>;
 using OwnerTable = Libs::Graphics::MultiLevelPageTable<PageOwners, 20, 44, 14>;
 
+// MSVC's STL and libc++ don't need these traits for std::any_of, but libstdc++ does, so only a
+// Linux build would notice them missing.
+static_assert(std::is_same_v<std::iterator_traits<PageOwners::const_iterator>::iterator_category,
+                             std::input_iterator_tag>);
+
 void Check(bool value, const char* text) {
 	if (!value) {
 		std::fprintf(stderr, "ImagePageTableTests: failed: %s\n", text);

tests/ResourceMaterializationTests.cpp

@@ -1314,6 +1314,62 @@ void TestSignedStorageSpecialization() {
   }
 }
 
+// One descriptor serves a store, an atomic that a user-data condition guards, and an image no
+// block lists. Unreached, the atomic reads a zero descriptor instead of the RGBA8 format the
+// others use; reached, it rejects that format.
+void TestUnreachedImageIgnoresSharedDescriptor(bool atomic64) {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  using Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind;
+  using Libs::Graphics::Prospero::BufferFormat;
+  Program program;
+  program.stage = Libs::Graphics::ShaderType::Compute;
+  program.srt_plan_complete = program.resource_tracking_complete = true;
+  program.user_data_count = 2u;
+  auto &entry = AddValueBlock(program);
+  auto &guarded = AddValueBlock(program);
+  auto &done = AddValueBlock(program);
+  const auto user = [&](uint32_t index) {
+    return Value(&entry.AppendNewInst(ValueOpcode::GetUserData, {Value(ScalarReg(index))}));
+  };
+  DescriptorSource source;
+  source.dword_count = 8u;
+  source.dwords.fill(Value(0u));
+  source.dwords[0] = Value(0x100u);
+  source.dwords[1] = user(0);
+  source.dwords[3] = Value(uint32_t(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
+  program.descriptor_sources.push_back(source);
+  const auto dimension = Libs::Graphics::ShaderRecompiler::Decoder::ImageDimension::Dim2D;
+  for (const bool atomic : {false, true, false})
+    program.info.images.push_back({.source = 0u, .resource_class = ImageResourceClass::Storage,
+        .dimension = dimension, .written = true, .atomic = atomic, .atomic64 = atomic && atomic64});
+  for (uint32_t image = 0; image < 2u; ++image)
+    program.memory_info.push_back({.kind = ResourceKind::Image, .resource = image});
+  entry.AppendNewInst(ValueOpcode::ImageWrite, {Value(0u), Value(0u), Value(0u), Value(true)})
+      .SetFlags(MemoryFlags{.index = 0});
+  guarded.AppendNewInst(atomic64 ? ValueOpcode::ImageAtomicUMax64 : ValueOpcode::ImageAtomicUMax32,
+                        {Value(0u), Value(0u), Value(0u), Value(true)})
+      .SetFlags(MemoryFlags{.index = 1});
+  entry.condition = Value(&entry.AppendNewInst(ValueOpcode::INotEqual32, {user(1), Value(0u)}));
+  entry.terminator.kind = TerminatorKind::ConditionalBranch;
+  entry.terminator.true_block = &guarded;
+  entry.terminator.false_block = &done;
+  guarded.terminator.kind = TerminatorKind::Branch;
+  guarded.terminator.true_block = &done;
+  done.terminator.kind = TerminatorKind::Return;
+  auto plan = Extract(program);
+  std::array<uint32_t, 2> data{uint32_t(BufferFormat::k8_8_8_8UNorm) << 20u, 0u};
+  const SrtRuntime runtime{.user_data = data};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            snapshot.images[0].dwords[1] == data[0] && snapshot.images[2].dwords[1] == data[0] &&
+            snapshot.images[1] == DescriptorValue{.dword_count = 8u},
+        "unreached atomic image kept the shared descriptor or an unlisted image lost it");
+  data[1] = 1u;
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
+        "reached atomic image accepted a format its width rejects");
+}
+
 void TestIndirectImageSamplerClasses(bool mixed) {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   using Libs::Graphics::Prospero::BufferFormat;
@@ -1451,6 +1507,7 @@ int main() {
   TestEmptyScalarBufferSpecialization();
   TestMixedSamplerVariantsShareRuntimeDescriptor();
   TestSignedStorageSpecialization();
+  for (const bool atomic64 : {false, true}) TestUnreachedImageIgnoresSharedDescriptor(atomic64);
   for (const bool mixed : {false, true}) TestIndirectImageSamplerClasses(mixed);
   TestShaderCacheRecords();
   std::puts("ResourceMaterializationTests: all cases passed");

tests/ResourceTrackingTests.cpp

@@ -1282,15 +1282,18 @@ void TestIndexedPointerImageTable() {
           {image, sampler, fixture->ImageAddress(), parent}, fixture->AddMemory(image_memory, 0x8bcu));
     return fixture;
   };
+  const auto nested_geometry = [](const auto& plan, uint32_t count) {
+    const auto& indirect = *plan.descriptor_sources[plan.info.images[0].source].indirect_descriptor;
+    return indirect.table_maximum == 1u && indirect.pointer_table &&
+           indirect.pointer_table->indexed && indirect.pointer_table->count == count &&
+           indirect.pointer_table->stride == 368u && indirect.pointer_table->offset == 64u;
+  };
   auto fixture = make();
   fixture->PlanAndTrack();
   const auto plan = ExtractResourcePlan(fixture->program);
   Check(plan.info.images.size() == 1u && plan.info.buffers.size() == 1u,
         "indexed nested table expanded IR candidates or discarded the GPU pointer load");
-  const auto& indirect = *plan.descriptor_sources[plan.info.images[0].source].indirect_descriptor;
-  Check(indirect.table_maximum == 1u && indirect.pointer_table &&
-            indirect.pointer_table->indexed && indirect.pointer_table->count == 3u &&
-            indirect.pointer_table->stride == 368u && indirect.pointer_table->offset == 64u,
+  Check(nested_geometry(plan, 3u),
         "nested table did not preserve separate parent and child geometry");
   constexpr uint64_t high = uint64_t{1} << 32u;
   constexpr uint32_t table = 0x2000u, objects = 0x10000u, row_bytes = 1536u;
@@ -1389,8 +1392,13 @@ void TestIndexedPointerImageTable() {
         "explicit null pointer bypassed the specialization reader");
   auto reversed = make(Case::ReverseSigned);
   reversed->PlanAndTrack();
+  // An inclusive child limit bounds one more child record, never the exclusive count.
+  auto inclusive = make(Case::InclusiveChild);
+  inclusive->PlanAndTrack();
+  Check(nested_geometry(ExtractResourcePlan(inclusive->program), 4u),
+        "an inclusive child limit did not bound one more child record");
   for (const auto variant : {Case::MissingParent, Case::WrongParent, Case::MissingChild,
-          Case::InclusiveChild, Case::MismatchedPair, Case::DifferentLaneMask,
+          Case::MismatchedPair, Case::DifferentLaneMask,
           Case::ExtraUse, Case::UnboundedChoice, Case::MissingOuterWitness, Case::WrongOuterWitness,
           Case::CyclicChoice, Case::ExcessiveDomain}) {
     auto invalid = make(variant);
@@ -2497,11 +2505,15 @@ void TestDescendingImageRange() {
   enum class Variant {
     Valid, WrongStep, IncrementBypass, MissingNonzero, NegativeInitial,
     Bitmap, BitmapAddsBit, BitmapWrongWord, BitmapUnknownInitial,
-    BitmapByteOffset, BitmapScalarOffset, BitmapOuterClobber, BitmapWrongLoadMask
+    BitmapByteOffset, BitmapScalarOffset, BitmapOuterClobber, BitmapWrongLoadMask,
+    Bitmap8, Bitmap8MisplacedLoad, Bitmap8OtherBuffer, Bitmap8OuterClobber, Bitmap8WrongUpperWord,
+    Bitmap8NarrowMask, Bitmap8OtherGuard
   };
   const auto make_plan = [](Variant variant) {
     Fixture fixture;
     const bool bitmap = variant >= Variant::Bitmap;
+    // Eight-word records are two four-word loads of one buffer record.
+    const uint32_t bitmap_words = variant >= Variant::Bitmap8 ? 8u : 4u;
     auto *entry = fixture.block;
     auto *header = fixture.AddBlock();
     auto *guard = fixture.AddBlock();
@@ -2524,11 +2536,11 @@ void TestDescendingImageRange() {
     branch(latch, header);
     auto &induction = header->AppendNewInst(ValueOpcode::Phi);
     induction.SetFlags(Type::U32);
-    std::array<Inst *, 4> outer_words {}, remaining_words {};
-    std::array<Value, 4> loaded_words {};
+    std::array<Inst *, 8> outer_words {}, remaining_words {};
+    std::array<Value, 8> loaded_words {};
     if (bitmap)
-      for (auto& word : outer_words)
-        word = &header->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U32));
+      for (uint32_t word = 0; word < bitmap_words; ++word)
+        outer_words[word] = &header->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U32));
     const auto bound = fixture.Emit(ValueOpcode::SGreaterThanEqual32,
         {Value(&induction), Value(0u)}, 0, header);
     choose(header, guard, variant == Variant::IncrementBypass ? latch : exit, bound);
@@ -2545,14 +2557,27 @@ void TestDescendingImageRange() {
       read.idxen = true;
       read.offen = variant == Variant::BitmapByteOffset;
       read.data_dwords = 4u;
-      const auto packed = fixture.Emit(ValueOpcode::LoadBufferU32x4,
-          {buffer, lane,
-           Value(variant == Variant::BitmapByteOffset ? 4u : 0u),
-           Value(variant == Variant::BitmapScalarOffset ? 4u : 0u), load_active},
-          fixture.AddMemory(read, 0x10));
-      for (uint32_t word = 0; word < 4u; ++word) {
+      std::array<Value, 2> packed;
+      for (uint32_t part = 0; part < bitmap_words / 4u; ++part) {
+        read.offset = part * (variant == Variant::Bitmap8MisplacedLoad ? 32u : 16u);
+        const auto source = variant == Variant::Bitmap8OtherBuffer && part != 0u
+            ? fixture.Buffer({fixture.UserData(8), fixture.UserData(9), fixture.UserData(10),
+                              fixture.UserData(11)}) : buffer;
+        const auto guard = variant == Variant::Bitmap8OtherGuard && part != 0u
+            ? fixture.Emit(ValueOpcode::LogicalAnd,
+                  {load_active, fixture.Emit(ValueOpcode::ULessThan32, {lane, Value(16u)})})
+            : load_active;
+        packed[part] = fixture.Emit(ValueOpcode::LoadBufferU32x4,
+            {source, lane,
+             Value(variant == Variant::BitmapByteOffset ? 4u : 0u),
+             Value(variant == Variant::BitmapScalarOffset ? 4u : 0u), guard},
+            fixture.AddMemory(read, 0x10 + part * 8u));
+      }
+      for (uint32_t word = 0; word < bitmap_words; ++word) {
+        const bool wrong_word = variant == Variant::BitmapWrongWord ||
+                                (variant == Variant::Bitmap8WrongUpperWord && word >= 4u);
         auto component = fixture.Emit(ValueOpcode::CompositeExtractU32x4,
-            {packed, Value(variant == Variant::BitmapWrongWord ? (word + 1u) % 4u : word)}, 0, entry);
+            {packed[word / 4u], Value((wrong_word ? word + 1u : word) % 4u)}, 0, entry);
         component = fixture.Emit(ValueOpcode::SelectU32,
             {load_active, component, fixture.Emit(ValueOpcode::UndefU32)});
         loaded_words[word] = component;
@@ -2565,7 +2590,7 @@ void TestDescendingImageRange() {
         remaining_words[word]->AddPhiOperand(header, Value(outer_words[word]));
       }
       auto selected = fixture.Emit(ValueOpcode::UndefU32, {}, 0, guard);
-      for (uint32_t word = 0; word < 4u; ++word) {
+      for (uint32_t word = 0; word < bitmap_words; ++word) {
         selected = fixture.Emit(ValueOpcode::SelectU32,
             {fixture.Emit(ValueOpcode::IEqual32, {Value(&induction), Value(word)}, 0, guard),
              Value(remaining_words[word]), selected}, 0, guard);
@@ -2628,32 +2653,39 @@ void TestDescendingImageRange() {
               {fixture.Emit(ValueOpcode::ShiftLeftLogical32, {Value(1u), bit})})});
       if (variant == Variant::BitmapAddsBit)
         next = fixture.Emit(ValueOpcode::BitwiseOr32, {next, Value(1u << 21u)});
-      for (uint32_t word = 0; word < 4u; ++word) {
+      for (uint32_t word = 0; word < bitmap_words; ++word) {
+        // Relative register writes index with M0 & 0xff; a narrower mask aliases upper words.
+        const auto index = bitmap_words == 4u ? Value(&induction)
+            : fixture.Emit(ValueOpcode::BitwiseAnd32,
+                  {Value(&induction), Value(variant == Variant::Bitmap8NarrowMask ? 3u : 255u)});
         const auto selected_word = fixture.Emit(ValueOpcode::LogicalAnd,
-            {active, fixture.Emit(ValueOpcode::IEqual32, {Value(&induction), Value(word)})});
+            {active, fixture.Emit(ValueOpcode::IEqual32, {index, Value(word)})});
         remaining_words[word]->AddPhiOperand(body, fixture.Emit(ValueOpcode::SelectU32,
             {selected_word, next, Value(remaining_words[word])}));
         auto carried = Value(remaining_words[word]);
-        if (variant == Variant::BitmapOuterClobber)
+        if (variant == Variant::BitmapOuterClobber || variant == Variant::Bitmap8OuterClobber)
           carried = fixture.Emit(ValueOpcode::SelectU32,
               {fixture.Emit(ValueOpcode::IEqual32,
                    {Value(&induction), Value(word + 1u)}, 0, latch),
-               loaded_words[(word + 1u) % 4u], carried}, 0, latch);
+               loaded_words[(word + 1u) % bitmap_words], carried}, 0, latch);
         outer_words[word]->AddPhiOperand(latch, carried);
       }
     }
     const auto step = fixture.Emit(ValueOpcode::IAdd32,
         {Value(&induction), Value(variant == Variant::WrongStep ? UINT32_MAX - 1u : UINT32_MAX)}, 0, latch);
-    induction.AddPhiOperand(entry, Value(variant == Variant::NegativeInitial ? UINT32_MAX : 3u));
+    induction.AddPhiOperand(entry,
+        Value(variant == Variant::NegativeInitial ? UINT32_MAX : bitmap_words - 1u));
     induction.AddPhiOperand(latch, step);
     fixture.PlanAndTrack();
     Check(ordinal.Instruction()->Arg(1).Resolve() == bit.Resolve(),
           "descending range retained impossible relative-register stores");
     const auto &indirect = fixture.program.descriptor_sources[
         fixture.program.info.images[0].source].indirect_descriptor;
-    Check(indirect && indirect->selector.has_value() == (variant == Variant::Bitmap) &&
-              (!indirect->selector || indirect->selector->bitset) &&
-              indirect->key_count == Value(128u) &&
+    const bool selects = variant == Variant::Bitmap || variant == Variant::Bitmap8;
+    Check(indirect && indirect->selector.has_value() == selects &&
+              (!indirect->selector || (indirect->selector->bitset &&
+                                       indirect->selector->stride == bitmap_words * 4u)) &&
+              indirect->key_count == Value(bitmap_words * 32u) &&
               !indirect->selector_first.IsEmpty(),
           "descending bitscan range lost its bound or accepted an unproved bitmap subset");
     return ExtractResourcePlan(fixture.program);
@@ -2693,10 +2725,14 @@ void TestDescendingImageRange() {
   }
 
   auto bitmap_plan = make_plan(Variant::Bitmap);
+  auto wide_plan = make_plan(Variant::Bitmap8);
   for (const auto variant : {Variant::BitmapAddsBit, Variant::BitmapWrongWord,
                              Variant::BitmapUnknownInitial, Variant::BitmapByteOffset,
                              Variant::BitmapScalarOffset, Variant::BitmapOuterClobber,
-                             Variant::BitmapWrongLoadMask})
+                             Variant::BitmapWrongLoadMask, Variant::Bitmap8MisplacedLoad,
+                             Variant::Bitmap8OtherBuffer, Variant::Bitmap8OuterClobber,
+                             Variant::Bitmap8WrongUpperWord, Variant::Bitmap8NarrowMask,
+                             Variant::Bitmap8OtherGuard})
     make_plan(variant);
   // Every descriptor is distinct, so retained keys are visible in the snapshot as well as
   // its lookup table. An unrelated table slot must never be read, even when it is malformed.
@@ -2711,23 +2747,25 @@ void TestDescendingImageRange() {
   const SrtRuntime bitmap_runtime {.user_data = bitmap_data, .userdata = &memory,
       .read_specialization_memory = ReadLinearTestMemory};
   memory.watched_address = bitmap_address;
+  auto *checked_plan = &bitmap_plan;
   const auto check_bitmap = [&](std::initializer_list<uint32_t> ranks, uint32_t count) {
     bitmap_data[2] = count;
     memory.watched_reads = memory.descriptor_reads = 0u;
     std::vector<uint32_t> keys;
     for (const auto rank : ranks) keys.push_back(count - rank);
     std::ranges::sort(keys);
-    Check(MaterializeResources(bitmap_plan, bitmap_runtime, snapshot, specialization) &&
+    const auto dwords = bitmap_data[6] * (bitmap_data[5] >> 18u);
+    Check(MaterializeResources(*checked_plan, bitmap_runtime, snapshot, specialization) &&
               memory.watched_reads == uint32_t(bitmap_data[6] != 0u) &&
               (bitmap_data[6] == 0u || memory.watched_dwords ==
-                  std::min(bitmap_data[6] * 4u, MaxIndirectDescriptorProbes)) &&
+                  std::min(dwords, MaxIndirectDescriptorProbes)) &&
               memory.descriptor_reads == keys.size() &&
               snapshot.images.size() == std::max<size_t>(keys.size(), 1u),
           "bitmap selector lost bounded reads or retained unreachable descriptors");
-    for (uint32_t first = 0; first < bitmap_data[6] * 4u; first += MaxIndirectDescriptorProbes)
+    for (uint32_t first = 0; first < dwords; first += MaxIndirectDescriptorProbes)
       Check(std::ranges::find(snapshot.specialization_reads,
                 std::pair{uint64_t{bitmap_address} + first * 4u,
-                    uint64_t{std::min(bitmap_data[6] * 4u - first, MaxIndirectDescriptorProbes)} * 4u}) !=
+                    uint64_t{std::min(dwords - first, MaxIndirectDescriptorProbes)} * 4u}) !=
                 snapshot.specialization_reads.end(),
             "bitmap chunk was omitted from same-dispatch write-alias protection");
     if (keys.empty()) {
@@ -2779,8 +2817,8 @@ void TestDescendingImageRange() {
   memory.fail_address = UINT64_MAX;
 
   const auto valid_data = bitmap_data;
-  for (const auto [word, value] : std::array<std::pair<uint32_t, uint32_t>, 6>{{
-           {4u, bitmap_address + 1u}, {5u, 8u << 16u}, {5u, (16u << 16u) | (1u << 31u)},
+  for (const auto [word, value] : std::array<std::pair<uint32_t, uint32_t>, 7>{{
+           {4u, bitmap_address + 1u}, {5u, 8u << 16u}, {5u, 32u << 16u}, {5u, (16u << 16u) | (1u << 31u)},
            {7u, 0x5204u | (1u << 23u)}, {7u, 0x5204u | (1u << 28u)},
            {7u, 0x5204u | (1u << 30u)}}}) {
     bitmap_data[word] = value;
@@ -2820,6 +2858,23 @@ void TestDescendingImageRange() {
               return read.first >= memory.fail_address;
             }),
         "a complete bitmap domain read redundant chunks or lost candidate keys");
+
+  // Eight-word records keep their upper words' ranks, and a key below the table is never read.
+  checked_plan = &wide_plan;
+  bitmap_data = valid_data;
+  bitmap_data[5] = 32u << 16u;
+  bitmap_data[6] = 2u;
+  std::fill_n(memory.words.begin() + bitmap_word, 16u, 0u);
+  memory.words[bitmap_word] = 1u | (1u << 2u);
+  memory.words[bitmap_word + 12u] = 1u << 1u;
+  memory.fail_address = 0x2000u - 32u;
+  check_bitmap({0u, 2u, 129u}, 2u);
+  memory.fail_address = UINT64_MAX;
+  for (const auto stride : {16u, 64u}) {
+    bitmap_data[5] = stride << 16u;
+    Check(!MaterializeResources(wide_plan, bitmap_runtime, snapshot, specialization),
+          "an eight-word bitmap selector accepted a different record stride");
+  }
 }
 
 void TestUniformizedMaterialImageKeys() {
@@ -4804,7 +4859,8 @@ void TestSharedAppendSelectorBounds() {
     Valid, PayloadOob, MissingInit, InactiveInit, UnknownInit, DuplicateAtomic,
     WrongAtomicGuard, UnguardedCandidate, WrongStep, WrongLocalIndex, Multiwave,
     OuterReentry, InnerAppendCycle, EarlyCounterRead, WrongConsumerBound, ExtraWriter, CounterAlias,
-    PayloadWrap, CounterOob
+    PayloadWrap, CounterOob, Lane, LaneUnimpliedRead, LaneUnimpliedCounter, LaneCounterMask,
+    LaneOtherBypass, LaneInactiveSelect
   };
   const auto make_plan = [](Variant variant, uint32_t first_base = 1032u, uint32_t consumed_base = 8u) {
     Fixture fixture;
@@ -4837,6 +4893,7 @@ void TestSharedAppendSelectorBounds() {
     auto *entry = fixture.block;
     auto *initialize = fixture.AddBlock();
     auto *initialized = fixture.AddBlock();
+    const bool lane = variant >= Variant::Lane;
     const auto material = fixture.Buffer({fixture.UserData(0), fixture.UserData(1),
                                           fixture.UserData(2), fixture.UserData(3)});
     const auto table = fixture.Address(fixture.UserData(4), fixture.UserData(5));
@@ -4855,12 +4912,18 @@ void TestSharedAppendSelectorBounds() {
         {Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(1u)});
     const auto local = fixture.Emit(ValueOpcode::IAdd32,
         {x, fixture.Emit(ValueOpcode::ShiftLeftLogical32, {y, Value(3u)})});
-    auto zero_lane = fixture.Emit(ValueOpcode::IEqual32, {local, Value(0u)});
+    // The negated form skips the initializer only when every lane is nonzero.
+    const auto local_is = [&](uint32_t value) {
+      return fixture.Emit(ValueOpcode::LogicalNot,
+          {fixture.Emit(ValueOpcode::INotEqual32, {local, Value(value)})});
+    };
+    auto zero_lane = lane ? local_is(0u) : fixture.Emit(ValueOpcode::IEqual32, {local, Value(0u)});
     if (variant == Variant::InactiveInit) zero_lane = Value(false);
     if (variant == Variant::UnknownInit)
       zero_lane = fixture.Emit(ValueOpcode::LogicalAnd,
           {zero_lane, fixture.Emit(ValueOpcode::IEqual32, {fixture.UserData(8), Value(0u)})});
-    choose(entry, initialize, initialized, zero_lane, CFG::BranchCondition::ExecNonZero);
+    choose(entry, initialize, initialized, variant == Variant::LaneOtherBypass ? local_is(1u) : zero_lane,
+           CFG::BranchCondition::ExecNonZero);
     if (variant != Variant::MissingInit)
       fixture.Emit(ValueOpcode::WriteSharedU32x2,
           {Value(0u), Value(0u), Value(0u), zero_lane}, shared(0u), initialize);
@@ -4929,13 +4992,25 @@ void TestSharedAppendSelectorBounds() {
     if (variant == Variant::ExtraWriter)
       fixture.Emit(ValueOpcode::WriteSharedU32,
           {Value(0u), Value(UINT32_MAX), Value(true)}, shared(12u), previous);
-    if (final_counter.IsEmpty())
-      final_counter = fixture.Emit(ValueOpcode::LoadSharedU32,
-          {Value(0u), Value(true)}, shared(0u), previous);
+    // Lane variants consume under a lane mask, which their counter and record reads use.
+    Value lanes(true), other;
+    if (lane) {
+      lanes = fixture.Emit(ValueOpcode::ULessThan32, {y, fixture.UserData(8)}, 0, previous);
+      other = fixture.Emit(ValueOpcode::ULessThan32, {x, fixture.UserData(8)}, 0, previous);
+    }
+    auto *guarded = lane ? fixture.AddBlock() : previous;
     auto *header = fixture.AddBlock();
     auto *body = fixture.AddBlock();
     auto *latch = fixture.AddBlock();
     auto *exit = fixture.AddBlock();
+    if (lane) choose(previous, guarded, exit, lanes, CFG::BranchCondition::ExecNonZero);
+    previous = guarded;
+    if (final_counter.IsEmpty())
+      final_counter = fixture.Emit(ValueOpcode::LoadSharedU32,
+          {Value(0u), variant == Variant::LaneUnimpliedCounter ? other : lanes}, shared(0u), previous);
+    if (lane)
+      final_counter = fixture.Emit(ValueOpcode::ReadFirstLane,
+          {final_counter, variant == Variant::LaneCounterMask ? other : lanes}, 0, previous);
     branch(previous, header);
     auto &index = header->AppendNewInst(ValueOpcode::Phi);
     index.SetFlags(Type::U32);
@@ -4944,12 +5019,22 @@ void TestSharedAppendSelectorBounds() {
     choose(header, body, exit, in_list, CFG::BranchCondition::VccNonZero);
     fixture.block = body;
     const auto selected = fixture.Emit(ValueOpcode::LoadSharedU32,
-        {fixture.Emit(ValueOpcode::ShiftLeftLogical32, {Value(&index), Value(2u)}), Value(true)},
-        shared(consumed_base));
-    const auto offset32 = fixture.Emit(ValueOpcode::ShiftLeftLogical32, {selected, Value(5u)});
-    const auto offset = fixture.Emit(ValueOpcode::ReadFirstLane,
-        {fixture.Emit(ValueOpcode::IAdd32,
-             {fixture.Emit(ValueOpcode::ShiftLeftLogical32, {offset32, Value(2u)}), offset32}), Value(true)});
+        {fixture.Emit(ValueOpcode::ShiftLeftLogical32, {Value(&index), Value(2u)}),
+         variant == Variant::LaneUnimpliedRead ? other : lanes}, shared(consumed_base));
+    Value offset;
+    if (lane) {
+      // The record is selected from one lane and scaled on the scalar side.
+      const auto mask = variant == Variant::LaneInactiveSelect
+          ? fixture.Emit(ValueOpcode::LogicalAnd, {lanes, other}) : lanes;
+      offset = fixture.Emit(ValueOpcode::IMul32,
+          {fixture.Emit(ValueOpcode::ReadFirstLane, {selected, mask}), Value(160u)});
+    } else {
+      const auto offset32 = fixture.Emit(ValueOpcode::ShiftLeftLogical32, {selected, Value(5u)});
+      offset = fixture.Emit(ValueOpcode::ReadFirstLane,
+          {fixture.Emit(ValueOpcode::IAdd32,
+               {fixture.Emit(ValueOpcode::ShiftLeftLogical32, {offset32, Value(2u)}), offset32}),
+           Value(true)});
+    }
     MemoryInfo scalar;
     scalar.kind = ResourceKind::ScalarBuffer;
     scalar.offset = 144u;
@@ -4978,7 +5063,8 @@ void TestSharedAppendSelectorBounds() {
     fixture.PlanAndTrack({.compute = &compute});
     const auto &indirect = fixture.program.descriptor_sources[
         fixture.program.info.images[0].source].indirect_descriptor;
-    const bool proven = variant == Variant::Valid || variant == Variant::PayloadOob;
+    const bool proven = variant == Variant::Valid || variant == Variant::PayloadOob ||
+                        variant == Variant::Lane;
     Check(indirect && indirect->selector && indirect->selector->offset == 144u &&
               indirect->selector->stride == (proven ? 160u : 32u) &&
               indirect->key_count.IsEmpty() != proven &&
@@ -4995,6 +5081,11 @@ void TestSharedAppendSelectorBounds() {
       Variant::InnerAppendCycle, Variant::EarlyCounterRead, Variant::WrongConsumerBound, Variant::ExtraWriter,
       Variant::CounterAlias, Variant::PayloadWrap, Variant::CounterOob})
     make_plan(variant);
+  auto lane_plan = make_plan(Variant::Lane);
+  for (const auto variant : {Variant::LaneUnimpliedRead, Variant::LaneUnimpliedCounter,
+      Variant::LaneCounterMask, Variant::LaneOtherBypass, Variant::LaneInactiveSelect})
+    CheckFatal([&] { make_plan(variant); }, "not a valid runtime value",
+               "shared append provenance accepted an unimplied lane mask or initializer bypass");
 
   LinearTestMemory memory;
   for (uint32_t index = 0; index < 8u; ++index) memory.words[(144u + 160u * index) / 4u] = index + 1u;
@@ -5007,7 +5098,7 @@ void TestSharedAppendSelectorBounds() {
     memory.words[word + 3u] = Libs::Graphics::DstSel(4, 5, 6, 7) |
         (static_cast<uint32_t>(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
   }
-  std::array<uint32_t, 9> user_data {0x1000u, 160u << 16u, 8u, 0u, 0x2000u, 0u, 0x1800u, 0u, 0u};
+  std::array<uint32_t, 9> user_data {0x1000u, 160u << 16u, 8u, 0u, 0x2000u, 0u, 0x1800u, 0u, 8u};
   SrtRuntime runtime {.user_data = user_data, .userdata = &memory,
       .read_specialization_memory = ReadLinearTestMemory};
   FakeGuestMemory fake(ReadRuntimeMemory, &runtime);
@@ -5022,10 +5113,11 @@ void TestSharedAppendSelectorBounds() {
     memory.words[0x800u / 4u] = counts[0];
     memory.words[0x804u / 4u] = counts[1];
     const auto expected = counts[1];
-    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
-              snapshot.images.size() == std::max(expected, 1u) &&
-              (expected == 0u || snapshot.images.back().dwords[0] == 0x100u + expected),
-          "disjoint shared append lists or refreshed counts selected unrelated material keys");
+    for (const auto *tested : {&plan, &lane_plan})
+      Check(MaterializeResources(*tested, runtime, snapshot, specialization) &&
+                snapshot.images.size() == std::max(expected, 1u) &&
+                (expected == 0u || snapshot.images.back().dwords[0] == 0x100u + expected),
+            "disjoint shared append lists or refreshed counts selected unrelated material keys");
   }
   memory.words[0x800u / 4u] = 24u;
   memory.words[0x804u / 4u] = 2u;
@@ -5047,9 +5139,14 @@ void TestSharedAppendSelectorBounds() {
   Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
             snapshot.images.size() == 2u && snapshot.images.back().dwords[0] == 0x104u,
         "shared append materialization reused stale selector contents");
+  // Record 1 lies past a one-record V#, so its scalar key reads zero instead of key 4.
   user_data[2] = 1u;
-  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
-        "shared append materialization accepted a range outside its V# bounds");
+  memory.fail_address = 0x1000u + 144u + 160u;
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            snapshot.images.size() == 2u && snapshot.images[0].dwords[0] == 0x100u &&
+            snapshot.images[1].dwords[0] == 0x101u,
+        "shared append materialization did not read records past its V# as zero");
+  memory.fail_address = UINT64_MAX;
   user_data[2] = 8u;
   for (const auto too_large : {65537u, UINT32_MAX}) {
     for (uint32_t list = 0; list < 2u; ++list) {

tests/SaveDataMemoryTests.cpp

@@ -660,6 +660,47 @@ void TestClassicSaveParams() {
 	CHECK(params[0].mtime == nested_mtime);
 }
 
+void TestSearchSortKeys() {
+	Reset("SORTED");
+	const std::array names {DirName("alpha"), DirName("bravo"), DirName("charlie")};
+	constexpr std::array<uint64_t, 3> blocks {96, 48, 48};
+	constexpr std::array<uint32_t, 3> user_params {2, 3, 1};
+	constexpr std::array<int, 3>      hours_ahead {1, 3, 2};
+	for (size_t i = 0; i < names.size(); i++) {
+		struct SaveDataMount3 mount {};
+		mount.user_id    = 1;
+		mount.dir_name   = &names[i];
+		mount.mount_mode = 4;
+		mount.blocks     = blocks[i];
+		SaveDataMountResult result {};
+		CHECK(SaveDataMount3(&mount, &result) == OK);
+		CHECK(SaveDataSetParam(&result.mount_point, 4, &user_params[i], sizeof(user_params[i])) ==
+		      OK);
+		CHECK(SaveDataUmount2(0, &result.mount_point) == OK);
+		fs::last_write_time(fs::path("_SaveData/SORTED") / names[i].data / "sce_sys/param.bin",
+		                    fs::file_time_type::clock::now() + std::chrono::hours(hours_ahead[i]));
+	}
+	const auto search = [](SaveDataSortKey key, SaveDataSortOrder order) {
+		std::array<SceSaveDataDirName, 3> found {};
+		SaveDataDirNameSearchCond         cond {};
+		cond.user_id = 1;
+		cond.key     = key;
+		cond.order   = order;
+		SaveDataDirNameSearchResult result {};
+		result.dir_names     = found.data();
+		result.dir_names_num = found.size();
+		CHECK(SaveDataDirNameSearch(&cond, &result) == OK && result.set_num == 3);
+		return std::string(found[0].data) + " " + found[1].data + " " + found[2].data;
+	};
+	CHECK(search(SaveDataSortKey::DirName, SaveDataSortOrder::Descent) == "charlie bravo alpha");
+	CHECK(search(SaveDataSortKey::UserParam, SaveDataSortOrder::Ascent) == "charlie alpha bravo");
+	CHECK(search(SaveDataSortKey::Blocks, SaveDataSortOrder::Ascent) == "bravo charlie alpha");
+	CHECK(search(SaveDataSortKey::FreeBlocks, SaveDataSortOrder::Descent) == "alpha charlie bravo");
+	CHECK(search(SaveDataSortKey::Mtime, SaveDataSortOrder::Descent) == "bravo charlie alpha");
+	CHECK(fs::remove("_SaveData/SORTED/alpha/sce_sys/blocks.bin"));
+	CHECK(search(SaveDataSortKey::Blocks, SaveDataSortOrder::Ascent) == "alpha bravo charlie");
+}
+
 void RunChild(const fs::path& executable, const char* mode) {
 #ifdef _WIN32
 	CHECK(_spawnl(_P_WAIT, executable.string().c_str(), executable.string().c_str(), mode,
@@ -706,6 +747,7 @@ int main(int argc, char** argv) {
 	TestGuestSaveIsolation();
 	TestSaveAllocations();
 	TestClassicSaveParams();
+	TestSearchSortKeys();
 	CHECK(SaveDataTerminate() == OK);
 	fs::current_path(previous);
 	fs::remove_all(temp);

tests/ScalarProvenanceTests.cpp

@@ -300,8 +300,12 @@ void TestOperandMutationLifecycle() {
   auto *source = definition.TryInstruction();
   const auto inserted = fixture.Emit(ValueOpcode::BitFieldInsert,
       {definition, definition, Value(4u), Value(4u)});
-  Check(source->UseCount() == 2, "inline operands lost distinct reverse-use indices");
-  inserted.TryInstruction()->SetArg(1, Value(3u));
+  auto *consumer = inserted.TryInstruction();
+  Check(source->Uses() == std::vector<Use>({{consumer, 0}, {consumer, 1}}),
+        "inline operands lost distinct reverse-use indices");
+  consumer->SetArg(0, definition);
+  Check(source->UseCount() == 2, "re-setting an operand to its value duplicated its use");
+  consumer->SetArg(1, Value(3u));
   Check(source->UseCount() == 1, "inline operand mutation did not detach its old use");
   source->ReplaceUsesWith(Value(0xaau));
   uint32_t result = 0;
@@ -349,6 +353,35 @@ void TestOperandMutationLifecycle() {
         "Phi replacement retained incoming uses or changed the consumer value");
 }
 
+void TestReplacementKeepsUseOrder() {
+  Fixture fixture(2);
+  const auto base = fixture.Emit(ValueOpcode::IAdd32, {Value(1u), Value(2u)});
+  auto *before = fixture.Emit(ValueOpcode::IAdd32, {base, Value(3u)}).TryInstruction();
+  auto *identity = fixture.Emit(ValueOpcode::Identity, {base}).TryInstruction();
+  auto *pair = fixture.Emit(ValueOpcode::IAdd32, {Value(identity), Value(identity)}).TryInstruction();
+  auto *after = fixture.Emit(ValueOpcode::IAdd32, {Value(4u), base}).TryInstruction();
+  auto &large = fixture.program.value_storage.emplace_back(ValueOpcode::GetImageResource);
+  const auto last = large.NumArgs() - 1;
+  large.SetArg(last, Value(identity));
+  identity->ReplaceUsesWith(base, false);
+  Check(!identity->HasUses() && identity->NumArgs() == 0 && pair->Arg(0) == base &&
+            pair->Arg(1) == base && large.Arg(last) == base &&
+            base.TryInstruction()->Uses() ==
+                std::vector<Use>({{before, 0}, {after, 1}, {pair, 0}, {pair, 1}, {&large, last}}),
+        "identity removal lost or reordered the uses of its argument");
+
+  auto &phi = fixture.BlockAt(1).AppendNewInst(ValueOpcode::Phi);
+  phi.SetFlags(Type::U32);
+  phi.AddPhiOperand(&fixture.BlockAt(0), Value(after));
+  phi.AddPhiOperand(&fixture.BlockAt(1), Value(&phi));
+  auto *user = fixture.Emit(ValueOpcode::IAdd32, {Value(&phi), Value(&phi)}, 0, 1).TryInstruction();
+  phi.ReplaceUsesWith(Value(after));
+  Check(phi.GetOpcode() == ValueOpcode::Identity && phi.Arg(0) == Value(after) && !phi.HasUses() &&
+            user->Arg(0) == Value(after) && user->Arg(1) == Value(after) &&
+            after->Uses() == std::vector<Use>({{user, 0}, {user, 1}, {&phi, 0}}),
+        "self-referencing Phi replacement kept a use of itself or reordered the moved uses");
+}
+
 void TestControlDependentStandaloneLoadStaysTyped() {
   Fixture fixture(3);
   const auto memory = fixture.AddMemory(ResourceKind::ScalarAddress);
@@ -515,6 +548,34 @@ void TestFloatComparisonDescriptorInputs() {
   }
 }
 
+void TestSignedConversionDescriptorInputs() {
+  // 3.75 and -3.75 truncate, the largest float below 2^31 converts exactly, 2^31 and +inf
+  // saturate up, -2^31 and -1e10 saturate down, and -0.0, a denormal and both NaNs read zero.
+  constexpr std::array<std::pair<uint32_t, uint32_t>, 11> inputs = {{
+      {0x40700000u, 3u},          {0xc0700000u, 0xfffffffdu}, {0x4effffffu, 0x7fffff80u},
+      {0x4f000000u, 0x7fffffffu}, {0x7f800000u, 0x7fffffffu}, {0xcf000000u, 0x80000000u},
+      {0xd01502f9u, 0x80000000u}, {0x80000000u, 0u},          {0x00000001u, 0u},
+      {0x7fc00000u, 0u},          {0xffc00000u, 0u}}};
+  Fixture fixture;
+  const auto user = fixture.Emit(
+      ValueOpcode::GetUserData, {Value(static_cast<ScalarReg>(2))});
+  const auto converted = fixture.Emit(
+      ValueOpcode::ConvertS32F32, {fixture.Emit(ValueOpcode::BitCastF32U32, {user})});
+  Check(ValidateRuntimeValue(fixture.program, converted),
+        "signed float conversion was rejected as a runtime value");
+  fixture.program.descriptor_sources.push_back(
+      {.dwords = {converted}, .dword_count = 1});
+
+  auto plan = ExtractResourcePlan(fixture.program);
+  for (const auto [bits, expected] : inputs) {
+    const std::array user_data{bits};
+    DescriptorValue result;
+    Check(EvaluateSource(plan, {.user_data = user_data}, 0, result) &&
+              result.dwords[0] == expected,
+          "descriptor signed conversion disagrees with the native conversion");
+  }
+}
+
 void TestSharedIntegerRuntimeDependencies() {
   Fixture fixture;
   const auto lane = fixture.Emit(ValueOpcode::LaneId);
@@ -884,10 +945,12 @@ int main() {
     TestCarryAndBitFields();
     TestInvariantAndDivergentPhi();
     TestOperandMutationLifecycle();
+    TestReplacementKeepsUseOrder();
     TestControlDependentStandaloneLoadStaysTyped();
     TestRuntime64BitDescriptorOps();
     TestUniformFirstLaneSamplerLod();
     TestFloatComparisonDescriptorInputs();
+    TestSignedConversionDescriptorInputs();
     TestSharedIntegerRuntimeDependencies();
     TestConstantBufferBounds();
     TestReadLaneElimination();

tests/SrtWalkerTestMemory.h

@@ -6,12 +6,16 @@
 #include <xbyak/xbyak.h>
 
 #include <csignal>
+#include <cstddef>
 #include <cstdlib>
+#include <iterator>
 #include <utility>
 #include <vector>
 
 #if defined(__linux__)
 #include <ucontext.h>
+#elif defined(__APPLE__)
+#include <sys/ucontext.h>
 #elif defined(_WIN32)
 #include <windows.h>
 #endif
@@ -23,18 +27,20 @@ public:
   explicit FakeGuestMemory(Libs::Graphics::ShaderRecompiler::IR::SrtMemoryReader reader = nullptr,
                            void *userdata = nullptr) {
     Current() = {reader, userdata};
-#if defined(__linux__)
+#if defined(__linux__) || defined(__APPLE__)
     struct sigaction action {};
     action.sa_sigaction = Handle;
     action.sa_flags = SA_SIGINFO;
-    sigaction(SIGSEGV, &action, &m_previous);
+    for (size_t index = 0; index < std::size(kSignals); ++index)
+      sigaction(kSignals[index], &action, &m_previous[index]);
 #elif defined(_WIN32)
     m_handler = AddVectoredExceptionHandler(1, Handle);
 #endif
   }
   ~FakeGuestMemory() {
-#if defined(__linux__)
-    sigaction(SIGSEGV, &m_previous, nullptr);
+#if defined(__linux__) || defined(__APPLE__)
+    for (size_t index = 0; index < std::size(kSignals); ++index)
+      sigaction(kSignals[index], &m_previous[index], nullptr);
 #elif defined(_WIN32)
     RemoveVectoredExceptionHandler(m_handler);
 #endif
@@ -54,16 +60,30 @@ private:
     const auto [reader, userdata] = Current();
     return reader != nullptr && reader(userdata, address, {&word, 1}) ? word : 0u;
   }
-#if defined(__linux__)
+#if defined(__linux__) || defined(__APPLE__)
+#if defined(__APPLE__)
+  // macOS raises SIGBUS for some protection faults and SIGSEGV for others.
+  static constexpr int kSignals[] = {SIGSEGV, SIGBUS};
+#else
+  static constexpr int kSignals[] = {SIGSEGV};
+#endif
   static void Handle(int, siginfo_t *, void *context) {
-    auto &registers = static_cast<ucontext_t *>(context)->uc_mcontext.gregs;
-    const auto length =
-        Libs::Graphics::ShaderRecompiler::IR::SrtWalkerLoadLength(registers[REG_RIP]);
+    auto *native = static_cast<ucontext_t *>(context);
+#if defined(__APPLE__)
+    auto &rip = native->uc_mcontext->__ss.__rip;
+    auto &rax = native->uc_mcontext->__ss.__rax;
+    const auto r10 = native->uc_mcontext->__ss.__r10;
+#else
+    auto &rip = native->uc_mcontext.gregs[REG_RIP];
+    auto &rax = native->uc_mcontext.gregs[REG_RAX];
+    const auto r10 = native->uc_mcontext.gregs[REG_R10];
+#endif
+    const auto length = Libs::Graphics::ShaderRecompiler::IR::SrtWalkerLoadLength(rip);
     if (length == 0) std::abort();
-    registers[REG_RAX] = Read(registers[REG_R10]);
-    registers[REG_RIP] += length;
+    rax = Read(r10);
+    rip += length;
   }
-  struct sigaction m_previous {};
+  struct sigaction m_previous[std::size(kSignals)] {};
 #elif defined(_WIN32)
   static LONG CALLBACK Handle(EXCEPTION_POINTERS *info) {
     auto &context = *info->ContextRecord;

tests/shaderCfgTests.cpp

@@ -1938,7 +1938,7 @@ void TestVertexBufferGrouping() {
   Check(input.fetch_attrib_reg == 2, "a draw reread the vertex metadata of a mapped shader");
   ShaderMapUserData(regs.es_regs.data_addr, mapped);
   (void)PrepareProgram(regs, HW::Context{}, HW::UserConfig{}, input);
-  Check(!input.fetch_embedded, "remapping a shader kept its previous vertex metadata");
+  Check(!input.has_dynamic_vertex_fetch, "remapping a shader kept its previous vertex metadata");
 }
 
 void TestNggVertexEntryState() {
@@ -2001,18 +2001,21 @@ void TestNggVertexEntryState() {
 
 void TestNativeDataCacheInvalidation() {
   using namespace ShaderRecompiler;
-  // Captured f5edf70dbb1f6057:0x194 and 0x1a4, each followed by S_WAITCNT.
+  // Captured f5edf70dbb1f6057:0x194 and 0x1a4, each followed by S_WAITCNT. The S_DCACHE_INV
+  // after them emits no fence.
   const std::array cache_ops = {0xe1c40000u, 0u, 0xbf8c0000u,
-                               0xe1c80000u, 0u, 0xbf8c0000u, EncodeSopp(0x01)};
+                               0xe1c80000u, 0u, 0xbf8c0000u,
+                               0xf4800000u, 0xfa000000u, EncodeSopp(0x01)};
   Decoder::Program decoded;
   Decoder::DecodeProgram(cache_ops, decoded);
-  for (const uint32_t index : {0u, 2u}) {
+  for (const auto [index, opcode] : {std::pair{0u, Decoder::Opcode::BUFFER_GL0_INV},
+                                     std::pair{2u, Decoder::Opcode::BUFFER_GL1_INV},
+                                     std::pair{4u, Decoder::Opcode::S_DCACHE_INV}}) {
     const auto &inst = decoded.instructions[index];
-    Check(inst.opcode == (index == 0 ? Decoder::Opcode::BUFFER_GL0_INV
-                                    : Decoder::Opcode::BUFFER_GL1_INV) &&
+    Check(inst.opcode == opcode &&
               inst.word_count == 2 && inst.src_count == 0 && inst.data_dwords == 0 &&
               inst.dst.kind == Decoder::OperandKind::Unknown,
-          "RDNA2 cache invalidation acquired a buffer operand or lost its encoding");
+          "RDNA2 cache invalidation acquired an operand or lost its encoding");
   }
 
   auto options = MakeCompileOptions(ShaderType::Compute);
@@ -4715,6 +4718,73 @@ void TestImageAtomicWidthDecoder() {
   }
 }
 
+void TestImageGatherDecoder() {
+  using namespace ShaderRecompiler::Decoder;
+  struct Gather {
+    uint32_t opcode;
+    const char *name;
+    uint32_t address_vgprs;
+  };
+  // Every PS5 image_gather4 encoding, with the vaddr VGPR count of its 2D full-float form.
+  constexpr Gather gathers[] = {
+      {0x40, "image_gather4", 2},           {0x41, "image_gather4_cl", 3},
+      {0x44, "image_gather4_l", 3},         {0x45, "image_gather4_b", 3},
+      {0x46, "image_gather4_b_cl", 4},      {0x47, "image_gather4_lz", 2},
+      {0x48, "image_gather4_c", 3},         {0x49, "image_gather4_c_cl", 4},
+      {0x4c, "image_gather4_c_l", 4},       {0x4d, "image_gather4_c_b", 4},
+      {0x4e, "image_gather4_c_b_cl", 5},    {0x4f, "image_gather4_c_lz", 3},
+      {0x50, "image_gather4_o", 3},         {0x51, "image_gather4_cl_o", 4},
+      {0x54, "image_gather4_l_o", 4},       {0x55, "image_gather4_b_o", 4},
+      {0x56, "image_gather4_b_cl_o", 5},    {0x57, "image_gather4_lz_o", 3},
+      {0x58, "image_gather4_c_o", 4},       {0x59, "image_gather4_c_cl_o", 5},
+      {0x5c, "image_gather4_c_l_o", 5},     {0x5d, "image_gather4_c_b_o", 5},
+      {0x5e, "image_gather4_c_b_cl_o", 6},  {0x5f, "image_gather4_c_lz_o", 4},
+      {0x61, "image_gather4h", 2},          {0xc0, "image_gather4_a", 2},
+      {0xc1, "image_gather4_cl_a", 3},      {0xc5, "image_gather4_b_a", 3},
+      {0xc6, "image_gather4_b_cl_a", 4},    {0xc8, "image_gather4_c_a", 3},
+      {0xc9, "image_gather4_c_cl_a", 4},    {0xcd, "image_gather4_c_b_a", 4},
+      {0xce, "image_gather4_c_b_cl_a", 5},  {0xd0, "image_gather4_o_a", 3},
+      {0xd1, "image_gather4_cl_o_a", 4},    {0xd5, "image_gather4_b_o_a", 4},
+      {0xd6, "image_gather4_b_cl_o_a", 5},  {0xd8, "image_gather4_c_o_a", 4},
+      {0xd9, "image_gather4_c_cl_o_a", 5},  {0xdd, "image_gather4_c_b_o_a", 5},
+      {0xde, "image_gather4_c_b_cl_o_a", 6},
+  };
+  const auto expected_flags = [](std::string_view name) {
+    if (name == "image_gather4h") return uint32_t{ImageSampleFlagGatherHorizontal};
+    uint32_t flags = 0;
+    for (size_t at = name.find('_', 6); at != std::string_view::npos;) {
+      const auto end = name.find('_', at + 1);
+      const auto modifier = name.substr(at + 1, end - at - 1);
+      if (modifier == "c") flags |= ImageSampleFlagCompare;
+      if (modifier == "b") flags |= ImageSampleFlagBias;
+      if (modifier == "cl") flags |= ImageSampleFlagLodClamp;
+      if (modifier == "l") flags |= ImageSampleFlagLod;
+      if (modifier == "lz") flags |= ImageSampleFlagLevelZero;
+      if (modifier == "o") flags |= ImageSampleFlagOffset;
+      if (modifier == "a") flags |= ImageSampleFlagAdjust;
+      at = end;
+    }
+    return flags;
+  };
+  for (uint32_t opcode = 0; opcode < 0x100u; ++opcode) {
+    const auto *gather = std::ranges::find(gathers, opcode, &Gather::opcode);
+    const uint32_t words[] = {EncodeMimg0(opcode, 0x1), EncodeMimg1(8, 0, 1, 0)};
+    Instruction decoded;
+    DecodeInstruction(words, 0, decoded);
+    if (gather == std::end(gathers)) {
+      Check(decoded.opcode != Opcode::IMAGE_GATHER4,
+            "a non-gather MIMG encoding decoded as image_gather4");
+      continue;
+    }
+    Check(decoded.opcode == Opcode::IMAGE_GATHER4 && decoded.data_components == 4u &&
+              decoded.image_sample_flags == expected_flags(gather->name) &&
+              decoded.image_address_components == gather->address_vgprs &&
+              InstructionToString(decoded).find(std::string(gather->name) + " ") !=
+                  std::string::npos,
+          "a PS5 gather encoding lost its opcode, name, modifiers or address count");
+  }
+}
+
 void TestNewShaderDecoderArchitecture() {
   using namespace ShaderRecompiler::Decoder;
 
@@ -5657,7 +5727,7 @@ void TestImageAddressOperands() {
   const uint32_t captured[] = {0xf130010au, 0x00a30013u, 0x00161a19u};
   Decoder::Instruction gather;
   Decoder::DecodeInstruction(captured, 0, gather);
-  Check(gather.opcode == Decoder::Opcode::IMAGE_GATHER4_C_L &&
+  Check(gather.opcode == Decoder::Opcode::IMAGE_GATHER4 &&
             gather.opcode_id == 0x4cu && gather.word_count == 3u &&
             gather.image_dimension == Decoder::ImageDimension::Dim2D &&
             gather.dmask == 1u && gather.data_components == 4u &&
@@ -5671,7 +5741,7 @@ void TestImageAddressOperands() {
             gather.image_nsa_addr[2] == 22u && !gather.glc && !gather.slc &&
             !gather.image_r128,
         "captured IMAGE_GATHER4_C_L lost compare/LOD, NSA, or result metadata");
-  Check(Decoder::InstructionToString(gather).find("IMAGE_GATHER4_C_L") != std::string::npos,
+  Check(Decoder::InstructionToString(gather).find("image_gather4_c_l") != std::string::npos,
         "captured IMAGE_GATHER4_C_L is missing from the decoded dump");
   struct Case {
     uint32_t opcode;
@@ -5687,6 +5757,9 @@ void TestImageAddressOperands() {
       Case{0x20, 1, true, false, 2, 1},  // packed xy
       Case{0x3d, 1, true, false, 5, 4},  // offset, half bias, full dref, packed xy
       Case{0x20, 2, true, true, 3, 2},   // packed xy/z in nonconsecutive VGPRs
+      Case{0x22, 4, true, false, 4, 3},  // a VGPR per 1D gradient, then packed x/slice
+      Case{0x22, 1, true, false, 6, 3},  // packed 2D gradients and xy
+      Case{0x22, 2, true, false, 9, 6},  // packed first four 3D gradients, a VGPR per last two
       Case{0x4c, 1, false, true, 4, 4},  // full dref, xy, and LOD in NSA VGPRs
       Case{0x0e, 1, false, false, 1, 1}, // resource query mip
   };
@@ -10739,6 +10812,8 @@ void TestCapturedBufferAtomicsX2() {
        "BufferAtomicIAdd64", "OpAtomicIAdd", ValueOpcode::BufferAtomicIAdd64},
       {{0xe1680018u, 0x80000000u}, 0u, 241u, "BUFFER_ATOMIC_OR_X2",
        "BufferAtomicOr64", "OpAtomicOr", ValueOpcode::BufferAtomicOr64},
+      {{0xe1582000u, 0x8003092au}, 9u, 237u, "BUFFER_ATOMIC_UMIN_X2",
+       "BufferAtomicUMin64", "OpAtomicUMin", ValueOpcode::BufferAtomicUMin64},
       {{0xe1402000u, 0x80000913u}, 9u, 229u, "BUFFER_ATOMIC_SWAP_X2",
        "BufferAtomicSwap64", "OpAtomicExchange", ValueOpcode::BufferAtomicSwap64},
   };
@@ -11500,7 +11575,7 @@ void TestEmbeddedFetchPreservesSharedScalarLoad() {
                              reinterpret_cast<uint64_t>(attributes.data())};
   std::memcpy(user_data.data() + 12, tables, sizeof(tables));
   ShaderVertexInputInfo input{};
-  input.fetch_embedded = true;
+  input.has_dynamic_vertex_fetch = true;
   input.fetch_buffer_reg = 12;
   input.fetch_attrib_reg = 14;
   input.resources_num = 1;
@@ -12593,6 +12668,71 @@ void TestFinalSsaRejectsRegisterStatePseudos() {
 }
 #endif
 
+void TestSsaRewriteSealsFilledBlocks() {
+  using namespace ShaderRecompiler::IR;
+  constexpr ScalarReg s0{0};
+
+  const auto add_blocks = [](Program &program) {
+    std::array<Block *, 4> blocks{};
+    for (auto &block : blocks) {
+      block =
+          program.block_storage.emplace_back(std::make_unique<Block>()).get();
+      program.blocks.push_back(block);
+    }
+    return blocks;
+  };
+  const auto use = [&](Block *block) {
+    const auto value = IREmitter(block).GetScalarReg(s0);
+    return &block->AppendNewInst(ValueOpcode::ReferenceU32, {value});
+  };
+  const auto is_phi = [](Value value, const Block *first, Value first_value,
+                         const Block *second, Value second_value) {
+    const auto *phi = value.TryInstruction();
+    return phi != nullptr && phi->GetOpcode() == ValueOpcode::Phi &&
+           phi->NumArgs() == 2u && phi->PhiBlock(0) == first &&
+           phi->Arg(0) == first_value && phi->PhiBlock(1) == second &&
+           phi->Arg(1) == second_value;
+  };
+
+  {
+    // entry(s0 = 7) -> left(s0 = 9) | right -> join.
+    Program program;
+    const auto [entry, left, right, join] = add_blocks(program);
+    entry->AddBranch(left);
+    entry->AddBranch(right);
+    left->AddBranch(join);
+    right->AddBranch(join);
+    IREmitter(entry).SetScalarReg(s0, U32(Value(7u)));
+    IREmitter(left).SetScalarReg(s0, U32(Value(9u)));
+    const auto *right_use = use(right);
+    const auto *join_use = use(join);
+    RewriteToSsa(program.blocks);
+    Check(right_use->Arg(0) == Value(7u) && right->Instructions().size() == 2u,
+          "SSA read in a single-predecessor block went through a trivial Phi");
+    Check(is_phi(join_use->Arg(0), left, Value(9u), right, Value(7u)),
+          "SSA join Phi lost a predecessor value");
+  }
+
+  // entry(s0 = 1) -> header <-> latch(s0 += 1), header -> tail.
+  Program program;
+  const auto [entry, header, latch, tail] = add_blocks(program);
+  entry->AddBranch(header);
+  header->AddBranch(latch);
+  header->AddBranch(tail);
+  latch->AddBranch(header);
+  IREmitter(entry).SetScalarReg(s0, U32(Value(1u)));
+  const auto *header_use = use(header);
+  IREmitter latch_ir(latch);
+  const auto sum = latch_ir.IAdd(latch_ir.GetScalarReg(s0), U32(Value(1u)));
+  latch_ir.SetScalarReg(s0, sum);
+  RewriteToSsa(program.blocks);
+  const auto phi = header_use->Arg(0);
+  Check(is_phi(phi, entry, Value(1u), latch, sum),
+        "SSA loop header Phi lost its entry or back-edge value");
+  Check(sum.Instruction()->Arg(0) == phi && latch->Instructions().size() == 3u,
+        "SSA read after a loop header went through a trivial Phi");
+}
+
 void TestValuePhiValidation() {
   using namespace ShaderRecompiler::IR;
 
@@ -15160,7 +15300,7 @@ void TestBoundedScalarMaterialImageKeys() {
       }
       return true;
     }
-    for (const uint32_t key : {5567u, 5582u, 5573u, 5588u}) {
+    for (const uint32_t key : {0u, 5567u, 5582u, 5573u, 5588u}) {
       if (address != 0x100000u + key * 32u || words.size() != 8) continue;
       ++m.images;
       std::ranges::fill(words, 0u);
@@ -15222,9 +15362,17 @@ void TestBoundedScalarMaterialImageKeys() {
     Check(!MaterializeResources(plan, runtime, snapshot, specialization),
           "unavailable selected material key was accepted");
   }
+  // Row 1's 0x94 key straddles a 311-byte V#, so it reads zero instead of 5573.
   memory = {.stride = 1, .records = 311};
-  Check(!MaterializeResources(plan, runtime, snapshot, specialization) && memory.key_reads == 0,
-        "scalar material byte range escaped V# bounds");
+  Check(MaterializeResources(plan, runtime, snapshot, specialization),
+        "scalar material table past its V# was rejected");
+  for (uint32_t root = 0; root < plan.info.images.size(); ++root) {
+    const auto &selector =
+        *plan.descriptor_sources[plan.info.images[root].source].indirect_descriptor->selector;
+    const auto offset = specialization.images[root].indirect_mapping_offset;
+    Check(snapshot.flattened_srt[offset + 1u] == (selector.offset == 0x94u ? 0u : 5567u),
+          "scalar material key past its V# did not read as zero");
+  }
   memory = {.mixed = true};
   Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
             snapshot.images.size() == 4 && specialization.images.size() == 4,
@@ -16130,6 +16278,153 @@ void TestCapturedBufferLoadUbyteD16Decode() {
         "captured BUFFER_LOAD_UBYTE_D16 lost its preserved half, address, or coherence flags");
 }
 
+void TestCapturedGlobalLoadShortD16HiDecode() {
+  namespace Decoder = ShaderRecompiler::Decoder;
+  // Ghost of Yotei CS 0x79b0066f0cd1f3cd: global_load_short_d16_hi v16, v18, s[0:1] offset:196.
+  const uint32_t code[] = {0xdc9480c4u, 0x10000012u};
+  Decoder::Instruction inst;
+  Decoder::DecodeInstruction(code, 0, inst);
+  Check(inst.opcode == Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI &&
+            inst.family == Decoder::Family::FLAT && inst.word_count == 2 &&
+            inst.data_dwords == 1 && inst.data_bits == 16 && !inst.data_signed &&
+            inst.memory_segment == 2 && inst.offset == 196 && !inst.glc && !inst.slc &&
+            !inst.dlc,
+        "captured GLOBAL_LOAD_SHORT_D16_HI lost its halfword payload or address fields");
+  Check(inst.dst.kind == Decoder::OperandKind::Vgpr && inst.dst.reg == 16 &&
+            inst.dst.sdwa_sel == 5 && inst.dst.sdwa_dst_unused == 2,
+        "captured GLOBAL_LOAD_SHORT_D16_HI must write only the high half of v16");
+  Check(inst.src0.kind == Decoder::OperandKind::Vgpr && inst.src0.reg == 18 &&
+            inst.src1.kind == Decoder::OperandKind::Sgpr && inst.src1.reg == 0,
+        "captured GLOBAL_LOAD_SHORT_D16_HI lost its VGPR offset or s[0:1] base");
+  Check(Decoder::InstructionToString(inst).find("FLAT_LOAD_SHORT_D16_HI") != std::string::npos,
+        "captured GLOBAL_LOAD_SHORT_D16_HI diagnostic text lost the opcode name");
+}
+
+void TestFlatLoadShortD16HiLlvmEncodingAnchor() {
+  namespace Decoder = ShaderRecompiler::Decoder;
+  // llvm/test/MC/AMDGPU/flat-gfx10.s: flat_load_short_d16_hi v1, v[3:4]
+  // encodes as [0x00,0x00,0x94,0xdc,0x03,0x00,0x7d,0x01].
+  const uint32_t code[] = {0xdc940000u, 0x017d0003u};
+  Decoder::Instruction inst;
+  Decoder::DecodeInstruction(code, 0, inst);
+  Check(inst.opcode == Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI && inst.word_count == 2 &&
+            inst.memory_segment == 0 && inst.offset == 0 && inst.data_dwords == 1 &&
+            inst.data_bits == 16 && !inst.data_signed,
+        "LLVM flat_load_short_d16_hi encoding did not decode as a FLAT halfword load");
+  Check(inst.dst.kind == Decoder::OperandKind::Vgpr && inst.dst.reg == 1 &&
+            inst.dst.sdwa_sel == 5 && inst.dst.sdwa_dst_unused == 2 &&
+            inst.src0.kind == Decoder::OperandKind::Vgpr && inst.src0.reg == 3 &&
+            inst.src1.kind == Decoder::OperandKind::Vgpr && inst.src1.reg == 4,
+        "LLVM flat_load_short_d16_hi v1, v[3:4] lost its destination half or address pair");
+}
+
+void TestFlatLoadShortD16HiCompilesAndValidates() {
+  // The captured high-half load followed by a buffer store of the merged register so that the
+  // load stays live: global_load_short_d16_hi v16, v18, s[0:1] offset:196.
+  const uint32_t shader[] = {
+      0xdc9480c4u, 0x10000012u,
+      EncodeMubuf0(0x1c, 0, false), EncodeMubuf1(16, 12, 0),
+      EncodeSopp(0x01),
+  };
+  auto options = MakeCompileOptions(ShaderType::Compute);
+  options.dump_ir = true;
+  auto result = RecompileForTest(shader, options);
+  Check(result.decoded_dump.find("FLAT_LOAD_SHORT_D16_HI") != std::string::npos &&
+            result.decoded_dump.find("unsupported family=") == std::string::npos,
+        "GLOBAL_LOAD_SHORT_D16_HI did not reach the decoded program");
+  Check(result.ir_dump.find("LoadAddressU16") != std::string::npos &&
+            result.ir_dump.find("BitFieldInsert") != std::string::npos,
+        "GLOBAL_LOAD_SHORT_D16_HI did not become a 16-bit load merged into the old register");
+  CheckSpirvBinaryValidates(result.spirv);
+}
+
+void TestFlatLoadShortD16HiAddressFields() {
+  namespace Decoder = ShaderRecompiler::Decoder;
+  // Modifier bits of the first FLAT-family word.
+  constexpr uint32_t DlcBit = 1u << 12u;
+  constexpr uint32_t LdsBit = 1u << 13u;
+  constexpr uint32_t GlcBit = 1u << 16u;
+  constexpr uint32_t SlcBit = 1u << 17u;
+  constexpr uint32_t SegFlat = 0;
+  constexpr uint32_t SegScratch = 1;
+  constexpr uint32_t SegGlobal = 2;
+  constexpr uint32_t SegReserved = 3;
+  constexpr uint32_t SaddrOff = 0x7du;
+  const auto decode = [](uint32_t seg, uint32_t offset_field, uint32_t saddr, uint32_t flags = 0) {
+    const uint32_t code[] = {EncodeFlat0(0x25, seg, offset_field) | flags,
+                             EncodeFlat1(7, saddr, 0, 9)};
+    Decoder::Instruction inst;
+    Decoder::DecodeInstruction(code, 0, inst);
+    return inst;
+  };
+
+  struct OffsetCase {
+    uint32_t seg;
+    uint32_t field;
+    int32_t expected;
+  };
+  const OffsetCase offsets[] = {
+      // FLAT ignores OFFSET[11] and uses an unsigned 11-bit byte offset.
+      {SegFlat, 0x000, 0},
+      {SegFlat, 0x7ff, 2047},
+      {SegFlat, 0x800, 0},
+      {SegFlat, 0xfff, 2047},
+      // GLOBAL and SCRATCH sign-extend the 12-bit field.
+      {SegGlobal, 0x000, 0},
+      {SegGlobal, 0x7ff, 2047},
+      {SegGlobal, 0x800, -2048},
+      {SegGlobal, 0xfff, -1},
+      {SegScratch, 0x000, 0},
+      {SegScratch, 0x7ff, 2047},
+      {SegScratch, 0x800, -2048},
+      {SegScratch, 0xfff, -1},
+  };
+  for (const auto &test : offsets) {
+    const auto inst = decode(test.seg, test.field, SaddrOff);
+    Check(inst.opcode == Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI &&
+              inst.memory_segment == test.seg &&
+              inst.offset == static_cast<uint32_t>(test.expected),
+          "FLAT-family D16 high-half load decoded the wrong byte offset for its segment");
+    Check(inst.dst.kind == Decoder::OperandKind::Vgpr && inst.dst.reg == 7 &&
+              inst.dst.sdwa_sel == 5 && inst.dst.sdwa_dst_unused == 2 &&
+              inst.src0.kind == Decoder::OperandKind::Vgpr && inst.src0.reg == 9,
+          "FLAT-family D16 high-half load lost its destination half or address VGPR");
+  }
+
+  // FLAT takes a VGPR pair whatever the SADDR field says; GLOBAL takes either an SGPR base plus
+  // a VGPR offset or, with SADDR off, a VGPR pair.
+  const auto flat = decode(SegFlat, 0, 2);
+  Check(flat.opcode == Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI &&
+            flat.src1.kind == Decoder::OperandKind::Vgpr && flat.src1.reg == 10,
+        "FLAT D16 high-half load must use the VGPR pair and ignore SADDR");
+  const auto global_sgpr = decode(SegGlobal, 0, 2);
+  Check(global_sgpr.opcode == Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI &&
+            global_sgpr.src_count == 2 && global_sgpr.src0.kind == Decoder::OperandKind::Vgpr &&
+            global_sgpr.src0.reg == 9 && global_sgpr.src1.kind == Decoder::OperandKind::Sgpr &&
+            global_sgpr.src1.reg == 2,
+        "GLOBAL D16 high-half load with SADDR s[2:3] lost its scalar base");
+  const auto global_pair = decode(SegGlobal, 0, SaddrOff);
+  Check(global_pair.opcode == Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI &&
+            global_pair.src_count == 2 && global_pair.src0.kind == Decoder::OperandKind::Vgpr &&
+            global_pair.src0.reg == 9 && global_pair.src1.kind == Decoder::OperandKind::Vgpr &&
+            global_pair.src1.reg == 10,
+        "GLOBAL D16 high-half load with SADDR off must use the VGPR pair");
+
+  // DLC is only a cache hint and stays visible; LDS, GLC and SLC on a load and the reserved
+  // segment are not implemented and must keep failing loudly.
+  const auto dlc = decode(SegGlobal, 4, 2, DlcBit);
+  Check(dlc.opcode == Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI && dlc.dlc && !dlc.glc && !dlc.slc,
+        "GLOBAL D16 high-half load lost its DLC hint");
+  Check(decode(SegGlobal, 0, 2, LdsBit).opcode == Decoder::Opcode::UNSUPPORTED,
+        "FLAT-family D16 high-half load accepted LDS=1");
+  Check(decode(SegGlobal, 0, 2, GlcBit).opcode == Decoder::Opcode::UNSUPPORTED,
+        "FLAT-family D16 high-half load accepted GLC=1");
+  Check(decode(SegGlobal, 0, 2, SlcBit).opcode == Decoder::Opcode::UNSUPPORTED,
+        "FLAT-family D16 high-half load accepted SLC=1 on a load");
+  Check(decode(SegReserved, 0, 2).opcode == Decoder::Opcode::UNSUPPORTED,
+        "FLAT-family D16 high-half load accepted the reserved segment");
+}
+
 void TestCapturedBufferStoreShortD16HiDecode() {
   namespace Decoder = ShaderRecompiler::Decoder;
   const uint32_t code[] = {0xe06c6000u, 0x80010326u};
@@ -16867,6 +17162,7 @@ int main() {
   // here.
   TestScalarAshrI64Decoder();
   TestImageAtomicWidthDecoder();
+  TestImageGatherDecoder();
   TestDsReturningAtomic64Decoder();
   TestDsCompareStore32Decoder();
   TestDsFloatMinMaxDecoder();
@@ -16966,6 +17262,7 @@ int main() {
 #if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
   TestFinalSsaRejectsRegisterStatePseudos();
 #endif
+  TestSsaRewriteSealsFilledBlocks();
   TestValuePhiValidation();
   TestLaneMaskProjection();
   TestLateConstantPropagation();
@@ -17017,6 +17314,10 @@ int main() {
   TestPixelProgramCacheBindingIdentity();
   TestGraphicsPushConstantPlacement();
   TestCapturedBufferLoadUbyteD16Decode();
+  TestCapturedGlobalLoadShortD16HiDecode();
+  TestFlatLoadShortD16HiLlvmEncodingAnchor();
+  TestFlatLoadShortD16HiAddressFields();
+  TestFlatLoadShortD16HiCompilesAndValidates();
   TestCapturedBufferStoreShortD16HiDecode();
   TestCapturedBufferStoreFormatD16HiXDecode();
   TestCapturedBufferLoadD16XyDecode();