launcher: Raise DualSense speaker volume range for SpeakerFX (#1043)

controller: restore speaker midpoint and raise volume ceiling

3rdparty/CMakeLists.txt

@@ -25,7 +25,7 @@ FetchContent_MakeAvailable(opus)
 set(BUILD_TESTING "${kyty_build_testing_saved}")
 install(FILES "${opus_SOURCE_DIR}/COPYING" DESTINATION "licenses/opus")
 
-# The Windows guest red-zone patcher needs a decoder/encoder and a small
+# 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)
 FetchContent_Declare(
@@ -48,6 +48,7 @@ FetchContent_Declare(
 )
 
 FetchContent_MakeAvailable(xbyak zydis)
+target_compile_definitions(xbyak INTERFACE XBYAK_NO_EXCEPTION)
 
 # ZArchive provides random-access reads for compressed game dumps. Build only
 # the reader/writer library and its zstd dependency; the command-line tool is

3rdparty/cpuinfo/src/arm/mach/init.c

@@ -14,7 +14,7 @@
 #include <cpuinfo/internal-api.h>
 #include <cpuinfo/log.h>
 
-/* Polyfill recent CPUFAMILY_ARM_* values for older SDKs */
+/* Polyfill recent CPUFAMILY_ARM_* values for older platform headers */
 #ifndef CPUFAMILY_ARM_MONSOON_MISTRAL
 	#define CPUFAMILY_ARM_MONSOON_MISTRAL   0xE81E7EF6
 #endif

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18...4.3)
 
 if(CMAKE_VERSION VERSION_GREATER_EQUAL 4.0)
-	set(CMAKE_POLICY_VERSION_MINIMUM 3.5 CACHE STRING "Minimum policy version for bundled dependencies")
+	set(CMAKE_POLICY_VERSION_MINIMUM 3.10 CACHE STRING "Minimum policy version for bundled dependencies")
 endif()
 
 get_property(KYTY_GENERATOR_IS_MULTI_CONFIG GLOBAL PROPERTY GENERATOR_IS_MULTI_CONFIG)

README.md

@@ -1,8 +1,8 @@
 # KytyPS5
 
-[![Build KytyPS5 (Windows)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28Windows%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
-[![Build KytyPS5 (Linux)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28Linux%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
-[![Build KytyPS5 (macOS)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28macOS%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (Windows)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28Windows%29&label=Build%20KytyPS5%20%28Windows%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (Linux)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28Linux%29&label=Build%20KytyPS5%20%28Linux%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (macOS)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28macOS%29&label=Build%20KytyPS5%20%28macOS%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
 [![Platform](https://img.shields.io/badge/platform-Windows%20x64%20%7C%20Linux%20x64%20%7C%20macOS%20x86__64-0078D4.svg)](#system-requirements)
 [![Status](https://img.shields.io/badge/status-active%20development-orange.svg)](#current-status)
 [![License](https://img.shields.io/badge/license-GPL--2.0-blue.svg)](LICENSE)
@@ -72,8 +72,8 @@ or graphical glitches, so please include the version you tested when reporting a
       <img src="docs/screenshots/ps5-02.png" width="300" alt="Demon's Souls running in KytyPS5">
     </td>
     <td align="center">
-      <strong>UFC 5</strong><br>
-      <img src="docs/screenshots/ps5-06.png" width="300" alt="UFC 5 running in KytyPS5">
+      <strong>UFC 6</strong><br>
+      <img src="docs/screenshots/ps5-06.jpg" width="300" alt="UFC 6 running in KytyPS5">
     </td>
   </tr>
 </table>

src/common/emulatorConfig.cpp

@@ -23,6 +23,7 @@ void Load(const ConfigOptions& cfg) {
 	EXIT_IF(g_config == nullptr);
 	EXIT_IF(cfg.user_name.empty() || cfg.user_name.size() > MAX_USER_NAME_LENGTH);
 	EXIT_IF(!IsConfiguredUserIdValid(cfg.user_id));
+	EXIT_IF(cfg.controller_speaker_volume > 100 || cfg.controller_vibration_intensity > 100);
 
 	*g_config = cfg;
 }
@@ -51,6 +52,14 @@ const std::optional<ControllerColor>& GetControllerColor() {
 	return g_config->controller_color;
 }
 
+uint32_t GetControllerSpeakerVolume() {
+	return g_config->controller_speaker_volume;
+}
+
+uint32_t GetControllerVibrationIntensity() {
+	return g_config->controller_vibration_intensity;
+}
+
 PresentMode GetPresentMode() {
 	return g_config->present_mode;
 }
@@ -63,6 +72,10 @@ bool FullscreenEnabled() {
 	return g_config->fullscreen_enabled;
 }
 
+bool HideCursorEnabled() {
+	return g_config->hide_cursor_enabled;
+}
+
 bool VrEnabled() {
 	return g_config->vr_enabled;
 }

src/common/emulatorConfig.h

@@ -48,9 +48,12 @@ struct ConfigOptions {
 	int32_t                user_id                     = DEFAULT_USER_ID;
 	std::string            audio_input_device;
 	std::optional<ControllerColor> controller_color;
+	uint32_t               controller_speaker_volume      = 50;
+	uint32_t               controller_vibration_intensity = 100;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int32_t                gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
+	bool                   hide_cursor_enabled         = false;
 	bool                   vr_enabled                  = false;
 	bool                   amd_cpu_enabled             = false;
 	uint32_t               vblank_frequency            = 60;
@@ -86,9 +89,12 @@ const std::string& GetUserName();
 int32_t  GetUserId();
 const std::string& GetAudioInputDevice();
 const std::optional<ControllerColor>& GetControllerColor();
+uint32_t GetControllerSpeakerVolume();
+uint32_t GetControllerVibrationIntensity();
 PresentMode GetPresentMode();
 int32_t GetGpuIndex();
 bool     FullscreenEnabled();
+bool     HideCursorEnabled();
 bool     VrEnabled();
 bool     AmdCpuEnabled();
 uint32_t GetVblankFrequency();

src/graphics/guest_gpu/gpu_format.cpp

@@ -61,8 +61,8 @@ constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k32_32_32_32UInt, 16, 0, 16, true, true},
 	{BufferFormat::k32_32_32_32SInt, 16, 0, 16, true, false, true},
 	{BufferFormat::k32_32_32_32Float, 16, 0, 16, true, false},
-	{BufferFormat::k8Srgb, 1, 0, 0, true, false},
-	{BufferFormat::k8_8Srgb, 2, 0, 0, true, false},
+	{BufferFormat::k8Srgb, 1, 0, 1, true, false},
+	{BufferFormat::k8_8Srgb, 2, 0, 2, true, false},
 	{BufferFormat::k8_8_8_8Srgb, 4, 0, 4, true, false},
 	{BufferFormat::k9_9_9_5Float, 4, 0, 0, true, false},
 	{BufferFormat::k5_6_5UNorm, 2, 0, 2, true, false},

src/graphics/host_gpu/graphicContext.h

@@ -29,6 +29,7 @@ struct GraphicContext {
 	bool                               memory_budget_ext_enabled             = false;
 	bool                               compute_subgroup_size_control_enabled = false;
 	bool                               sample_rate_shading_enabled           = false;
+	bool                               shader_image_int64_atomics_enabled    = false;
 	// bool fp64_denorm_preserve = false; // Temporarily disabled.
 	bool                               attachment_feedback_loop_enabled      = false;
 	bool                               provoking_vertex_last_enabled         = false;

src/graphics/host_gpu/regionManager.h

@@ -114,6 +114,9 @@ public:
 	void ForEachModifiedRange(uint64_t vaddr, uint64_t size, Func&& func) {
 		const auto [start, end] = GetPageRange(vaddr, size);
 		auto&      bits         = GetBits<source>();
+		if (bits.FirstRangeFrom(start).first >= end) {
+			return;
+		}
 		RegionBits mask(bits, start, end);
 		if constexpr (clear) {
 			bits.UnsetRange(start, end);

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

@@ -258,15 +258,8 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
 		}
 		auto& buffer = m_slot_buffers[FindBuffer(vaddr, size)];
 
-		// Widen nearby CPU reads so they share one GPU drain.
-		constexpr uint64_t WindowSize   = 512 * 1024;
-		const auto         buffer_begin = buffer.CpuAddress();
-		const auto         buffer_end   = buffer_begin + buffer.Size();
-		const auto window_begin = std::max(Common::AlignDown(vaddr, WindowSize), buffer_begin);
-		const auto window_end = std::min(std::max(window_begin + WindowSize, vaddr + size), buffer_end);
-
-		if (DownloadBufferMemory<false>(buffer, window_begin, window_end - window_begin)) {
-			m_memory_tracker.UnmarkRegionAsGpuModified(window_begin, window_end - window_begin);
+		if (DownloadBufferMemory<false>(buffer, vaddr, size)) {
+			m_memory_tracker.UnmarkRegionAsGpuModified(vaddr, size);
 		}
 		if (is_write) {
 			m_memory_tracker.MarkRegionAsCpuModified(vaddr, size);

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

@@ -791,13 +791,6 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		if (const auto depth_id = ResolveDepthOverlap(requested, binding, cached_id)) {
 			return {depth_id};
 		}
-		if (requested.IsBlock() && !cached.info.IsBlock()) {
-			return {ExpandImage(requested, cached_id)};
-		}
-		if (requested.data.size == cached.info.data.size &&
-		    (requested.IsVolume() || cached.info.IsVolume())) {
-			return {ExpandImage(requested, cached_id)};
-		}
 		// Equal pitch does not imply equal mip placement: a changed extent can move
 		// a level into or out of the mip tail. These are separate guest layouts.
 		if (requested.tile_mode != cached.info.tile_mode ||
@@ -808,6 +801,19 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 			}
 			return {merged_id};
 		}
+		if (requested.IsBlock() && !cached.info.IsBlock()) {
+			return {ExpandImage(requested, cached_id)};
+		}
+		// Volume depth is not an array-layer count. A larger depth can retain the
+		// same block-slice layout while requiring a larger native image.
+		if ((requested.IsVolume() || cached.info.IsVolume()) &&
+		    (requested.data.size == cached.info.data.size ||
+		     (requested.type == cached.info.type && requested.resources == cached.info.resources &&
+		      requested.extent.width == cached.info.extent.width &&
+		      requested.extent.height == cached.info.extent.height &&
+		      requested.extent.depth > cached.info.extent.depth))) {
+			return {ExpandImage(requested, cached_id)};
+		}
 		// PPSA08394
 		// A view cannot change the native image type or grow its extent.
 		if (requested.data.size == cached.info.data.size &&
@@ -878,9 +884,7 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 ImageId TextureCache::ExpandImage(const ImageInfo& info, ImageId source_id) {
 	RefreshCopySource(source_id);
 	const auto expanded_id = InsertImage(info);
-	auto&      expanded    = m_slot_images[expanded_id];
 	auto&      source      = m_slot_images[source_id];
-	expanded.usage         = source.usage;
 	if (source.binding.is_bound || source.binding.is_target) {
 		source.binding.needs_rebind = true;
 	}
@@ -926,22 +930,12 @@ TextureCache::BuildTextureTransfer(const Image& image, BindingType binding,
 	const bool  upload           = direction == TransferDirection::Upload;
 	const bool  render_target    = binding == BindingType::RenderTarget;
 	const bool  video_out        = binding == BindingType::VideoOut;
-	auto        format           = info.guest_format;
 	uint32_t    layers           = info.TransferLayers();
 	bool        volume           = info.IsVolume();
-	bool        allow_depth_tile = upload;
 	const char* owner            = "TextureCache readback";
 
 	TextureTransfer transfer;
-	transfer.swap_bgra16 = info.bgra16 && (!upload || render_target || video_out);
-	if (render_target) {
-		format = ImageOps::RenderTargetTransferFormat(info.bytes_per_block);
-	}
-	if (video_out) {
-		allow_depth_tile = false;
-	} else if (render_target || binding == BindingType::Storage) {
-		allow_depth_tile = true;
-	}
+	transfer.swap_bgra16 = info.bgra16;
 	if (upload) {
 		if ((render_target || video_out) &&
 		    (info.resources.layers == 0 || info.data.size % info.resources.layers != 0 ||
@@ -963,9 +957,9 @@ TextureCache::BuildTextureTransfer(const Image& image, BindingType binding,
 		}
 	}
 
-	transfer.layout  = TextureCalcUploadLayout(format, info.extent.width, info.extent.height,
-	                                       info.resources.levels, layers, info.tile_mode,
-	                                       info.data.size, allow_depth_tile, volume, owner);
+	transfer.layout = TextureCalcUploadLayout(info.guest_format, info.extent.width,
+	                                         info.extent.height, info.resources.levels, layers,
+	                                         info.tile_mode, info.data.size, volume, owner);
 	transfer.regions = TextureBuildImageCopies(transfer.layout);
 	if (info.IsDepth()) {
 		for (auto& region: transfer.regions) {
@@ -1043,6 +1037,8 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 	auto info = destination.info;
 	if (image.depth_id) {
 		info.data            = image.info.data;
+		info.resources       = {image.stencil_subresources.level_count,
+		                        image.stencil_subresources.layer_count};
 		info.guest_format    = Prospero::BufferFormat::k8UInt;
 		info.bytes_per_block = 1;
 		if (info.IsTiled()) info.pitch = TileGetDepthPitch(info.extent.width, 1, 0);
@@ -1057,6 +1053,11 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 	auto copies = BuildDepthCopies(info, full_slice_size, image.depth_id
 	                                                        ? vk::ImageAspectFlagBits::eStencil
 	                                                        : vk::ImageAspectFlagBits::eDepth);
+	if (image.depth_id) {
+		for (auto& copy: copies) {
+			copy.imageSubresource.baseArrayLayer += image.stencil_subresources.base_layer;
+		}
+	}
 	TileManager::Result linear {source.Handle(), source_offset, source.Size() - source_offset};
 	if (info.IsTiled()) {
 		const auto tiles = BuildDepthTiles(info);
@@ -1133,7 +1134,7 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 		image.info.metadata = desc.info.metadata;
 		// Native color metadata must not retain a reused HTile/CMask/FMask clear flag.
 		m_surface_metas.erase(range.address);
-		if (range.size == 0 || desc.info.resources.levels != 1 || image.info.resources.levels != 1) {
+		if (range.size == 0 || desc.info.resources.levels != 1) {
 			return;
 		}
 	}
@@ -1181,10 +1182,12 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 			           {vk::ImageAspectFlagBits::eColor, view.base_level, view.level_count,
 			            image_first + slice, 1}, clear);
 		}
-		// Native expanded keys own consumption. Existing buffer tracking publishes this CPU
-		// write to future GPU readers; FillBuffer can fault and must run outside the texture lock.
+		// Publish the conversion's expanded keys without treating them as guest writes
+		// to overlapping image data. Invalidate the buffer before updating its backing.
 		if (desc.type != BindingType::VideoOut) {
-			m_buffer_cache.FillBuffer(address, slice_size, UINT32_MAX, false);
+			std::fill(bytes.begin(), bytes.end(), uint8_t {0xff});
+			m_buffer_cache.InvalidateMemory(address, slice_size);
+			LibKernel::Memory::WriteBacking(address, bytes.data(), bytes.size());
 		}
 	}
 }
@@ -1243,7 +1246,8 @@ ImageId TextureCache::AssociateStencil(ImageId depth_id, GuestRange stencil) {
 	}
 	auto& record = m_slot_images[association];
 	TouchImage(record);
-	record.depth_id = depth_id;
+	record.depth_id             = depth_id;
+	record.stencil_subresources = depth.stencil_subresources;
 	return association;
 }
 
@@ -1449,6 +1453,18 @@ vk::ImageView TextureCache::FindDepthTarget(ImageId id, const ImageDesc& desc) {
 	TouchImage(image);
 	image.MarkGpuModified();
 	image.usage.depth_target = true;
+	if (desc.info.HasStencil()) {
+		const auto layers = image.info.resources.layers;
+		EXIT_IF(layers == 0 || image.info.data.size % layers != 0 ||
+		        desc.info.data.address < image.info.data.address);
+		const auto slice_size = image.info.data.size / layers;
+		const auto offset     = desc.info.data.address - image.info.data.address;
+		EXIT_IF(slice_size == 0 || offset % slice_size != 0 || offset / slice_size >= layers ||
+		        desc.info.resources.layers > layers - offset / slice_size);
+		image.stencil_subresources = {0, desc.info.resources.levels,
+		                             static_cast<uint32_t>(offset / slice_size),
+		                             desc.info.resources.layers};
+	}
 	image.info.stencil = desc.info.stencil;
 	image.info.metadata = desc.info.metadata;
 	if (desc.info.HasMetadata()) {
@@ -1495,9 +1511,11 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	if (command.IsInvalid() || !GuestRange {address, size}.Valid()) {
 		EXIT("TextureCache: invalid image clear\n");
 	}
-	std::scoped_lock     lock {m_lock};
-	ImageId              selected {};
-	vk::ImageAspectFlags aspect {};
+	std::scoped_lock      lock {m_lock};
+	ImageId               selected {};
+	ImageId               stencil_id {};
+	vk::ImageAspectFlags  aspect {};
+	ImageSubresourceRange subresources;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
 		auto owner = m_slot_images.try_get(id);
 		if (owner == nullptr) {
@@ -1505,10 +1523,13 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		}
 		vk::ImageAspectFlags candidate {};
 		ImageId              candidate_id = id;
+		ImageSubresourceRange candidate_subresources {
+		    0, owner->info.resources.levels, 0, owner->info.TransferLayers()};
 		if (owner->depth_id && owner->info.data.address == address &&
 		    owner->info.data.size == size) {
-			candidate    = vk::ImageAspectFlagBits::eStencil;
-			candidate_id = owner->depth_id;
+			candidate              = vk::ImageAspectFlagBits::eStencil;
+			candidate_id           = owner->depth_id;
+			candidate_subresources = owner->stencil_subresources;
 			owner        = m_slot_images.try_get(candidate_id);
 			if (owner == nullptr || owner->backing.image == nullptr || !owner->info.HasStencil()) {
 				continue;
@@ -1521,11 +1542,13 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		if (!candidate) {
 			continue;
 		}
-		if (selected && selected != candidate_id) {
+		if (selected && (selected != candidate_id || subresources != candidate_subresources)) {
 			return false;
 		}
-		selected = candidate_id;
-		aspect   = candidate;
+		selected     = candidate_id;
+		stencil_id   = candidate == vk::ImageAspectFlagBits::eStencil ? id : ImageId {};
+		aspect       = candidate;
+		subresources = candidate_subresources;
 	}
 	if (!selected) {
 		return false;
@@ -1547,7 +1570,12 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		clear.depthStencil.stencil = stencil_clear;
 	}
 	ClearImage(command, selected, image.backing.format,
-	           {aspect, 0, image.info.resources.levels, 0, image.info.TransferLayers()}, clear);
+	           {aspect, subresources.base_level, subresources.level_count,
+	            subresources.base_layer, subresources.layer_count}, clear);
+	if (stencil_id) {
+		TrackImage(stencil_id);
+		CommitGpuWrite(m_slot_images[stencil_id]);
+	}
 	return true;
 }
 
@@ -1850,7 +1878,7 @@ void TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size) {
 	std::scoped_lock lock {m_lock};
 	for (const auto id: FindImagesInRegion(address, size, true)) {
 		auto& image = m_slot_images[id];
-		if (!image.Overlaps(address, size)) {
+		if (image.info.data.address != address) {
 			continue;
 		}
 		if (image.IsGpuModified()) {
@@ -1860,23 +1888,6 @@ void TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size) {
 	}
 }
 
-bool TextureCache::IsRegionGpuModified(uint64_t address, uint64_t size) {
-	if (!GuestRange {address, size}.Valid()) {
-		return false;
-	}
-	std::scoped_lock lock {m_lock};
-	for (const auto id: FindImagesInRegion(address, size, false)) {
-		const auto& image = m_slot_images[id];
-		// PPSA17168: S_LOAD_DWORD reads shader data at an address overlapping an old
-		// render target whose memory the CPU has reused. The cached image still retains
-		// its earlier GPU-modified flag.
-		if (!image.depth_id && image.IsGpuModified() && !image.IsDefinitelyCpuDirty()) {
-			return true;
-		}
-	}
-	return false;
-}
-
 void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
 	const auto page_begin = Common::AlignDown(address, TRACKER_PAGE_SIZE);
 	const auto page_end   = Common::AlignUp(address + size, TRACKER_PAGE_SIZE);

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

@@ -61,7 +61,6 @@ public:
 	                                        uint32_t packed_clear);
 	void               InvalidateMemory(uint64_t address, uint64_t size);
 	void               InvalidateMemoryFromGPU(uint64_t address, uint64_t size);
-	[[nodiscard]] bool IsRegionGpuModified(uint64_t address, uint64_t size);
 
 	[[nodiscard]] bool IsMeta(uint64_t address);
 	[[nodiscard]] bool IsMetaCleared(uint64_t address, uint32_t slice);

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

@@ -340,7 +340,9 @@ void Image::CopyImage(Image& source) {
 	EXIT_IF(source.backing.samples != backing.samples);
 	m_scheduler.EndRendering();
 	const uint32_t levels     = std::min(source.backing.mip_levels, backing.mip_levels);
-	const uint32_t base_depth = backing.image_type == vk::ImageType::e3D
+	const uint32_t base_depth = source.backing.image_type == backing.image_type
+	                                ? std::min(source.backing.extent.depth, backing.extent.depth)
+	                            : backing.image_type == vk::ImageType::e3D
 	                                ? backing.extent.depth
 	                                : source.backing.extent.depth;
 	const auto     source_aspect =
@@ -671,7 +673,9 @@ Prospero::BufferFormat RenderTargetTransferFormat(uint32_t bytes_per_element) {
 } // namespace ImageOps
 
 Image::Image(GraphicContext& graphics, CommandScheduler& scheduler, const ImageInfo& image_info)
-    : info(image_info), m_graphics(graphics), m_scheduler(scheduler) {
+    : info(image_info),
+      stencil_subresources {0, image_info.resources.levels, 0, image_info.resources.layers},
+      m_graphics(graphics), m_scheduler(scheduler) {
 	KYTY_PROFILER_FUNCTION();
 	ImageOps::Validate(info);
 	m_cpu_dirty =

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

@@ -148,6 +148,8 @@ public:
 	uint64_t         track_addr     = 0;
 	uint64_t         track_addr_end = 0;
 	ImageId          depth_id {};
+	// The current stencil plane's mapping into the depth image, also retained by its association.
+	ImageSubresourceRange stencil_subresources;
 	uint64_t         tick_accessed_last = 0;
 	size_t           lru_id             = 0;
 

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

@@ -358,10 +358,8 @@ vk::ImageView Image::FindView(const ImageViewInfo& view_info) {
 	}
 
 	vk::ImageViewUsageCreateInfo usage {};
-	usage.usage = image.usage;
-	if (!is_storage) {
-		usage.usage &= ~vk::ImageUsageFlagBits::eStorage;
-	}
+	usage.usage = is_storage ? vk::ImageUsageFlagBits::eStorage
+	                         : image.usage & ~vk::ImageUsageFlagBits::eStorage;
 	vk::ImageViewMinLodCreateInfoEXT min_lod {};
 	if (normalized.min_lod != 0) {
 		min_lod.minLod = static_cast<float>(normalized.base_level) +

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

@@ -90,8 +90,11 @@ IsSupportedSampledDepthResource(const ShaderRecompiler::IR::ImageResource& resou
 	if (resource.numeric_class != Prospero::TextureNumericClass::Float && resource.depth_compare) {
 		return false;
 	}
-	return resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None && resource.read &&
-	       !resource.written && !resource.atomic;
+	return ((resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None &&
+	         resource.mip_count == 1u) ||
+	        (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::Dynamic &&
+	         resource.mip_count != 0u)) &&
+	       resource.read && !resource.written && !resource.atomic;
 }
 
 inline void ValidateStorageColorView(vk::Format image_format, vk::Format view_format,
@@ -114,7 +117,7 @@ IsSupportedStorageImageResource(const ShaderRecompiler::IR::ImageResource& resou
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DArray) &&
 	       ((resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None &&
 	         resource.mip_count == 1u) ||
-	        (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage &&
+	        (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::Dynamic &&
 	         resource.mip_count != 0u)) &&
 	       resource.written &&
 	       (!resource.atomic ||

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

@@ -175,8 +175,7 @@ SurfaceFormatInfo TextureGetSurfaceFormatInfo(Prospero::BufferFormat format) {
 TextureUploadLayout TextureCalcUploadLayout(Prospero::BufferFormat format, uint32_t width,
                                             uint32_t height, uint32_t levels, uint32_t depth,
                                             Prospero::TileMode tile_mode, uint64_t upload_size,
-                                            bool allow_depth_tile, bool volume_texture,
-                                            const char* owner) {
+                                            bool volume_texture, const char* owner) {
 	TextureUploadLayout layout {};
 	layout.surface.description = {
 	    format,
@@ -204,8 +203,7 @@ TextureUploadLayout TextureCalcUploadLayout(Prospero::BufferFormat format, uint3
 			     static_cast<uint32_t>(format));
 		}
 	} else {
-		if ((tile_mode == Prospero::TileMode::kDepth && !allow_depth_tile) ||
-		    !TileGetTiledTextureLayout(description, layout.surface)) {
+		if (!TileGetTiledTextureLayout(description, layout.surface)) {
 			EXIT("%s: unsupported typed tiled upload: fmt=%u tile=%u "
 			     "size=%" PRIu64 " extent=%ux%u levels=%u\n",
 			     owner, static_cast<uint32_t>(format), static_cast<uint32_t>(tile_mode),

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

@@ -47,8 +47,7 @@ RenderTargetFormatInfo TextureGetRenderTargetFormat(Prospero::ChannelLayout layo
 TextureUploadLayout    TextureCalcUploadLayout(Prospero::BufferFormat format, uint32_t width,
                                                uint32_t height, uint32_t levels, uint32_t depth,
                                                Prospero::TileMode tile_mode, uint64_t upload_size,
-                                               bool allow_depth_tile, bool volume_texture,
-                                               const char* owner);
+                                               bool volume_texture, const char* owner);
 std::vector<vk::BufferImageCopy> TextureBuildImageCopies(const TextureUploadLayout& layout);
 bool TextureBuildGpuTileInfos(uint64_t tiled_size, const std::vector<vk::BufferImageCopy>& regions,
                               const TextureUploadLayout& layout, uint32_t levels,

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

@@ -130,7 +130,7 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	const auto aligned_offset = Common::AlignDown(offset, alignment);
 	const auto adjustment     = offset - aligned_offset;
 	const auto max_range      = graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange;
-	if (adjustment % sizeof(uint32_t) != 0 || adjustment >= 256 || size > max_range - adjustment) {
+	if (adjustment >= 256 || size > max_range - adjustment) {
 		EXIT("storage buffer offset adjustment is unsupported\n");
 	}
 	buffer_offset = static_cast<uint32_t>(adjustment);
@@ -312,13 +312,14 @@ void ValidateStorageTexture(const ShaderRecompiler::IR::ImageResource& resource,
 	                              resource.written && !resource.read && !resource.atomic;
 	const auto numeric_class = Prospero::SampledTextureNumericClass(format);
 	const bool raw_float_atomic = format == Prospero::BufferFormat::k32Float && uint_resource &&
-	                              resource.atomic;
+	                              resource.atomic && !resource.atomic64;
 	const bool format_ok =
 	    raw_sint_storage || raw_float_atomic ||
 	    (numeric_class != Prospero::TextureNumericClass::Unsupported &&
 	     numeric_class != Prospero::TextureNumericClass::Sint &&
 	     uint_resource == (numeric_class == Prospero::TextureNumericClass::Uint) &&
-	     (!resource.atomic || format == Prospero::BufferFormat::k32UInt));
+	     (!resource.atomic || format == (resource.atomic64 ? Prospero::BufferFormat::k32_32UInt
+	                                                       : Prospero::BufferFormat::k32UInt)));
 	if (resource_ok && descriptor_ok && encoding_ok && format_ok && size != 0) {
 		return;
 	}
@@ -354,7 +355,8 @@ static TextureCache::ImageDesc NullTextureDesc(const ShaderRecompiler::IR::Image
 			desc.info.guest_format = Prospero::BufferFormat::k32Float;
 			break;
 		case Prospero::TextureNumericClass::Uint:
-			desc.info.guest_format = Prospero::BufferFormat::k32UInt;
+			desc.info.guest_format = resource.atomic64 ? Prospero::BufferFormat::k32_32UInt
+			                                         : Prospero::BufferFormat::k32UInt;
 			break;
 		case Prospero::TextureNumericClass::Sint:
 			desc.info.guest_format = Prospero::BufferFormat::k32SInt;
@@ -365,10 +367,10 @@ static TextureCache::ImageDesc NullTextureDesc(const ShaderRecompiler::IR::Image
 	desc.info.type            = Prospero::ImageType::kColor2D;
 	desc.info.extent          = {1, 1, 1};
 	desc.info.resources       = {1, 1};
-	desc.info.bytes_per_block = 4;
+	desc.info.bytes_per_block = Prospero::NumBytesPerElement(desc.info.guest_format);
 	desc.info.samples         = 1;
 	desc.info.mip_layout[0]   = {0, 0, 1, 1};
-	desc.view_info.format     = desc.info.pixel_format;
+	desc.view_info.format     = resource.atomic64 ? vk::Format::eR64Uint : desc.info.pixel_format;
 	desc.view_info.type       = vk::ImageViewType::e2D;
 	desc.view_info.aspect     = vk::ImageAspectFlagBits::eColor;
 	desc.view_info.usage      = binding == TextureCache::BindingType::Storage
@@ -538,6 +540,9 @@ static bool ResolveTextureMipView(const TileSurfaceDescription& description, boo
 
 TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   resource,
                                               const ShaderRecompiler::IR::DescriptorValue& value) {
+	if (resource.atomic64 && !m_context.GetGraphics().shader_image_int64_atomics_enabled) {
+		EXIT("64-bit image atomics require shaderImageInt64Atomics\n");
+	}
 	auto descriptor = DecodeNativeDescriptor<ShaderTextureResource>(value);
 	const bool storage = resource.written;
 	if (storage) {
@@ -563,7 +568,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	const auto physical_levels = multisampled ? 1u : static_cast<uint32_t>(max_mip) + 1u;
 	// IMAGE_STORE addresses BASE_LEVEL; only IMAGE_STORE_MIP selects other view mips.
 	const bool single_storage_mip =
-	    storage && resource.mip_mode != ShaderRecompiler::IR::ImageMipMode::DynamicStorage;
+	    storage && resource.mip_mode != ShaderRecompiler::IR::ImageMipMode::Dynamic;
 	const auto view_levels = multisampled || single_storage_mip
 	                             ? 1u
 	                             : static_cast<uint32_t>(last_level - base_level) + 1u;
@@ -650,8 +655,9 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 			pixel_format = depth_format->depth_attachment_format;
 		}
 	}
-	const auto storage_view_format = storage && (resource.atomic ||
-	                                            format == Prospero::BufferFormat::k32SInt)
+	const auto storage_view_format = resource.atomic64 ? vk::Format::eR64Uint
+	                                 : storage && (resource.atomic ||
+	                                               format == Prospero::BufferFormat::k32SInt)
 	                                     ? vk::Format::eR32Uint
 	                                     : SrgbStorageViewFormat(pixel_format);
 	const auto view_format         = storage && storage_view_format != vk::Format::eUndefined
@@ -773,8 +779,6 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
 	prepared.gds = {nullptr, 0, VK_WHOLE_SIZE};
 	prepared.flattened_srt = {};
 	prepared.shader_data_buffer = {};
-	prepared.buffer_sources.clear();
-	prepared.buffers.clear();
 	prepared.images.resize(program.info.images.size());
 	prepared.samplers.clear();
 	prepared.shader_data.clear();
@@ -787,7 +791,8 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
 	}
 	prepared.samplers.reserve(program.info.samplers.size());
 	for (uint32_t i = 0; i < program.info.samplers.size(); i++) {
-		prepared.samplers.push_back(NativeSampler(m_context, program, i, snapshot.samplers[i]));
+		prepared.samplers.push_back(NativeSampler(
+		    m_context, program, i, snapshot.samplers[program.info.samplers[i].snapshot_index]));
 	}
 	prepared.shader_data.reserve(program.bindings.ShaderDataDwords());
 	for (const auto reg: program.bindings.user_data_registers) {
@@ -887,7 +892,7 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 		auto& binding = images[i];
 		binding.mip_views.clear();
 		const auto& resource = program.info.images[i];
-		if (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage) {
+		if (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::Dynamic) {
 			EXIT_IF(resource.mip_count == 0u ||
 			        resource.mip_count != binding.desc.view_info.level_count);
 			binding.mip_views.reserve(resource.mip_count);
@@ -895,6 +900,8 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 				auto desc = binding.desc;
 				desc.view_info.base_level += mip;
 				desc.view_info.level_count = 1;
+				// The shader selects the mip after applying the guest minimum LOD.
+				desc.view_info.min_lod = 0;
 				binding.mip_views.push_back(texture_cache.FindTexture(binding.image_id, desc));
 			}
 			binding.image_view = binding.mip_views.front();

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

@@ -94,8 +94,10 @@ void PipelineCacheLog(fmt::format_string<Args...> format, Args&&... args) {
 }
 
 bool ReadShaderGuestMemory(void*, uint64_t address, std::span<uint32_t> values) {
+	// Scalar and unformatted buffer dependencies use the same backing as native raw loads.
+	// Image synchronization belongs to formatted buffer bindings, not these reads.
 	return !values.empty() &&
-	       Libs::LibKernel::Memory::TryReadGpuCleanBacking(address, values.data(), values.size_bytes());
+	       Libs::LibKernel::Memory::TryReadBufferBacking(address, values.data(), values.size_bytes());
 }
 
 void DumpShaderSpirv(const char* stage_name, uint64_t shader_hash,
@@ -658,6 +660,19 @@ ShaderProgram PipelineCache::GetComputeProgram(const HW::ComputeShaderInfo& regs
                                                ShaderComputeInputInfo&      input_info) {
 	input_info.host_subgroup_size = m_graphics.SupportsComputeWave64() ? 64u : 32u;
 	const auto        params      = PrepareProgram(regs, sh, input_info);
+	// Use one effective size for the cache key, LDS declaration, and access bounds.
+	const auto max_lds_dwords =
+	    m_graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize / 4u;
+	if (input_info.lds_size_dwords > max_lds_dwords) {
+		static std::atomic_bool warned = false;
+		if (!warned.exchange(true, std::memory_order_relaxed)) {
+			PipelineCacheLog("GPU warning: game compute shader requests {} bytes of LDS, but "
+			                 "the Vulkan device limit is {} bytes. Clamping LDS; rendering may "
+			                 "be incorrect.",
+			                 input_info.lds_size_dwords * 4u, max_lds_dwords * 4u);
+		}
+	}
+	input_info.lds_size_dwords = std::min(input_info.lds_size_dwords, max_lds_dwords);
 	uint32_t          push_data_cursor = 0;
 	return m_program_cache->Get(params, input_info, push_data_cursor);
 }

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -272,22 +272,79 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		return;
 	}
 
-	const bool large_workgroup =
-	    (input_info.threads_num[0] * input_info.threads_num[1] * input_info.threads_num[2] >= 512);
-	const bool                   has_sampler = !program.info.samplers.empty();
-	static std::atomic<uint32_t> dispatch_log_count {0};
-	if ((large_workgroup || has_sampler) &&
-	    dispatch_log_count.fetch_add(1, std::memory_order_relaxed) < 512) {
+	if (use_thread_dimensions) {
+		auto groups_from_threads = [](uint32_t threads, uint32_t group_size) {
+			return (threads == 0
+			            ? 0u
+			            : (threads + std::max(group_size, 1u) - 1u) / std::max(group_size, 1u));
+		};
+
+		const uint32_t old_x = thread_group_x;
+		const uint32_t old_y = thread_group_y;
+		const uint32_t old_z = thread_group_z;
+		thread_group_x       = groups_from_threads(thread_group_x, cs_regs.cs_regs.num_thread_x);
+		thread_group_y       = groups_from_threads(thread_group_y, cs_regs.cs_regs.num_thread_y);
+		thread_group_z       = groups_from_threads(thread_group_z, cs_regs.cs_regs.num_thread_z);
+
+		static std::atomic<uint32_t> log_count {0};
+		if (log_count.fetch_add(1, std::memory_order_relaxed) < 32) {
+			LOGF("GraphicsRenderDispatchDirect: use-thread-dimensions %ux%ux%u / %ux%ux%u -> "
+			     "groups %ux%ux%u\n",
+			     old_x, old_y, old_z, std::max(cs_regs.cs_regs.num_thread_x, 1u),
+			     std::max(cs_regs.cs_regs.num_thread_y, 1u),
+			     std::max(cs_regs.cs_regs.num_thread_z, 1u), thread_group_x, thread_group_y,
+			     thread_group_z);
+		}
+	}
+
+	buffer.EndRendering();
+	auto& pipeline =
+	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& bindings = m_compute_bindings;
+	PrepareBindings(input_info.stage, bindings);
+	if (program.bindings.dispatch_thread_dword != ShaderRecompiler::IR::PushData::NoStart) {
+		std::copy(std::begin(input_info.dispatch_threads_num), std::end(input_info.dispatch_threads_num),
+		          bindings.shader_data.begin() + program.bindings.dispatch_thread_dword);
+	}
+	FindBuffers(bindings);
+	if (program.info.uses_dma) {
+		m_context.PrepareBda();
+	}
+	RebindImages(bindings);
+	RebindBuffers(bindings);
+
+	auto              vk_buffer        = buffer.Handle();
+	PreparedBindings* descriptor_stage = &bindings;
+	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
+	               std::span {&descriptor_stage, 1u});
+	bool has_storage_writes = HasShaderBufferWrites(input_info.stage);
+	has_storage_writes =
+	    std::any_of(program.info.images.begin(), program.info.images.end(),
+	                [](const auto& image) {
+		                return image.written &&
+		                       image.resource_class ==
+		                           ShaderRecompiler::IR::ImageResourceClass::Storage;
+	                }) ||
+	    has_storage_writes;
+	if (has_storage_writes) {
+		// A host fence used to serialize every dispatch. Preserve its read-before-write ordering
+		// while allowing the queue to execute asynchronously.
+		ShaderWriteHazardBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
+	}
+	if (Config::GraphicsDebugDumpEnabled()) {
 		const auto sampled_images = std::count_if(
 		    program.info.images.begin(), program.info.images.end(), [](const auto& image) {
 			    return image.resource_class == ShaderRecompiler::IR::ImageResourceClass::Sampled;
 		    });
 		const uint32_t frame_num = static_cast<uint32_t>(m_context.GetGpu().GetFrameNum());
 		LOGF("GraphicsRenderDispatchDirect: frame=%u shader=0x%016" PRIx64
+		     " hash=0x%016" PRIx64 " tick=%" PRIu64
 		     " groups=%ux%ux%u mode=0x%08" PRIx32 " local=%ux%ux%u "
 		     "buffers=%zu textures=%zu sampled=%zu storage=%zu samplers=%zu push=%u\n",
-		     frame_num, sh_ctx.GetCs().cs_regs.data_addr, thread_group_x, thread_group_y,
-		     thread_group_z, mode, input_info.threads_num[0], input_info.threads_num[1],
+		     frame_num, sh_ctx.GetCs().cs_regs.data_addr, program.shader_hash,
+		     m_context.GetCommandScheduler().CurrentTick(),
+		     thread_group_x, thread_group_y, thread_group_z, mode,
+		     input_info.threads_num[0], input_info.threads_num[1],
 		     input_info.threads_num[2], program.info.buffers.size(), program.info.images.size(),
 		     sampled_images, program.info.images.size() - sampled_images,
 		     program.info.samplers.size(),
@@ -321,7 +378,8 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 			     static_cast<uint32_t>(r.TileMode()));
 		}
 		for (uint32_t i = 0; i < program.info.samplers.size(); i++) {
-			const auto r = DecodeNativeDescriptor<ShaderSamplerResource>(resources.samplers[i]);
+			const auto r = DecodeNativeDescriptor<ShaderSamplerResource>(
+			    resources.samplers[program.info.samplers[i].snapshot_index]);
 			LOGF("  CS sampler[%u]: source=%u clamp=%u/%u/%u filter=%u/%u/%u mip=%u "
 			     "lod=%u-%u bias=%d\n",
 			     i, program.info.samplers[i].source, static_cast<uint32_t>(r.ClampX()),
@@ -333,61 +391,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		}
 	}
 
-	if (use_thread_dimensions) {
-		auto groups_from_threads = [](uint32_t threads, uint32_t group_size) {
-			return (threads == 0
-			            ? 0u
-			            : (threads + std::max(group_size, 1u) - 1u) / std::max(group_size, 1u));
-		};
-
-		const uint32_t old_x = thread_group_x;
-		const uint32_t old_y = thread_group_y;
-		const uint32_t old_z = thread_group_z;
-		thread_group_x       = groups_from_threads(thread_group_x, cs_regs.cs_regs.num_thread_x);
-		thread_group_y       = groups_from_threads(thread_group_y, cs_regs.cs_regs.num_thread_y);
-		thread_group_z       = groups_from_threads(thread_group_z, cs_regs.cs_regs.num_thread_z);
-
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1, std::memory_order_relaxed) < 32) {
-			LOGF("GraphicsRenderDispatchDirect: use-thread-dimensions %ux%ux%u / %ux%ux%u -> "
-			     "groups %ux%ux%u\n",
-			     old_x, old_y, old_z, std::max(cs_regs.cs_regs.num_thread_x, 1u),
-			     std::max(cs_regs.cs_regs.num_thread_y, 1u),
-			     std::max(cs_regs.cs_regs.num_thread_z, 1u), thread_group_x, thread_group_y,
-			     thread_group_z);
-		}
-	}
-
-	buffer.EndRendering();
-	auto& pipeline =
-	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
-	auto& bindings = m_compute_bindings;
-	PrepareBindings(input_info.stage, bindings);
-	FindBuffers(bindings);
-	if (program.info.uses_dma) {
-		m_context.PrepareBda();
-	}
-	RebindImages(bindings);
-	RebindBuffers(bindings);
-
-	auto              vk_buffer        = buffer.Handle();
-	PreparedBindings* descriptor_stage = &bindings;
-	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
-	               std::span {&descriptor_stage, 1u});
-	bool has_storage_writes = HasShaderBufferWrites(input_info.stage);
-	has_storage_writes =
-	    std::any_of(program.info.images.begin(), program.info.images.end(),
-	                [](const auto& image) {
-		                return image.written &&
-		                       image.resource_class ==
-		                           ShaderRecompiler::IR::ImageResourceClass::Storage;
-	                }) ||
-	    has_storage_writes;
-	if (has_storage_writes) {
-		// A host fence used to serialize every dispatch. Preserve its read-before-write ordering
-		// while allowing the queue to execute asynchronously.
-		ShaderWriteHazardBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
-	}
 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
 	vk_buffer.dispatch(thread_group_x, thread_group_y, thread_group_z);
 

src/graphics/presentation/systemOverlay.cpp

@@ -29,9 +29,9 @@ namespace Libs::Graphics {
 
 namespace {
 
-namespace CoreIme   = Libs::Ime;
-namespace DialogIme = Libs::Dialog::ImeDialog;
-namespace ErrorDialog = Libs::Dialog::ErrorDialog;
+namespace CoreIme      = Libs::Ime;
+namespace DialogIme    = Libs::Dialog::ImeDialog;
+namespace SystemDialog = Libs::Dialog::SystemDialog;
 
 namespace Ime {
 
@@ -104,7 +104,7 @@ bool HostQueueExternalInput(uint64_t generation, ExternalInput input) {
 
 } // namespace Ime
 
-enum class OverlayKind : uint8_t { None, Ime, Error };
+enum class OverlayKind : uint8_t { None, Ime, Dialog };
 
 struct OverlaySession {
 	OverlayKind kind       = OverlayKind::None;
@@ -114,14 +114,14 @@ struct OverlaySession {
 };
 
 struct OverlaySnapshot {
-	OverlaySession            session;
-	Ime::HostSnapshot         ime;
-	ErrorDialog::HostSnapshot error;
+	OverlaySession             session;
+	Ime::HostSnapshot          ime;
+	SystemDialog::HostSnapshot dialog;
 };
 
 bool GetOverlaySnapshot(OverlaySnapshot* snapshot) {
-	if (ErrorDialog::GetHostSnapshot(&snapshot->error)) {
-		snapshot->session = {OverlayKind::Error, snapshot->error.generation};
+	if (SystemDialog::GetHostSnapshot(&snapshot->dialog)) {
+		snapshot->session = {OverlayKind::Dialog, snapshot->dialog.generation};
 		return true;
 	}
 	if (Ime::GetHostSnapshot(&snapshot->ime)) {
@@ -234,7 +234,7 @@ void RefreshVisibility() {
 	bool capture_keyboard   = false;
 	bool text_input         = false;
 	bool multiline          = false;
-	if (snapshot.session.kind == OverlayKind::Error) {
+	if (snapshot.session.kind == OverlayKind::Dialog) {
 		capture_controller = true;
 		capture_keyboard   = true;
 	} else if (snapshot.session.kind == OverlayKind::Ime) {
@@ -429,14 +429,14 @@ void InitializeSystemOverlayInput(SDL_Window* window) {
 	}
 	CoreIme::SetVisibilityCallback(OnCoreVisibilityChanged);
 	DialogIme::SetVisibilityCallback(OnDialogVisibilityChanged);
-	ErrorDialog::SetVisibilityCallback(RefreshVisibility);
+	SystemDialog::SetVisibilityCallback(RefreshVisibility);
 	RefreshVisibility();
 }
 
 void ShutdownSystemOverlayInput() {
 	CoreIme::SetVisibilityCallback(nullptr);
 	DialogIme::SetVisibilityCallback(nullptr);
-	ErrorDialog::SetVisibilityCallback(nullptr);
+	SystemDialog::SetVisibilityCallback(nullptr);
 	{
 		std::scoped_lock lock(g_visibility_mutex);
 		g_input_lifecycle_active = false;
@@ -466,9 +466,9 @@ void ShutdownSystemOverlayInput() {
 SystemOverlayVisualState GetSystemOverlayVisualState() noexcept {
 	const auto core   = CoreIme::GetVisualState();
 	const auto dialog = DialogIme::GetVisualState();
-	const auto error  = ErrorDialog::GetVisualState();
-	return {core.active || dialog.active || error.active,
-	        core.revision + dialog.revision + error.revision};
+	const auto system = SystemDialog::GetVisualState();
+	return {core.active || dialog.active || system.active,
+	        core.revision + dialog.revision + system.revision};
 }
 
 bool ProcessSystemOverlayInput(const SDL_Event& event) {
@@ -521,10 +521,11 @@ bool ProcessSystemOverlayInput(const SDL_Event& event) {
 	if (keyboard_event && !g_input_keyboard) {
 		return false;
 	}
-	if (keyboard_event && session.kind == OverlayKind::Error) {
+	if (keyboard_event && session.kind == OverlayKind::Dialog) {
 		if (event.type == SDL_EVENT_KEY_DOWN && !event.key.repeat &&
-		    (event.key.key == SDLK_RETURN || event.key.key == SDLK_KP_ENTER)) {
-			ErrorDialog::HostAccept(generation);
+		    (event.key.key == SDLK_RETURN || event.key.key == SDLK_KP_ENTER ||
+		     event.key.key == SDLK_ESCAPE)) {
+			SystemDialog::HostClose(generation);
 		}
 		return true;
 	}
@@ -904,7 +905,7 @@ struct SystemOverlay::Impl {
 		ImGui::End();
 	}
 
-	void DrawError(const ErrorDialog::HostSnapshot& snapshot, vk::Extent2D extent) {
+	void DrawDialog(const SystemDialog::HostSnapshot& snapshot, vk::Extent2D extent) {
 		const ImVec2 display(static_cast<float>(extent.width), static_cast<float>(extent.height));
 		const float  scale = std::max(std::min(display.x / 1280.0f, display.y / 720.0f), 0.5f);
 		ImGui::GetBackgroundDrawList()->AddRectFilled({0.0f, 0.0f}, display,
@@ -922,17 +923,23 @@ struct SystemOverlay::Impl {
 		constexpr ImGuiWindowFlags flags = ImGuiWindowFlags_NoDecoration | ImGuiWindowFlags_NoMove |
 		                                   ImGuiWindowFlags_NoSavedSettings |
 		                                   ImGuiWindowFlags_AlwaysAutoResize;
-		ImGui::Begin("##SystemError", nullptr, flags);
-		ImGui::TextUnformatted("Error");
+		const bool                 signin = snapshot.kind == SystemDialog::Kind::Signin;
+		ImGui::Begin("##SystemDialog", nullptr, flags);
+		ImGui::TextUnformatted(signin ? "Sign in to PlayStation Network" : "Error");
 		ImGui::Separator();
 		const auto error_code = static_cast<uint32_t>(snapshot.error_code);
-		ImGui::TextWrapped("%s", error_code == 0x80550006u
+		ImGui::TextWrapped("%s", signin ? "PlayStation Network sign-in is unavailable."
+		                         : error_code == 0x80550006u
 		                             ? "You are not signed in to PlayStation Network."
 		                             : "An error has occurred.");
-		ImGui::TextDisabled("Error code: 0x%08X", error_code);
+		if (!signin) {
+			ImGui::TextDisabled("Error code: 0x%08X", error_code);
+		}
 		const float button_width = std::min(140.0f * scale, ImGui::GetContentRegionAvail().x);
 		ImGui::SetCursorPosX((ImGui::GetWindowSize().x - button_width) * 0.5f);
-		const bool accepted = ImGui::Button("OK", {button_width, 44.0f * scale});
+		const bool accepted =
+		    ImGui::Button(signin ? "Cancel" : "OK", {button_width, 44.0f * scale}) ||
+		    (signin && ImGui::IsKeyPressed(ImGuiKey_GamepadFaceRight, false));
 		if (focus_pending) {
 			ImGui::SetItemDefaultFocus();
 			focus_pending = false;
@@ -941,7 +948,7 @@ struct SystemOverlay::Impl {
 		ImGui::PopFont();
 		ImGui::PopStyleVar(2);
 		if (accepted) {
-			ErrorDialog::HostAccept(snapshot.generation);
+			SystemDialog::HostClose(snapshot.generation);
 		}
 	}
 
@@ -955,7 +962,8 @@ struct SystemOverlay::Impl {
 		if (session != snapshot.session) {
 			session       = snapshot.session;
 			focus_pending = true;
-			shift         = (snapshot.ime.option & Ime::OPTION_NO_AUTO_CAPITALIZE) == 0;
+			shift         = session.kind == OverlayKind::Ime &&
+			                (snapshot.ime.option & Ime::OPTION_NO_AUTO_CAPITALIZE) == 0;
 			symbol_mode   = false;
 			panel_offset  = {};
 			right_stick   = {};
@@ -981,8 +989,8 @@ struct SystemOverlay::Impl {
 			ImGui::EndFrame();
 			return false;
 		}
-		if (snapshot.session.kind == OverlayKind::Error) {
-			DrawError(snapshot.error, frame_extent);
+		if (snapshot.session.kind == OverlayKind::Dialog) {
+			DrawDialog(snapshot.dialog, frame_extent);
 		} else {
 			DrawIme(snapshot.ime, frame_extent);
 		}

src/graphics/presentation/window/hostInput.cpp

@@ -72,6 +72,7 @@ struct Binding {
 };
 
 constexpr int              MOUSE_POLL_INTERVAL_MS = 33;
+constexpr uint64_t         CURSOR_IDLE_HIDE_MS    = 2000;
 constexpr std::string_view MOUSE_SENSITIVITY      = "MouseSensitivity=";
 
 struct MouseJoystickState {
@@ -81,7 +82,8 @@ struct MouseJoystickState {
 };
 
 MouseJoystickState g_mouse;
-SDL_Window*        g_mouse_window = nullptr;
+SDL_Window*        g_mouse_window   = nullptr;
+uint64_t           g_cursor_hide_at = 0;
 
 std::size_t ControlFromName(std::string_view name) {
 	const auto info = std::find_if(CONTROL_INFO.begin(), CONTROL_INFO.end(),
@@ -380,11 +382,31 @@ int PollMouse(uint64_t now_ms) {
 	return MOUSE_POLL_INTERVAL_MS;
 }
 
+bool IsCursorActivity(const SDL_Event& event) {
+	SDL_MouseID which;
+	switch (event.type) {
+		case SDL_EVENT_MOUSE_MOTION:
+			if (event.motion.xrel == 0.0f && event.motion.yrel == 0.0f) {
+				return false;
+			}
+			which = event.motion.which;
+			break;
+		case SDL_EVENT_MOUSE_BUTTON_DOWN:
+		case SDL_EVENT_MOUSE_BUTTON_UP: which = event.button.which; break;
+		case SDL_EVENT_MOUSE_WHEEL: which = event.wheel.which; break;
+		default: return false;
+	}
+	return which != SDL_TOUCH_MOUSEID && which != SDL_PEN_MOUSEID;
+}
+
 } // namespace
 
 void HostInputInit(SDL_Window* window) {
 	GetInputMap();
 	g_mouse_window = window;
+	if (Config::HideCursorEnabled()) {
+		g_cursor_hide_at = SDL_GetTicks() + CURSOR_IDLE_HIDE_MS;
+	}
 }
 
 void HostInputShutdown() {
@@ -393,7 +415,11 @@ void HostInputShutdown() {
 		CenterMouseStick();
 		g_mouse = {};
 	}
-	g_mouse_window = nullptr;
+	g_mouse_window   = nullptr;
+	g_cursor_hide_at = 0;
+	if (Config::HideCursorEnabled()) {
+		SDL_ShowCursor();
+	}
 }
 
 void HostInputKey(int key_code, bool down) {
@@ -429,20 +455,36 @@ void HostInputToggleMouseToJoystick() {
 }
 
 bool HostInputWaitEvent(SDL_Event* event) {
-	bool has_event;
+	int timeout = -1;
 	if (!g_mouse.enabled || SDL_GetKeyboardFocus() != g_mouse_window) {
 		g_mouse.next_poll = 0;
 		CenterMouseStick();
-		has_event = SDL_WaitEvent(event);
-		if (!has_event) {
-			EXIT("%s\n", SDL_GetError());
-		}
 	} else {
 		if (g_mouse.next_poll == 0) {
 			SDL_GetRelativeMouseState(nullptr, nullptr);
 			g_mouse.next_poll = SDL_GetTicks() + MOUSE_POLL_INTERVAL_MS;
 		}
-		has_event = SDL_WaitEventTimeout(event, PollMouse(SDL_GetTicks()));
+		timeout = PollMouse(SDL_GetTicks());
+	}
+
+	if (g_cursor_hide_at != 0) {
+		const auto now_ms = SDL_GetTicks();
+		const int  cursor_timeout =
+		    now_ms < g_cursor_hide_at ? static_cast<int>(g_cursor_hide_at - now_ms) : 0;
+		timeout = timeout < 0 ? cursor_timeout : std::min(timeout, cursor_timeout);
+	}
+	const bool has_event = SDL_WaitEventTimeout(event, timeout);
+
+	if (Config::HideCursorEnabled()) {
+		const auto now_ms = SDL_GetTicks();
+		if (has_event && !g_mouse.enabled && IsCursorActivity(*event) &&
+		    SDL_GetWindowFromEvent(event) == g_mouse_window) {
+			SDL_ShowCursor();
+			g_cursor_hide_at = now_ms + CURSOR_IDLE_HIDE_MS;
+		} else if (g_cursor_hide_at != 0 && now_ms >= g_cursor_hide_at) {
+			SDL_HideCursor();
+			g_cursor_hide_at = 0;
+		}
 	}
 
 	if (has_event && event->type == SDL_EVENT_WINDOW_FOCUS_LOST &&

src/graphics/presentation/window/vulkanWindow.cpp

@@ -47,6 +47,12 @@ struct VulkanExtensions {
 	std::vector<vk::LayerProperties>     available_layers;
 };
 
+vk::PhysicalDeviceVulkan11Features WindowContext::RequiredVulkan11Features() noexcept {
+	vk::PhysicalDeviceVulkan11Features features {};
+	features.storageBuffer16BitAccess = VK_TRUE;
+	return features;
+}
+
 vk::PhysicalDeviceVulkan12Features WindowContext::RequiredVulkan12Features() noexcept {
 	vk::PhysicalDeviceVulkan12Features features {};
 	features.samplerMirrorClampToEdge  = VK_TRUE;
@@ -55,6 +61,8 @@ vk::PhysicalDeviceVulkan12Features WindowContext::RequiredVulkan12Features() noe
 	features.shaderOutputViewportIndex = VK_TRUE;
 	features.bufferDeviceAddress       = VK_TRUE;
 	features.shaderBufferInt64Atomics  = VK_TRUE;
+	features.storageBuffer8BitAccess   = VK_TRUE;
+	features.shaderSampledImageArrayNonUniformIndexing = VK_TRUE;
 	return features;
 }
 
@@ -206,6 +214,7 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		depth_clip_control.pNext = &depth_clip_enable;
 
 		vk::PhysicalDeviceVulkan12Features features12 {};
+		vk::PhysicalDeviceVulkan11Features features11 {};
 #if defined(__APPLE__)
 		features12.pNext = &depth_clip_control;
 #else
@@ -213,6 +222,8 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		fragment_barycentric.pNext = &depth_clip_control;
 		features12.pNext           = &fragment_barycentric;
 #endif
+		features11.pNext       = features12.pNext;
+		features12.pNext       = &features11;
 		features13.pNext       = &features12;
 		device_features2.pNext = &features13;
 
@@ -256,6 +267,10 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 		check_feature(features12.samplerMirrorClampToEdge, "samplerMirrorClampToEdge",
 		              required_features12.samplerMirrorClampToEdge);
+		check_feature(features11.storageBuffer16BitAccess, "storageBuffer16BitAccess",
+		              WindowContext::RequiredVulkan11Features().storageBuffer16BitAccess);
+		check_feature(features12.storageBuffer8BitAccess, "storageBuffer8BitAccess",
+		              required_features12.storageBuffer8BitAccess);
 		check_feature(features12.timelineSemaphore, "timelineSemaphore",
 		              required_features12.timelineSemaphore);
 		check_feature(features12.shaderOutputLayer, "shaderOutputLayer",
@@ -266,6 +281,9 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		              required_features12.bufferDeviceAddress);
 		check_feature(features12.shaderBufferInt64Atomics, "shaderBufferInt64Atomics",
 		              required_features12.shaderBufferInt64Atomics);
+		check_feature(features12.shaderSampledImageArrayNonUniformIndexing,
+		              "shaderSampledImageArrayNonUniformIndexing",
+		              required_features12.shaderSampledImageArrayNonUniformIndexing);
 		check_feature(features13.robustImageAccess, "robustImageAccess");
 		check_feature(features13.dynamicRendering, "dynamicRendering",
 		              required_features13.dynamicRendering);
@@ -460,7 +478,15 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		provoking_vertex.pNext = supported_features2.pNext;
 		supported_features2.pNext = &provoking_vertex;
 	}
+	const bool image_atomic_int64_extension =
+	    HasExtension(device_extensions, VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME);
+	vk::PhysicalDeviceShaderImageAtomicInt64FeaturesEXT image_atomic_int64 {};
+	if (image_atomic_int64_extension) {
+		image_atomic_int64.pNext = supported_features2.pNext;
+		supported_features2.pNext = &image_atomic_int64;
+	}
 	physical_device.getFeatures2(&supported_features2);
+	graphics.shader_image_int64_atomics_enabled = image_atomic_int64.shaderImageInt64Atomics;
 
 	auto features12 = WindowContext::RequiredVulkan12Features();
 	features12.shaderSharedInt64Atomics = supported_features12.shaderSharedInt64Atomics;
@@ -470,6 +496,9 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	workgroup_layout.pNext = &depth_clip_control;
 	features12.pNext = workgroup_layout_extension ? static_cast<void*>(&workgroup_layout)
 	                                             : static_cast<void*>(&depth_clip_control);
+	auto features11 = WindowContext::RequiredVulkan11Features();
+	features11.pNext = features12.pNext;
+	features12.pNext = &features11;
 	if (!features12.shaderSharedInt64Atomics || !workgroup_layout.workgroupMemoryExplicitLayout) {
 		Log::WriteToConsoleAndLog(fmt::format(
 		    "WARNING: Native 64-bit LDS atomics are unavailable: shaderSharedInt64Atomics={}, "
@@ -552,8 +581,7 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	device_features.shaderInt64 = VK_TRUE;
 	device_features.shaderFloat64 =
 	    supported_features2.features.shaderFloat64 &&
-	    float_controls.shaderSignedZeroInfNanPreserveFloat64 &&
-	    float_controls.shaderRoundingModeRTEFloat32;
+	    float_controls.shaderSignedZeroInfNanPreserveFloat64;
 	// if (device_features.shaderFloat64 && !float_controls.shaderDenormPreserveFloat64) {
 	// 	Log::WriteToConsoleAndLog(
 	// 	    "WARNING: Vulkan device does not guarantee FP64 denormal preservation; "
@@ -605,6 +633,11 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		provoking_vertex.transformFeedbackPreservesProvokingVertex = VK_FALSE;
 		create_info.pNext = &provoking_vertex;
 	}
+	if (graphics.shader_image_int64_atomics_enabled) {
+		image_atomic_int64.pNext = const_cast<void*>(create_info.pNext);
+		image_atomic_int64.sparseImageInt64Atomics = VK_FALSE;
+		create_info.pNext = &image_atomic_int64;
+	}
 	create_info.pQueueCreateInfos       = &queue_create_info;
 	create_info.queueCreateInfoCount    = 1;
 	create_info.enabledExtensionCount   = static_cast<uint32_t>(device_extensions.size());
@@ -969,6 +1002,7 @@ void WindowContext::CreateVulkan() {
 			graphic_ctx.memory_budget_ext_enabled = true;
 		}
 		for (const auto* extension: {VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
+		                             VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME,
 		                             VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME,
 		                             VK_EXT_MESH_SHADER_EXTENSION_NAME,
 		                             VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME,

src/graphics/presentation/window/window.cpp

@@ -306,14 +306,20 @@ static void GameEventController([[maybe_unused]] const EventController& f) {
 
 	if (f.added) {
 		auto* pad = SDL_OpenGamepad(f.id);
-		EXIT_NOT_IMPLEMENTED(pad == nullptr);
+		if (pad == nullptr) {
+			LOGF("Controller: ignoring gamepad %d that could not be opened: %s\n", f.id,
+			     SDL_GetError());
+			return;
+		}
 		int id = SDL_GetJoystickID(SDL_GetGamepadJoystick(pad));
 		Controller::Connect(id);
 	}
 
 	if (f.removed) {
-		Controller::Disconnect(f.id);
-		SDL_CloseGamepad(SDL_GetGamepadFromID(f.id));
+		if (auto* pad = SDL_GetGamepadFromID(f.id); pad != nullptr) {
+			Controller::Disconnect(f.id);
+			SDL_CloseGamepad(pad);
+		}
 	}
 
 	if (f.down || f.up) {
@@ -356,11 +362,13 @@ static void GameEventDidEnterForeground(WindowLoopState& game) {
 	SetPause(game, false);
 }
 
-void WindowContext::Resize(uint32_t new_width, uint32_t new_height) {
-	EXIT_IF(new_width == 0 || new_height == 0);
+void WindowContext::Resize(int new_width, int new_height) {
+	if (new_width <= 0 || new_height <= 0) {
+		return;
+	}
 	Common::LockGuard lock(mutex);
-	graphic_ctx.screen_width  = new_width;
-	graphic_ctx.screen_height = new_height;
+	graphic_ctx.screen_width  = static_cast<uint32_t>(new_width);
+	graphic_ctx.screen_height = static_cast<uint32_t>(new_height);
 }
 
 void WindowContext::ProcessWindowEvent(const SDL_WindowEvent& event) {

src/graphics/presentation/window/windowInternal.h

@@ -34,6 +34,7 @@ struct WindowContext {
 	~WindowContext();
 	KYTY_CLASS_NO_COPY(WindowContext);
 
+	[[nodiscard]] static vk::PhysicalDeviceVulkan11Features RequiredVulkan11Features() noexcept;
 	[[nodiscard]] static vk::PhysicalDeviceVulkan12Features RequiredVulkan12Features() noexcept;
 	[[nodiscard]] static vk::PhysicalDeviceVulkan13Features RequiredVulkan13Features() noexcept;
 	[[nodiscard]] static uint32_t InitialWindowFlags(bool fullscreen) noexcept;
@@ -42,7 +43,7 @@ struct WindowContext {
 	void                                                    RefreshSurfaceCapabilities();
 	void                                                    UpdateIcon();
 	void                                                    UpdateTitle();
-	void                                                    Resize(uint32_t width, uint32_t height);
+	void                                                    Resize(int width, int height);
 	void ProcessWindowEvent(const SDL_WindowEvent& event);
 	void ProcessDisplayEvent(const SDL_DisplayEvent& event);
 	void ProcessEvent(double time_seconds);

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -617,7 +617,6 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 		}
 	}
 	IR::TrackResources(ir, decoded, native_cfg);
-	IR::EliminateDeadCode(ir.blocks);
 	TranslateResult result;
 	result.program = std::move(ir);
 	if (options.dump_ir) {
@@ -649,6 +648,11 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 				})) {
 					continue;
 				}
+				const auto resource = inst.Flags<uint32_t>();
+				first               = resource < ir.info.buffers.size() &&
+				                              ir.info.buffers[resource].indirect_root == resource
+				                          ? 1u
+				                          : 0u;
 			} else if (op == IR::ValueOpcode::GetImageResource) {
 				const auto resource = inst.Flags<uint32_t>();
 				first = resource < ir.info.images.size() &&

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

@@ -41,7 +41,7 @@ void ValidateNativeProgram(const IR::Program& program) {
 			Fail(program, "native shader plan has an invalid image class");
 		}
 		present[static_cast<size_t>(*kind)] = true;
-		const auto dynamic = program.info.images[i].mip_mode == IR::ImageMipMode::DynamicStorage;
+		const auto dynamic = program.info.images[i].mip_mode == IR::ImageMipMode::Dynamic;
 		const auto count   = dynamic ? program.info.images[i].mip_count : 1u;
 		if (count == 0u || (!dynamic && program.info.images[i].mip_count != 1u)) {
 			Fail(program, "native shader plan has an invalid image mip descriptor count");
@@ -85,9 +85,13 @@ void ValidateNativeProgram(const IR::Program& program) {
 	}
 	const auto has_shader_data_storage = present[static_cast<size_t>(Kind::ShaderData)];
 	const auto shader_data_dwords = program.bindings.ShaderDataDwords();
+	const auto user_data_dwords = program.bindings.user_data_registers.size();
+	const bool has_dispatch_threads = program.bindings.dispatch_thread_dword != IR::PushData::NoStart;
 	if ((program.bindings.UsesPushData() &&
 	     !IR::PushData::CanFit(program.bindings.push_data_start_dword, shader_data_dwords)) ||
-	    program.bindings.memory_offset_dword != program.bindings.user_data_registers.size() ||
+	    (has_dispatch_threads && (program.stage != ShaderType::Compute ||
+	                              program.bindings.dispatch_thread_dword != user_data_dwords)) ||
+	    program.bindings.memory_offset_dword != user_data_dwords + (has_dispatch_threads ? 3u : 0u) ||
 	    program.bindings.memory_offset_count != buffers.size() ||
 	    has_shader_data_storage != (shader_data_dwords != 0 && !program.bindings.UsesPushData()) ||
 	    !std::is_sorted(program.bindings.user_data_registers.begin(),
@@ -218,7 +222,8 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					Fail(program, "buffer operation has invalid memory metadata");
 				}
 				const auto& memory = program.memory_info[memory_index];
-				if (memory.kind == IR::ResourceKind::IndirectBuffer) {
+				if (memory.kind == IR::ResourceKind::IndirectBuffer &&
+				    inst.GetOpcode() != IR::ValueOpcode::ReadConstBuffer) {
 					requirements.subgroup_local_invocation_id = true;
 				}
 				if (memory.kind == IR::ResourceKind::Buffer) {
@@ -226,6 +231,9 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					if (memory.resource >= program.info.buffers.size()) {
 						Fail(program, "buffer operation has invalid resource metadata");
 					}
+					const auto bits = StorageBufferElementBits(program, memory);
+					requirements.buffer_u8 |= bits == 8u;
+					requirements.buffer_u16 |= bits == 16u;
 					if ((program.info.buffers[memory.resource].packed_stride & (1u << 20u)) != 0u) {
 						if (program.stage != ShaderType::Compute) {
 							Fail(program, "buffer ADD_TID is only valid for compute shaders");
@@ -263,6 +271,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				}
 			}
 			switch (inst.GetOpcode()) {
+				case IR::ValueOpcode::StoreCompletion: requirements.subgroup_barrier = true; break;
 				case IR::ValueOpcode::BvhIntersect: requirements.bvh = true; break;
 				case IR::ValueOpcode::Ballot: requirements.subgroup_ballot = true; break;
 				case IR::ValueOpcode::DppMoveU32:
@@ -288,6 +297,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					break;
 				}
 				case IR::ValueOpcode::SwizzleU32:
+				case IR::ValueOpcode::PermuteU32:
 				case IR::ValueOpcode::BpermuteU32: {
 					requirements.subgroup_ballot              = true;
 					requirements.subgroup_shuffle             = true;

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

@@ -83,7 +83,7 @@ uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
 	}
 	const auto pointer = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpAccessChain, result_ptr_type, pointer, variable_id,
-	                          ConstantU32(state, array_index));
+	                          array_index);
 	return pointer;
 }
 
@@ -124,13 +124,21 @@ uint32_t ImageType(EmitterState& state, const IR::ImageResource& image) {
 		sampled = 2;
 		if (image.atomic) {
 			EXIT_IF(image.numeric_class != Prospero::TextureNumericClass::Uint);
-			format = spv::ImageFormatR32ui;
+			format = image.atomic64 ? spv::ImageFormatR64ui : spv::ImageFormatR32ui;
+			if (image.atomic64) {
+				state.builder.RequireExtension("SPV_EXT_shader_image_int64");
+				state.builder.RequireCapability(spv::CapabilityInt64);
+				state.builder.RequireCapability(spv::CapabilityInt64Atomics);
+				state.builder.RequireCapability(spv::CapabilityInt64ImageEXT);
+			}
 		}
 	} else {
 		EXIT("invalid image resource class");
 	}
 	const auto& info = ImageDimensionInfoFor(image.dimension);
-	return state.builder.Type(spv::OpTypeImage, ImageScalarType(state, image.numeric_class),
+	const auto scalar_type = image.atomic64 ? TypeScalarU64(state)
+	                                        : ImageScalarType(state, image.numeric_class);
+	return state.builder.Type(spv::OpTypeImage, scalar_type,
 	                          info.spirv_dimension, image.depth_compare ? 1u : 0u, info.arrayed,
 	                          info.multisampled, sampled, format);
 }
@@ -144,20 +152,12 @@ uint32_t ImageViewSizeType(EmitterState& state, ImageDimension dimension) {
 	}
 }
 
-uint32_t LoadSampledImageDescriptor(EmitterState& state, uint32_t resource) {
-	const auto& image_resource = state.program.info.images.at(resource);
-	EXIT_IF(image_resource.resource_class != IR::ImageResourceClass::Sampled);
-	const auto kind = IR::DescriptorBindingForImage(image_resource);
-	EXIT_IF(!kind.has_value());
-	const auto array_index  = ResourceForDescriptor(state, *kind, resource);
-	const auto variable     = state.image_variables[IR::ImageBindingIndex(*kind)];
-	const auto pointer_type = state.builder.Type(
-	    spv::OpTypePointer, spv::StorageClassUniformConstant, ImageType(state, image_resource));
-	const auto pointer =
-	    DescriptorElementPointer(state, pointer_type, variable, array_index, *kind, resource,
-	                             "sampled image descriptor array was not emitted");
+uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip,
+                             uint32_t array_index) {
+	const auto pointer = ImageDescriptorPointer(state, resource, mip, array_index);
 	const auto image = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpLoad, ImageType(state, image_resource), image, pointer);
+	state.builder.AddFunction(spv::OpLoad, ImageType(state, state.program.info.images.at(resource)),
+	                          image, pointer);
 	return image;
 }
 
@@ -168,35 +168,45 @@ uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler) {
 	const auto pointer_type =
 	    state.builder.Type(spv::OpTypePointer, spv::StorageClassUniformConstant, sampler_type);
 	const auto pointer = DescriptorElementPointer(
-	    state, pointer_type, state.sampler_variable, array_index,
+	    state, pointer_type, state.sampler_variable, ConstantU32(state, array_index),
 	    IR::DescriptorBindingKind::Samplers, sampler, "sampler descriptor array was not emitted");
 	const auto sampler_id = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpLoad, sampler_type, sampler_id, pointer);
 	return sampler_id;
 }
 
-uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler) {
+uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id, uint32_t mip,
+                          uint32_t array_index) {
 	const auto& image_resource = state.program.info.images.at(resource);
-	const auto  image          = LoadSampledImageDescriptor(state, resource);
-	const auto  sampler_id     = LoadSamplerDescriptor(state, sampler);
+	EXIT_IF(image_resource.resource_class != IR::ImageResourceClass::Sampled);
+	const auto  image          = LoadImageDescriptor(state, resource, mip, array_index);
 	const auto  sampled_image = state.builder.AllocateId();
 	const auto  sampled_type =
 	    state.builder.Type(spv::OpTypeSampledImage, ImageType(state, image_resource));
 	state.builder.AddFunction(spv::OpSampledImage, sampled_type, sampled_image, image, sampler_id);
+	if (array_index != 0u) {
+		state.builder.RequireExtension("SPV_EXT_descriptor_indexing");
+		state.builder.RequireCapability(spv::CapabilityShaderNonUniform);
+		state.builder.RequireCapability(spv::CapabilitySampledImageArrayNonUniformIndexing);
+		state.builder.AddAnnotation(spv::OpDecorate, sampled_image, spv::DecorationNonUniform);
+	}
 	return sampled_image;
 }
 
-uint32_t StorageImageDescriptorPointer(EmitterState& state, uint32_t resource) {
+uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t mip,
+                                uint32_t array_index) {
 	const auto& image = state.program.info.images.at(resource);
-	EXIT_IF(image.resource_class != IR::ImageResourceClass::Storage);
+	EXIT_IF(mip >= image.mip_count);
 	const auto kind = IR::DescriptorBindingForImage(image);
 	EXIT_IF(!kind.has_value());
-	const auto array_index  = ResourceForDescriptor(state, *kind, resource);
+	if (array_index == 0u) {
+		array_index = ConstantU32(state, ResourceForDescriptor(state, *kind, resource) + mip);
+	}
 	const auto pointer_type = state.builder.Type(
 	    spv::OpTypePointer, spv::StorageClassUniformConstant, ImageType(state, image));
 	const auto variable = state.image_variables[IR::ImageBindingIndex(*kind)];
 	return DescriptorElementPointer(state, pointer_type, variable, array_index, *kind, resource,
-	                                "storage image descriptor array was not emitted");
+	                                "image descriptor array was not emitted");
 }
 
 void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_lod, uint32_t coord,
@@ -206,46 +216,15 @@ void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_
 	if (!image.atomic) {
 		state.builder.RequireCapability(spv::CapabilityStorageImageWriteWithoutFormat);
 	}
-	const auto kind = IR::DescriptorBindingForImage(image);
-	EXIT_IF(!kind.has_value());
-	const auto array_index = ResourceForDescriptor(state, *kind, resource);
-	const auto image_type  = ImageType(state, image);
-	const auto pointer_type =
-	    state.builder.Type(spv::OpTypePointer, spv::StorageClassUniformConstant, image_type);
-	const auto variable = state.image_variables[IR::ImageBindingIndex(*kind)];
-	const auto LoadAt   = [&](uint32_t array_index) {
-		const auto pointer =
-		    DescriptorElementPointer(state, pointer_type, variable, array_index, *kind, resource,
-		                             "storage image descriptor array was not emitted");
-		const auto descriptor = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpLoad, image_type, descriptor, pointer);
-		return descriptor;
+	const auto EmitWrite = [&](uint32_t mip) {
+		state.builder.AddFunction(spv::OpImageWrite, LoadImageDescriptor(state, resource, mip),
+		                          coord, texel);
 	};
-	if (image.mip_mode != IR::ImageMipMode::DynamicStorage) {
-		state.builder.AddFunction(spv::OpImageWrite, LoadAt(array_index), coord, texel);
+	if (image.mip_mode != IR::ImageMipMode::Dynamic) {
+		EmitWrite(0);
 		return;
 	}
-	if (image.mip_count == 0u) {
-		ExitDescriptorBindingFailure(state, *kind, resource,
-		                             "dynamic storage image has no mip descriptors");
-	}
-
-	const auto            merge_label = state.builder.AllocateId();
-	std::vector<uint32_t> labels(image.mip_count);
-	std::vector<uint32_t> words {spv::OpSwitch, mip_lod, merge_label};
-	for (uint32_t mip = 0; mip < image.mip_count; mip++) {
-		labels[mip] = state.builder.AllocateId();
-		words.push_back(mip);
-		words.push_back(labels[mip]);
-	}
-	state.builder.AddFunction(spv::OpSelectionMerge, merge_label, spv::SelectionControlMaskNone);
-	state.builder.AddFunction(words);
-	for (uint32_t mip = 0; mip < image.mip_count; mip++) {
-		EmitLabel(state, labels[mip]);
-		state.builder.AddFunction(spv::OpImageWrite, LoadAt(array_index + mip), coord, texel);
-		state.builder.AddFunction(spv::OpBranch, merge_label);
-	}
-	EmitLabel(state, merge_label);
+	EmitIndexSwitch(state, mip_lod, image.mip_count, 0, EmitWrite);
 }
 
 spv::ExecutionModel ExecutionModelForStage(ShaderType stage) {

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

@@ -226,13 +226,13 @@ void DefineBvhIntersect(EmitterState& s) {
 	if (!s.requirements.bvh) return;
 	const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeScalarU64(s);
 	const auto vec4 = TypeU32Vector(s, 4), vec3 = TypeF32Vector(s, 3);
-	const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, u, f, vec3, vec3, vec3);
+	const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, wide, f, vec3, vec3, vec3);
 	s.bvh_intersect_function = s.builder.AllocateId();
 	s.builder.AddName(s.bvh_intersect_function, "bvh_intersect");
 	s.builder.AddFunction(spv::OpFunction, vec4, s.bvh_intersect_function,
 	                      spv::FunctionControlMaskNone, signature);
 	std::array<uint32_t, 6> args;
-	const std::array types {vec4, u, f, vec3, vec3, vec3};
+	const std::array types {vec4, wide, f, vec3, vec3, vec3};
 	for (uint32_t i = 0; i < args.size(); ++i) {
 		args[i] = s.builder.AllocateId();
 		s.builder.AddFunction(spv::OpFunctionParameter, types[i], args[i]);
@@ -253,7 +253,7 @@ void DefineBvhIntersect(EmitterState& s) {
 	const auto shr = [&](uint32_t value, uint32_t shift) {
 		return Binary(s, spv::OpShiftRightLogical, u, value, ConstantU32(s, shift));
 	};
-	const auto kind = and_bits(node, 7);
+	const auto kind = and_bits(Unary(s, spv::OpUConvert, u, node), 7);
 	const auto triangle = Binary(s, spv::OpULessThan, b, kind, ConstantU32(s, 4));
 	const auto full = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 5));
 	const auto half = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 4));
@@ -262,7 +262,7 @@ void DefineBvhIntersect(EmitterState& s) {
 	const auto grow = and_bits(shr(words[1], 23), 0xff);
 	const auto base = Binary(s, spv::OpShiftLeftLogical, wide,
 	    DeviceAddressFromWords(s, words[0], and_bits(words[1], 0xff)), ConstantDeviceAddress(s, 8));
-	const auto index = Unary(s, spv::OpUConvert, wide, shr(node, 3));
+	const auto index = Binary(s, spv::OpShiftRightLogical, wide, node, ConstantDeviceAddress(s, 3));
 	const auto last_index = Binary(s, spv::OpIAdd, wide, index,
 	    Select(s, wide, full, ConstantDeviceAddress(s, 1), ConstantDeviceAddress(s, 0)));
 	const auto address = Binary(s, spv::OpIAdd, wide, base,
@@ -325,17 +325,23 @@ void DefineBvhIntersect(EmitterState& s) {
 uint32_t EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto& s = ctx.state;
 	const auto* ray = ctx.ImageAddress(inst.Arg(1));
-	if (ray == nullptr || ray->NumArgs() < 11 || s.bvh_intersect_function == 0)
-		ctx.Fail(inst, "BVH intersection requires eleven address components and its shared function");
+	const auto node_words = inst.Flags<uint32_t>();
+	if (ray == nullptr || (node_words != 1u && node_words != 2u) ||
+	    ray->NumArgs() < node_words + 10u || s.bvh_intersect_function == 0)
+		ctx.Fail(inst, "BVH intersection requires its node pointer, ray components and shared function");
 	const auto vector = [&](uint32_t first) {
 		Vec3 values;
 		for (uint32_t i = 0; i < 3; ++i)
 			values[i] = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, first + i));
 		return Vector(s, TypeF32Vector(s, 3), values);
 	};
-	const auto descriptor = ctx.Arg(inst, 0), node = ctx.Arg(*ray, 0);
-	const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, 1));
-	const auto origin = vector(2), direction = vector(5), inverse = vector(8);
+	const auto descriptor = ctx.Arg(inst, 0);
+	const auto node = node_words == 2u
+	                      ? DeviceAddressFromWords(s, ctx.Arg(*ray, 0), ctx.Arg(*ray, 1))
+	                      : Unary(s, spv::OpUConvert, TypeScalarU64(s), ctx.Arg(*ray, 0));
+	const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, node_words));
+	const auto origin = vector(node_words + 1u), direction = vector(node_words + 4u),
+	           inverse = vector(node_words + 7u);
 	return EmitValueOrDefaultIfCondition(s, ctx.Arg(inst, 2), TypeU32Vector(s, 4),
 	    ConstantU32CompositeZero(s, 4), [&] {
 		const auto result = s.builder.AllocateId();

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

@@ -366,18 +366,23 @@ void EmitAuxPositionExport(ValueEmitContext& ctx, uint32_t data, const IR::Expor
 			state.builder.AddFunction(spv::OpStore, state.point_size_variable, f32);
 			continue;
 		}
-		auto StoreDistance = [&](uint32_t variable, uint32_t index) {
+		auto StoreDistance = [&](IR::StageOutputKind kind, uint32_t variable, uint32_t index) {
 			if (index == UINT32_MAX) {
 				return;
 			}
-			const auto pointer = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpAccessChain,
-			                          TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
-			                          pointer, variable, ConstantU32(state, index));
+			uint32_t pointer;
+			if (state.program.stage == ShaderType::Mesh) {
+				pointer = MeshOutputPointer(state, kind, index);
+			} else {
+				pointer = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpAccessChain,
+				                          TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
+				                          pointer, variable, ConstantU32(state, index));
+			}
 			state.builder.AddFunction(spv::OpStore, pointer, f32);
 		};
-		StoreDistance(state.clip_distance_variable, output.clip_distance);
-		StoreDistance(state.cull_distance_variable, output.cull_distance);
+		StoreDistance(IR::StageOutputKind::ClipDistance, state.clip_distance_variable, output.clip_distance);
+		StoreDistance(IR::StageOutputKind::CullDistance, state.cull_distance_variable, output.cull_distance);
 	}
 }
 
@@ -574,6 +579,14 @@ uint32_t EmitIdentity(ValueEmitContext&, uint32_t value) {
 
 void EmitVoid(ValueEmitContext&) {}
 
+void EmitStoreCompletion(EmitterState& state) {
+	state.builder.AddFunction(
+	    spv::OpControlBarrier, ConstantU32(state, spv::ScopeSubgroup),
+	    ConstantU32(state, spv::ScopeDevice),
+	    ConstantU32(state, spv::MemorySemanticsReleaseMask | spv::MemorySemanticsUniformMemoryMask |
+	                          spv::MemorySemanticsImageMemoryMask));
+}
+
 void EmitBarrier(EmitterState& state) {
 	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
 	if (!tessellation && ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
@@ -623,6 +636,14 @@ uint32_t EmitGetBuiltin(ValueEmitContext& ctx, IR::Value kind, IR::Value index)
 	return EmitBuiltinU32(ctx.state, static_cast<IR::StageInputKind>(kind.U32()), index.U32());
 }
 
+uint32_t EmitGetDispatchThreadExtent(ValueEmitContext& ctx, const IR::Inst& inst) {
+	const auto start = ctx.state.program.bindings.dispatch_thread_dword;
+	if (start == IR::PushData::NoStart || !inst.Arg(0).IsImmediate() || inst.Arg(0).U32() >= 3u) {
+		ctx.Fail(inst, "invalid dispatch thread extent");
+	}
+	return EmitShaderDataDwordLoad(ctx.state, start + inst.Arg(0).U32());
+}
+
 uint32_t EmitUndefU1(EmitterState& state, const IR::Inst& inst) {
 	const auto result = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpUndef, TypeId(state, inst.GetType()), result);

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

@@ -3,9 +3,7 @@
 #include "graphics/shader/recompiler/frontend/decode/ImageOps.h"
 
 #include <algorithm>
-#include <atomic>
 #include <bit>
-#include <cstdio>
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 namespace {
@@ -86,6 +84,46 @@ uint32_t ZeroF32(EmitterState& state) {
 	return ConstantF32(state, 0);
 }
 
+uint32_t GatherMip(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
+                   const IR::Inst& address, const ImageSampleLayout& layout, uint32_t mip_count) {
+	auto& state = ctx.state;
+	const auto Extract = [&](uint32_t word, uint32_t first, uint32_t count) {
+		return EmitBitFieldUExtract(state, word, ConstantU32(state, first),
+		                           ConstantU32(state, count));
+	};
+	const auto LodFixed = [&](uint32_t word, uint32_t first) {
+		return Binary(state, spv::OpFMul, TypeF32(state),
+		              Unary(state, spv::OpConvertUToF, TypeF32(state), Extract(word, first, 12)),
+		              ConstantF32Value(state, 1.0f / 256.0f));
+	};
+	const auto sampler_control = ctx.Arg(inst, 6);
+	const auto preclamp = Binary(state, spv::OpINotEqual, TypeBool(state),
+	                            Extract(sampler_control, 28, 1), ConstantU32(state, 0));
+	const auto half = ConstantF32Value(state, 0.5f);
+	const auto zero = ZeroF32(state);
+	const auto base = Unary(state, spv::OpConvertUToF, TypeF32(state),
+	                        Extract(ctx.Arg(inst, 4), 12, 4));
+	const auto minimum = Binary(state, spv::OpFSub, TypeF32(state),
+	                            LodFixed(ctx.Arg(inst, 3), 8), base);
+	const auto last = ConstantF32Value(state, static_cast<float>(mip_count - 1u));
+	auto lod = Binary(state, spv::OpFAdd, TypeF32(state),
+	                  AddressF32(ctx, mem, address, layout.lod),
+	                  Select(state, TypeF32(state), preclamp, half, zero));
+	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(
+	    state, lod, LodFixed(ctx.Arg(inst, 5), 0), LodFixed(ctx.Arg(inst, 5), 12));
+	// T# minimum LOD is in physical mip space and applies after the S# view-relative clamp.
+	lod = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(state, lod, minimum);
+	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(state, lod, zero, last);
+	lod = Binary(state, spv::OpFAdd, TypeF32(state), lod,
+	             Select(state, TypeF32(state), preclamp, zero, half));
+	// The clamped value is nonnegative, so conversion performs the integer mip rounding.
+	const auto mip = EmitUMin32(state, Unary(state, spv::OpConvertFToU, TypeU32(state), lod),
+	                           ConstantU32(state, mip_count - 1u));
+	const auto mip_none = Binary(state, spv::OpIEqual, TypeBool(state),
+	                            Extract(sampler_control, 26, 2), ConstantU32(state, 0));
+	return Select(state, TypeU32(state), mip_none, ConstantU32(state, 0), mip);
+}
+
 uint32_t CubeAxis(EmitterState& state, uint32_t value) {
 	return Binary(state, spv::OpFSub, TypeF32(state), value, ConstantF32(state, 0x3f800000u));
 }
@@ -252,7 +290,7 @@ uint32_t QueryDimensions(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
                          const IR::Inst& address) {
 	const auto  dimension = ctx.state.program.info.images.at(mem.resource).dimension;
 	const auto& info      = ImageDimensionInfoFor(dimension);
-	const auto  image     = LoadSampledImageDescriptor(ctx.state, mem.resource);
+	const auto  image     = LoadImageDescriptor(ctx.state, mem.resource);
 	const auto  size      = ctx.state.builder.AllocateId();
 	if (info.multisampled != 0u) {
 		ctx.state.builder.AddFunction(spv::OpImageQuerySize,
@@ -433,7 +471,7 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
                                     uint32_t coord, Prospero::TextureNumericClass numeric_class) {
 	auto& state = ctx.state;
 	state.builder.RequireCapability(spv::CapabilityImageQuery);
-	const auto image = LoadSampledImageDescriptor(state, mem.resource);
+	const auto image = LoadImageDescriptor(state, mem.resource);
 	const auto width = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpImageQuerySizeLod, TypeU32(state), width, image,
 	                          ConstantU32(state, 0));
@@ -446,7 +484,8 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
 	                                 Binary(state, spv::OpFMul, TypeF32(state), coord, width_f32),
 	                                 ConstantF32(state, 0x3f000000u)));
 
-	const auto sampled     = MakeSampledImage(state, mem.resource, mem.sampler);
+	const auto sampled = MakeSampledImage(state, mem.resource,
+	                                     LoadSamplerDescriptor(state, mem.sampler));
 	const auto vector_type = ImageVectorType(state, numeric_class, 4);
 	const auto scalar_type = ImageScalarType(state, numeric_class);
 	const auto component =
@@ -558,7 +597,8 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (op == IR::ValueOpcode::ImageQueryLod) {
 		state.builder.RequireCapability(spv::CapabilityImageQuery);
 		const auto dimension = image.dimension;
-		const auto sampled   = MakeSampledImage(state, mem.resource, mem.sampler);
+		const auto sampled = MakeSampledImage(state, mem.resource,
+		                                     LoadSamplerDescriptor(state, mem.sampler));
 		const auto lod       = state.builder.AllocateId();
 		state.builder.AddFunction(
 		    spv::OpImageQueryLod, TypeF32Vector(state, 2), lod, sampled,
@@ -588,7 +628,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		    EmitValueOrDefaultIfCondition(
 		        state, condition, TypeU32Vector(state, 4), ConstantU32CompositeZero(state, 4),
 		        [&]() {
-			        const auto descriptor = LoadSampledImageDescriptor(state, mem.resource);
+			        const auto descriptor = LoadImageDescriptor(state, mem.resource);
 			        const auto color      = state.builder.AllocateId();
 			        const auto coord      = CoordU32(ctx, mem, *address, dimension);
 			        if (dimension_info.multisampled != 0u) {
@@ -612,7 +652,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto dimension  = image.dimension;
 		EmitIfCondition(state, ctx.Arg(inst, 3), [&]() {
 			const auto mip_lod =
-			    state.program.info.images[mem.resource].mip_mode == IR::ImageMipMode::DynamicStorage
+			    state.program.info.images[mem.resource].mip_mode == IR::ImageMipMode::Dynamic
 			        ? LodU32(ctx, mem, *address, dimension)
 			        : 0u;
 			const auto coord = CoordU32(ctx, mem, *address, dimension);
@@ -635,13 +675,6 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		if (op == IR::ValueOpcode::ImageGatherRaw) {
 			const auto coord = CoordF32(ctx, mem, *address, layout.coord,
 			                            dimension_info.coordinate_components, image.cube);
-			if (HasFlag(mem, Decoder::ImageSampleFlagLod)) {
-				static std::atomic_flag warned = ATOMIC_FLAG_INIT;
-				if (!warned.test_and_set(std::memory_order_relaxed)) {
-					std::fputs("Warning: approximating IMAGE_GATHER4_L at mip level 0; explicit LOD is ignored.\n",
-					           stderr);
-				}
-			}
 			if (dimension == ImageDimension::Dim1D) {
 				if (dref || !HasFlag(mem, Decoder::ImageSampleFlagLevelZero) ||
 				    HasFlag(mem, Decoder::ImageSampleFlagOffset) ||
@@ -658,44 +691,48 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				ctx.Fail(inst, "has an unsupported 1D-array gather");
 				return;
 			}
-			const auto            sampled = MakeSampledImage(state, mem.resource, mem.sampler);
-			const auto            sample  = state.builder.AllocateId();
-			std::vector<uint32_t> words;
+			const auto result_numeric_class = dref ? Prospero::TextureNumericClass::Float
+			                                       : numeric_class;
+			const auto result_type = ImageVectorType(state, result_numeric_class, 4);
+			uint32_t component_or_dref;
 			if (dref) {
-				auto dref_value = ZeroF32(state);
-				if (layout.dref != NoImageComponent) {
-					dref_value = AddressF32(ctx, mem, *address, layout.dref);
-				}
-				words = {spv::OpImageDrefGather,
-				         TypeF32Vector(state, 4),
-				         sample,
-				         sampled,
-				         coord,
-				         dref_value};
+				component_or_dref = layout.dref != NoImageComponent
+				                        ? AddressF32(ctx, mem, *address, layout.dref)
+				                        : ZeroF32(state);
 			} else {
-				uint32_t component = 0;
-				if (ImageConversionFormat(state, mem).format == Prospero::BufferFormat::kInvalid) {
-					component = ImageGatherComponent(mem.dmask);
-				}
-				words = {spv::OpImageGather,
-				         ImageVectorType(state, numeric_class, 4),
-				         sample,
-				         sampled,
-				         coord,
-				         ConstantU32(state, component)};
+				component_or_dref = ConstantU32(
+				    state, ImageConversionFormat(state, mem).format == Prospero::BufferFormat::kInvalid
+				               ? ImageGatherComponent(mem.dmask) : 0u);
 			}
+			uint32_t operand_mask = 0;
+			uint32_t offset = 0;
 			if (HasFlag(mem, Decoder::ImageSampleFlagGatherHorizontal)) {
-				words.push_back(spv::ImageOperandsConstOffsetsMask);
-				words.push_back(HorizontalOffsets(state, dimension));
+				operand_mask = spv::ImageOperandsConstOffsetsMask;
+				offset = HorizontalOffsets(state, dimension);
 			} else if (layout.offset != NoImageComponent) {
-				words.push_back(spv::ImageOperandsOffsetMask);
-				words.push_back(PackedOffset(ctx, mem, *address, layout, dimension));
-			}
-			state.builder.AddFunction(words);
-			auto result_numeric_class = numeric_class;
-			if (dref) {
-				result_numeric_class = Prospero::TextureNumericClass::Float;
+				operand_mask = spv::ImageOperandsOffsetMask;
+				offset = PackedOffset(ctx, mem, *address, layout, dimension);
 			}
+			const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
+			const auto EmitGather = [&](uint32_t mip) {
+				const auto sampled = MakeSampledImage(state, mem.resource, sampler_id, mip);
+				const auto sample = state.builder.AllocateId();
+				std::vector<uint32_t> words {
+				    static_cast<uint32_t>(dref ? spv::OpImageDrefGather : spv::OpImageGather), result_type,
+				    sample, sampled, coord, component_or_dref};
+				if (operand_mask != 0u) {
+					words.push_back(operand_mask);
+					words.push_back(offset);
+				}
+				state.builder.AddFunction(words);
+				return sample;
+			};
+			const auto sample = image.mip_mode == IR::ImageMipMode::Dynamic && image.mip_count > 1u
+			                        ? EmitIndexSwitch(
+			                              state, GatherMip(ctx, inst, mem, *address, layout,
+			                                               image.mip_count),
+			                              image.mip_count, result_type, EmitGather)
+			                        : EmitGather(0);
 			ctx.Define(inst, ResultVector(ctx, UnpackImageGather(ctx, mem, sample),
 			                              result_numeric_class, false, mem, true));
 			return;
@@ -738,13 +775,14 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			operand_mask |= spv::ImageOperandsBiasMask;
 			operands.push_back(AddressF32(ctx, mem, *address, layout.bias));
 		}
-		const auto EmitSample = [&](uint32_t resource) {
+		const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
+		const auto EmitSample = [&](uint32_t resource, uint32_t array_index = 0u) {
 			const auto& candidate = state.program.info.images[resource];
 			const auto coord =
 			    CoordF32(ctx, mem, *address, layout.coord,
 			             ImageDimensionInfoFor(candidate.dimension).coordinate_components,
 			             candidate.cube);
-			const auto            sampled = MakeSampledImage(state, resource, mem.sampler);
+			const auto sampled = MakeSampledImage(state, resource, sampler_id, 0u, array_index);
 			const auto            sample  = state.builder.AllocateId();
 			std::vector<uint32_t> sample_operands {result_type, sample, sampled, coord};
 			if (dref) {
@@ -770,7 +808,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto* source = image.source < state.program.descriptor_sources.size()
 		                         ? &state.program.descriptor_sources[image.source]
 		                         : nullptr;
-		if (handle == nullptr || source == nullptr || !source->indirect_image.has_value() ||
+		if (handle == nullptr || source == nullptr || !source->indirect_descriptor.has_value() ||
 		    handle->NumArgs() == 0u) {
 			ctx.Fail(inst, "has invalid indirect image key provenance");
 			return;
@@ -781,55 +819,42 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			ctx.Fail(inst, "has no indirect image runtime mapping");
 			return;
 		}
-		const auto LoadMapping = [&](uint32_t index) {
-			const auto pointer = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state),
-			                          pointer, state.flattened_srt_variable, ConstantU32(state, 0),
-			                          index);
-			const auto value = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpLoad, TypeU32(state), value, pointer);
-			return value;
-		};
-		const auto mapping  = ConstantU32(state, image.indirect_mapping_offset);
-		auto       low      = ConstantU32(state, 0u);
-		auto       high     = LoadMapping(mapping);
-		auto       selected = ConstantU32(state, 0u);
-		for (uint32_t iteration = 0; iteration < image.indirect_search_iterations;
-		     iteration++) {
-			const auto active = Binary(state, spv::OpULessThan, TypeBool(state), low, high);
-			const auto mid    = Binary(state, spv::OpShiftRightLogical, TypeU32(state),
-			                           Binary(state, spv::OpIAdd, TypeU32(state), low, high),
-			                           ConstantU32(state, 1u));
-			const auto probe  = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), probe, active, mid,
-			                          ConstantU32(state, 0u));
-			const auto entry = Binary(state, spv::OpIAdd, TypeU32(state), mapping,
-			                          Binary(state, spv::OpIAdd, TypeU32(state),
-			                                 Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
-			                                        probe, ConstantU32(state, 1u)),
-			                                 ConstantU32(state, 1u)));
-			const auto mapped_key = LoadMapping(entry);
-			const auto candidate  = LoadMapping(
-			    Binary(state, spv::OpIAdd, TypeU32(state), entry, ConstantU32(state, 1u)));
-			const auto equal = Binary(state, spv::OpIEqual, TypeBool(state), mapped_key, key);
-			const auto match = Binary(state, spv::OpLogicalAnd, TypeBool(state), active, equal);
-			const auto next_selected = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), next_selected, match,
-			                          candidate, selected);
-			selected              = next_selected;
-			const auto less = Binary(state, spv::OpULessThan, TypeBool(state), mapped_key, key);
-			const auto take_upper = Binary(state, spv::OpLogicalAnd, TypeBool(state), active, less);
-			const auto take_lower = Binary(state, spv::OpLogicalAnd, TypeBool(state), active,
-			                               Unary(state, spv::OpLogicalNot, TypeBool(state), less));
-			const auto next_low   = state.builder.AllocateId();
-			state.builder.AddFunction(
-			    spv::OpSelect, TypeU32(state), next_low, take_upper,
-			    Binary(state, spv::OpIAdd, TypeU32(state), mid, ConstantU32(state, 1u)), low);
-			low                  = next_low;
-			const auto next_high = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), next_high, take_lower, mid,
-			                          high);
-			high = next_high;
+		const auto selected = EmitIndirectResourceIndex(
+		    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
+		const auto kind = *IR::DescriptorBindingForImage(image);
+		const auto& binding = *IR::FindBinding(state.program.bindings, kind);
+		uint32_t first_child = 0;
+		bool homogeneous = true;
+		for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
+			const auto resource = image.indirect_resources[ordinal];
+			const auto& candidate = state.program.info.images[resource];
+			if (candidate.dimension != image.dimension || candidate.cube != image.cube ||
+			    IR::DescriptorBindingForImage(candidate) != kind ||
+			    candidate.mip_count != 1u) {
+				homogeneous = false;
+				break;
+			}
+			if (ordinal == 1u) first_child = ResourceForDescriptor(state, kind, resource);
+			const auto child_slot = first_child + ordinal - 1u;
+			if (ordinal > 1u && (child_slot >= binding.resources.size() ||
+			                     binding.resources[child_slot] != resource)) {
+				homogeneous = false;
+				break;
+			}
+		}
+		if (homogeneous) {
+			const auto child_index = Binary(state, spv::OpIAdd, TypeU32(state), selected,
+			                                ConstantU32(state, first_child - 1u));
+			const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
+			                            ConstantU32(state, 0u));
+			const auto array_index = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpSelect, TypeU32(state), array_index, is_root,
+			                          ConstantU32(state, ResourceForDescriptor(state, kind, mem.resource)),
+			                          child_index);
+			const auto sample = EmitSample(mem.resource, array_index);
+			const auto result = dref ? sample : UnpackImageTexel(ctx, mem, sample);
+			ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
+			return;
 		}
 		const auto            default_label = state.builder.AllocateId();
 		const auto            merge_label   = state.builder.AllocateId();
@@ -865,12 +890,16 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 	}
 	if (image_info.access == IR::ImageAccess::Atomic) {
 		const auto dimension = image.dimension;
-		ctx.Define(inst, EmitValueOrZeroIfCondition(state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
+		const auto result_type = TypeId(state, inst.GetType());
+		const auto zero = image.atomic64 ? ConstantU64(state, 0) : ConstantU32(state, 0);
+		ctx.Define(inst, EmitValueOrDefaultIfCondition(
+		                     state, ctx.Arg(inst, inst.NumArgs() - 1), result_type, zero, [&]() {
 			           const auto pointer      = state.builder.AllocateId();
 			           const auto pointer_type = state.builder.Type(
-			               spv::OpTypePointer, spv::StorageClassImage, TypeU32(state));
+			               spv::OpTypePointer, spv::StorageClassImage,
+			               image.atomic64 ? TypeScalarU64(state) : TypeU32(state));
 			           state.builder.AddFunction(spv::OpImageTexelPointer, pointer_type, pointer,
-			                                     StorageImageDescriptorPointer(state, mem.resource),
+			                                     ImageDescriptorPointer(state, mem.resource),
 			                                     CoordU32(ctx, mem, *address, dimension),
 			                                     ConstantU32(state, 0));
 			           if (op == IR::ValueOpcode::ImageAtomicFMin32 ||

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

@@ -179,6 +179,7 @@ inline constexpr auto EmitFPFloor32     = EmitGlsl<GLSLstd450Floor, IR::Type::F3
 inline constexpr auto EmitFPCeil32      = EmitGlsl<GLSLstd450Ceil, IR::Type::F32, uint32_t>;
 inline constexpr auto EmitFPTrunc32     = EmitGlsl<GLSLstd450Trunc, IR::Type::F32, uint32_t>;
 inline constexpr auto EmitFPFract32     = EmitGlsl<GLSLstd450Fract, IR::Type::F32, uint32_t>;
+inline constexpr auto EmitFPFract64     = EmitGlsl<GLSLstd450Fract, IR::Type::F64, uint32_t>;
 uint32_t              EmitFPSin(EmitterState& state, uint32_t arg0);
 uint32_t              EmitFPCos(EmitterState& state, uint32_t arg0);
 #undef EMIT_NATIVE
@@ -188,16 +189,17 @@ void                  EmitVoid(ValueEmitContext&);
 inline constexpr auto EmitReference    = EmitVoid;
 inline constexpr auto EmitReferenceU32 = EmitVoid;
 inline constexpr auto EmitControlNop   = EmitVoid;
-inline constexpr auto EmitWaitcnt      = EmitVoid;
 inline constexpr auto EmitSendmsg      = EmitVoid;
 inline constexpr auto EmitTtraceData   = EmitVoid;
 inline constexpr auto EmitInstPrefetch = EmitVoid;
 void                  EmitBarrier(EmitterState& state);
+void                  EmitStoreCompletion(EmitterState& state);
 void                  EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetUserData(EmitterState& state, IR::ScalarReg reg);
+uint32_t              EmitGetDispatchThreadExtent(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetBuiltin(ValueEmitContext& ctx, IR::Value kind, IR::Value index);
 uint32_t              EmitUndefU1(EmitterState& state, const IR::Inst& inst);
 inline constexpr auto EmitUndefU8  = EmitUndefU1;
@@ -230,10 +232,12 @@ void                  EmitStoreMemory(ValueEmitContext& ctx, const IR::Inst& ins
 uint32_t              EmitAtomic32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
+void                  EmitSharedAtomicMaskedOr32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitSharedIncDec(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitPermuteU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBpermuteU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadConst(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitReadConstBuffer(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -317,6 +321,7 @@ inline constexpr auto EmitImageAtomicSMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicSMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMax32    = EmitImage;
+inline constexpr auto EmitImageAtomicUMax64    = EmitImage;
 inline constexpr auto EmitImageAtomicAnd32     = EmitImage;
 inline constexpr auto EmitImageAtomicOr32      = EmitImage;
 inline constexpr auto EmitImageAtomicXor32     = EmitImage;

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

@@ -19,6 +19,7 @@
 #include <map>
 #include <spirv/unified1/GLSL.std.450.h>
 #include <string>
+#include <type_traits>
 #include <unordered_map>
 #include <utility>
 #include <vector>
@@ -60,6 +61,7 @@ const ImageDimensionInfo& ImageDimensionInfoFor(ImageDimension dimension);
 struct SpirvRequirements {
 	bool bvh                          = false;
 	bool subgroup_ballot              = false;
+	bool subgroup_barrier             = false;
 	bool subgroup_shuffle             = false;
 	bool subgroup_local_invocation_id = false;
 	bool compute_derivatives          = false;
@@ -68,6 +70,8 @@ struct SpirvRequirements {
 	bool function_scratch             = false;
 	bool pixel_valid_mask             = false;
 	bool buffer_int64_atomics         = false;
+	bool buffer_u8                    = false;
+	bool buffer_u16                   = false;
 	bool shared_int64_atomics         = false;
 	bool coherent_buffers             = false;
 	bool float64                      = false;
@@ -106,6 +110,8 @@ struct EmitterState {
 	uint32_t                                         lane_count              = 1;
 	uint32_t                                         lane_half               = 0;
 	uint32_t                                         storage_buffer_variable = 0;
+	uint32_t                                         storage_buffer_u8_variable = 0;
+	uint32_t                                         storage_buffer_u16_variable = 0;
 	uint32_t                                         storage_buffer_u64_variable = 0;
 	std::array<uint32_t, IR::ShaderInfo::MaxBuffers> memory_byte_offsets {};
 	uint32_t                                         bda_pagetable_variable  = 0;
@@ -166,10 +172,9 @@ uint32_t TypeI32Vector(EmitterState& state, uint32_t components);
 uint32_t TypeF32Vector(EmitterState& state, uint32_t components);
 uint32_t TypePointer(EmitterState& state, spv::StorageClass storage_class, uint32_t pointee);
 uint32_t TypeFunction(EmitterState& state);
-uint32_t TypeStorageBufferPointer(EmitterState& state);
-uint32_t TypeStorageBufferElementPointer(EmitterState& state);
-uint32_t TypeStorageBufferU64Pointer(EmitterState& state);
-uint32_t TypeStorageBufferU64ElementPointer(EmitterState& state);
+uint32_t TypeStorageBufferElement(EmitterState& state, uint32_t bits);
+uint32_t TypeStorageBufferPointer(EmitterState& state, uint32_t bits = 32);
+uint32_t TypeStorageBufferElementPointer(EmitterState& state, uint32_t bits = 32);
 uint32_t TypePhysicalU32Pointer(EmitterState& state);
 uint32_t TypePushConstantElementPointer(EmitterState& state);
 uint32_t TypeU32ArrayPointer(EmitterState& state, spv::StorageClass storage_class, uint32_t dwords);
@@ -314,13 +319,16 @@ uint32_t ImageType(EmitterState& state, const IR::ImageResource& image);
 
 uint32_t ImageViewSizeType(EmitterState& state, ImageDimension dimension);
 
-uint32_t LoadSampledImageDescriptor(EmitterState& state, uint32_t resource);
+uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t mip = 0,
+                                uint32_t array_index = 0);
+
+uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip = 0,
+                             uint32_t array_index = 0);
 
 uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler);
 
-uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler);
-
-uint32_t StorageImageDescriptorPointer(EmitterState& state, uint32_t resource);
+uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id,
+                          uint32_t mip = 0, uint32_t array_index = 0);
 
 void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_lod, uint32_t coord,
                            uint32_t texel);
@@ -346,7 +354,8 @@ uint32_t DefineInterfaceVariable(EmitterState& state, uint32_t type, spv::Storag
 void     DefineModule(EmitterState& state);
 void     DefineTessellationInterfaces(EmitterState& state);
 void     DefineTessellationExecutionModes(EmitterState& state);
-void     DefineMeshOutputs(EmitterState& state);
+void     DefineMeshOutputs(EmitterState& state, uint32_t clip_distance_count,
+                           uint32_t cull_distance_count);
 void     EmitMeshEntryPoint(EmitterState& state);
 void     EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t MeshOutputPointer(EmitterState& state, IR::StageOutputKind kind, uint32_t index = 0);
@@ -390,6 +399,7 @@ uint32_t EmitShaderDataDwordLoad(EmitterState& state, uint32_t dword_index);
 uint32_t StorageBufferPackedStride(const EmitterState& state, const IR::MemoryInfo& mem);
 
 Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::MemoryInfo& mem);
+uint32_t StorageBufferElementBits(const IR::Program& program, const IR::MemoryInfo& mem);
 
 void EmitMemoryOffsets(EmitterState& state);
 
@@ -400,21 +410,19 @@ struct MemoryResourceAccess {
 	IR::ResourceKind      kind             = IR::ResourceKind::None;
 	uint32_t              object_pointer   = 0;
 	uint32_t              length           = 0;
-	uint32_t              index_offset     = 0;
 	uint32_t              byte_offset      = 0;
-	bool                  add_index_offset = false;
+	uint32_t              element_bits     = 32;
 	spv::MemoryAccessMask memory_access    = spv::MemoryAccessMaskNone;
 };
 
+uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t mapping_offset,
+                                   uint32_t search_iterations, uint32_t default_resource);
 MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::MemoryInfo& mem);
 
 MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState&         state,
                                                         const IR::MemoryInfo& mem,
                                                         uint32_t variable, uint32_t pointer_type);
 
-uint32_t EmitMemoryElementIndex(EmitterState& state, const MemoryResourceAccess& access,
-                                uint32_t raw_index);
-
 uint32_t EmitMemoryElementInBounds(EmitterState& state, const MemoryResourceAccess& access,
                                    uint32_t index);
 
@@ -514,6 +522,45 @@ uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value);
 uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high);
 
 // These templates accept local lambdas from several emitter translation units.
+template <typename Fn>
+auto EmitIndexSwitch(EmitterState& state, uint32_t index, uint32_t count,
+                        uint32_t result_type, Fn&& emit) {
+	constexpr bool has_result = !std::is_void_v<std::invoke_result_t<Fn, uint32_t>>;
+	EXIT_IF(count == 0u);
+	const auto merge_label = state.builder.AllocateId();
+	std::vector<uint32_t> labels(count);
+	std::vector<uint32_t> words {spv::OpSwitch, index, merge_label};
+	for (uint32_t item = 0; item < count; item++) {
+		labels[item] = state.builder.AllocateId();
+		words.push_back(item);
+		words.push_back(labels[item]);
+	}
+	if constexpr (has_result) {
+		words[2] = labels.front();
+	}
+	state.builder.AddFunction(spv::OpSelectionMerge, merge_label, spv::SelectionControlMaskNone);
+	state.builder.AddFunction(words);
+	std::vector<uint32_t> phi;
+	if constexpr (has_result) {
+		phi = {spv::OpPhi, result_type, state.builder.AllocateId()};
+	}
+	for (uint32_t item = 0; item < count; item++) {
+		EmitLabel(state, labels[item]);
+		if constexpr (has_result) {
+			phi.push_back(emit(item));
+			phi.push_back(state.current_label);
+		} else {
+			emit(item);
+		}
+		state.builder.AddFunction(spv::OpBranch, merge_label);
+	}
+	EmitLabel(state, merge_label);
+	if constexpr (has_result) {
+		state.builder.AddFunction(phi);
+		return phi[2];
+	}
+}
+
 template <typename Fn>
 void EmitIfCondition(EmitterState& state, uint32_t condition, Fn&& fn) {
 	const auto then_label  = state.builder.AllocateId();

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

@@ -4,6 +4,8 @@
 #include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 
 #include <algorithm>
+#include <bit>
+#include <utility>
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 namespace {
@@ -233,7 +235,15 @@ uint32_t LoadBda(ValueEmitContext& ctx, uint32_t address, uint32_t active, uint3
 
 uint32_t ByteAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem) {
 	if (mem.kind == IR::ResourceKind::Buffer) {
-		return BufferByteAddress(ctx, inst, mem);
+		const auto resource = ResourceForDescriptor(ctx.state, IR::DescriptorBindingKind::Buffers,
+		                                            mem.resource);
+		const auto prefix = ctx.state.memory_byte_offsets[resource];
+		const auto address = Binary(ctx.state, spv::OpIAdd, TypeU32(ctx.state),
+		                            BufferByteAddress(ctx, inst, mem), prefix);
+		// The host binding prefix must not wrap an out-of-range guest address into the buffer.
+		return Select(ctx.state, TypeU32(ctx.state),
+		              Binary(ctx.state, spv::OpUGreaterThanEqual, TypeBool(ctx.state), address, prefix),
+		              address, ConstantU32(ctx.state, UINT32_MAX));
 	}
 	if (mem.kind == IR::ResourceKind::Lds || mem.kind == IR::ResourceKind::Gds) {
 		if (mem.offset == 0u) {
@@ -249,20 +259,14 @@ uint32_t ByteAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memo
 }
 
 uint32_t DwordIndex(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem) {
+	auto address = ByteAddress(ctx, inst, mem);
+	if (mem.kind == IR::ResourceKind::Lds || mem.kind == IR::ResourceKind::Gds) {
+		// RDNA2 DS region addresses retain bits [15:2] after adding the byte offset.
+		address = Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state), address,
+		                 ConstantU32(ctx.state, 0xffffu));
+	}
 	return Binary(ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state),
-	              ByteAddress(ctx, inst, mem), ConstantU32(ctx.state, 2));
-}
-
-struct PreparedMemoryElement {
-	MemoryResourceAccess resource;
-	uint32_t             index = 0;
-};
-
-PreparedMemoryElement PrepareMemoryElement(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
-                                           uint32_t raw_index) {
-	auto       resource = PrepareMemoryResourceAccess(ctx.state, mem);
-	const auto index    = EmitMemoryElementIndex(ctx.state, resource, raw_index);
-	return {.resource = resource, .index = index};
+	              address, ConstantU32(ctx.state, 2));
 }
 
 uint32_t LoadWordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess& resource,
@@ -273,7 +277,7 @@ uint32_t LoadSubwordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess&
 
 uint32_t LoadWordPrepared(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
                           const MemoryResourceAccess& resource) {
-	const auto index = EmitMemoryElementIndex(ctx.state, resource, DwordIndex(ctx, inst, mem));
+	const auto index = DwordIndex(ctx, inst, mem);
 	return EmitValueOrZeroIfCondition(
 	    ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index),
 	    [&]() { return LoadWordInBounds(ctx, resource, index); });
@@ -289,10 +293,10 @@ uint32_t LoadWord(ValueEmitContext& ctx, const IR::Inst& inst, IR::MemoryInfo me
 uint32_t LoadSubwordPrepared(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
                              const MemoryResourceAccess& resource, uint32_t bits,
                              bool sign_extend) {
-	const auto address   = ByteAddress(ctx, inst, mem);
-	const auto raw_index = Binary(ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state), address,
-	                              ConstantU32(ctx.state, 2));
-	const auto index     = EmitMemoryElementIndex(ctx.state, resource, raw_index);
+	const auto address = ByteAddress(ctx, inst, mem);
+	const auto index = Binary(
+	    ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state), address,
+	    ConstantU32(ctx.state, std::countr_zero(resource.element_bits / 8u)));
 	return EmitValueOrZeroIfCondition(
 	    ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index), [&]() {
 		    return LoadSubwordInBounds(ctx, resource, address, index, bits, sign_extend);
@@ -310,15 +314,23 @@ uint32_t LoadWordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess& res
 
 uint32_t LoadSubwordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess& resource,
                              uint32_t address, uint32_t index, uint32_t bits, bool sign_extend) {
-	const auto word = LoadWordInBounds(ctx, resource, index);
-	const auto byte  = Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state), address,
-	                          ConstantU32(ctx.state, 3));
-	const auto shift = Binary(ctx.state, spv::OpShiftLeftLogical, TypeU32(ctx.state), byte,
-	                          ConstantU32(ctx.state, 3));
-	const auto value =
-	    Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state),
-	           Binary(ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state), word, shift),
-	           ConstantU32(ctx.state, bits == 8u ? 0xffu : 0xffffu));
+	uint32_t value;
+	if (resource.element_bits < 32u) {
+		const auto loaded = ctx.state.builder.AllocateId();
+		ctx.state.builder.AddFunction(spv::OpLoad, TypeStorageBufferElement(ctx.state, bits), loaded,
+		                              EmitMemoryElementPointer(ctx.state, resource, index),
+		                              resource.memory_access);
+		value = Unary(ctx.state, spv::OpUConvert, TypeU32(ctx.state), loaded);
+	} else {
+		const auto word = LoadWordInBounds(ctx, resource, index);
+		const auto byte = Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state), address,
+		                         ConstantU32(ctx.state, 3));
+		const auto shift = Binary(ctx.state, spv::OpShiftLeftLogical, TypeU32(ctx.state), byte,
+		                          ConstantU32(ctx.state, 3));
+		value = Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state),
+		               Binary(ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state), word, shift),
+		               ConstantU32(ctx.state, bits == 8u ? 0xffu : 0xffffu));
+	}
 	if (!sign_extend) return value;
 	const auto left = Binary(ctx.state, spv::OpShiftLeftLogical, TypeU32(ctx.state), value,
 	                         ConstantU32(ctx.state, 32u - bits));
@@ -339,14 +351,6 @@ Prospero::BufferFormat BufferFormat(const ValueEmitContext& ctx, const IR::Memor
 	                 : StorageBufferFormat(ctx.state, mem);
 }
 
-IR::MemoryInfo RebaseFormattedComponent(IR::MemoryInfo mem, const Format::BufferFormatInfo& info,
-                                        uint32_t component) {
-	mem.offset += Format::GetFormatComponentByteOffset(info, component);
-	mem.data_dwords     = 1u;
-	mem.component_index = component;
-	return mem;
-}
-
 IR::MemoryInfo RebaseRawComponent(IR::MemoryInfo mem, uint32_t component) {
 	mem.offset += component * 4u;
 	mem.data_dwords     = 1u;
@@ -381,19 +385,17 @@ uint32_t FormattedConstant(ValueEmitContext& ctx, const Format::BufferFormatInfo
 }
 
 template <typename LoadWordFn, typename LoadSubwordFn>
-uint32_t LoadFormattedComponent(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
-                                const Format::BufferFormatInfo& info,
-                                uint32_t output_component, LoadWordFn&& load_word,
+uint32_t LoadFormattedComponent(ValueEmitContext& ctx, const Format::BufferFormatInfo& info,
+                                FormattedSource source, LoadWordFn&& load_word,
                                 LoadSubwordFn&& load_subword) {
-	const auto source = ResolveFormattedSource(ctx, mem, info, output_component);
 	if (source.kind != FormattedSourceKind::Memory) {
 		return FormattedConstant(ctx, info, source.kind);
 	}
 	const auto component = source.component;
 	const auto bits      = info.component_bits[component];
-	uint32_t   raw       = 0;
-	if (info.packed_bitfield) {
-		raw = load_word(component);
+	auto raw = info.packed_bitfield || bits == 32u ? load_word(component)
+	                                                : load_subword(component, bits);
+	if (info.packed_bitfield || (bits < 32u && IsSignedFormatComponent(info.type))) {
 		const auto type =
 		    IsSignedFormatComponent(info.type) ? TypeI32(ctx.state) : TypeU32(ctx.state);
 		const auto source_value =
@@ -402,49 +404,28 @@ uint32_t LoadFormattedComponent(ValueEmitContext& ctx, const IR::MemoryInfo& mem
 		ctx.state.builder.AddFunction(IsSignedFormatComponent(info.type) ? spv::OpBitFieldSExtract
 		                                                                 : spv::OpBitFieldUExtract,
 		                              type, extracted, source_value,
-		                              ConstantU32(ctx.state, info.component_bit_offset[component]),
+		                              ConstantU32(ctx.state, info.packed_bitfield
+		                                                         ? info.component_bit_offset[component]
+		                                                         : 0u),
 		                              ConstantU32(ctx.state, bits));
 		raw = type == TypeI32(ctx.state)
 		          ? Unary(ctx.state, spv::OpBitcast, TypeU32(ctx.state), extracted)
 		          : extracted;
-	} else if (bits == 32u) {
-		raw = load_word(component);
-	} else {
-		raw = load_subword(component, bits, IsSignedFormatComponent(info.type));
 	}
 	return NormalizeFormatComponent(ctx.state, info, component, raw);
 }
 
-uint32_t FormattedLoadPrepared(ValueEmitContext& ctx, const IR::Inst& inst,
-                               const IR::MemoryInfo& mem, uint32_t output_component,
-                               const MemoryResourceAccess& resource) {
-	const auto info = Format::GetFormatInfo(BufferFormat(ctx, mem));
-	if (info.type == Format::ComponentType::Unknown) {
-		return LoadWordPrepared(ctx, inst, RebaseRawComponent(mem, output_component), resource);
-	}
-	return LoadFormattedComponent(
-	    ctx, mem, info, output_component,
-	    [&](uint32_t component) {
-		    return LoadWordPrepared(ctx, inst, RebaseFormattedComponent(mem, info, component),
-		                            resource);
-	    },
-	    [&](uint32_t component, uint32_t bits, bool sign_extend) {
-		    return LoadSubwordPrepared(ctx, inst, RebaseFormattedComponent(mem, info, component),
-		                               resource, bits, sign_extend);
-	    });
-}
-
-uint32_t FormattedLoad(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem) {
-	return EmitValueOrZeroIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
-		const auto resource = PrepareMemoryResourceAccess(ctx.state, mem);
-		return FormattedLoadPrepared(ctx, inst, mem, 0u, resource);
-	});
-}
-
 void StoreSubwordInBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
                           const MemoryResourceAccess& resource, uint32_t address, uint32_t index,
                           uint32_t bits, uint32_t data) {
 	const auto pointer = EmitMemoryElementPointer(ctx.state, resource, index);
+	if (resource.element_bits < 32u) {
+		ctx.state.builder.AddFunction(
+		    spv::OpStore, pointer,
+		    Unary(ctx.state, spv::OpUConvert, TypeStorageBufferElement(ctx.state, bits), data),
+		    resource.memory_access);
+		return;
+	}
 	const auto shift   = Binary(ctx.state, spv::OpShiftLeftLogical, TypeU32(ctx.state),
 	                            Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state), address,
 	                                   ConstantU32(ctx.state, 3)),
@@ -472,10 +453,10 @@ void StoreSubwordInBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
 
 void StoreSubwordPrepared(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
                           const MemoryResourceAccess& resource, uint32_t bits, uint32_t data) {
-	const auto address   = ByteAddress(ctx, inst, mem);
-	const auto raw_index = Binary(ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state), address,
-	                              ConstantU32(ctx.state, 2));
-	const auto index     = EmitMemoryElementIndex(ctx.state, resource, raw_index);
+	const auto address = ByteAddress(ctx, inst, mem);
+	const auto index = Binary(
+	    ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state), address,
+	    ConstantU32(ctx.state, std::countr_zero(resource.element_bits / 8u)));
 	EmitIfCondition(
 	    ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index), [&]() {
 		    StoreSubwordInBounds(ctx, mem, resource, address, index, bits, data);
@@ -542,16 +523,58 @@ void StoreLocalFlat(ValueEmitContext& ctx, const IR::Inst& inst) {
 
 void StoreWordPrepared(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
                        const MemoryResourceAccess& resource, uint32_t data) {
-	const auto index = EmitMemoryElementIndex(ctx.state, resource, DwordIndex(ctx, inst, mem));
+	const auto index = DwordIndex(ctx, inst, mem);
 	EmitIfCondition(ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index), [&]() {
 		StoreWordInBounds(ctx, resource, index, data);
 	});
 }
 
 void StoreWord(ValueEmitContext& ctx, const IR::Inst& inst, IR::MemoryInfo mem) {
-	EmitIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
-		const auto resource = PrepareMemoryResourceAccess(ctx.state, mem);
-		StoreWordPrepared(ctx, inst, mem, resource, ctx.Arg(inst, inst.NumArgs() - 2));
+	auto& state = ctx.state;
+	EmitIfCondition(state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
+		const auto data = ctx.Arg(inst, inst.NumArgs() - 2);
+		const auto store = [&](uint32_t resource) {
+			if (mem.kind == IR::ResourceKind::Buffer) {
+				const auto& buffer = state.program.info.buffers[resource];
+				if (buffer.packed_stride == 0u &&
+				    buffer.source < state.program.descriptor_sources.size() &&
+				    state.program.descriptor_sources[buffer.source].indirect_descriptor) return;
+			}
+			mem.resource = resource;
+			const auto access = PrepareMemoryResourceAccess(state, mem);
+			StoreWordPrepared(ctx, inst, mem, access, data);
+		};
+		if (mem.kind != IR::ResourceKind::Buffer ||
+		    state.program.info.buffers[mem.resource].indirect_root != mem.resource) {
+			store(mem.resource);
+			return;
+		}
+		const auto& buffer = state.program.info.buffers[mem.resource];
+		const auto* handle = inst.Arg(0).ResolveInstruction();
+		if (handle == nullptr || buffer.indirect_resources.empty() ||
+		    state.flattened_srt_variable == 0 || buffer.indirect_search_iterations == 0u) {
+			ctx.Fail(inst, "has no indirect buffer runtime mapping");
+			return;
+		}
+		const auto selected = EmitIndirectResourceIndex(
+		    state, ctx.Arg(*handle, 0), buffer.indirect_mapping_offset,
+		    buffer.indirect_search_iterations, UINT32_MAX);
+		const auto merge = state.builder.AllocateId();
+		std::vector<uint32_t> labels(buffer.indirect_resources.size());
+		std::vector<uint32_t> branches {spv::OpSwitch, selected, merge};
+		for (uint32_t i = 0; i < labels.size(); ++i) {
+			labels[i] = state.builder.AllocateId();
+			branches.push_back(i);
+			branches.push_back(labels[i]);
+		}
+		state.builder.AddFunction(spv::OpSelectionMerge, merge, spv::SelectionControlMaskNone);
+		state.builder.AddFunction(branches);
+		for (uint32_t i = 0; i < labels.size(); ++i) {
+			EmitLabel(state, labels[i]);
+			store(buffer.indirect_resources[i]);
+			state.builder.AddFunction(spv::OpBranch, merge);
+		}
+		EmitLabel(state, merge);
 	});
 }
 
@@ -559,10 +582,11 @@ template <typename Fn>
 uint32_t EmitAtomicAccess(ValueEmitContext& ctx, const IR::Inst& inst,
                           const IR::MemoryInfo& mem, Fn&& operation) {
 	return EmitValueOrZeroIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
-		const auto access = PrepareMemoryElement(ctx, mem, DwordIndex(ctx, inst, mem));
+		const auto index = DwordIndex(ctx, inst, mem);
+		const auto resource = PrepareMemoryResourceAccess(ctx.state, mem);
 		return EmitValueOrZeroIfCondition(
-		    ctx.state, EmitMemoryElementInBounds(ctx.state, access.resource, access.index), [&]() {
-			    return operation(EmitMemoryElementPointer(ctx.state, access.resource, access.index));
+		    ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index), [&]() {
+			    return operation(EmitMemoryElementPointer(ctx.state, resource, index));
 		    });
 	});
 }
@@ -597,83 +621,59 @@ uint32_t AtomicDecrement(EmitterState& state, uint32_t old, uint32_t limit) {
 struct PreparedFormattedMemory {
 	Format::BufferFormatInfo info;
 	MemoryResourceAccess     resource;
-	std::array<uint32_t, 4>  addresses {};
-	std::array<uint32_t, 4>  indices {};
+	uint32_t                 base = 0;
 	uint32_t                 in_bounds = 0;
 };
 
-enum class FormattedAccess { Load, Store };
-
 PreparedFormattedMemory PrepareFormattedMemory(ValueEmitContext& ctx, const IR::Inst& inst,
-                                               const IR::MemoryInfo&       mem,
+                                               const IR::MemoryInfo& mem,
                                                const MemoryResourceAccess& resource,
-                                               const Format::BufferFormatInfo& info,
-                                               uint32_t components, FormattedAccess access) {
+                                               const Format::BufferFormatInfo& info) {
 	PreparedFormattedMemory plan;
-	plan.info     = info;
-	plan.resource = resource;
-	std::array<bool, 4> required_components {};
-	if (access == FormattedAccess::Load) {
-		for (uint32_t output = 0; output < components; output++) {
-			const auto source = ResolveFormattedSource(ctx, mem, plan.info, output);
-			if (source.kind == FormattedSourceKind::Memory) {
-				required_components[source.component] = true;
-			}
-		}
-	} else {
-		for (uint32_t component = 0; component < std::min(components, plan.info.component_count);
-		     component++) {
-			required_components[component] = true;
-		}
-	}
-	bool first_bound = true;
-	for (uint32_t component = 0; component < plan.info.component_count; component++) {
-		if (!required_components[component]) continue;
-		const auto byte_offset = Format::GetFormatComponentByteOffset(plan.info, component);
-		bool       reused      = false;
-		for (uint32_t previous = 0; previous < component; previous++) {
-			if (required_components[previous] &&
-			    Format::GetFormatComponentByteOffset(plan.info, previous) == byte_offset) {
-				plan.addresses[component] = plan.addresses[previous];
-				plan.indices[component]   = plan.indices[previous];
-				reused                    = true;
-				break;
-			}
-		}
-		if (reused) continue;
-		const auto component_mem  = RebaseFormattedComponent(mem, plan.info, component);
-		plan.addresses[component] = ByteAddress(ctx, inst, component_mem);
-		const auto raw_index      = Binary(ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state),
-		                                   plan.addresses[component], ConstantU32(ctx.state, 2));
-		plan.indices[component]   = EmitMemoryElementIndex(ctx.state, resource, raw_index);
-		const auto component_bound =
-		    EmitMemoryElementInBounds(ctx.state, resource, plan.indices[component]);
-		if (first_bound) {
-			plan.in_bounds = component_bound;
-			first_bound    = false;
-		} else {
-			plan.in_bounds = AndCondition(ctx.state, plan.in_bounds, component_bound);
-		}
-	}
-	if (first_bound) plan.in_bounds = ConstantBool(ctx.state, true);
+	plan.info      = info;
+	plan.resource  = resource;
+	const auto start = ByteAddress(ctx, inst, mem);
+	// Check the whole, unaligned format span before aligning the actual memory address.
+	const auto end = Binary(ctx.state, spv::OpIAdd, TypeU32(ctx.state), start,
+	                        ConstantU32(ctx.state, info.byte_size - 1u));
+	const auto last_index = Binary(ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state), end,
+	                               ConstantU32(ctx.state, std::countr_zero(resource.element_bits / 8u)));
+	plan.in_bounds = AndCondition(
+	    ctx.state, Binary(ctx.state, spv::OpUGreaterThanEqual, TypeBool(ctx.state), end, start),
+	    EmitMemoryElementInBounds(ctx.state, plan.resource, last_index));
+	plan.base = Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state), start,
+	                   ConstantU32(ctx.state, ~(std::min(4u, info.byte_size) - 1u)));
 	return plan;
 }
 
+std::pair<uint32_t, uint32_t> FormattedComponentAddress(EmitterState& state,
+                                                       const PreparedFormattedMemory& plan,
+                                                       uint32_t component) {
+	// Components advance sequentially from one address; do not reapply buffer swizzling.
+	const auto offset = Format::GetFormatComponentByteOffset(plan.info, component);
+	const auto address = offset == 0u ? plan.base : Binary(
+	    state, spv::OpIAdd, TypeU32(state), plan.base, ConstantU32(state, offset));
+	return {address, Binary(state, spv::OpShiftRightLogical, TypeU32(state), address,
+	                        ConstantU32(state, std::countr_zero(plan.resource.element_bits / 8u)))};
+}
+
 uint32_t LoadFormattedInBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
                                const PreparedFormattedMemory& plan, uint32_t output_component) {
 	return LoadFormattedComponent(
-	    ctx, mem, plan.info, output_component,
+	    ctx, plan.info, ResolveFormattedSource(ctx, mem, plan.info, output_component),
 	    [&](uint32_t component) {
-		    return LoadWordInBounds(ctx, plan.resource, plan.indices[component]);
+		    return LoadWordInBounds(ctx, plan.resource,
+		                            FormattedComponentAddress(ctx.state, plan, component).second);
 	    },
-	    [&](uint32_t component, uint32_t bits, bool sign_extend) {
-		    return LoadSubwordInBounds(ctx, plan.resource, plan.addresses[component],
-		                               plan.indices[component], bits, sign_extend);
+	    [&](uint32_t component, uint32_t bits) {
+		    const auto [address, index] = FormattedComponentAddress(ctx.state, plan, component);
+		    return LoadSubwordInBounds(ctx, plan.resource, address, index, bits, false);
 	    });
 }
 
 uint32_t ConstructU32Composite(EmitterState& state, uint32_t components,
                                const std::array<uint32_t, 4>& values) {
+	if (components == 1u) return values[0];
 	const auto            result = state.builder.AllocateId();
 	std::vector<uint32_t> words {spv::OpCompositeConstruct, TypeU32Composite(state, components),
 	                             result};
@@ -692,6 +692,21 @@ uint32_t FormattedOutOfBoundsValue(ValueEmitContext& ctx, const IR::MemoryInfo&
 	return ConstructU32Composite(ctx.state, components, values);
 }
 
+uint32_t FormattedLoad(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem) {
+	return EmitValueOrZeroIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
+		const auto resource = PrepareMemoryResourceAccess(ctx.state, mem);
+		const auto info = Format::GetFormatInfo(BufferFormat(ctx, mem));
+		if (info.type == Format::ComponentType::Unknown) {
+			return LoadWordPrepared(ctx, inst, RebaseRawComponent(mem, 0u), resource);
+		}
+		const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info);
+		return EmitValueOrDefaultIfCondition(
+		    ctx.state, plan.in_bounds, TypeU32(ctx.state),
+		    FormattedOutOfBoundsValue(ctx, mem, plan, 1u),
+		    [&]() { return LoadFormattedInBounds(ctx, mem, plan, 0u); });
+	});
+}
+
 uint32_t EncodeFormattedStoreComponent(ValueEmitContext& ctx,
                                        const Format::BufferFormatInfo& info,
                                        uint32_t component, uint32_t data) {
@@ -722,8 +737,7 @@ void StoreFormattedPrepared(ValueEmitContext& ctx, const IR::Inst& inst,
                              const Format::BufferFormatInfo& info, uint32_t data,
                              uint32_t components) {
 	// RDNA2 formatted stores transfer the entire format, zero-filling missing VGPRs.
-	const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info, info.component_count,
-	                                         FormattedAccess::Store);
+	const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info);
 	EmitIfCondition(ctx.state, plan.in_bounds, [&]() {
 		auto packed = ConstantU32(ctx.state, 0);
 		for (uint32_t component = 0; component < info.component_count; component++) {
@@ -743,14 +757,19 @@ void StoreFormattedPrepared(ValueEmitContext& ctx, const IR::Inst& inst,
 				    ctx.state, packed, value,
 				    ConstantU32(ctx.state, info.component_bit_offset[component]),
 				    ConstantU32(ctx.state, bits));
-			} else if (bits == 8u || bits == 16u) {
-				StoreSubwordInBounds(ctx, mem, resource, plan.addresses[component],
-				                     plan.indices[component], bits, value);
 			} else {
-				StoreWordInBounds(ctx, resource, plan.indices[component], value);
+				const auto [address, index] = FormattedComponentAddress(ctx.state, plan, component);
+				if (bits == 8u || bits == 16u) {
+					StoreSubwordInBounds(ctx, mem, plan.resource, address, index, bits, value);
+				} else {
+					StoreWordInBounds(ctx, plan.resource, index, value);
+				}
 			}
 		}
-		if (info.packed_bitfield) StoreWordInBounds(ctx, resource, plan.indices[0], packed);
+		if (info.packed_bitfield) {
+			StoreWordInBounds(ctx, plan.resource,
+			                  FormattedComponentAddress(ctx.state, plan, 0u).second, packed);
+		}
 	});
 }
 
@@ -767,78 +786,193 @@ void FormattedStore(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memor
 	});
 }
 
-uint32_t LoadIndirectBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t components) {
-	auto&       state   = ctx.state;
-	const auto& handle  = *inst.Arg(0).ResolveInstruction();
-	const auto  word1   = ctx.Arg(handle, 1);
-	const auto  records = ctx.Arg(handle, 2);
-	const auto  word3   = ctx.Arg(handle, 3);
-	const auto  field   = [&](uint32_t word, uint32_t first, uint32_t count) {
-		return EmitBitFieldUExtract(state, word, ConstantU32(state, first),
-		                            ConstantU32(state, count));
+uint32_t LoadIndirectScalarBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto&       state  = ctx.state;
+	const auto& handle = *inst.Arg(0).ResolveInstruction();
+	const auto  word1  = ctx.Arg(handle, 1);
+	const auto  stride = EmitBitFieldUExtract(state, word1, ConstantU32(state, 16),
+	                                         ConstantU32(state, 14));
+	const auto size = Binary(
+	    state, spv::OpIMul, TypeScalarU64(state),
+	    Unary(state, spv::OpUConvert, TypeScalarU64(state),
+	          EmitUMax32(state, stride, ConstantU32(state, 1))),
+	    Unary(state, spv::OpUConvert, TypeScalarU64(state), ctx.Arg(handle, 2)));
+	const auto offset = Binary(
+	    state, spv::OpIAdd, TypeScalarU64(state),
+	    Unary(state, spv::OpUConvert, TypeScalarU64(state),
+	          Binary(state, spv::OpBitwiseAnd, TypeU32(state), ctx.Arg(inst, 1),
+	                 ConstantU32(state, ~3u))),
+	    ConstantDeviceAddress(state, ctx.Memory(inst).offset & ~3u));
+	const auto end = Binary(state, spv::OpIAdd, TypeScalarU64(state), offset,
+	                        ConstantDeviceAddress(state, sizeof(uint32_t)));
+	return EmitValueOrZeroIfCondition(
+	    state, Binary(state, spv::OpULessThanEqual, TypeBool(state), end, size), [&]() {
+		    const auto base = DeviceAddressFromWords(
+		        state, Binary(state, spv::OpBitwiseAnd, TypeU32(state), ctx.Arg(handle, 0),
+		                      ConstantU32(state, ~3u)),
+		        Binary(state, spv::OpBitwiseAnd, TypeU32(state), word1, ConstantU32(state, 0xffffu)));
+		    return LoadBdaDword(ctx, Binary(state, spv::OpIAdd, TypeScalarU64(state), base, offset));
+	    });
+}
+
+struct IndirectBufferAccess {
+	uint32_t address;
+	uint32_t offset;
+	uint32_t stride;
+	uint32_t swizzle;
+	uint32_t format;
+	uint32_t selector;
+	uint32_t mode;
+	uint32_t records;
+	uint32_t index_in_bounds;
+	uint32_t scalar_in_bounds;
+	uint32_t raw_records;
+	uint32_t raw_index_in_bounds;
+};
+
+IndirectBufferAccess PrepareIndirectBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto&       state  = ctx.state;
+	const auto& handle = *inst.Arg(0).ResolveInstruction();
+	const auto word1   = ctx.Arg(handle, 1);
+	const auto records = ctx.Arg(handle, 2);
+	const auto word3   = ctx.Arg(handle, 3);
+	const auto field = [&](uint32_t word, uint32_t first, uint32_t count) {
+		return EmitBitFieldUExtract(state, word, ConstantU32(state, first), ConstantU32(state, count));
 	};
 	const auto nonzero = [&](uint32_t value) {
 		return Binary(state, spv::OpINotEqual, TypeBool(state), value, ConstantU32(state, 0));
 	};
-	const auto has_dword = [&](uint32_t offset, uint32_t size) {
-		return AndCondition(
-		    state,
-		    Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), size, ConstantU32(state, 4)),
-		    Binary(state, spv::OpULessThanEqual, TypeBool(state), offset,
-		           Binary(state, spv::OpISub, TypeU32(state), size, ConstantU32(state, 4))));
-	};
-	const auto stride       = field(word1, 16, 14);
-	const auto swizzle      = AndCondition(state, nonzero(stride), nonzero(field(word1, 31, 1)));
-	const auto index_stride = field(word3, 21, 2);
-	const auto add_tid      = nonzero(field(word3, 23, 1));
-	const auto index =
-	    Binary(state, spv::OpIAdd, TypeU32(state), ctx.Arg(inst, 1),
-	           Select(state, TypeU32(state), add_tid, BufferLane(state), ConstantU32(state, 0)));
+	const auto stride  = field(word1, 16, 14);
+	const auto swizzle = AndCondition(state, nonzero(stride), nonzero(field(word1, 31, 1)));
+	const auto index = Binary(
+	    state, spv::OpIAdd, TypeU32(state), ctx.Arg(inst, 1),
+	    Select(state, TypeU32(state), nonzero(field(word3, 23, 1)), BufferLane(state),
+	           ConstantU32(state, 0)));
 	const auto soffset = ctx.Arg(inst, 3);
-	const auto base    = DeviceAddressFromWords(state, ctx.Arg(handle, 0), field(word1, 0, 16));
-	const auto valid_format    = nonzero(field(word3, 12, 7));
-	const auto mode            = field(word3, 28, 2);
-	const auto index_in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index, records);
-	const auto scalar_in_bounds =
-	    Binary(state, spv::OpULessThanEqual, TypeBool(state), soffset, records);
+	const auto address = CalculateBufferAddress(state, index, ctx.Arg(inst, 2), soffset,
+	                                            ctx.Memory(inst).offset, stride, swizzle,
+	                                            field(word3, 21, 2));
+	const auto base = DeviceAddressFromWords(state, ctx.Arg(handle, 0), field(word1, 0, 16));
 	const auto raw_records = Binary(state, spv::OpISub, TypeU32(state), records, soffset);
-	const auto raw_index_in_bounds =
-	    Binary(state, spv::OpULessThan, TypeBool(state), index, raw_records);
+	return {
+	    .address = Binary(state, spv::OpIAdd, TypeScalarU64(state), base,
+	                      Unary(state, spv::OpUConvert, TypeScalarU64(state), address.byte)),
+	    .offset = address.offset,
+	    .stride = stride,
+	    .swizzle = swizzle,
+	    .format = field(word3, 12, 7),
+	    .selector = ctx.Memory(inst).formatted ? field(word3, 0, 3) : 0u,
+	    .mode = field(word3, 28, 2),
+	    .records = records,
+	    .index_in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index, records),
+	    .scalar_in_bounds = Binary(state, spv::OpULessThanEqual, TypeBool(state), soffset, records),
+	    .raw_records = raw_records,
+	    .raw_index_in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index, raw_records),
+	};
+}
+
+uint32_t IndirectBufferInBounds(EmitterState& state, const IndirectBufferAccess& buffer,
+                                uint32_t displacement, uint32_t bytes, bool formatted) {
+	const auto offset = displacement == 0u ? buffer.offset : Binary(
+	    state, spv::OpIAdd, TypeU32(state), buffer.offset, ConstantU32(state, displacement));
+	const auto has_payload = [&](uint32_t size) {
+		return AndCondition(
+		    state, Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), size,
+		                  ConstantU32(state, bytes)),
+		    Binary(state, spv::OpULessThanEqual, TypeBool(state), offset,
+		           Binary(state, spv::OpISub, TypeU32(state), size, ConstantU32(state, bytes))));
+	};
+	const auto structured = AndCondition(
+	    state, buffer.index_in_bounds,
+	    formatted ? has_payload(buffer.stride)
+	              : Binary(state, spv::OpULessThan, TypeBool(state), offset, buffer.stride));
+	// OOB_SELECT=3 reduces NUM_RECORDS by SOFFSET before its offset/index checks.
+	const auto raw = AndCondition(
+	    state, buffer.scalar_in_bounds,
+	    Select(state, TypeBool(state), buffer.swizzle,
+	           AndCondition(state, buffer.raw_index_in_bounds, has_payload(buffer.stride)),
+	           has_payload(buffer.raw_records)));
+	auto in_bounds = Select(
+	    state, TypeBool(state),
+	    Binary(state, spv::OpIEqual, TypeBool(state), buffer.mode, ConstantU32(state, 0)),
+	    structured, buffer.index_in_bounds);
+	in_bounds = Select(
+	    state, TypeBool(state),
+	    Binary(state, spv::OpULessThan, TypeBool(state), buffer.mode, ConstantU32(state, 2)),
+	    in_bounds,
+	    Select(state, TypeBool(state),
+	           Binary(state, spv::OpIEqual, TypeBool(state), buffer.mode, ConstantU32(state, 2)),
+	           Binary(state, spv::OpINotEqual, TypeBool(state), buffer.records, ConstantU32(state, 0)),
+	           raw));
+	return in_bounds;
+}
+
+uint32_t LoadIndirectBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t components) {
+	auto& state = ctx.state;
+	const auto buffer = PrepareIndirectBuffer(ctx, inst);
+	const auto valid_format = Binary(state, spv::OpINotEqual, TypeBool(state), buffer.format,
+	                                  ConstantU32(state, 0));
+	const auto base = Binary(state, spv::OpBitwiseAnd, TypeScalarU64(state), buffer.address,
+	                         ConstantDeviceAddress(state, ~uint64_t {3}));
 	std::array<uint32_t, 4> values {};
-	for (uint32_t component = 0; component < components; component++) {
-		const auto address = CalculateBufferAddress(state, index, ctx.Arg(inst, 2), soffset,
-		                                            ctx.Memory(inst).offset + component * 4u,
-		                                            stride, swizzle, index_stride);
-		// RDNA2 raw DWORD vectors check each component, unlike formatted accesses.
-		const auto structured_bounds =
-		    AndCondition(state, index_in_bounds,
-		                 Binary(state, spv::OpULessThan, TypeBool(state), address.offset, stride));
-		// OOB_SELECT=3 reduces NUM_RECORDS by SOFFSET before the offset/index check.
-		const auto raw_bounds = AndCondition(
-		    state, scalar_in_bounds,
-		    Select(state, TypeBool(state), swizzle,
-		           AndCondition(state, raw_index_in_bounds, has_dword(address.offset, stride)),
-		           has_dword(address.offset, raw_records)));
-		auto in_bounds =
-		    Select(state, TypeBool(state),
-		           Binary(state, spv::OpIEqual, TypeBool(state), mode, ConstantU32(state, 0)),
-		           structured_bounds, index_in_bounds);
-		in_bounds = Select(
-		    state, TypeBool(state),
-		    Binary(state, spv::OpULessThan, TypeBool(state), mode, ConstantU32(state, 2)),
-		    in_bounds,
-		    Select(state, TypeBool(state),
-		           Binary(state, spv::OpIEqual, TypeBool(state), mode, ConstantU32(state, 2)),
-		           nonzero(records), raw_bounds));
-		const auto guest =
-		    Binary(state, spv::OpIAdd, TypeScalarU64(state), base,
-		           Unary(state, spv::OpUConvert, TypeScalarU64(state), address.byte));
-		values[component] = LoadBda(ctx, guest, AndCondition(state, valid_format, in_bounds), 32u);
+	for (uint32_t component = 0; component < components; ++component) {
+		const auto address = component == 0u ? base : Binary(
+		    state, spv::OpIAdd, TypeScalarU64(state), base,
+		    ConstantDeviceAddress(state, component * 4u));
+		values[component] = EmitValueOrZeroIfCondition(
+		    state, AndCondition(state, valid_format,
+		                        IndirectBufferInBounds(state, buffer, component * 4u, 4u, false)),
+		    [&]() { return LoadBdaDword(ctx, address); });
 	}
 	return ConstructU32Composite(state, components, values);
 }
 
-uint32_t LoadWideBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t components) {
+uint32_t LoadIndirectFormattedX(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& state = ctx.state;
+	return EmitValueOrZeroIfCondition(state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
+		const auto buffer = PrepareIndirectBuffer(ctx, inst);
+		return EmitIndexSwitch(
+		    state, buffer.format, static_cast<uint32_t>(Prospero::BufferFormat::k32_32_32_32Float) + 1u,
+		    TypeU32(state), [&](uint32_t format) {
+			    const auto info = Format::GetFormatInfo(static_cast<Prospero::BufferFormat>(format));
+			    if (info.type == Format::ComponentType::Unknown) return ConstantU32(state, 0);
+			    const auto constant = Select(
+			        state, TypeU32(state),
+			        Binary(state, spv::OpIEqual, TypeBool(state), buffer.selector, ConstantU32(state, 1)),
+			        FormattedConstant(ctx, info, FormattedSourceKind::One), ConstantU32(state, 0));
+			    return EmitValueOrDefaultIfCondition(
+			        state, Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), buffer.selector,
+			                      ConstantU32(state, 4)),
+			        TypeU32(state), constant, [&]() {
+				        const auto in_bounds = IndirectBufferInBounds(state, buffer, 0u, info.byte_size, true);
+				        const auto base = Binary(
+				            state, spv::OpBitwiseAnd, TypeScalarU64(state), buffer.address,
+				            ConstantDeviceAddress(state, ~(uint64_t(std::min(4u, info.byte_size)) - 1u)));
+				        const auto load = [&](uint32_t component, uint32_t bits) {
+					        const auto offset = Format::GetFormatComponentByteOffset(info, component);
+					        const auto address = offset == 0u ? base : Binary(
+					            state, spv::OpIAdd, TypeScalarU64(state), base,
+					            ConstantDeviceAddress(state, offset));
+					        return LoadBda(ctx, address, in_bounds, bits);
+				        };
+				        const auto emit_component = [&](uint32_t component) {
+					        return LoadFormattedComponent(
+					            ctx, info, {FormattedSourceKind::Memory, component},
+					            [&](uint32_t source) { return load(source, 32u); }, load);
+				        };
+				        if (info.component_count == 1u) return emit_component(0u);
+				        const auto component = Binary(
+				            state, spv::OpUMod, TypeU32(state),
+				            Binary(state, spv::OpISub, TypeU32(state), buffer.selector, ConstantU32(state, 4)),
+				            ConstantU32(state, info.component_count));
+				        return EmitIndexSwitch(state, component, info.component_count, TypeU32(state),
+				                               emit_component);
+			        });
+		    });
+	});
+}
+
+uint32_t LoadBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t components) {
 	auto& state = ctx.state;
 	return EmitValueOrDefaultIfCondition(
 	    state, ctx.Arg(inst, inst.NumArgs() - 1), TypeU32Composite(state, components),
@@ -851,8 +985,7 @@ uint32_t LoadWideBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t co
 		    const auto info = Format::GetFormatInfo(
 		        mem.formatted ? BufferFormat(ctx, mem) : Prospero::BufferFormat::kInvalid);
 		    if (info.type != Format::ComponentType::Unknown) {
-			    const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info, components,
-			                                             FormattedAccess::Load);
+			    const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info);
 			    return EmitValueOrDefaultIfCondition(
 			        state, plan.in_bounds, TypeU32Composite(state, components),
 			        FormattedOutOfBoundsValue(ctx, mem, plan, components), [&]() {
@@ -907,9 +1040,8 @@ uint32_t LoadWideShared(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t co
 			                               ? base
 			                               : Binary(state, spv::OpIAdd, TypeU32(state), base,
 			                                        ConstantU32(state, component * 4u));
-			    const auto raw_index = Binary(state, spv::OpShiftRightLogical, TypeU32(state),
+			    const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state),
 			                                  address, ConstantU32(state, 2));
-			    const auto index     = EmitMemoryElementIndex(state, resource, raw_index);
 			    values[component]    = EmitValueOrZeroIfCondition(
 			        state, EmitMemoryElementInBounds(state, resource, index),
 			        [&]() { return LoadWordInBounds(ctx, resource, index); });
@@ -928,9 +1060,8 @@ void StoreWideShared(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t compo
 			const auto address = component == 0u ? base
 			                                     : Binary(state, spv::OpIAdd, TypeU32(state), base,
 			                                              ConstantU32(state, component * 4u));
-			const auto raw_index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), address,
+			const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), address,
 			                              ConstantU32(state, 2));
-			const auto index = EmitMemoryElementIndex(state, resource, raw_index);
 			EmitIfCondition(state, EmitMemoryElementInBounds(state, resource, index),
 			                [&]() {
 				                StoreWordInBounds(ctx, resource, index, ctx.Arg(inst, component + 1u));
@@ -988,7 +1119,7 @@ void DefineGetBdaPointer(EmitterState& state) {
 	                                  ConstantDeviceAddress(state, BufferCache::CACHING_PAGEBITS));
 	const auto page          = Unary(state, spv::OpUConvert, TypeU32(state), page64);
 	const auto entry_pointer = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferU64ElementPointer(state),
+	state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state, 64),
 	                          entry_pointer, state.bda_pagetable_variable, ConstantU32(state, 0),
 	                          page);
 	const auto base = state.builder.AllocateId();
@@ -1036,25 +1167,18 @@ uint32_t EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst) {
 	return EmitValueOrDefaultIfCondition(
 	    state, ctx.Arg(inst, inst.NumArgs() - 1), TypeU64(state), ConstantU64(state, 0), [&]() {
 		    const auto resource = PrepareStorageBufferResourceAccess(
-		        state, mem, state.storage_buffer_u64_variable, TypeStorageBufferU64Pointer(state));
-		    const auto byte_address = Binary(state, spv::OpIAdd, TypeU32(state),
-		                                     ByteAddress(ctx, inst, mem), resource.byte_offset);
+		        state, mem, state.storage_buffer_u64_variable, TypeStorageBufferPointer(state, 64));
+		    const auto byte_address = ByteAddress(ctx, inst, mem);
 		    const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), byte_address,
 		                              ConstantU32(state, 3u));
 		    return EmitValueOrDefaultIfCondition(
 		        state, EmitMemoryElementInBounds(state, resource, index), TypeU64(state),
 		        ConstantU64(state, 0), [&]() {
-			        const auto value = Unary(state, spv::OpBitcast, TypeScalarU64(state),
-			                                 ctx.Arg(inst, inst.NumArgs() - 2));
-			        const auto old   = state.builder.AllocateId();
-			        state.builder.AddFunction(
-			            SpirvAtomicOpcode(inst.GetOpcode()), TypeScalarU64(state), old,
-			            EmitStorageBufferElementPointer(state, resource, index,
-			                                            TypeStorageBufferU64ElementPointer(state)),
-			            ConstantU32(state, spv::ScopeDevice),
-			            ConstantU32(state, spv::MemorySemanticsMaskNone), value);
+			        const auto pointer = EmitStorageBufferElementPointer(
+			            state, resource, index, TypeStorageBufferElementPointer(state, 64));
+			        const auto old = EmitAtomicOperation(ctx, inst, pointer, spv::ScopeDevice);
 			        EmitAtomicMemoryBarrier(state, mem.kind);
-			        return Unary(state, spv::OpBitcast, TypeU64(state), old);
+			        return old;
 		        });
 	    });
 }
@@ -1098,16 +1222,17 @@ void EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto& mem       = ctx.Memory(inst);
 	const bool  max_value = inst.GetOpcode() == IR::ValueOpcode::SharedAtomicFMax32;
 	EmitIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
-		const auto access = PrepareMemoryElement(ctx, mem, DwordIndex(ctx, inst, mem));
+		const auto index = DwordIndex(ctx, inst, mem);
+		const auto resource = PrepareMemoryResourceAccess(ctx.state, mem);
 		EmitIfCondition(
-		    ctx.state, EmitMemoryElementInBounds(ctx.state, access.resource, access.index), [&]() {
+		    ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index), [&]() {
 			    ctx.state.builder.AddFunction(spv::OpStore, ctx.scratch_u32_variable,
 			                                  ctx.Arg(inst, 1));
 			    const auto data = ctx.state.builder.AllocateId();
 			    ctx.state.builder.AddFunction(spv::OpLoad, TypeU32(ctx.state), data,
 			                                  ctx.scratch_u32_variable);
 			    AtomicUpdate(
-			        ctx.state, EmitMemoryElementPointer(ctx.state, access.resource, access.index),
+			        ctx.state, EmitMemoryElementPointer(ctx.state, resource, index),
 			        mem.kind, [&](uint32_t old) {
 				        const auto old_f =
 				            Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), old);
@@ -1151,10 +1276,9 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
 		           Binary(state, spv::OpIAdd, TypeU32(state), base, size),
 		           ConstantU32(state, 0xc000u)));
 	}
-	const auto raw_index =
+	const auto index =
 	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), address, ConstantU32(state, 2));
 	const auto access = PrepareMemoryResourceAccess(state, mem);
-	const auto index  = EmitMemoryElementIndex(state, access, raw_index);
 	const auto exec   = ctx.Arg(inst, 1);
 	const auto ballot = ctx.Ballot(inst.Arg(1));
 	const auto low    = state.builder.AllocateId();
@@ -1213,6 +1337,10 @@ uint32_t EmitReadConst(ValueEmitContext& ctx, const IR::Inst& inst) {
 void EmitReadConstBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto mem = ctx.Memory(inst);
 	if (mem.planning_only) return;
+	if (mem.kind == IR::ResourceKind::IndirectBuffer) {
+		ctx.Define(inst, LoadIndirectScalarBuffer(ctx, inst));
+		return;
+	}
 	auto& state        = ctx.state;
 	mem.kind           = IR::ResourceKind::ScalarBuffer;
 	auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), ctx.Arg(inst, 1),
@@ -1221,7 +1349,10 @@ void EmitReadConstBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
 		index = Binary(state, spv::OpIAdd, TypeU32(state), index, ConstantU32(state, mem.offset >> 2u));
 	}
 	const auto access    = PrepareMemoryResourceAccess(state, mem);
-	const auto element   = EmitMemoryElementIndex(state, access, index);
+	// Scalar buffer addressing aligns the base and each offset independently.
+	const auto element = Binary(
+	    state, spv::OpIAdd, TypeU32(state), index,
+	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), access.byte_offset, ConstantU32(state, 2)));
 	const auto condition = EmitMemoryElementInBounds(state, access, element);
 	ctx.Define(inst, EmitValueOrZeroIfCondition(state, condition, [&]() {
 		           return EmitNative<spv::OpLoad, IR::Type::U32>(
@@ -1239,8 +1370,9 @@ void EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst) {
 	uint32_t   value;
 	if (mem.kind == IR::ResourceKind::FlatLocal)
 		value = LoadLocalFlat(ctx, inst);
-	else if (buffer_components > 1u)
-		value = LoadWideBuffer(ctx, inst, buffer_components);
+	else if (buffer_components > 1u ||
+	         (buffer_components == 1u && mem.kind == IR::ResourceKind::IndirectBuffer && !mem.formatted))
+		value = LoadBuffer(ctx, inst, buffer_components);
 	else if (shared_components > 1u)
 		value = LoadWideShared(ctx, inst, shared_components);
 	else if (mem.kind == IR::ResourceKind::ScalarAddress)
@@ -1250,7 +1382,8 @@ void EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst) {
 		value = LoadBda(ctx, GuestAddress(ctx, inst, mem), ctx.Arg(inst, inst.NumArgs() - 1),
 		                address_info.data_bits);
 	else if (op == IR::ValueOpcode::LoadBufferU32 && mem.formatted)
-		value = FormattedLoad(ctx, inst, mem);
+		value = mem.kind == IR::ResourceKind::IndirectBuffer ? LoadIndirectFormattedX(ctx, inst)
+		                                                     : FormattedLoad(ctx, inst, mem);
 	else if (inst.GetType() == IR::Type::U8)
 		value = LoadSubword(ctx, inst, mem, 8, false);
 	else if (inst.GetType() == IR::Type::U16)
@@ -1288,6 +1421,15 @@ uint32_t EmitSharedIncDec(ValueEmitContext& ctx, const IR::Inst& inst) {
 	return EmitAtomicUpdate(ctx, inst, ctx.Memory(inst), replacement);
 }
 
+void EmitSharedAtomicMaskedOr32(ValueEmitContext& ctx, const IR::Inst& inst) {
+	const auto keep = Unary(ctx.state, spv::OpNot, TypeU32(ctx.state), ctx.Arg(inst, 1));
+	EmitAtomicUpdate(ctx, inst, ctx.Memory(inst),
+	                 [keep](EmitterState& state, uint32_t old, uint32_t value) {
+		                 return Binary(state, spv::OpBitwiseOr, TypeU32(state),
+		                               Binary(state, spv::OpBitwiseAnd, TypeU32(state), old, keep), value);
+	                 });
+}
+
 uint32_t EmitSwizzleU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto& state = ctx.state;
 	state.builder.AddFunction(spv::OpStore, ctx.scratch_u32_variable, ctx.Arg(inst, 0));
@@ -1297,6 +1439,52 @@ uint32_t EmitSwizzleU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	return EmitDsMaskedLaneRead(state, source, target, ctx.Arg(inst, 2));
 }
 
+uint32_t EmitPermuteU32(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto&      state   = ctx.state;
+	auto       lane    = EmitSubgroupLocalInvocationId(state);
+	if (state.lane_count == 2) {
+		lane = Binary(state, spv::OpBitwiseAnd, TypeU32(state), lane, ConstantU32(state, 31));
+	}
+	const auto address = ctx.Arg(inst, 1);
+	const auto word    = Binary(state, spv::OpShiftRightLogical, TypeU32(state), lane,
+	                            ConstantU32(state, 5));
+	const auto ballot_word = [&](uint32_t predicate) {
+		const auto ballot = state.builder.AllocateId();
+		const auto result = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpGroupNonUniformBallot, TypeU32Vector(state, 4), ballot,
+		                          ConstantU32(state, spv::ScopeSubgroup), predicate);
+		state.builder.AddFunction(spv::OpVectorExtractDynamic, TypeU32(state), result, ballot, word);
+		return result;
+	};
+	// RDNA2 permutes independently within each 32-lane half. Intersect the source
+	// address bit ballots to find this destination's enabled writers without LDS.
+	auto writers = ballot_word(ctx.Arg(inst, 2));
+	for (uint32_t bit = 0; bit < 5; ++bit) {
+		const auto address_bit = Binary(state, spv::OpBitwiseAnd, TypeU32(state), address,
+		                                ConstantU32(state, 1u << (bit + 2)));
+		const auto mask = ballot_word(Binary(state, spv::OpINotEqual, TypeBool(state),
+		                                      address_bit, ConstantU32(state, 0)));
+		const auto lane_bit = Binary(state, spv::OpBitwiseAnd, TypeU32(state), lane,
+		                             ConstantU32(state, 1u << bit));
+		const auto selected = Select(
+		    state, TypeU32(state),
+		    Binary(state, spv::OpINotEqual, TypeBool(state), lane_bit, ConstantU32(state, 0)),
+		    mask, Unary(state, spv::OpNot, TypeU32(state), mask));
+		writers = Binary(state, spv::OpBitwiseAnd, TypeU32(state), writers, selected);
+	}
+	const auto active = Binary(state, spv::OpINotEqual, TypeBool(state), writers,
+	                           ConstantU32(state, 0));
+	const auto base = Binary(state, spv::OpBitwiseAnd, TypeU32(state), lane,
+	                         ConstantU32(state, ~31u));
+	const auto source = Select(state, TypeU32(state), active,
+	                           Binary(state, spv::OpBitwiseOr, TypeU32(state), base,
+	                                  EmitFindUMsb32(state, writers)), lane);
+	const auto result = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpGroupNonUniformShuffle, TypeU32(state), result,
+	                          ConstantU32(state, spv::ScopeSubgroup), ctx.Arg(inst, 0), source);
+	return Select(state, TypeU32(state), active, result, ConstantU32(state, 0));
+}
+
 uint32_t EmitBpermuteU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto&      state  = ctx.state;
 	const auto source = ctx.Arg(inst, 0);

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

@@ -31,6 +31,51 @@ uint32_t EmitBinaryU32(EmitterState& state, spv::Op opcode, uint32_t lhs, uint32
 	return ret;
 }
 
+uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t mapping_offset,
+                                   uint32_t search_iterations, uint32_t default_resource) {
+	const auto LoadMapping = [&](uint32_t index) {
+		const auto pointer = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state),
+		                          pointer, state.flattened_srt_variable, ConstantU32(state, 0),
+		                          index);
+		const auto value = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpLoad, TypeU32(state), value, pointer);
+		return value;
+	};
+	const auto mapping = ConstantU32(state, mapping_offset);
+	const auto count = LoadMapping(mapping);
+	const auto entry_at = [&](uint32_t index) {
+		return Binary(state, spv::OpIAdd, TypeU32(state), mapping,
+		              Binary(state, spv::OpIAdd, TypeU32(state),
+		                     Binary(state, spv::OpShiftLeftLogical, TypeU32(state), index,
+		                            ConstantU32(state, 1u)), ConstantU32(state, 1u)));
+	};
+	auto low = ConstantU32(state, 0u);
+	auto high = count;
+	for (uint32_t iteration = 0; iteration < search_iterations; ++iteration) {
+		const auto active = Binary(state, spv::OpULessThan, TypeBool(state), low, high);
+		const auto mid = Binary(state, spv::OpShiftRightLogical, TypeU32(state),
+		                        Binary(state, spv::OpIAdd, TypeU32(state), low, high),
+		                        ConstantU32(state, 1u));
+		const auto probe = Select(state, TypeU32(state), active, mid, ConstantU32(state, 0u));
+		const auto less = Binary(state, spv::OpULessThan, TypeBool(state),
+		                         LoadMapping(entry_at(probe)), key);
+		low = Select(state, TypeU32(state),
+		             Binary(state, spv::OpLogicalAnd, TypeBool(state), active, less),
+		             Binary(state, spv::OpIAdd, TypeU32(state), mid, ConstantU32(state, 1u)), low);
+		high = Select(state, TypeU32(state),
+		              Binary(state, spv::OpLogicalAnd, TypeBool(state), active,
+		                     Unary(state, spv::OpLogicalNot, TypeBool(state), less)), mid, high);
+	}
+	const auto in_range = Binary(state, spv::OpULessThan, TypeBool(state), low, count);
+	const auto entry = entry_at(Select(state, TypeU32(state), in_range, low, ConstantU32(state, 0u)));
+	const auto match = Binary(state, spv::OpLogicalAnd, TypeBool(state), in_range,
+	                          Binary(state, spv::OpIEqual, TypeBool(state), LoadMapping(entry), key));
+	const auto resource = LoadMapping(Binary(state, spv::OpIAdd, TypeU32(state), entry,
+	                                         ConstantU32(state, 1u)));
+	return Select(state, TypeU32(state), match, resource, ConstantU32(state, default_resource));
+}
+
 uint32_t StorageBufferPackedStride(const EmitterState& state, const IR::MemoryInfo& mem) {
 	if (mem.resource >= state.program.info.buffers.size()) {
 		ExitDescriptorBindingFailure(state, IR::DescriptorBindingKind::Buffers, mem.resource,
@@ -47,6 +92,15 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
 	return state.program.info.buffers[mem.resource].descriptor_format;
 }
 
+uint32_t StorageBufferElementBits(const IR::Program& program, const IR::MemoryInfo& mem) {
+	if (!mem.formatted) return mem.data_bits;
+	const auto format = mem.typed ? Format::DecodeTBufferFormat(mem.data_format, mem.number_format)
+	                              : program.info.buffers[mem.resource].descriptor_format;
+	const auto info = Format::GetFormatInfo(format);
+	return info.type == Format::ComponentType::Unknown || info.packed_bitfield
+	           ? 32u : info.component_bits[0];
+}
+
 void EmitMemoryOffsets(EmitterState& state) {
 	for (uint32_t i = 0; i < state.program.bindings.memory_offset_count; i++) {
 		const auto word =
@@ -149,24 +203,18 @@ MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::
 			EXIT("physical address memory must use the BDA emitter\n");
 		case IR::ResourceKind::ScalarBuffer:
 		case IR::ResourceKind::Buffer: {
+			const auto bits = mem.kind == IR::ResourceKind::Buffer
+			                      ? StorageBufferElementBits(state.program, mem) : 32u;
+			const auto variable = bits == 8u ? state.storage_buffer_u8_variable
+			                      : bits == 16u ? state.storage_buffer_u16_variable
+			                                    : state.storage_buffer_variable;
 			access = PrepareStorageBufferResourceAccess(
-			    state, mem, state.storage_buffer_variable, TypeStorageBufferPointer(state));
-			access.index_offset = EmitBinaryU32(state, spv::OpShiftRightLogical, access.byte_offset,
-			                                    ConstantU32(state, 2u));
-			access.add_index_offset = true;
+			    state, mem, variable, TypeStorageBufferPointer(state, bits));
+			access.element_bits = bits;
 			return access;
 		}
 		default: EXIT("unsupported memory resource kind: %u\n", static_cast<unsigned>(mem.kind));
 	}
-	access.length = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpArrayLength, TypeU32(state), access.length,
-	                          access.object_pointer, 0);
-	return access;
-}
-
-uint32_t EmitMemoryElementIndex(EmitterState& state, const MemoryResourceAccess& access,
-                                uint32_t raw_index) {
-	return access.add_index_offset ? EmitAddU32(state, raw_index, access.index_offset) : raw_index;
 }
 
 uint32_t EmitMemoryElementInBounds(EmitterState& state, const MemoryResourceAccess& access,
@@ -195,7 +243,7 @@ uint32_t EmitMemoryElementPointer(EmitterState& state, const MemoryResourceAcces
 		return pointer;
 	}
 	return EmitStorageBufferElementPointer(state, access, index,
-	                                       TypeStorageBufferElementPointer(state));
+	                                       TypeStorageBufferElementPointer(state, access.element_bits));
 }
 
 uint32_t EmitStorageBufferElementPointer(EmitterState& state,
@@ -330,6 +378,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::BufferAtomicSMax32:
 		case IR::ValueOpcode::SharedAtomicSMax32: return spv::OpAtomicSMax;
 		case IR::ValueOpcode::ImageAtomicUMax32:
+		case IR::ValueOpcode::ImageAtomicUMax64:
 		case IR::ValueOpcode::BufferAtomicUMax32:
 		case IR::ValueOpcode::SharedAtomicUMax32: return spv::OpAtomicUMax;
 		case IR::ValueOpcode::ImageAtomicAnd32:
@@ -352,6 +401,8 @@ uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32
                              uint32_t scope) {
 	const auto opcode = SpirvAtomicOpcode(inst.GetOpcode());
 	const auto old    = ctx.state.builder.AllocateId();
+	const bool wide   = inst.GetType() == IR::Type::U64;
+	const auto type   = wide ? TypeScalarU64(ctx.state) : TypeU32(ctx.state);
 	if (opcode == spv::OpAtomicCompareExchange) {
 		const auto desired    = ctx.Arg(inst, inst.NumArgs() - 3);
 		const auto comparator = ctx.Arg(inst, inst.NumArgs() - 2);
@@ -360,12 +411,13 @@ uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32
 		    ConstantU32(ctx.state, scope), ConstantU32(ctx.state, spv::MemorySemanticsMaskNone),
 		    ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), desired, comparator);
 	} else {
-		const auto value = ctx.Arg(inst, inst.NumArgs() - 2);
-		ctx.state.builder.AddFunction(opcode, TypeU32(ctx.state), old, pointer,
+		auto value = ctx.Arg(inst, inst.NumArgs() - 2);
+		if (wide) value = Unary(ctx.state, spv::OpBitcast, type, value);
+		ctx.state.builder.AddFunction(opcode, type, old, pointer,
 		                              ConstantU32(ctx.state, scope),
 		                              ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), value);
 	}
-	return old;
+	return wide ? Unary(ctx.state, spv::OpBitcast, TypeU64(ctx.state), old) : old;
 }
 
 void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind) {

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

@@ -25,43 +25,58 @@ uint32_t MeshLoad(EmitterState& state, uint32_t variable, spv::StorageClass stor
 }
 
 uint32_t MeshOutputType(EmitterState& state, IR::StageOutputKind kind) {
-	return kind == IR::StageOutputKind::Layer ? TypeU32(state) : TypeF32Vector(state, 4);
+	switch (kind) {
+		case IR::StageOutputKind::Layer: return TypeU32(state);
+		case IR::StageOutputKind::ClipDistance:
+		case IR::StageOutputKind::CullDistance: return TypeF32(state);
+		case IR::StageOutputKind::Position:
+		case IR::StageOutputKind::Parameter: return TypeF32Vector(state, 4);
+		default: EXIT("unsupported mesh output kind=%u\n", static_cast<uint32_t>(kind));
+	}
 }
 
 } // namespace
 
-void DefineMeshOutputs(EmitterState& state) {
+void DefineMeshOutputs(EmitterState& state, uint32_t clip_distance_count,
+                       uint32_t cull_distance_count) {
 	const auto& mesh = state.input_info.vertex->mesh;
 	for (auto& output: state.outputs) {
-		if (output.kind != IR::StageOutputKind::Position &&
-		    output.kind != IR::StageOutputKind::Parameter &&
-		    output.kind != IR::StageOutputKind::Layer) {
-			EXIT("unsupported mesh output kind=%u\n", static_cast<uint32_t>(output.kind));
+		const auto type = MeshOutputType(state, output.kind);
+		const bool clip = output.kind == IR::StageOutputKind::ClipDistance;
+		const bool cull = output.kind == IR::StageOutputKind::CullDistance;
+		auto& variable = clip ? state.clip_distance_variable
+		                 : cull ? state.cull_distance_variable : output.variable_id;
+		if (variable == 0) {
+			const auto element_type = clip || cull
+			                              ? state.builder.Type(spv::OpTypeArray, type,
+			                                                   ConstantU32(state, clip ? clip_distance_count
+			                                                                           : cull_distance_count))
+			                              : type;
+			variable = MeshArray(state, spv::StorageClassOutput, element_type,
+			                     output.kind == IR::StageOutputKind::Layer ? mesh.max_primitives
+			                                                             : mesh.max_vertices);
+			state.interface_variables.push_back(variable);
+			state.builder.AddName(variable, output.debug_name.c_str());
+			if (output.kind == IR::StageOutputKind::Parameter) {
+				state.builder.AddAnnotation(spv::OpDecorate, variable,
+				                            spv::DecorationLocation, output.location);
+			} else {
+				const auto builtin = clip ? spv::BuiltInClipDistance
+				                     : cull ? spv::BuiltInCullDistance
+				                     : output.kind == IR::StageOutputKind::Layer ? spv::BuiltInLayer
+				                                                                 : spv::BuiltInPosition;
+				state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationBuiltIn, builtin);
+			}
+			if (output.kind == IR::StageOutputKind::Layer) {
+				state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationPerPrimitiveEXT);
+			}
 		}
-		const auto type    = MeshOutputType(state, output.kind);
-		output.variable_id = MeshArray(
-		    state, spv::StorageClassOutput, type,
-		    output.kind == IR::StageOutputKind::Layer ? mesh.max_primitives : mesh.max_vertices);
+		output.variable_id = variable;
 		// Only Layer is read by another invocation, through the primitive's provoking vertex.
 		const bool shared = output.kind == IR::StageOutputKind::Layer;
 		output.mesh_data_variable =
 		    MeshArray(state, shared ? spv::StorageClassWorkgroup : spv::StorageClassPrivate, type,
 		              shared ? mesh.max_vertices : state.lane_count);
-		state.interface_variables.push_back(output.variable_id);
-		state.builder.AddName(output.variable_id, output.debug_name.c_str());
-		if (output.kind == IR::StageOutputKind::Parameter) {
-			state.builder.AddAnnotation(spv::OpDecorate, output.variable_id,
-			                            spv::DecorationLocation, output.location);
-		} else {
-			state.builder.AddAnnotation(spv::OpDecorate, output.variable_id, spv::DecorationBuiltIn,
-			                            output.kind == IR::StageOutputKind::Layer
-			                                ? spv::BuiltInLayer
-			                                : spv::BuiltInPosition);
-		}
-		if (output.kind == IR::StageOutputKind::Layer) {
-			state.builder.AddAnnotation(spv::OpDecorate, output.variable_id,
-			                            spv::DecorationPerPrimitiveEXT); // PerPrimitiveEXT
-		}
 	}
 	state.mesh_allocation = MeshArray(state, spv::StorageClassWorkgroup, TypeU32(state), 2);
 	state.mesh_primitive_data =
@@ -148,8 +163,16 @@ void EmitMeshEntryPoint(EmitterState& state) {
 				const auto value =
 				    MeshLoad(state, output.mesh_data_variable, spv::StorageClassPrivate, type,
 				             ConstantU32(state, half));
-				const auto pointer =
-				    MeshElement(state, output.variable_id, spv::StorageClassOutput, type, index);
+				uint32_t pointer;
+				if (output.kind == IR::StageOutputKind::ClipDistance ||
+				    output.kind == IR::StageOutputKind::CullDistance) {
+					pointer = state.builder.AllocateId();
+					state.builder.AddFunction(
+					    spv::OpAccessChain, TypePointer(state, spv::StorageClassOutput, type), pointer,
+					    output.variable_id, index, ConstantU32(state, output.index));
+				} else {
+					pointer = MeshElement(state, output.variable_id, spv::StorageClassOutput, type, index);
+				}
 				state.builder.AddFunction(spv::OpStore, pointer, value);
 			}
 		});

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

@@ -91,7 +91,8 @@ uint32_t TypeU32Vector(EmitterState& state, uint32_t components) {
 }
 
 uint32_t TypeU32Composite(EmitterState& state, uint32_t components) {
-	EXIT_IF(components < 2u || components > 4u);
+	EXIT_IF(components < 1u || components > 4u);
+	if (components == 1u) return TypeU32(state);
 	return components == 2u ? TypeU32Pair(state) : TypeU32Vector(state, components);
 }
 
@@ -122,46 +123,31 @@ uint32_t TypeFunction(EmitterState& state) {
 	return state.function_type;
 }
 
-uint32_t StorageRuntimeArrayType(EmitterState& state) {
-	return state.builder.DecoratedType(
-	    spv::OpTypeRuntimeArray,
-	    {{spv::OpDecorate, {spv::DecorationArrayStride, sizeof(uint32_t)}}}, TypeU32(state));
+uint32_t TypeStorageBufferElement(EmitterState& state, uint32_t bits) {
+	return bits == 32u ? TypeU32(state) : bits == 64u ? TypeScalarU64(state)
+	                                                   : state.builder.Type(spv::OpTypeInt, bits, 0);
 }
 
-uint32_t StorageBufferType(EmitterState& state) {
+uint32_t StorageRuntimeArrayType(EmitterState& state, uint32_t bits) {
+	return state.builder.DecoratedType(
+	    spv::OpTypeRuntimeArray,
+	    {{spv::OpDecorate, {spv::DecorationArrayStride, bits / 8u}}},
+	    TypeStorageBufferElement(state, bits));
+}
+
+uint32_t StorageBufferType(EmitterState& state, uint32_t bits = 32) {
 	return state.builder.DecoratedType(spv::OpTypeStruct,
 	                                   {{spv::OpMemberDecorate, {0, spv::DecorationOffset, 0}},
 	                                    {spv::OpDecorate, {spv::DecorationBlock}}},
-	                                   StorageRuntimeArrayType(state));
+	                                   StorageRuntimeArrayType(state, bits));
 }
 
-uint32_t TypeStorageBufferPointer(EmitterState& state) {
-	return TypePointer(state, spv::StorageClassStorageBuffer, StorageBufferType(state));
+uint32_t TypeStorageBufferPointer(EmitterState& state, uint32_t bits) {
+	return TypePointer(state, spv::StorageClassStorageBuffer, StorageBufferType(state, bits));
 }
 
-uint32_t TypeStorageBufferElementPointer(EmitterState& state) {
-	return TypePointer(state, spv::StorageClassStorageBuffer, TypeU32(state));
-}
-
-uint32_t StorageU64RuntimeArrayType(EmitterState& state) {
-	return state.builder.DecoratedType(
-	    spv::OpTypeRuntimeArray,
-	    {{spv::OpDecorate, {spv::DecorationArrayStride, sizeof(uint64_t)}}}, TypeScalarU64(state));
-}
-
-uint32_t StorageBufferU64Type(EmitterState& state) {
-	return state.builder.DecoratedType(spv::OpTypeStruct,
-	                                   {{spv::OpMemberDecorate, {0, spv::DecorationOffset, 0}},
-	                                    {spv::OpDecorate, {spv::DecorationBlock}}},
-	                                   StorageU64RuntimeArrayType(state));
-}
-
-uint32_t TypeStorageBufferU64Pointer(EmitterState& state) {
-	return TypePointer(state, spv::StorageClassStorageBuffer, StorageBufferU64Type(state));
-}
-
-uint32_t TypeStorageBufferU64ElementPointer(EmitterState& state) {
-	return TypePointer(state, spv::StorageClassStorageBuffer, TypeScalarU64(state));
+uint32_t TypeStorageBufferElementPointer(EmitterState& state, uint32_t bits) {
+	return TypePointer(state, spv::StorageClassStorageBuffer, TypeStorageBufferElement(state, bits));
 }
 
 uint32_t TypePhysicalU32Pointer(EmitterState& state) {
@@ -243,27 +229,34 @@ void DefineDescriptors(EmitterState& state) {
 			case IR::DescriptorBindingKind::Buffers:
 				state.storage_buffer_variable =
 				    Define(ArrayType(StorageBufferType(state)), "buffers");
+				if (state.requirements.buffer_u8) {
+					state.storage_buffer_u8_variable =
+					    Define(ArrayType(StorageBufferType(state, 8)), "buffers_u8");
+				}
+				if (state.requirements.buffer_u16) {
+					state.storage_buffer_u16_variable =
+					    Define(ArrayType(StorageBufferType(state, 16)), "buffers_u16");
+				}
 				if (state.requirements.buffer_int64_atomics) {
 					state.storage_buffer_u64_variable =
-					    Define(ArrayType(StorageBufferU64Type(state)), "buffers_u64");
-					state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_variable,
-					                            spv::DecorationAliased);
-					state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_u64_variable,
-					                            spv::DecorationAliased);
+					    Define(ArrayType(StorageBufferType(state, 64)), "buffers_u64");
 				}
-				if (state.requirements.coherent_buffers) {
+				for (const auto variable: {state.storage_buffer_variable, state.storage_buffer_u8_variable,
+				                           state.storage_buffer_u16_variable, state.storage_buffer_u64_variable}) {
+					if (variable == 0) continue;
+					if (state.storage_buffer_u8_variable != 0 || state.storage_buffer_u16_variable != 0 ||
+					    state.storage_buffer_u64_variable != 0) {
+						state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationAliased);
+					}
 					// RDNA2 stores publish to L2 even without GLC; every alias of the buffer
 					// must participate in visibility for cache-bypassing polling loads.
-					state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_variable,
-					                            spv::DecorationCoherent);
-					if (state.storage_buffer_u64_variable != 0) {
-						state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_u64_variable,
-						                            spv::DecorationCoherent);
+					if (state.requirements.coherent_buffers) {
+						state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationCoherent);
 					}
 				}
 				break;
 			case IR::DescriptorBindingKind::BdaPagetable:
-				state.bda_pagetable_variable = Define(StorageBufferU64Type(state), "bda_pagetable");
+				state.bda_pagetable_variable = Define(StorageBufferType(state, 64), "bda_pagetable");
 				break;
 			case IR::DescriptorBindingKind::FaultBuffer:
 				state.fault_buffer_variable = Define(StorageBufferType(state), "fault_buffer");
@@ -333,8 +326,9 @@ uint32_t ConstantU64(EmitterState& state, uint64_t value) {
 }
 
 uint32_t ConstantU32CompositeZero(EmitterState& state, uint32_t components) {
-	EXIT_IF(components < 2u || components > 4u);
+	EXIT_IF(components < 1u || components > 4u);
 	const auto            zero = ConstantU32(state, 0);
+	if (components == 1u) return zero;
 	std::vector<uint32_t> values(components, zero);
 	return state.builder.Constant(spv::OpConstantComposite, TypeU32Composite(state, components),
 	                              values);
@@ -541,7 +535,7 @@ void DefineOutputs(EmitterState& state) {
 		}
 	}
 	if (state.program.stage == ShaderType::Mesh) {
-		DefineMeshOutputs(state);
+		DefineMeshOutputs(state, clip_distance_count, cull_distance_count);
 		return;
 	}
 	if (state.program.stage == ShaderType::Vertex && clip_distance_count + cull_distance_count < 8u &&
@@ -667,6 +661,13 @@ void DefineModule(EmitterState& state) {
 
 	state.builder.RequireCapability(spv::CapabilityShader);
 	state.builder.RequireCapability(spv::CapabilitySignedZeroInfNanPreserve);
+	if (state.requirements.buffer_u8) {
+		state.builder.RequireExtension("SPV_KHR_8bit_storage");
+		state.builder.RequireCapability(spv::CapabilityStorageBuffer8BitAccess);
+	}
+	if (state.requirements.buffer_u16) {
+		state.builder.RequireCapability(spv::CapabilityStorageBuffer16BitAccess);
+	}
 	if (state.program.info.uses_dma) {
 		state.builder.RequireCapability(spv::CapabilityInt64);
 		state.builder.RequireCapability(spv::CapabilityPhysicalStorageBufferAddresses);
@@ -701,8 +702,9 @@ void DefineModule(EmitterState& state) {
 	if (state.requirements.image_gather_extended) {
 		state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
 	}
-	if (state.lane_count == 2 || state.requirements.subgroup_ballot ||
-	    state.requirements.subgroup_shuffle || state.requirements.subgroup_local_invocation_id) {
+	if (state.lane_count == 2 || state.requirements.subgroup_barrier ||
+	    state.requirements.subgroup_ballot || state.requirements.subgroup_shuffle ||
+	    state.requirements.subgroup_local_invocation_id) {
 		state.builder.RequireCapability(spv::CapabilityGroupNonUniform);
 	}
 	if (state.lane_count == 2 || state.requirements.subgroup_ballot) {
@@ -737,15 +739,13 @@ void DefineModule(EmitterState& state) {
 	if (state.requirements.float64) {
 		EXIT_NOT_IMPLEMENTED(state.program.stage == ShaderType::Compute &&
 		                     state.input_info.compute->float_mode != 0xc0);
-		// MODE=0xc0 uses round-to-nearest-even and preserves FP64 input/output denormals.
+		// Use native rounding for MODE=0xc0, consistent with ordinary FP32 arithmetic.
 		state.builder.RequireCapability(spv::CapabilityFloat64);
-		state.builder.RequireCapability(spv::CapabilityRoundingModeRTE);
 		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSignedZeroInfNanPreserve,
 		                               64u);
 		// FP64 denormal preservation is temporarily disabled.
 		// state.builder.RequireCapability(spv::CapabilityDenormPreserve);
 		// state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeDenormPreserve, 64u);
-		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeRoundingModeRTE, 32u);
 	}
 	if (const auto* cs = ShaderWorkgroupInput(state.program.stage, state.input_info)) {
 		uint32_t    local_x = state.requirements.compute_derivatives ? 2u : 1u;

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

@@ -315,6 +315,7 @@ void EmitBlock(ValueEmitContext& ctx, const IR::Block* block, EmitInstruction&&
 			auto& lane          = half == 0 ? ctx : *ctx.other_half;
 			ctx.state.lane_half = half;
 			if (half == 0 || (inst.GetOpcode() != IR::ValueOpcode::Barrier &&
+			                  inst.GetOpcode() != IR::ValueOpcode::StoreCompletion &&
 			                  inst.GetOpcode() != IR::ValueOpcode::MeshAllocate)) {
 				emit_instruction(lane, inst);
 			}

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

@@ -228,7 +228,8 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
 		case 0x09u: return Opcode::IMAGE_STORE_MIP;
 		case 0x0eu: return Opcode::IMAGE_GET_RESINFO;
 		case 0x60u: return Opcode::IMAGE_GET_LOD;
-		case 0xe6u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
+		case 0xe6u:
+		case 0xe7u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
 		default: return Opcode::UNSUPPORTED;
 	}
 }
@@ -258,6 +259,7 @@ uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
 
 	switch (opcode) {
 		case 0xe6u: return 11u;
+		case 0xe7u: return 12u;
 		case 0x0eu: return 1u;
 		case 0x01u:
 		case 0x09u: return ImageCoordComponents(dimension) + 1u;
@@ -360,19 +362,29 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 		SetUnsupported(inst, Family::MIMG, opcode,
 		               "MIMG image gather requires exactly one dmask bit");
 	}
-	if (inst.opcode == Opcode::IMAGE_ATOMIC_CMPSWAP && inst.dmask != 0x3u) {
-		SetUnsupported(inst, Family::MIMG, opcode,
-		               "MIMG image compare-and-swap requires 32-bit DMASK 0x3");
+	if (atomic != nullptr) {
+		const bool compare_swap = inst.opcode == Opcode::IMAGE_ATOMIC_CMPSWAP;
+		const auto mask32       = compare_swap ? 0x3u : 0x1u;
+		const auto mask64       = compare_swap ? 0xfu : 0x3u;
+		if (inst.dmask != mask32 && inst.dmask != mask64) {
+			SetUnsupported(inst, Family::MIMG, opcode, "MIMG image atomic has invalid DMASK");
+		} else if (inst.dmask == mask64) {
+			inst.data_bits = 64u;
+			if (inst.opcode != Opcode::IMAGE_ATOMIC_UMAX) {
+				SetUnsupported(inst, Family::MIMG, opcode,
+				               "MIMG 64-bit image atomic opcode is not implemented");
+			}
+		}
 	}
 	const bool supports_d16 = sample != nullptr || gather != 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");
 	}
-	if (opcode == 0xe6u &&
+	if (inst.opcode == Opcode::IMAGE_BVH_INTERSECT_RAY &&
 	    (a16 || !r128 || inst.dmask != 0xfu || (nsa_dwords != 0u && nsa_dwords != 3u))) {
 		SetUnsupported(inst, Family::MIMG, opcode,
-		               "BVH intersection requires eleven full-float ray DWORDs and R128/dmask:0xf");
+		               "BVH intersection requires full-float ray DWORDs and R128/dmask:0xf");
 	}
 
 	DecodeVectorGpr(vdata, inst.dst);

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

@@ -92,7 +92,8 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x05u, Opcode::DS_MIN_I32, 1, 32},          {0x06u, Opcode::DS_MAX_I32, 1, 32},
     {0x07u, Opcode::DS_MIN_U32, 1, 32},          {0x08u, Opcode::DS_MAX_U32, 1, 32},
     {0x09u, Opcode::DS_AND_B32, 1, 32},          {0x0au, Opcode::DS_OR_B32, 1, 32},
-    {0x0bu, Opcode::DS_XOR_B32, 1, 32},          {0x0du, Opcode::DS_WRITE_B32, 1, 32},
+    {0x0bu, Opcode::DS_XOR_B32, 1, 32},          {0x0cu, Opcode::DS_MSKOR_B32, 1, 32},
+    {0x0du, Opcode::DS_WRITE_B32, 1, 32},
     {0x0eu, Opcode::DS_WRITE2_B32, 2, 32},       {0x0fu, Opcode::DS_WRITE2ST64_B32, 2, 32},
     {0x12u, Opcode::DS_MIN_F32, 1, 32},          {0x13u, Opcode::DS_MAX_F32, 1, 32},
     {0x1eu, Opcode::DS_WRITE_B8, 1, 8},          {0x1fu, Opcode::DS_WRITE_B16, 1, 16},
@@ -116,7 +117,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0xa6u, Opcode::DS_READ_U16_D16, 1, 16},
     {0xa7u, Opcode::DS_READ_U16_D16_HI, 1, 16},
     {0xb0u, Opcode::DS_WRITE_ADDTID_B32, 1, 32}, {0xb1u, Opcode::DS_READ_ADDTID_B32, 1, 32},
-    {0xb3u, Opcode::DS_BPERMUTE_B32, 1, 32},
+    {0xb2u, Opcode::DS_PERMUTE_B32, 1, 32},     {0xb3u, Opcode::DS_BPERMUTE_B32, 1, 32},
     {0xdeu, Opcode::DS_WRITE_B96, 3, 32},        {0xdfu, Opcode::DS_WRITE_B128, 4, 32},
     {0xfeu, Opcode::DS_READ_B96, 3, 32},         {0xffu, Opcode::DS_READ_B128, 4, 32},
 };
@@ -192,6 +193,7 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:
+		case Opcode::DS_MSKOR_B32:
 		case Opcode::DS_WRXCHG_RTN_B32: return true;
 		default: return false;
 	}
@@ -203,8 +205,10 @@ uint32_t DsSourceCount(Opcode opcode) {
 		case Opcode::DS_WRITE2ST64_B32:
 		case Opcode::DS_WRITE2_B64:
 		case Opcode::DS_WRITE2ST64_B64:
+		case Opcode::DS_MSKOR_B32:
 		case Opcode::DS_MIN_F32:
 		case Opcode::DS_MAX_F32: return 3u;
+		case Opcode::DS_PERMUTE_B32:
 		case Opcode::DS_BPERMUTE_B32: return 2u;
 		case Opcode::DS_READ_ADDTID_B32:
 		case Opcode::DS_CONSUME:
@@ -407,7 +411,8 @@ void DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 		SetUnsupported(inst, Family::DS, opcode, "DS swizzle FFT mode is not implemented");
 	}
 	if (inst.gds &&
-	    (inst.opcode == Opcode::DS_SWIZZLE_B32 || inst.opcode == Opcode::DS_BPERMUTE_B32 ||
+	    (inst.opcode == Opcode::DS_SWIZZLE_B32 || inst.opcode == Opcode::DS_PERMUTE_B32 ||
+	     inst.opcode == Opcode::DS_BPERMUTE_B32 ||
 	     inst.opcode == Opcode::DS_WRITE_ADDTID_B32 ||
 	     inst.opcode == Opcode::DS_READ_ADDTID_B32)) {
 		SetUnsupported(inst, Family::DS, opcode, "DS lane operation is available only for LDS");

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

@@ -86,7 +86,7 @@ constexpr OpcodeMap SOPK_OPCODE_LIST[] = {
     {0x08u, Opcode::S_CMP_LE_I32}, {0x09u, Opcode::S_CMP_EQ_U32}, {0x0au, Opcode::S_CMP_LG_U32},
     {0x0bu, Opcode::S_CMP_GT_U32}, {0x0cu, Opcode::S_CMP_GE_U32}, {0x0du, Opcode::S_CMP_LT_U32},
     {0x0eu, Opcode::S_CMP_LE_U32}, {0x0fu, Opcode::S_ADD_I32},    {0x10u, Opcode::S_MULK_I32},
-    {0x13u, Opcode::S_SETREG_B32}, {0x17u, Opcode::S_WAITCNT},    {0x18u, Opcode::S_WAITCNT},
+    {0x13u, Opcode::S_SETREG_B32}, {0x17u, Opcode::S_WAITCNT_VSCNT}, {0x18u, Opcode::S_WAITCNT},
     {0x19u, Opcode::S_WAITCNT},    {0x1au, Opcode::S_WAITCNT},
     {0x1bu, Opcode::S_SUBVECTOR_LOOP_BEGIN}, {0x1cu, Opcode::S_SUBVECTOR_LOOP_END},
 };
@@ -234,6 +234,14 @@ void DecodeSopk(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 			inst.src_count         = 1;
 			return;
 		}
+		case Opcode::S_WAITCNT_VSCNT:
+			inst.dst.kind       = OperandKind::Null;
+			inst.src1           = inst.src0;
+			inst.src1.value     = word & 0xffffu;
+			inst.src1.signed_val = static_cast<int32_t>(inst.src1.value);
+			DecodeScalarSource(sdst, pc, inst.src0);
+			inst.src_count = 2;
+			return;
 		case Opcode::S_SETREG_B32:
 			inst.dst.kind        = OperandKind::Null;
 			inst.src1.kind       = OperandKind::LiteralConstant;

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

@@ -535,6 +535,11 @@ std::string InstructionToString(const Instruction& inst) {
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: s_setreg_b32 {}, {}", inst.pc,
 			                                               OperandToString(inst.src0).c_str(),
 			                                               OperandToString(inst.src1).c_str()));
+		case Opcode::S_WAITCNT_VSCNT:
+			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: {} {}, {}", inst.pc,
+			                                               magic_enum::enum_name(inst.opcode),
+			                                               OperandToString(inst.src0),
+			                                               OperandToString(inst.src1)));
 		case Opcode::S_NOP:
 		case Opcode::S_WAITCNT:
 		case Opcode::S_WAITCNT_DEPCTR:
@@ -692,10 +697,12 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:
+		case Opcode::DS_MSKOR_B32:
 		case Opcode::DS_WRXCHG_RTN_B32:
 		case Opcode::DS_MIN_F32:
 		case Opcode::DS_MAX_F32:
 		case Opcode::DS_SWIZZLE_B32:
+		case Opcode::DS_PERMUTE_B32:
 		case Opcode::DS_BPERMUTE_B32:
 		case Opcode::DS_READ_I8:
 		case Opcode::DS_READ_U8:

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

@@ -186,6 +186,7 @@ enum class Opcode {
 	V_RCP_F32,
 	V_RCP_IFLAG_F32,
 	V_FRACT_F32,
+	V_FRACT_F64,
 	V_TRUNC_F32,
 	V_CEIL_F32,
 	V_RNDNE_F32,
@@ -445,6 +446,7 @@ enum class Opcode {
 	V_CMP_EQ_I64,
 	V_CMP_LT_I64,
 	V_CMP_LE_I64,
+	V_CMP_NE_I64,
 	V_CMP_LT_U64,
 	V_CMP_EQ_U64,
 	V_CMP_LE_U64,
@@ -556,10 +558,12 @@ enum class Opcode {
 	DS_OR_RTN_B32,
 	DS_XOR_B32,
 	DS_XOR_RTN_B32,
+	DS_MSKOR_B32,
 	DS_WRXCHG_RTN_B32,
 	DS_MIN_F32,
 	DS_MAX_F32,
 	DS_SWIZZLE_B32,
+	DS_PERMUTE_B32,
 	DS_BPERMUTE_B32,
 	DS_CONSUME,
 	DS_APPEND,
@@ -625,6 +629,7 @@ enum class Opcode {
 
 	S_NOP,
 	S_WAITCNT,
+	S_WAITCNT_VSCNT,
 	S_WAITCNT_DEPCTR,
 	S_BARRIER,
 	S_BRANCH,

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

@@ -123,6 +123,7 @@ constexpr OpcodeMap VOP1_OPCODE_LIST[] = {
     {0x39u, Opcode::V_FFBH_U32},
     {0x3au, Opcode::V_FFBL_B32},
     {0x3bu, Opcode::V_FFBH_I32},
+    {0x3eu, Opcode::V_FRACT_F64},
     {0x3fu, Opcode::V_FREXP_EXP_I32_F32},
     {0x40u, Opcode::V_FREXP_MANT_F32},
     {0x42u, Opcode::V_MOVRELD_B32},
@@ -182,6 +183,7 @@ constexpr OpcodeMap VOP3_ENCODED_VOP1_OPCODE_LIST[] = {
     {0x39u, Opcode::V_FFBH_U32},
     {0x3au, Opcode::V_FFBL_B32},
     {0x3bu, Opcode::V_FFBH_I32},
+    {0x3eu, Opcode::V_FRACT_F64},
     {0x3fu, Opcode::V_FREXP_EXP_I32_F32},
     {0x40u, Opcode::V_FREXP_MANT_F32},
     {0x42u, Opcode::V_MOVRELD_B32},
@@ -254,6 +256,7 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0xc5u, Opcode::V_CMP_NE_U32},         {0xc6u, Opcode::V_CMP_GE_U32},
     {0xc7u, Opcode::V_CMP_T_U32},          {0xa2u, Opcode::V_CMP_EQ_I64, false},
     {0xa1u, Opcode::V_CMP_LT_I64, false},  {0xa3u, Opcode::V_CMP_LE_I64, false},
+    {0xa5u, Opcode::V_CMP_NE_I64, false},
     {0xb5u, Opcode::V_CMPX_NE_I64, false}, {0xd1u, Opcode::V_CMPX_LT_U32},
     {0xd2u, Opcode::V_CMPX_EQ_U32},        {0xd3u, Opcode::V_CMPX_LE_U32},
     {0xd4u, Opcode::V_CMPX_GT_U32},        {0xd5u, Opcode::V_CMPX_NE_U32},
@@ -472,11 +475,18 @@ bool IsNativeVop3B16BinaryOpcode(Opcode opcode) {
 bool IsVop1FloatResultOpcode(Opcode opcode);
 bool IsVopcCompareExec(Opcode opcode);
 
+bool IsVop1Float64Opcode(Opcode opcode) {
+	return opcode == Opcode::V_CVT_F64_I32 || opcode == Opcode::V_CVT_F32_F64 ||
+	       opcode == Opcode::V_CVT_F64_F32 || opcode == Opcode::V_CVT_F64_U32 ||
+	       opcode == Opcode::V_RCP_F64 || opcode == Opcode::V_FRACT_F64;
+}
+
 bool IsVop1FloatSourceOpcode(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::V_CVT_F32_F64:
 		case Opcode::V_CVT_F64_F32:
 		case Opcode::V_RCP_F64:
+		case Opcode::V_FRACT_F64:
 		case Opcode::V_MOV_B32:
 		case Opcode::V_CVT_F32_F16:
 		case Opcode::V_CVT_U32_F32:
@@ -551,6 +561,7 @@ constexpr Vop1SdwaRule VOP1_SDWA_RULES[] = {
     // partial destination may combine with any source selector.
     {Opcode::V_NOT_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelBytes() | SdwaSelWords(),
      SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), false},
+    {Opcode::V_BFREV_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_FFBL_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_FFBH_U32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_CVT_F32_F16, SdwaSelWords() | SdwaSelFull(), 0, 0, true},
@@ -647,6 +658,10 @@ bool SupportsVop1Clamp(Opcode opcode) {
 }
 
 bool ValidateVop1Sdwa(Instruction& inst, uint32_t opcode, uint32_t modifier) {
+	if (IsVop1Float64Opcode(inst.opcode)) {
+		SetUnsupported(inst, Family::VOP1, opcode, "FP64 instructions do not support SDWA");
+		return false;
+	}
 	const auto dst_sel   = (modifier >> 8u) & 0x7u;
 	const auto dst_u     = (modifier >> 11u) & 0x3u;
 	const auto clamp     = (modifier >> 13u) & 0x1u;
@@ -735,9 +750,7 @@ void ApplyDppModifier(Operand& operand, uint32_t modifier, uint32_t encoding) {
 
 void DecodeVop1Dpp(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
                    uint32_t opcode, uint32_t vdst, Instruction& inst) {
-	if (inst.opcode == Opcode::V_CVT_F64_I32 || inst.opcode == Opcode::V_CVT_F32_F64 ||
-	    inst.opcode == Opcode::V_CVT_F64_F32 || inst.opcode == Opcode::V_CVT_F64_U32 ||
-	    inst.opcode == Opcode::V_RCP_F64) {
+	if (IsVop1Float64Opcode(inst.opcode)) {
 		SetUnsupported(inst, Family::VOP1, opcode, "FP64 instructions do not support DPP");
 		return;
 	}
@@ -1152,7 +1165,8 @@ VopcSdwaFields DecodeVopcSdwaFields(uint32_t modifier) {
 }
 
 bool SupportsVopcSdwa(Opcode opcode) {
-	return opcode != Opcode::UNSUPPORTED && opcode != Opcode::V_CMP_EQ_F64 &&
+	return opcode != Opcode::UNSUPPORTED && opcode != Opcode::V_CMP_NE_I64 &&
+	       opcode != Opcode::V_CMP_EQ_F64 &&
 	       opcode != Opcode::V_CMP_LE_F64 && opcode != Opcode::V_CMPX_LE_F64 &&
 	       opcode != Opcode::V_CMPX_GE_F64;
 }

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

@@ -31,6 +31,7 @@ void Translator::EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueO
 	const bool signed_64 = opcode == IR::ValueOpcode::SLessThan64 ||
 	                       opcode == IR::ValueOpcode::SLessThanEqual64 ||
 	                       inst.opcode == Decoder::Opcode::V_CMP_EQ_I64 ||
+	                       inst.opcode == Decoder::Opcode::V_CMP_NE_I64 ||
 	                       inst.opcode == Decoder::Opcode::V_CMPX_NE_I64;
 	const auto read = [&](const Decoder::Operand& operand) {
 		// RDNA2 expands signed 64-bit integer literals by sign extension.

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

@@ -181,8 +181,12 @@ void Translator::EmitControlNop() {
 	ir.Emit(IR::ValueOpcode::ControlNop);
 }
 
-void Translator::EmitWaitcnt() {
-	ir.Emit(IR::ValueOpcode::Waitcnt);
+void Translator::S_WAITCNT_VSCNT(const Decoder::Instruction& inst) {
+	const auto count = inst.src1.value & 63u;
+	if (inst.src0.kind != Decoder::OperandKind::Null || (count != 0u && count != 63u)) {
+		EXIT("unsupported partial or register-based S_WAITCNT_VSCNT at 0x%08x", inst.pc);
+	}
+	if (count == 0u) ir.Emit(IR::ValueOpcode::StoreCompletion);
 }
 
 void Translator::S_BARRIER() {
@@ -252,13 +256,14 @@ void Translator::ScalarSelect64(const Decoder::Instruction& inst,
 	const auto selected_mask_valid =
 	    IR::U1(ir.Emit(IR::ValueOpcode::SelectU1,
 	                   {condition, ReadMaskValid(inst.src0), ReadMaskValid(false_source)}));
-	if (IsExecOrVcc(inst.dst)) {
-		WriteMask(inst.dst, selected_mask, true);
-		return;
-	}
 	WriteU32Pair(inst.dst,
 	             {ir.Select(condition, lhs[0], rhs[0]), ir.Select(condition, lhs[1], rhs[1])});
-	if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
+	// Scalar selects preserve the raw pair even when EXEC/VCC holds an address.
+	if (inst.dst.kind == Decoder::OperandKind::ExecLo) {
+		ir.SetExec(selected_mask);
+	} else if (inst.dst.kind == Decoder::OperandKind::VccLo) {
+		ir.SetVcc(selected_mask);
+	} else if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
 		const auto dst = static_cast<IR::ScalarReg>(inst.dst.reg);
 		ir.SetThreadBitScalarReg(dst, selected_mask);
 		ir.SetScalarMaskTag(dst, selected_mask_valid);

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

@@ -483,10 +483,11 @@ void Translator::S_BITCMP_B64(const Decoder::Instruction& inst, bool expected) {
 	WriteCompareResult(inst.dst, ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
 }
 
-void Translator::V_ALIGNBIT_B32(const Decoder::Instruction& inst) {
-	const auto hi      = ReadU32(inst.src0);
-	const auto lo      = ReadU32(inst.src1);
-	const auto shift   = ir.BitwiseAnd(ReadU32(inst.src2), IR::U32(IR::Value(31u)));
+void Translator::V_ALIGN_B32(const Decoder::Instruction& inst, bool byte_offset) {
+	const auto hi = ReadU32(inst.src0);
+	const auto lo = ReadU32(inst.src1);
+	auto shift    = ir.BitwiseAnd(ReadU32(inst.src2), IR::U32(IR::Value(byte_offset ? 3u : 31u)));
+	if (byte_offset) shift = ir.ShiftLeftLogical(shift, IR::U32(IR::Value(3u)));
 	const auto lo_part = ir.ShiftRightLogical(lo, shift);
 	const auto inverse =
 	    ir.BitwiseAnd(ir.ISub(IR::U32(IR::Value(32u)), shift), IR::U32(IR::Value(31u)));
@@ -496,21 +497,6 @@ void Translator::V_ALIGNBIT_B32(const Decoder::Instruction& inst) {
 	WriteOperand(DestinationOperand(inst), ir.BitwiseOr(lo_part, hi_part));
 }
 
-void Translator::V_ALIGNBYTE_B32(const Decoder::Instruction& inst) {
-	const auto hi           = ReadU32(inst.src0);
-	const auto lo           = ReadU32(inst.src1);
-	const auto byte_offset  = ir.BitwiseAnd(ReadU32(inst.src2), IR::U32(IR::Value(31u)));
-	const auto bit_offset   = ir.ShiftLeftLogical(byte_offset, IR::U32(IR::Value(3u)));
-	const auto concatenated = ir.ConstructU64(lo, hi);
-	const auto shifted =
-	    IR::U64(ir.Emit(IR::ValueOpcode::ShiftRightLogical64,
-	                    {concatenated, ir.BitwiseAnd(bit_offset, IR::U32(IR::Value(63u)))}));
-	const auto in_range =
-	    IR::U1(ir.Emit(IR::ValueOpcode::ULessThan32, {byte_offset, IR::Value(8u)}));
-	WriteOperand(DestinationOperand(inst),
-	             ir.Select(in_range, ExtractU64(shifted)[0], IR::U32(IR::Value(0u))));
-}
-
 void Translator::V_LSHL_ADD_U32(const Decoder::Instruction& inst) {
 	const auto shift  = ir.BitwiseAnd(ReadU32(inst.src1), IR::U32(IR::Value(31u)));
 	const auto result = ir.IAdd(ir.ShiftLeftLogical(ReadU32(inst.src0), shift), ReadU32(inst.src2));

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

@@ -239,9 +239,6 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 			case Decoder::Opcode::DS_READ_ADDTID_B32: return MakeM0Operand();
 			case Decoder::Opcode::DS_WRITE_ADDTID_B32:
 				return index == 0u ? decoded.src1 : MakeM0Operand();
-			case Decoder::Opcode::DS_MIN_F32:
-			case Decoder::Opcode::DS_MAX_F32:
-				return index == 0u ? decoded.src1 : index == 1u ? decoded.src0 : decoded.src2;
 			case Decoder::Opcode::DS_WRITE_B8:
 			case Decoder::Opcode::DS_WRITE_B16:
 			case Decoder::Opcode::DS_WRITE_B8_D16_HI:
@@ -255,12 +252,7 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 			case Decoder::Opcode::DS_WRITE2_B64:
 			case Decoder::Opcode::DS_WRITE2ST64_B64:
 				return index == 0u ? decoded.src1 : index == 1u ? decoded.src0 : decoded.src2;
-			default:
-				if (decoded.opcode >= Decoder::Opcode::DS_ADD_U32 &&
-				    decoded.opcode <= Decoder::Opcode::DS_WRXCHG_RTN_B32) {
-					return index == 0u ? decoded.src1 : decoded.src0;
-				}
-				return decoded.src0;
+			default: return decoded.src0;
 		}
 	}
 	return DecodedSourceAt(decoded, index);
@@ -568,7 +560,9 @@ void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode
 		                  ir.GetExec()},
 		                 flags);
 	} else {
-		result = ir.Emit(opcode, {resource, address, ReadU32(data), ir.GetExec()}, flags);
+		const IR::Value value =
+		    memory.data_bits == 64u ? IR::Value(ReadU64(data)) : IR::Value(ReadU32(data));
+		result = ir.Emit(opcode, {resource, address, value, ir.GetExec()}, flags);
 	}
 	if (inst.glc) {
 		WriteOperand(inst.dst, result);
@@ -578,12 +572,13 @@ void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode
 void Translator::DS_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                            bool returns_value) {
 	const auto memory  = MemoryInfoFromDecoded(inst);
-	const auto address = ReadU32(MemorySourceAt(inst, 1));
-	const auto data_src = MemorySourceAt(inst, 0);
+	const auto address = ReadU32(inst.src0);
 	const IR::Value value =
-	    memory.data_dwords == 2u ? IR::Value(ReadU64(data_src)) : IR::Value(ReadU32(data_src));
-	const auto result  = ir.Emit(opcode, {address, value, ir.GetExec()},
-	                             AddMemoryInfo(memory, inst.pc));
+	    memory.data_dwords == 2u ? IR::Value(ReadU64(inst.src1)) : IR::Value(ReadU32(inst.src1));
+	const auto flags = AddMemoryInfo(memory, inst.pc);
+	const auto result = inst.src_count == 3u
+	                        ? ir.Emit(opcode, {address, value, ReadU32(inst.src2), ir.GetExec()}, flags)
+	                        : ir.Emit(opcode, {address, value, ir.GetExec()}, flags);
 	if (returns_value) {
 		WriteOperand(inst.dst, result);
 	}
@@ -706,8 +701,16 @@ void Translator::IMAGE_GATHER(const Decoder::Instruction& inst) {
 	const auto resource = GetImageResource(memory);
 	const auto sampler  = GetSamplerResource(memory);
 	const auto address  = MakeImageAddress(inst, MemorySourceAt(inst, 0));
-	const auto result   = ir.Emit(IR::ValueOpcode::ImageGatherRaw, {resource, sampler, address},
-	                              AddMemoryInfo(memory, inst.pc));
+	const bool has_lod = (memory.image_sample_flags & Decoder::ImageSampleFlagLod) != 0u;
+	// LOD selection consumes these words on the GPU after descriptor handles are stripped.
+	const auto result = ir.Emit(
+	    IR::ValueOpcode::ImageGatherRaw,
+	    {resource, sampler, address,
+	     has_lod ? resource.Instruction()->Arg(1) : IR::Value(0u),
+	     has_lod ? resource.Instruction()->Arg(3) : IR::Value(0u),
+	     has_lod ? sampler.Instruction()->Arg(1) : IR::Value(0u),
+	     has_lod ? sampler.Instruction()->Arg(2) : IR::Value(0u)},
+	    AddMemoryInfo(memory, inst.pc));
 	for (uint32_t index = 0; index < memory.data_dwords; index++) {
 		WriteOperand(OffsetOperand(inst.dst, index),
 		             ir.Emit(IR::ValueOpcode::CompositeExtractU32x4, {result, IR::Value(index)}));
@@ -847,14 +850,6 @@ void Translator::DS_WRITE2(const Decoder::Instruction& inst) {
 	}
 }
 
-void Translator::DS_MINMAX_F32(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
-	const auto memory = MemoryInfoFromDecoded(inst);
-	ir.Emit(opcode,
-	        {ReadU32(MemorySourceAt(inst, 1)), ReadU32(MemorySourceAt(inst, 0)),
-	         ReadU32(MemorySourceAt(inst, 2)), ir.GetExec()},
-	        AddMemoryInfo(memory, inst.pc));
-}
-
 void Translator::DS_APPEND_CONSUME(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
 	const auto memory = MemoryInfoFromDecoded(inst);
 	WriteOperand(inst.dst, ir.Emit(opcode,
@@ -884,9 +879,10 @@ void Translator::DS_SWIZZLE_B32(const Decoder::Instruction& inst) {
 	                                ReadU32(MemorySourceAt(inst, 1)), ir.GetExec()}));
 }
 
-void Translator::DS_BPERMUTE_B32(const Decoder::Instruction& inst) {
+void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 	const auto address = ir.IAdd(ReadU32(inst.src0), IR::U32(IR::Value(inst.offset)));
-	WriteOperand(inst.dst, ir.Emit(IR::ValueOpcode::BpermuteU32,
+	WriteOperand(inst.dst, ir.Emit(backward ? IR::ValueOpcode::BpermuteU32
+	                                       : IR::ValueOpcode::PermuteU32,
 	                               {ReadU32(inst.src1), address, ir.GetExec()}));
 }
 
@@ -894,7 +890,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
 		case Decoder::Opcode::IMAGE_BVH_INTERSECT_RAY: {
 			const auto result = ir.Emit(IR::ValueOpcode::BvhIntersect,
-			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()});
+			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()},
+			    inst.image_address_components - 10u);
 			for (uint32_t component = 0; component < 4; ++component) {
 				WriteOperand(OffsetOperand(inst.dst, component),
 				    ir.Emit(IR::ValueOpcode::CompositeExtractU32x4, {result, IR::Value(component)}));
@@ -1036,6 +1033,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicXor32, false);
 		case Decoder::Opcode::DS_XOR_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicXor32, true);
+		case Decoder::Opcode::DS_MSKOR_B32:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicMaskedOr32, false);
 		case Decoder::Opcode::DS_WRXCHG_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicSwap32, true);
 
@@ -1052,7 +1051,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::IMAGE_ATOMIC_SMAX:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMax32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMAX:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMax32);
+			return IMAGE_ATOMIC(inst, inst.data_bits == 64u ? IR::ValueOpcode::ImageAtomicUMax64
+			                                                : IR::ValueOpcode::ImageAtomicUMax32);
 		case Decoder::Opcode::IMAGE_ATOMIC_AND:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicAnd32);
 		case Decoder::Opcode::IMAGE_ATOMIC_OR:
@@ -1097,11 +1097,12 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::IMAGE_GATHER4H: return IMAGE_GATHER(inst);
 
 		case Decoder::Opcode::DS_MIN_F32:
-			return DS_MINMAX_F32(inst, IR::ValueOpcode::SharedAtomicFMin32);
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicFMin32, false);
 		case Decoder::Opcode::DS_MAX_F32:
-			return DS_MINMAX_F32(inst, IR::ValueOpcode::SharedAtomicFMax32);
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicFMax32, false);
 		case Decoder::Opcode::DS_SWIZZLE_B32: return DS_SWIZZLE_B32(inst);
-		case Decoder::Opcode::DS_BPERMUTE_B32: return DS_BPERMUTE_B32(inst);
+		case Decoder::Opcode::DS_PERMUTE_B32: return DS_PERMUTE(inst, false);
+		case Decoder::Opcode::DS_BPERMUTE_B32: return DS_PERMUTE(inst, true);
 		case Decoder::Opcode::DS_CONSUME:
 			return DS_APPEND_CONSUME(inst, IR::ValueOpcode::DataConsume);
 		case Decoder::Opcode::DS_APPEND:

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

@@ -18,7 +18,8 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_CSELECT_B64: ScalarSelect64(inst, inst.src1); return;
 		case O::S_CMOV_B64: ScalarSelect64(inst, inst.dst); return;
 		case O::S_SETREG_B32: EmitControlNop(); return;
-		case O::S_WAITCNT: EmitWaitcnt(); return;
+		case O::S_WAITCNT_VSCNT: S_WAITCNT_VSCNT(inst); return;
+		case O::S_WAITCNT: return;
 
 		case O::S_AND_SAVEEXEC_B32:
 			S_SAVEEXEC(inst, IR::ValueOpcode::LogicalAnd, false, false, false);
@@ -224,7 +225,7 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_SLEEP:
 		case O::S_SETPRIO:
 		case O::S_TRAP: EmitControlNop(); return;
-		case O::S_WAITCNT_DEPCTR: EmitWaitcnt(); return;
+		case O::S_WAITCNT_DEPCTR: return;
 		case O::S_BARRIER: S_BARRIER(); return;
 		case O::S_SENDMSG: S_SENDMSG(inst); return;
 		case O::S_TTRACEDATA: S_TTRACEDATA(); return;

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

@@ -896,7 +896,8 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
 		case Decoder::Opcode::V_FMA_F64: include_vector(inst.src2, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_ADD_F64:
 		case Decoder::Opcode::V_MUL_F64: include_vector(inst.src1, 2u); [[fallthrough]];
-		case Decoder::Opcode::V_RCP_F64: include_vector(inst.dst, 2u); [[fallthrough]];
+		case Decoder::Opcode::V_RCP_F64:
+		case Decoder::Opcode::V_FRACT_F64: include_vector(inst.dst, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_CVT_F32_F64: include_vector(inst.src0, 2u); break;
 		case Decoder::Opcode::V_CMP_EQ_F64:
 		case Decoder::Opcode::V_CMP_LE_F64:
@@ -905,6 +906,7 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
 		case Decoder::Opcode::V_CMP_EQ_I64:
 		case Decoder::Opcode::V_CMP_LT_I64:
 		case Decoder::Opcode::V_CMP_LE_I64:
+		case Decoder::Opcode::V_CMP_NE_I64:
 		case Decoder::Opcode::V_CMP_LT_U64:
 		case Decoder::Opcode::V_CMP_EQ_U64:
 		case Decoder::Opcode::V_CMP_LE_U64:
@@ -1104,6 +1106,16 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				                                  IR::U32(IR::Value(total_threads)));
 			}
 		}
+		if (options.stage == ShaderType::Compute &&
+		    options.input_info.compute->dispatch_thread_dimensions) {
+			for (uint32_t axis = 0; axis < 3u; axis++) {
+				const auto extent = IR::U32(entry_ir.Emit(
+				    IR::ValueOpcode::GetDispatchThreadExtent, {IR::Value(axis)}));
+				initial_exec = entry_ir.LogicalAnd(
+				    initial_exec,
+				    entry_ir.ULessThan(builtin(IR::StageInputKind::GlobalInvocationId, axis), extent));
+			}
+		}
 		entry_ir.SetExec(initial_exec);
 		const auto initial_mask = entry_ir.Emit(IR::ValueOpcode::Ballot, {initial_exec});
 		entry_ir.SetExecLo(entry_ir.CompositeExtract(initial_mask, 0));

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

@@ -105,11 +105,10 @@ private:
 	void DS_READ2(const Decoder::Instruction& inst);
 	void DS_WRITE(const Decoder::Instruction& inst);
 	void DS_WRITE2(const Decoder::Instruction& inst);
-	void DS_MINMAX_F32(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
 	void DS_APPEND_CONSUME(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
 	void DS_ADDTID(const Decoder::Instruction& inst, bool write);
 	void DS_SWIZZLE_B32(const Decoder::Instruction& inst);
-	void DS_BPERMUTE_B32(const Decoder::Instruction& inst);
+	void DS_PERMUTE(const Decoder::Instruction& inst, bool backward);
 
 	IR::F32 SelectF32(IR::U1 condition, IR::F32 true_value, IR::F32 false_value);
 	IR::U32 ConvertF32ToU32Saturated(IR::F32 value, float upper_bound, float safe_upper,
@@ -208,8 +207,7 @@ private:
 	void    V_BFI_B32(const Decoder::Instruction& inst);
 	void    S_BITCMP_B32(const Decoder::Instruction& inst, bool expected);
 	void    S_BITCMP_B64(const Decoder::Instruction& inst, bool expected);
-	void    V_ALIGNBIT_B32(const Decoder::Instruction& inst);
-	void    V_ALIGNBYTE_B32(const Decoder::Instruction& inst);
+	void    V_ALIGN_B32(const Decoder::Instruction& inst, bool byte_offset);
 	void    V_LSHL_ADD_U32(const Decoder::Instruction& inst);
 	void    V_ADD_LSHL_U32(const Decoder::Instruction& inst);
 	void    V_XAD_U32(const Decoder::Instruction& inst);
@@ -229,7 +227,7 @@ private:
 	void ScalarMinMax32(const Decoder::Instruction& inst, IR::ValueOpcode value_opcode,
 	                    IR::ValueOpcode compare_opcode);
 	void EmitControlNop();
-	void EmitWaitcnt();
+	void S_WAITCNT_VSCNT(const Decoder::Instruction& inst);
 	void S_BARRIER();
 	void S_SENDMSG(const Decoder::Instruction& inst);
 	void S_TTRACEDATA();

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

@@ -106,6 +106,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMP_GT_U64:
 			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan64, IR::Type::U64, false, false);
 			return;
+		case O::V_CMP_NE_I64:
 		case O::V_CMP_NE_U64:
 			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, false, false);
 			return;
@@ -453,6 +454,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_RCP_F64: return FloatUnary(inst, IR::ValueOpcode::FPRecip64);
 		case O::V_RCP_IFLAG_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecipIFlag32);
 		case O::V_FRACT_F32: return FloatUnary(inst, IR::ValueOpcode::FPFract32);
+		case O::V_FRACT_F64: return FloatUnary(inst, IR::ValueOpcode::FPFract64);
 		case O::V_TRUNC_F32: return FloatUnary(inst, IR::ValueOpcode::FPTrunc32);
 		case O::V_CEIL_F32: return FloatUnary(inst, IR::ValueOpcode::FPCeil32);
 		case O::V_RNDNE_F32: return FloatUnary(inst, IR::ValueOpcode::FPRoundEven32);
@@ -591,8 +593,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_BFE_U32: return V_BFE_U32(inst, false);
 		case O::V_BFE_I32: return V_BFE_U32(inst, true);
 		case O::V_BFI_B32: return V_BFI_B32(inst);
-		case O::V_ALIGNBIT_B32: return V_ALIGNBIT_B32(inst);
-		case O::V_ALIGNBYTE_B32: return V_ALIGNBYTE_B32(inst);
+		case O::V_ALIGNBIT_B32: return V_ALIGN_B32(inst, false);
+		case O::V_ALIGNBYTE_B32: return V_ALIGN_B32(inst, true);
 		case O::V_LSHL_ADD_U32: return V_LSHL_ADD_U32(inst);
 		case O::V_ADD_LSHL_U32: return V_ADD_LSHL_U32(inst);
 		case O::V_XAD_U32: return V_XAD_U32(inst);

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

@@ -190,6 +190,18 @@ Value ResolveInvariantPhi(const ResourcePlan& program, Value value) {
 	return invariant;
 }
 
+Value ResolveActiveU32(Value value, Value active) {
+	for (uint32_t depth = 0; depth <= 32; ++depth) {
+		value = value.Resolve();
+		const auto* inst = value.TryInstruction();
+		if (inst == nullptr || inst->GetOpcode() != ValueOpcode::SelectU32 ||
+		    inst->Arg(0).Resolve() != active)
+			return value;
+		value = inst->Arg(1);
+	}
+	return {};
+}
+
 bool HasShaderMemoryWrites(const Program& program) {
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
@@ -399,7 +411,8 @@ void ValidateProgram(const Program& program, bool require_ssa) {
 					                        ValueOpcodeName(inst.GetOpcode())));
 				}
 				const auto& memory = program.memory_info[memory_index];
-				if (memory.kind != ResourceKind::ScalarBuffer) {
+				if (memory.kind != ResourceKind::ScalarBuffer &&
+				    memory.kind != ResourceKind::IndirectBuffer) {
 					return Fail(fmt::format("{} has an invalid scalar-memory resource kind",
 					                        ValueOpcodeName(inst.GetOpcode())));
 				}
@@ -457,7 +470,7 @@ void ValidateProgram(const Program& program, bool require_ssa) {
 				}
 				if (memory.kind == ResourceKind::IndirectBuffer &&
 				    !memory.SupportsIndirectBufferLoad(inst.GetOpcode())) {
-					return Fail("indirect buffer requires a raw DWORD x2/x3/x4 load");
+					return Fail("indirect buffer requires a scalar, raw DWORD x1/x2/x3/x4, or formatted X load");
 				}
 				if (buffer_components > 1u &&
 				    (!vector_buffer || memory.data_bits != 32u ||

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

@@ -73,9 +73,13 @@ struct MemoryInfo {
 	bool                    planning_only                                         = false;
 
 	[[nodiscard]] bool SupportsIndirectBufferLoad(ValueOpcode opcode) const {
-		return !formatted && !typed && data_bits == 32u &&
-		       (opcode == ValueOpcode::LoadBufferU32x2 || opcode == ValueOpcode::LoadBufferU32x3 ||
-		        opcode == ValueOpcode::LoadBufferU32x4);
+		return !typed && data_bits == 32u &&
+		       (formatted ? opcode == ValueOpcode::LoadBufferU32
+		                  : opcode == ValueOpcode::ReadConstBuffer ||
+		                        opcode == ValueOpcode::LoadBufferU32 ||
+		                        opcode == ValueOpcode::LoadBufferU32x2 ||
+		                        opcode == ValueOpcode::LoadBufferU32x3 ||
+		                        opcode == ValueOpcode::LoadBufferU32x4);
 	}
 
 	bool operator==(const MemoryInfo& other) const = default;
@@ -97,6 +101,7 @@ struct ExportInfo {
 
 struct BufferResource {
 	static constexpr uint32_t NoImageAlias = UINT32_MAX;
+	static constexpr uint32_t NoIndirectBuffer = UINT32_MAX;
 
 	uint32_t               source             = 0;
 	uint32_t               first_use_pc       = 0;
@@ -110,11 +115,15 @@ struct BufferResource {
 	bool                   atomic             = false;
 	bool                   formatted          = false;
 	bool                   scalar             = false;
+	uint32_t               indirect_root              = NoIndirectBuffer;
+	uint32_t               indirect_mapping_offset    = 0;
+	uint32_t               indirect_search_iterations = 0;
+	std::vector<uint32_t>  indirect_resources;
 
 	bool operator==(const BufferResource& other) const = default;
 };
 
-enum class ImageMipMode { None, DynamicStorage };
+enum class ImageMipMode { None, Dynamic };
 
 constexpr uint32_t ShaderImageIdentitySwizzle = 0x00000facu;
 
@@ -133,6 +142,7 @@ struct ImageResource {
 	bool                          read              = false;
 	bool                          written           = false;
 	bool                          atomic            = false;
+	bool                          atomic64          = false;
 	bool                          depth_compare     = false;
 	bool                          cube              = false;
 	bool                          r128              = false;
@@ -147,9 +157,12 @@ struct ImageResource {
 struct SamplerResource {
 	uint32_t source                = 0;
 	uint32_t first_use_pc          = 0;
+	// Native filtering/border variants share the original sampler's runtime descriptor.
+	uint32_t snapshot_index        = 0;
 	bool     force_point_filtering = false;
 	bool     depth_compare         = false;
 	bool     integer_border        = false;
+	bool     gather_lod            = false;
 
 	bool operator==(const SamplerResource& other) const = default;
 };
@@ -278,7 +291,7 @@ struct StageOutput {
 inline constexpr uint32_t FirstImageBinding           = 1u;
 inline constexpr uint32_t FirstComparisonImageBinding = 22u;
 inline constexpr uint32_t FirstStorageImageBinding    = 29u;
-inline constexpr uint32_t ImageBindingCount           = 43u;
+inline constexpr uint32_t ImageBindingCount           = 48u;
 
 enum class DescriptorBindingKind : uint32_t {
 	Buffers  = 0u,
@@ -291,8 +304,8 @@ enum class DescriptorBindingKind : uint32_t {
 	Count,
 };
 
-static_assert(static_cast<uint32_t>(DescriptorBindingKind::Samplers) == 44u);
-static_assert(static_cast<uint32_t>(DescriptorBindingKind::Count) == 50u);
+static_assert(static_cast<uint32_t>(DescriptorBindingKind::Samplers) == 49u);
+static_assert(static_cast<uint32_t>(DescriptorBindingKind::Count) == 55u);
 
 struct PushData {
 	static constexpr uint32_t DwordCount = 32;
@@ -369,7 +382,7 @@ DescriptorBindingForImage(const ImageResource& image) {
 			if (image.numeric_class != Prospero::TextureNumericClass::Uint) {
 				return std::nullopt;
 			}
-			base = AtomicUintBinding;
+			base = AtomicUintBinding + (image.atomic64 ? 5u : 0u);
 		} else {
 			switch (image.numeric_class) {
 				case Prospero::TextureNumericClass::Float: base = StorageFloatBinding; break;
@@ -417,6 +430,7 @@ struct DescriptorBinding {
 
 struct BindingLayout {
 	uint32_t                       push_data_start_dword = PushData::NoStart;
+	uint32_t                       dispatch_thread_dword = PushData::NoStart;
 	uint32_t                       memory_offset_dword = 0;
 	uint32_t                       memory_offset_count = 0;
 	std::vector<uint32_t>          user_data_registers;
@@ -468,21 +482,24 @@ struct BlockInfo {
 };
 
 struct DescriptorSource {
-	struct IndirectImage {
+	struct IndirectDescriptor {
 		uint32_t material_source = UINT32_MAX;
 		uint32_t table_source    = 0;
 		uint32_t selector_stride = 0;
 		uint32_t selector_offset = 0;
 		uint32_t table_offset    = 0;
+		uint32_t selector_shift  = 0;
 		Value    key_count;
+		Value                 selector_first;
 		Value    selector_mask;
+		std::vector<uint32_t> sources;
 
-		bool operator==(const IndirectImage& other) const = default;
+		bool operator==(const IndirectDescriptor& other) const = default;
 	};
 
 	std::array<Value, 8>         dwords {};
 	uint32_t                     dword_count = 0;
-	std::optional<IndirectImage> indirect_image;
+	std::optional<IndirectDescriptor> indirect_descriptor;
 
 	bool operator==(const DescriptorSource& other) const = default;
 };
@@ -605,6 +622,7 @@ void  ValidateProgram(const Program& program, bool require_ssa);
 void  ResolveControlFlowIdentities(Program& program);
 bool  EquivalentValue(const ResourcePlan& program, Value left, Value right);
 Value ResolveInvariantPhi(const ResourcePlan& program, Value value);
+Value ResolveActiveU32(Value value, Value active);
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR
 

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

@@ -152,6 +152,8 @@ public:
 	}
 
 private:
+	friend void EliminateDeadCode(const std::vector<Block*>& blocks);
+
 	void AddUse(Inst* used, size_t operand);
 	void RemoveUse(Inst* used, size_t operand);
 	void ClearArgs();
@@ -161,6 +163,7 @@ private:
 
 	ValueOpcode         opcode;
 	uint8_t             num_args;
+	bool                live = false;
 	mutable uint32_t    evaluation_index = UINT32_MAX;
 	uint64_t            flags;
 	Block*              parent = nullptr;

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

@@ -93,6 +93,7 @@ bool HasSideEffects(ValueOpcode opcode) {
 		case ValueOpcode::SetAttribute:
 		case ValueOpcode::SetTessellationAttribute:
 		case ValueOpcode::MeshAllocate:
+		case ValueOpcode::StoreCompletion:
 		case ValueOpcode::Barrier: return true;
 		default: return false;
 	}
@@ -177,7 +178,8 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::SharedAtomicAnd32:
 		case ValueOpcode::SharedAtomicOr32:
 		case ValueOpcode::SharedAtomicOr64:
-		case ValueOpcode::SharedAtomicXor32: return SharedAccess::Atomic;
+		case ValueOpcode::SharedAtomicXor32:
+		case ValueOpcode::SharedAtomicMaskedOr32: return SharedAccess::Atomic;
 		case ValueOpcode::DataAppend: return SharedAccess::Append;
 		case ValueOpcode::DataConsume: return SharedAccess::Consume;
 		default: return SharedAccess::None;
@@ -227,6 +229,7 @@ ImageOpcodeInfo ImageOpcodeInfoOf(ValueOpcode opcode) {
 		case ValueOpcode::ImageAtomicUMin32:
 		case ValueOpcode::ImageAtomicSMax32:
 		case ValueOpcode::ImageAtomicUMax32:
+		case ValueOpcode::ImageAtomicUMax64:
 		case ValueOpcode::ImageAtomicAnd32:
 		case ValueOpcode::ImageAtomicOr32:
 		case ValueOpcode::ImageAtomicXor32:

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

@@ -8,6 +8,7 @@ VALUE_OPCODE(ConditionRef, U1, U1)
 
 // Translation and register-state pseudo operations.
 VALUE_OPCODE(GetUserData, U32, ScalarReg)
+VALUE_OPCODE(GetDispatchThreadExtent, U32, U32)
 VALUE_OPCODE(GetShaderBase, U64)
 VALUE_OPCODE(MeshDrawParameter, U32, U32)
 VALUE_OPCODE(MeshAllocate, Void, U32)
@@ -202,6 +203,7 @@ VALUE_OPCODE(FPFloor32, F32, F32)
 VALUE_OPCODE(FPCeil32, F32, F32)
 VALUE_OPCODE(FPTrunc32, F32, F32)
 VALUE_OPCODE(FPFract32, F32, F32)
+VALUE_OPCODE(FPFract64, F64, F64)
 
 VALUE_OPCODE(LaneId, U32)
 VALUE_OPCODE(Ballot, U32x4, U1)
@@ -209,6 +211,7 @@ VALUE_OPCODE(ReadFirstLane, U32, U32, U1)
 VALUE_OPCODE(ReadLane, U32, U32, U32)
 VALUE_OPCODE(WriteLane, U32, U32, U32, U32)
 VALUE_OPCODE(Permlane16U32, U32, U32, U32, U32, U1)
+VALUE_OPCODE(PermuteU32, U32, U32, U32, U1)
 VALUE_OPCODE(BpermuteU32, U32, U32, U32, U1)
 
 // Typed shader resources and memory operations. Descriptor state is represented by SSA values;
@@ -287,6 +290,7 @@ VALUE_OPCODE(SharedAtomicAnd32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicOr32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicOr64, Void, U32, U64, U1)
 VALUE_OPCODE(SharedAtomicXor32, U32, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicMaskedOr32, Void, U32, U32, U32, U1)
 VALUE_OPCODE(DataAppend, U32, U32, U1)
 VALUE_OPCODE(DataConsume, U32, U32, U1)
 VALUE_OPCODE(SwizzleU32, U32, U32, U32, U1)
@@ -297,7 +301,7 @@ VALUE_OPCODE(ImageQueryLod, U32x4, ImageResource, SamplerResource, ImageAddress)
 VALUE_OPCODE(ImageRead, U32x4, ImageResource, ImageAddress, U1)
 VALUE_OPCODE(ImageWrite, Void, ImageResource, ImageAddress, U32x4, U1)
 VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress)
-VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress)
+VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U32, U32, U32, U32)
 VALUE_OPCODE(ImageAtomicCompareSwap32, U32, ImageResource, ImageAddress, U32, U32, U1)
 VALUE_OPCODE(ImageAtomicSwap32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicIAdd32, U32, ImageResource, ImageAddress, U32, U1)
@@ -305,6 +309,7 @@ VALUE_OPCODE(ImageAtomicSMin32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMin32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicSMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMax32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicUMax64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicAnd32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicOr32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicXor32, U32, ImageResource, ImageAddress, U32, U1)
@@ -315,7 +320,7 @@ VALUE_OPCODE(GetInterpolationParameter, U32, U32, U32, U32)
 VALUE_OPCODE(SetAttribute, Void, U32x4, U1)
 
 VALUE_OPCODE(ControlNop, Void)
-VALUE_OPCODE(Waitcnt, Void)
+VALUE_OPCODE(StoreCompletion, Void)
 VALUE_OPCODE(Barrier, Void)
 VALUE_OPCODE(Sendmsg, Void)
 VALUE_OPCODE(TtraceData, Void)

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

@@ -15,11 +15,16 @@ namespace {
 	std::abort();
 }
 
-std::vector<uint32_t> CollectUserData(const Program& program) {
+void CollectShaderData(const Program& program, BindingLayout& layout) {
 	std::array<bool, NumScalarRegs> registers {};
+	bool uses_dispatch_threads = false;
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
-			if (inst.GetOpcode() != ValueOpcode::GetUserData || !inst.HasUses()) {
+			if (!inst.HasUses()) {
+				continue;
+			}
+			uses_dispatch_threads |= inst.GetOpcode() == ValueOpcode::GetDispatchThreadExtent;
+			if (inst.GetOpcode() != ValueOpcode::GetUserData) {
 				continue;
 			}
 			if (inst.Arg(0).GetType() != Type::ScalarReg) {
@@ -32,13 +37,16 @@ std::vector<uint32_t> CollectUserData(const Program& program) {
 			registers[index] = true;
 		}
 	}
-	std::vector<uint32_t> result;
 	for (uint32_t index = 0; index < registers.size(); index++) {
 		if (registers[index]) {
-			result.push_back(index);
+			layout.user_data_registers.push_back(index);
 		}
 	}
-	return result;
+	layout.memory_offset_dword = static_cast<uint32_t>(layout.user_data_registers.size());
+	if (uses_dispatch_threads) {
+		layout.dispatch_thread_dword = layout.memory_offset_dword;
+		layout.memory_offset_dword += 3u;
+	}
 }
 
 void AddBinding(BindingLayout& layout, DescriptorBindingKind kind,
@@ -74,6 +82,9 @@ bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffe
 			} else if (memory.kind == ResourceKind::Buffer || memory.kind == ResourceKind::ScalarBuffer) {
 				EXIT_IF(memory.resource >= program.info.buffers.size());
 				live_buffers.at(memory.resource) = true;
+				for (const auto child: program.info.buffers[memory.resource].indirect_resources) {
+					live_buffers.at(child) = true;
+				}
 			}
 		}
 	}
@@ -86,13 +97,15 @@ bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffe
 }
 
 bool UsesFlattenedSrt(const Program& program) {
+	const auto uses_mapping = [](const auto& resource) {
+		return resource.indirect_search_iterations != 0u;
+	};
 	return std::ranges::any_of(program.blocks, [](const Block* block) {
 		return std::ranges::any_of(*block, [](const Inst& inst) {
 			return inst.GetOpcode() == ValueOpcode::ReadConst;
 		});
-	}) || std::ranges::any_of(program.info.images, [](const ImageResource& image) {
-		return image.indirect_search_iterations != 0u;
-	});
+	}) || std::ranges::any_of(program.info.buffers, uses_mapping) ||
+	       std::ranges::any_of(program.info.images, uses_mapping);
 }
 
 void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
@@ -104,8 +117,7 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 	BindingLayout next;
 	std::vector<uint32_t> buffers;
 	const bool            uses_gds = CollectMemoryResources(program, buffers);
-	next.user_data_registers = CollectUserData(program);
-	next.memory_offset_dword = static_cast<uint32_t>(next.user_data_registers.size());
+	CollectShaderData(program, next);
 	next.memory_offset_count       = static_cast<uint32_t>(buffers.size());
 	next.push_data_start_dword =
 	    PushData::StartFor(push_data_start_dword, next.ShaderDataDwords());
@@ -126,7 +138,7 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 			EXIT("shader binding layout failed: image %u has an unmapped binding class", i);
 		}
 		auto&      resources = image_groups[group];
-		const auto dynamic   = program.info.images[i].mip_mode == ImageMipMode::DynamicStorage;
+		const auto dynamic   = program.info.images[i].mip_mode == ImageMipMode::Dynamic;
 		const auto count     = dynamic ? program.info.images[i].mip_count : 1u;
 		if (count == 0u || (!dynamic && program.info.images[i].mip_count != 1u)) {
 			EXIT("shader binding layout failed: image %u has invalid specialized mip count %u", i,

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

@@ -199,6 +199,11 @@ void FoldInstruction(Block& block, Block::iterator instruction,
                       std::unordered_set<Inst*>& lowered_ancillary) {
 	auto& inst = *instruction;
 	switch (inst.GetOpcode()) {
+		case ValueOpcode::StoreBufferU32: {
+			const auto data = ResolveActiveU32(Arg(inst, 4), Arg(inst, 5));
+			if (!data.IsEmpty() && data != inst.Arg(4)) inst.SetArg(4, data);
+			return;
+		}
 		case ValueOpcode::Phi: FoldPhi(inst); return;
 		case ValueOpcode::SelectU1:
 			if (!FoldSelect(inst) && IsImmediate(Arg(inst, 2), Type::U1) &&
@@ -503,6 +508,17 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 			}
 			return;
 		}
+		case ValueOpcode::BitReverse32: {
+			const auto value = Arg(inst, 0);
+			if (IsImmediate(value, Type::U32)) {
+				uint32_t source = value.U32(), reversed = 0;
+				for (uint32_t bit = 0; bit < 32u; ++bit, source >>= 1u) {
+					reversed = (reversed << 1u) | (source & 1u);
+				}
+				Replace(inst, Value(reversed));
+			}
+			return;
+		}
 		case ValueOpcode::BitCount32: {
 			const auto value = Arg(inst, 0);
 			if (IsImmediate(value, Type::U32)) {

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

@@ -1,5 +1,7 @@
 #include "graphics/shader/recompiler/ir/passes/DeadCodeElimination.h"
 
+#include <algorithm>
+
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
 void RemoveIdentities(const BlockList& blocks) {
@@ -18,23 +20,32 @@ void RemoveIdentities(const BlockList& blocks) {
 }
 
 void EliminateDeadCode(const BlockList& blocks) {
-	bool changed;
-	do {
-		changed = false;
-		for (auto block = blocks.rbegin(); block != blocks.rend(); block++) {
-			auto& instructions = (*block)->Instructions();
-			auto  inst         = instructions.end();
-			while (inst != instructions.begin()) {
-				--inst;
-				if (inst->HasUses() || inst->MayHaveSideEffects()) {
-					continue;
-				}
-				inst->Invalidate();
-				inst    = instructions.erase(inst);
-				changed = true;
+	std::vector<Inst*> pending;
+	for (auto* block: blocks) {
+		for (auto& inst: *block) {
+			// Retained resource-planning instructions live outside the block list.
+			inst.live = inst.MayHaveSideEffects() || std::ranges::any_of(inst.Uses(),
+			    [](const Use& use) { return use.user->Parent() == nullptr; });
+			if (inst.live) pending.push_back(&inst);
+		}
+	}
+	while (!pending.empty()) {
+		const auto* inst = pending.back();
+		pending.pop_back();
+		for (size_t arg = 0; arg < inst->NumArgs(); ++arg) {
+			auto* input = inst->Arg(arg).TryInstruction();
+			if (input != nullptr && input->Parent() != nullptr && !input->live) {
+				input->live = true;
+				pending.push_back(input);
 			}
 		}
-	} while (changed);
+	}
+	// Unlink every dead operand before freeing nodes in a dead Phi cycle.
+	for (auto* block: blocks)
+		for (auto& inst: *block)
+			if (!inst.live) inst.Invalidate();
+	for (auto* block: blocks)
+		std::erase_if(block->Instructions(), [](const Inst& inst) { return !inst.live; });
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -20,7 +20,7 @@ namespace Libs::Graphics::ShaderRecompiler::IR {
 namespace {
 
 constexpr uint64_t AddressMask            = 0x0000ffffffffffffull;
-constexpr uint64_t MaxIndirectImageProbes = 65536u;
+constexpr uint64_t MaxIndirectDescriptorProbes = 65536u;
 
 bool SpecializationFail(std::string_view message) {
 	std::fprintf(stderr, "shader resource specialization failed: %.*s\n",
@@ -109,8 +109,7 @@ Prospero::BufferFormat ImageConversionFormat(Prospero::BufferFormat format) {
 
 enum class SamplerClass : uint8_t { Float, Integer, PointInteger };
 
-template <typename Image>
-SamplerClass ClassifySampler(const Image& image) {
+SamplerClass ClassifySampler(const ImageResource& image) {
 	if (image.numeric_class == Prospero::TextureNumericClass::Sint ||
 	    image.conversion_format != Prospero::BufferFormat::kInvalid) {
 		return SamplerClass::PointInteger;
@@ -124,8 +123,8 @@ bool DescriptorIsCube(const DescriptorValue& descriptor) {
 	       Prospero::ImageType::kCube;
 }
 
-uint32_t StorageMipCount(const ImageResource& image, const DescriptorValue& descriptor) {
-	if (image.mip_mode != ImageMipMode::DynamicStorage || NullImageDescriptor(descriptor)) {
+uint32_t ImageMipCount(const ImageResource& image, const DescriptorValue& descriptor) {
+	if (image.mip_mode != ImageMipMode::Dynamic || NullImageDescriptor(descriptor)) {
 		return 1;
 	}
 	const auto base = (descriptor.dwords[3] >> 12u) & 0xfu;
@@ -225,135 +224,193 @@ bool ReadScalarTable(uint64_t base, uint64_t size, uint64_t dynamic_offset,
 	       runtime.read_specialization_memory(runtime.userdata, address, prefix);
 }
 
-bool MaterializeIndirectImage(const ResourcePlan& program,
-                              const DescriptorSource::IndirectImage& indirect,
-                              const DescriptorValue& material_value,
-                              const DescriptorValue& table_value, uint32_t image_index,
-                              const SrtRuntime& runtime, SrtWalker& clean,
-                              ResourceSnapshot& snapshot,
-                              ResourceSpecialization& specialization) {
+bool NormalizeIndirectStoreBuffer(DescriptorValue& value) {
+	ShaderBufferResource descriptor;
+	if (!DecodeBufferDescriptor(value, descriptor)) return false;
+	if (descriptor.Type() != 0u || descriptor.Format() == Prospero::BufferFormat::kInvalid ||
+	    descriptor.Base48() == 0u || descriptor.GetSize() == 0u) {
+		value.dwords.fill(0u);
+	} else if (descriptor.Stride() != 1u || descriptor.SwizzleEnabled() || descriptor.AddTid() ||
+	           descriptor.OutOfBounds() != 3u) {
+		return SpecializationFail("indirect buffer stores require linear byte bounds");
+	}
+	return true;
+}
+
+template <typename Specialization, typename Normalize>
+bool MaterializeIndirectDescriptor(const ResourcePlan&                         program,
+                                   const DescriptorSource::IndirectDescriptor& indirect,
+                                   uint32_t resource_index, uint32_t dword_count,
+                                   const SrtRuntime& runtime, SrtWalker& clean,
+                                   ResourceSnapshot&             snapshot,
+                                   std::vector<DescriptorValue>& descriptors,
+                                   std::vector<Specialization>&  specializations,
+                                   uint32_t maximum_resources, Normalize&& normalize) {
+	const auto  descriptor_bytes = dword_count * sizeof(uint32_t);
+	const auto& sources = indirect.sources;
+	auto& keys = program.material_keys;
+	const auto read_keys = [&](const ShaderBufferResource& material, uint64_t first,
+	                           uint64_t step, uint64_t count) {
+		keys.resize(count);
+		if (step == 4u) {
+			return ReadScalarTable(material.Base48(), material.GetSize(), first, runtime, keys);
+		}
+		for (auto& key: keys) {
+			if (!ReadScalarTable(material.Base48(), material.GetSize(), first, runtime, {&key, 1}))
+				return false;
+			first += step;
+		}
+		return true;
+	};
 	uint64_t table_base = 0;
 	uint64_t table_size = UINT64_MAX; // Scalar addresses have no buffer descriptor bounds.
-	ShaderBufferResource table;
-	if (table_value.dword_count == 2u) {
-		table_base = (static_cast<uint64_t>(table_value.dwords[1]) << 32u) | table_value.dwords[0];
-	} else if (DecodeBufferDescriptor(table_value, table)) {
-		table_base = table.Base48();
-		table_size = table.GetSize();
-	} else {
-		return false;
-	}
-	auto& keys = program.material_keys;
-	keys.clear();
-	if (indirect.material_source == UINT32_MAX) {
-		uint32_t key_count = 0;
-		const bool evaluated = clean.Evaluate(indirect.key_count, key_count);
-		if (std::bit_cast<int32_t>(key_count) <= 0) key_count = 0;
-		if (table_value.dword_count != 2u || !evaluated ||
-		    key_count > MaxIndirectImageProbes ||
-		    uint64_t {indirect.table_offset} + uint64_t {key_count} * 32u > UINT32_MAX + 1ull) {
+	if (sources.empty()) {
+		keys.clear();
+		DescriptorValue material_value;
+		DescriptorValue table_value;
+		if ((indirect.material_source != UINT32_MAX &&
+		     !clean.EvaluateDescriptor(indirect.material_source, material_value)) ||
+		    !clean.EvaluateDescriptor(indirect.table_source, table_value)) {
 			return false;
 		}
-		keys.resize(key_count);
-		std::iota(keys.begin(), keys.end(), 0u);
-	} else if (!indirect.selector_mask.IsEmpty()) {
-		uint32_t mask = 0;
-		uint32_t count = 0;
-		if (material_value.dword_count != 2u || table_value.dword_count != 2u ||
-		    !clean.Evaluate(indirect.selector_mask, mask) ||
-		    !clean.Evaluate(indirect.key_count, count) || count == 0u || count > 32u) {
+		ShaderBufferResource table;
+		if (table_value.dword_count == 2u) {
+			table_base = (static_cast<uint64_t>(table_value.dwords[1]) << 32u) | table_value.dwords[0];
+		} else if (DecodeBufferDescriptor(table_value, table)) {
+			table_base = table.Base48();
+			table_size = table.GetSize();
+		} else {
 			return false;
 		}
-		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>(indirect.selector_offset) +
-			                    static_cast<uint64_t>(index) * indirect.selector_stride;
-			if (offset > UINT32_MAX) return false;
-			uint32_t key = 0;
-			if (!ReadScalarTable(material_base, UINT64_MAX, static_cast<uint32_t>(offset),
-			                     runtime, {&key, 1})) return false;
-			keys.push_back(key);
-			mask &= mask - 1u;
-		}
-		std::ranges::sort(keys);
-		keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
-	} else {
-		ShaderBufferResource material;
-		if (!DecodeBufferDescriptor(material_value, material) || table_value.dword_count != 4u ||
-		    material.Stride() != indirect.selector_stride) {
-			return false;
-		}
-		// The first aligned offset includes the immediate added after shader U32 arithmetic.
-		const auto step = std::max<uint64_t>(4u,
-		    std::gcd<uint64_t>(indirect.selector_stride, uint64_t {1} << 32u));
-		const uint64_t first = indirect.selector_offset;
-		const auto size = material.GetSize();
-		const auto limit = std::min(first + (uint64_t {1} << 32u) - step, size >= 4u ? size - 4u : 0u);
-		const auto probe_count = size >= 4u && first <= limit ? (limit - first) / step + 1u : 0u;
-		if (probe_count > MaxIndirectImageProbes) {
-			return false;
-		}
-		keys.reserve(static_cast<size_t>(probe_count) + 1u);
-		keys.push_back(0u);
-		for (uint64_t probe = 0, offset = first; probe < probe_count; ++probe, offset += step) {
-			uint32_t key = 0;
-			if (!ReadScalarTable(material.Base48(), size, offset, runtime, {&key, 1})) {
+		if (indirect.material_source == UINT32_MAX) {
+			uint32_t key_count = 0;
+			const bool evaluated = clean.Evaluate(indirect.key_count, key_count);
+			if (std::bit_cast<int32_t>(key_count) <= 0) key_count = 0;
+			if (table_value.dword_count != 2u || !evaluated ||
+			    key_count > MaxIndirectDescriptorProbes ||
+			    uint64_t {indirect.table_offset} + uint64_t {key_count} * descriptor_bytes >
+			        UINT32_MAX + 1ull) {
 				return false;
 			}
-			keys.push_back(key);
+			keys.resize(key_count);
+			std::iota(keys.begin(), keys.end(), 0u);
+		} else if (!indirect.selector_first.IsEmpty()) {
+			ShaderBufferResource material;
+			uint32_t             first = 0, count = 0;
+			if (!DecodeBufferDescriptor(material_value, material) || material.Type() != 0u ||
+			    (indirect.selector_shift != 0u && (material.Base48() & 3u) != 0u) ||
+			    material.SwizzleEnabled() ||
+			    material.AddTid() || material.OutOfBounds() != 0u ||
+			    uint64_t {indirect.selector_offset} + 4u > material.Stride() ||
+			    !clean.Evaluate(indirect.selector_first, first) ||
+			    !clean.Evaluate(indirect.key_count, count) || count > MaxIndirectDescriptorProbes ||
+			    uint64_t {first} + count > material.NumRecords())
+				return false;
+			if (count != 0u && (uint64_t {first} + count - 1u) * material.Stride() +
+			                           indirect.selector_offset + 4u >
+			                       uint64_t {UINT32_MAX} + 1u)
+				return false;
+			if (!read_keys(material, uint64_t {first} * material.Stride() + indirect.selector_offset,
+			               material.Stride(), count)) return false;
+		} else if (!indirect.selector_mask.IsEmpty()) {
+			uint32_t mask = 0;
+			uint32_t count = 0;
+			if (material_value.dword_count != 2u || table_value.dword_count != 2u ||
+			    !clean.Evaluate(indirect.selector_mask, mask) ||
+			    !clean.Evaluate(indirect.key_count, count) || count == 0u || count > 32u) {
+				return false;
+			}
+			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>(indirect.selector_offset) +
+				                    static_cast<uint64_t>(index) * indirect.selector_stride;
+				if (offset > UINT32_MAX) return false;
+				uint32_t key = 0;
+				if (!ReadScalarTable(material_base, UINT64_MAX, static_cast<uint32_t>(offset),
+				                     runtime, {&key, 1})) return false;
+				keys.push_back(key);
+				mask &= mask - 1u;
+			}
+		} else {
+			ShaderBufferResource material;
+			if (!DecodeBufferDescriptor(material_value, material) || table_value.dword_count != 4u) {
+				return false;
+			}
+			// The first aligned offset includes the immediate added after shader U32 arithmetic.
+			const auto step = std::max<uint64_t>(4u,
+			    std::gcd<uint64_t>(indirect.selector_stride, uint64_t {1} << 32u));
+			const uint64_t first = indirect.selector_offset;
+			const auto size = material.GetSize();
+			const auto limit = std::min(first + (uint64_t {1} << 32u) - step, size >= 4u ? size - 4u : 0u);
+			const auto probe_count = size >= 4u && first <= limit ? (limit - first) / step + 1u : 0u;
+			if (probe_count > MaxIndirectDescriptorProbes) {
+				return false;
+			}
+			if (!read_keys(material, first, step, probe_count)) return false;
+			if (indirect.selector_shift == 0u) keys.push_back(0u);
+		}
+		if (indirect.material_source != UINT32_MAX) {
+			if (indirect.selector_shift != 0u) {
+				for (auto& key: keys) key >>= indirect.selector_shift;
+				keys.push_back(0u); // An out-of-range material load returns zero.
+			}
+			std::ranges::sort(keys);
+			keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
 		}
-		std::ranges::sort(keys);
-		keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
 	}
 
-	const auto children_begin = snapshot.images.size();
+	const auto children_begin = descriptors.size();
 	const auto mapping_offset = snapshot.flattened_srt.size();
-	const auto root_image = specialization.images[image_index];
-	snapshot.flattened_srt.resize(mapping_offset + 1u + keys.size() * 2u);
-	snapshot.flattened_srt[mapping_offset] = static_cast<uint32_t>(keys.size());
-	for (uint32_t entry = 0; entry < keys.size(); ++entry) {
-		const auto key = keys[entry];
+	const auto root_resource  = specializations[resource_index];
+	const auto key_count = sources.empty() ? keys.size() : sources.size();
+	snapshot.flattened_srt.resize(mapping_offset + 1u + key_count * 2u);
+	snapshot.flattened_srt[mapping_offset] = static_cast<uint32_t>(key_count);
+	for (uint32_t entry = 0; entry < key_count; ++entry) {
+		const auto key = sources.empty() ? keys[entry] : entry;
 		DescriptorValue candidate;
-		candidate.dword_count = 8u;
-		const auto table_offset = (key << 5u) + indirect.table_offset;
-		if (!ReadScalarTable(table_base, table_size, table_offset, runtime, candidate.dwords)) {
+		candidate.dword_count = dword_count;
+		if (sources.empty()) {
+			const auto table_offset =
+			    static_cast<uint32_t>(key * descriptor_bytes) + indirect.table_offset;
+			if (!ReadScalarTable(table_base, table_size, table_offset, runtime,
+			                     std::span(candidate.dwords).first(dword_count))) {
+				return false;
+			}
+		} else if (!clean.EvaluateDescriptor(sources[entry], candidate)) {
 			return false;
 		}
-		if (NullImageDescriptor(candidate) ||
-		    !ValidImageDescriptor(candidate, program.info.images[image_index].r128)) {
-			candidate.dwords.fill(0);
-		}
+		if (!normalize(candidate)) return false;
 		uint32_t ordinal = 0;
 		if (entry == 0) {
-			snapshot.images[image_index] = candidate;
-		} else if (snapshot.images[image_index] != candidate) {
-			const auto found = std::find(snapshot.images.begin() + children_begin,
-			                             snapshot.images.end(), candidate);
-			ordinal = static_cast<uint32_t>(found - snapshot.images.begin() - children_begin + 1u);
-			if (found == snapshot.images.end()) {
-				if (snapshot.images.size() >= ShaderInfo::MaxImages) {
+			descriptors[resource_index] = candidate;
+		} else if (descriptors[resource_index] != candidate) {
+			const auto found =
+			    std::find(descriptors.begin() + children_begin, descriptors.end(), candidate);
+			ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
+			if (found == descriptors.end()) {
+				if (descriptors.size() >= maximum_resources) {
 					return false;
 				}
-				snapshot.images.push_back(candidate);
-				auto child = root_image;
-				child.indirect_root = image_index;
-				specialization.images.push_back(child);
+				descriptors.push_back(candidate);
+				auto child          = root_resource;
+				child.indirect_root = resource_index;
+				specializations.push_back(child);
 			}
 		}
 		snapshot.flattened_srt[mapping_offset + 1u + entry * 2u] = key;
 		snapshot.flattened_srt[mapping_offset + 2u + entry * 2u] = ordinal;
 	}
-	if (snapshot.images.size() == children_begin) {
+	if (descriptors.size() == children_begin) {
 		snapshot.flattened_srt.resize(mapping_offset);
 	} else {
-		auto& root = specialization.images[image_index];
-		root.indirect_root = image_index;
+		auto& root                      = specializations[resource_index];
+		root.indirect_root              = resource_index;
 		root.indirect_mapping_offset = static_cast<uint32_t>(mapping_offset);
-		root.indirect_search_iterations = std::bit_width(keys.size());
+		root.indirect_search_iterations = std::bit_width(key_count);
 	}
 	return true;
 }
@@ -370,10 +427,21 @@ struct SamplerPlan {
 	uint32_t                                                  sampler_count = 0;
 };
 
+template <typename T, typename Keep>
+void CompactImages(std::vector<T>& images, Keep&& keep) {
+	size_t count = 0;
+	for (size_t index = 0; index < images.size(); ++index) {
+		if (!keep(index)) continue;
+		if (count != index) images[count] = std::move(images[index]);
+		++count;
+	}
+	images.resize(count);
+}
+
 struct ImageRemap {
 	explicit ImageRemap(const ResourceSpecialization& specialization)
-	    : source_count(static_cast<uint32_t>(specialization.images.size())) {
-		EXIT_IF(specialization.images.size() > indices.size());
+	    : indices(specialization.images.size()),
+	      source_count(static_cast<uint32_t>(specialization.images.size())) {
 		for (uint32_t index = 0; index < source_count; index++) {
 			indices[index] = specialization.images[index].fmask ? UINT32_MAX : count++;
 		}
@@ -390,55 +458,17 @@ struct ImageRemap {
 		if (count == source_count) {
 			return;
 		}
-		for (uint32_t index = 0; index < source_count; index++) {
-			if (indices[index] != UINT32_MAX && indices[index] != index) {
-				images[indices[index]] = std::move(images[index]);
-			}
-		}
-		images.resize(count);
+		CompactImages(images, [&](size_t index) { return indices[index] != UINT32_MAX; });
 	}
 
 private:
-	std::array<uint32_t, ShaderInfo::MaxImages> indices;
+	std::vector<uint32_t> indices;
 	uint32_t                                    source_count;
 	uint32_t                                    count = 0;
 };
 
-template <typename Images>
-bool BuildSamplerPlan(const ShaderInfo& base, const Images& images, SamplerPlan& plan);
-
 static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSnapshot& snapshot,
                                         ResourceSpecialization& specialization) {
-	specialization.buffers.clear();
-	specialization.buffers.reserve(program.info.buffers.size());
-	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
-		auto&                descriptor_value = snapshot.buffers[i];
-		ShaderBufferResource descriptor;
-		if (!DecodeBufferDescriptor(descriptor_value, descriptor)) {
-			return SpecializationFail(fmt::format("buffer descriptor {} has invalid width", i));
-		}
-		if (descriptor.Type() != 0) {
-			descriptor_value.dwords.fill(0);
-			descriptor = {};
-		}
-		auto       packed_stride = descriptor.PackedStride();
-		const auto stride        = packed_stride & 0x3fffu;
-		const bool swizzle       = stride != 0u && ((packed_stride >> 14u) & 1u) != 0u;
-		if (stride == 0u) {
-			packed_stride &= ~((1u << 14u) | (3u << 16u));
-		} else if (!swizzle) {
-			packed_stride &= ~(3u << 16u);
-		}
-		specialization.buffers.push_back({
-		    .packed_stride     = packed_stride,
-		    .descriptor_format = program.info.buffers[i].formatted
-		                             ? descriptor.Format()
-		                             : Prospero::BufferFormat::kInvalid,
-		    .descriptor_swizzle =
-		        program.info.buffers[i].formatted ? descriptor.DstSelXYZW() : DstSel(4, 5, 6, 7),
-		    .zero_stride_oob = descriptor.OutOfBounds() == 0u && stride == 0u,
-		});
-	}
 	for (uint32_t i = 0; i < specialization.images.size(); i++) {
 		const auto& descriptor = snapshot.images[i];
 		auto&       image      = specialization.images[i];
@@ -451,10 +481,10 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 		    (base.atomic && base.resource_class != ImageResourceClass::Storage)) {
 			return SpecializationFail(fmt::format("image resource {} has an invalid class", i));
 		}
-		image.mip_count = StorageMipCount(base, descriptor);
+		image.mip_count = ImageMipCount(base, descriptor);
 		if (image.mip_count == 0u) {
 			return SpecializationFail(
-			    fmt::format("storage image descriptor {} has an invalid mip range", i));
+			    fmt::format("image descriptor {} has an invalid mip range", i));
 		}
 		if (NullImageDescriptor(descriptor)) {
 			image.numeric_class = base.atomic ? Prospero::TextureNumericClass::Uint
@@ -476,8 +506,10 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 		image.cube      = DescriptorIsCube(descriptor);
 		const auto format =
 		    static_cast<Prospero::BufferFormat>((descriptor.dwords[1] >> 20u) & 0x1ffu);
-		if (base.atomic && format != Prospero::BufferFormat::k32UInt &&
-		    format != Prospero::BufferFormat::k32Float) {
+		if (base.atomic &&
+		    (base.atomic64 ? format != Prospero::BufferFormat::k32_32UInt
+		                   : format != Prospero::BufferFormat::k32UInt &&
+		                         format != Prospero::BufferFormat::k32Float)) {
 			return SpecializationFail(
 			    fmt::format("atomic image descriptor {} uses unsupported format {}", i,
 			                static_cast<uint32_t>(format)));
@@ -577,30 +609,21 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			}
 		}
 	}
-	SamplerPlan sampler_plan;
-	if (!BuildSamplerPlan(program.info, specialization.images, sampler_plan)) {
-		return SpecializationFail("specialized sampler layout exceeds its resource limit");
-	}
-	for (uint32_t index = static_cast<uint32_t>(program.info.samplers.size());
-	     index < sampler_plan.sampler_count; index++) {
-		snapshot.samplers.push_back(snapshot.samplers[sampler_plan.bindings[index].source]);
-	}
-	ImageRemap(specialization).Apply(snapshot.images);
+	CompactImages(snapshot.images, [&](size_t index) { return !specialization.images[index].fmask; });
 	return true;
 }
 
-template <typename Images>
-bool BuildSamplerPlan(const ShaderInfo& base, const Images& images, SamplerPlan& plan) {
+bool BuildSamplerPlan(const ShaderInfo& base, SamplerPlan& plan) {
 	if (base.samplers.size() > plan.mapping.size()) {
 		return false;
 	}
 	std::array<uint8_t, ShaderInfo::MaxSamplers> usage {};
 	plan.sampler_count = static_cast<uint32_t>(base.samplers.size());
 	for (const auto& pair: base.sampled_pairs) {
-		if (pair.image >= images.size() || pair.sampler >= base.samplers.size()) {
+		if (pair.image >= base.images.size() || pair.sampler >= base.samplers.size()) {
 			return false;
 		}
-		usage[pair.sampler] |= 1u << static_cast<uint32_t>(ClassifySampler(images[pair.image]));
+		usage[pair.sampler] |= 1u << static_cast<uint32_t>(ClassifySampler(base.images[pair.image]));
 	}
 	for (uint32_t index = 0; index < base.samplers.size(); index++) {
 		auto& mapping = plan.mapping[index];
@@ -632,6 +655,7 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 		}
 	}
 	std::vector<ResourceBlock> blocks(program.blocks.size());
+	std::vector<std::vector<uint32_t>> predecessors(blocks.size());
 	for (uint32_t i = 0; i < blocks.size(); i++) {
 		auto&                 block      = blocks[i];
 		const auto&           info       = program.block_info[i];
@@ -655,6 +679,7 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 				return {};
 			}
 			block.successors.push_back(found->second);
+			predecessors[found->second].push_back(i);
 		}
 		for (const auto& inst: *program.blocks[i]) {
 			const auto op     = inst.GetOpcode();
@@ -684,7 +709,30 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 		block.sources.erase(std::unique(block.sources.begin(), block.sources.end()),
 		                    block.sources.end());
 	}
-	for (auto& block: blocks) block.condition = predicate(block.condition);
+	// Host branches only matter when they guard a later resource access. Leave
+	// data-only branches (for example a mutable flag checked before a trap) on the GPU.
+	std::vector<uint8_t> resource_reachable(blocks.size());
+	std::vector<uint32_t> pending;
+	for (uint32_t i = 0; i < blocks.size(); ++i) {
+		if (!blocks[i].sources.empty() || !blocks[i].srt_reads.empty()) {
+			resource_reachable[i] = 1;
+			pending.push_back(i);
+		}
+	}
+	while (!pending.empty()) {
+		const auto index = pending.back();
+		pending.pop_back();
+		for (const auto predecessor: predecessors[index]) {
+			if (resource_reachable[predecessor]) continue;
+			resource_reachable[predecessor] = 1;
+			pending.push_back(predecessor);
+		}
+	}
+	for (auto& block: blocks) {
+		block.condition = std::ranges::any_of(block.successors, [&](uint32_t successor) {
+			return resource_reachable[successor] != 0;
+		}) ? predicate(block.condition) : Value {};
+	}
 	if (std::ranges::none_of(
 	        blocks, [](const ResourceBlock& block) { return !block.condition.IsEmpty(); })) {
 		return {};
@@ -692,11 +740,33 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 	return blocks;
 }
 
+// ResolveComputeBufferFill proves complete workgroups and matching requested extents, so every
+// dispatch-bound predicate accepted here is true for every invocation covered by the fill.
+static bool IsFullDispatchPredicate(Value value, uint32_t depth = 0) {
+	value = value.Resolve();
+	if (value == Value(true)) return true;
+	const auto* inst = value.TryInstruction();
+	if (inst == nullptr || depth > 32) return false;
+	if (inst->GetOpcode() == ValueOpcode::LogicalAnd) {
+		return IsFullDispatchPredicate(inst->Arg(0), depth + 1) &&
+		       IsFullDispatchPredicate(inst->Arg(1), depth + 1);
+	}
+	if (inst->GetOpcode() != ValueOpcode::ULessThan32) return false;
+	const auto* global = inst->Arg(0).Resolve().TryInstruction();
+	const auto* extent = inst->Arg(1).Resolve().TryInstruction();
+	if (global == nullptr || global->GetOpcode() != ValueOpcode::GetBuiltin ||
+	    global->Arg(0).Resolve() != Value(static_cast<uint32_t>(StageInputKind::GlobalInvocationId)) ||
+	    extent == nullptr || extent->GetOpcode() != ValueOpcode::GetDispatchThreadExtent)
+		return false;
+	const auto axis = global->Arg(1).Resolve();
+	return axis.IsImmediate() && axis.U32() < 3 && extent->Arg(0).Resolve() == axis;
+}
+
 // Nonnegative affine coefficients for constant, local and workgroup coordinates. Reject modular
 // arithmetic that could wrap; runtime coverage also bounds the largest invocation index.
-static std::optional<std::array<uint64_t, 3>> FillIndex(Value value, uint32_t axis,
+static std::optional<std::array<uint64_t, 3>> FillIndex(Value value, uint32_t axis, Value guard,
                                                       uint32_t depth = 0) {
-	value = value.Resolve();
+	value = ResolveActiveU32(value, guard);
 	if (depth > 32 || value.GetType() != Type::U32) {
 		return {};
 	}
@@ -720,8 +790,8 @@ static std::optional<std::array<uint64_t, 3>> FillIndex(Value value, uint32_t ax
 	    op != ValueOpcode::ShiftLeftLogical32) {
 		return {};
 	}
-	auto left  = FillIndex(inst->Arg(0), axis, depth + 1);
-	auto right = FillIndex(inst->Arg(1), axis, depth + 1);
+	auto left  = FillIndex(inst->Arg(0), axis, guard, depth + 1);
+	auto right = FillIndex(inst->Arg(1), axis, guard, depth + 1);
 	if (!left || !right) {
 		return {};
 	}
@@ -777,6 +847,7 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 	UniformFillPlan result;
 	result.fill.resource = memory.resource;
 	Value data;
+	Value guard(true);
 	if (store->GetOpcode() == ValueOpcode::ImageWrite) {
 		if (program.info.images.size() != 1 || memory.dmask != 1 || memory.data_bits != 32 ||
 		    memory.image_has_mip || memory.image_sample_flags != 0 || memory.image_r128 ||
@@ -787,7 +858,7 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 		const auto* address = store->Arg(1).ResolveInstruction();
 		if (address == nullptr || address->GetOpcode() != ValueOpcode::MakeImageAddress) return {};
 		for (uint32_t axis = 0; axis < 3; ++axis) {
-			const auto index = FillIndex(address->Arg(axis), axis);
+			const auto index = FillIndex(address->Arg(axis), axis, guard);
 			if (!index || (*index)[0] != 0) return {};
 			if (axis < 2) {
 				if ((*index)[1] != 1 || (*index)[2] == 0) return {};
@@ -809,20 +880,22 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 		                             ValueOpcode::StoreBufferU32x3, ValueOpcode::StoreBufferU32x4};
 		const auto           store_op = std::ranges::find(stores, op);
 		if (store_op == stores.end() || store->Arg(2).Resolve() != Value(0u) ||
-		    store->Arg(3).Resolve() != Value(0u) || store->Arg(5).Resolve() != Value(true))
+		    store->Arg(3).Resolve() != Value(0u))
 			return {};
+		guard = store->Arg(5).Resolve();
+		if (!IsFullDispatchPredicate(guard)) return {};
 		if (!memory.formatted || memory.typed || !memory.idxen || memory.offen || memory.offset != 0 ||
 		    memory.data_bits != 32 ||
 		    memory.data_dwords != static_cast<uint32_t>(store_op - stores.begin() + 1))
 			return {};
-		const auto address = FillIndex(store->Arg(1), 0);
+		const auto address = FillIndex(store->Arg(1), 0, guard);
 		if (!address || (*address)[0] != 0 || (*address)[1] != 1 || (*address)[2] == 0) return {};
 		result.fill.kind = UniformFillKind::Buffer;
 		result.fill.group_stride[0] = static_cast<uint32_t>((*address)[2]);
 		result.fill.words = memory.data_dwords;
 		data = store->Arg(4);
 	}
-	data = data.Resolve();
+	data = ResolveActiveU32(data, guard);
 	const auto*          vector = data.TryInstruction();
 	constexpr std::array composites {ValueOpcode::CompositeConstructU32x2,
 	                                 ValueOpcode::CompositeConstructU32x3,
@@ -831,7 +904,7 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 	    (vector == nullptr || vector->GetOpcode() != composites[result.fill.words - 2]))
 		return {};
 	for (uint32_t i = 0; i < result.fill.words; ++i) {
-		const auto word = result.fill.words == 1 ? data : vector->Arg(i);
+		const auto word = ResolveActiveU32(result.fill.words == 1 ? data : vector->Arg(i), guard);
 		if (word.GetType() != Type::U32 ||
 		    !ValidateRuntimeValue(program, word, RuntimeValueType::Integer))
 			return {};
@@ -858,6 +931,11 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		if (source == nullptr) {
 			return value;
 		}
+		if (source->GetOpcode() == ValueOpcode::ReadFirstLane &&
+		    ValidateRuntimeValue(program, source->Arg(0))) {
+			// Uniform values need no new EXEC context; preserve their shared evaluation memo.
+			return Clone(source->Arg(0));
+		}
 		if (source->GetOpcode() == ValueOpcode::Phi) {
 			const auto invariant = ResolveInvariantPhi(program, value);
 			if (!invariant.IsEmpty() && invariant != value) {
@@ -867,6 +945,10 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		if (const auto found = cloned.find(source); found != cloned.end()) {
 			return Value(found->second);
 		}
+		if (source->GetOpcode() == ValueOpcode::LoadBufferU32) {
+			plan.requires_specialization_memory = true;
+			plan.capture_specialization_reads   = true;
+		}
 		auto& target =
 		    plan.value_storage.emplace_back(source->GetOpcode(), source->Flags<uint64_t>());
 		cloned.emplace(source, &target);
@@ -886,10 +968,13 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	for (const auto& source: program.descriptor_sources) {
 		auto& target          = plan.descriptor_sources.emplace_back();
 		target.dword_count    = source.dword_count;
-		target.indirect_image = source.indirect_image;
-		if (target.indirect_image.has_value()) {
-			target.indirect_image->key_count = Clone(target.indirect_image->key_count);
-			target.indirect_image->selector_mask = Clone(target.indirect_image->selector_mask);
+		target.indirect_descriptor = source.indirect_descriptor;
+		if (target.indirect_descriptor.has_value()) {
+			target.indirect_descriptor->key_count = Clone(target.indirect_descriptor->key_count);
+			target.indirect_descriptor->selector_first =
+			    Clone(target.indirect_descriptor->selector_first);
+			target.indirect_descriptor->selector_mask =
+			    Clone(target.indirect_descriptor->selector_mask);
 		}
 		for (uint32_t dword = 0; dword < source.dword_count; dword++) {
 			target.dwords[dword] = Clone(source.dwords[dword]);
@@ -936,28 +1021,35 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		plan.uniform_fill.values[i] = Clone(plan.uniform_fill.values[i]);
 	}
 	plan.clean_flat_slots.resize(plan.srt_reads.size());
-	bool masked_image = false;
-	for (const auto& image: plan.info.images) {
-		const auto* source = Source(plan, image.source);
-		if (source == nullptr || !source->indirect_image.has_value()) {
-			continue;
-		}
+	bool capture_indirect_reads = false;
+	for (const auto& source: plan.descriptor_sources) {
+		if (!source.indirect_descriptor.has_value()) continue;
+		const auto& indirect                = *source.indirect_descriptor;
 		plan.requires_specialization_memory = true;
-		masked_image |= !source->indirect_image->selector_mask.IsEmpty();
-		MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->material_source),
-		                   plan.clean_flat_slots, source->indirect_image->selector_mask);
-		MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->table_source),
-		                   plan.clean_flat_slots);
+		capture_indirect_reads |= indirect.material_source != UINT32_MAX || !indirect.sources.empty();
+		MarkCleanFlatSlots(plan, Source(plan, indirect.material_source), plan.clean_flat_slots,
+		                   indirect.selector_mask);
+		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.selector_first);
+		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.key_count);
+		if (indirect.sources.empty()) {
+			MarkCleanFlatSlots(plan, Source(plan, indirect.table_source), plan.clean_flat_slots);
+		} else {
+			for (const auto candidate: indirect.sources) {
+				MarkCleanFlatSlots(plan, Source(plan, candidate), plan.clean_flat_slots);
+			}
+		}
 	}
-	plan.capture_specialization_reads = masked_image || !plan.control_flow.empty();
-	// Writable descriptor addresses must come from clean backing for the alias proof.
-	for (const auto& buffer: plan.info.buffers) {
-		if (buffer.written) MarkCleanFlatSlots(plan, Source(plan, buffer.source), plan.clean_flat_slots);
+	plan.capture_specialization_reads |= capture_indirect_reads || !plan.control_flow.empty();
+	if (plan.capture_specialization_reads) {
+		// Clean writable addresses prove that shader writes cannot overlap captured reads.
+		for (const auto& buffer: plan.info.buffers) {
+			if (buffer.written) MarkCleanFlatSlots(plan, Source(plan, buffer.source), plan.clean_flat_slots);
+		}
+		for (const auto& image: plan.info.images) {
+			if (image.written) MarkCleanFlatSlots(plan, Source(plan, image.source), plan.clean_flat_slots);
+		}
 	}
-	for (const auto& image: plan.info.images) {
-		if (image.written) MarkCleanFlatSlots(plan, Source(plan, image.source), plan.clean_flat_slots);
-	}
-	if (masked_image) plan.resource_tracking_complete &= !program.has_address_writes;
+	if (capture_indirect_reads) plan.resource_tracking_complete &= !program.has_address_writes;
 	return plan;
 }
 
@@ -979,7 +1071,8 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		observed.read_memory = CaptureOrdinaryRead;
 	}
 	SrtWalker clean(program, CleanRuntime(observed));
-	SrtWalker walker(program, observed, program.clean_flat_slots, &clean);
+	SrtWalker walker(program, observed, program.clean_flat_slots,
+	                 capture_reads || program.requires_specialization_memory ? &clean : nullptr);
 	if (!walker.RefreshFlatBuffer(snapshot.flattened_srt)) {
 		return false;
 	}
@@ -1001,18 +1094,58 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			return false;
 		}
 		if (active.empty() || active[source]) {
-			return (written ? clean : walker).EvaluateDescriptor(source, value);
+			return (capture_reads && written ? clean : walker).EvaluateDescriptor(source, value);
 		}
 		value = {};
 		value.dword_count = program.descriptor_sources[source].dword_count;
 		return true;
 	};
 	snapshot.buffers.resize(program.info.buffers.size());
-	for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
-		if (!evaluate(program.info.buffers[i].source, snapshot.buffers[i],
-		              program.info.buffers[i].written)) {
-			return false;
+	specialization.buffers.assign(program.info.buffers.size(), {});
+	for (uint32_t i = 0; i < snapshot.buffers.size(); ++i) {
+		const auto base_index =
+		    i < program.info.buffers.size() ? i : specialization.buffers[i].indirect_root;
+		if (base_index >= program.info.buffers.size()) return false;
+		const auto& base = program.info.buffers[base_index];
+		if (i < program.info.buffers.size()) {
+			const auto* source = Source(program, base.source);
+			if (source == nullptr) return false;
+			if (source->indirect_descriptor.has_value()) {
+				snapshot.buffers[i] = {.dword_count = 4u};
+				if (active.empty() || active[base.source]) {
+					if (!MaterializeIndirectDescriptor(
+					        program, *source->indirect_descriptor, i, 4u, observed, clean, snapshot,
+					        snapshot.buffers, specialization.buffers, ShaderInfo::MaxBuffers,
+					        NormalizeIndirectStoreBuffer))
+						return false;
+				}
+			} else if (!evaluate(base.source, snapshot.buffers[i], base.written)) {
+				return false;
+			}
 		}
+		auto&                descriptor_value = snapshot.buffers[i];
+		ShaderBufferResource descriptor;
+		if (!DecodeBufferDescriptor(descriptor_value, descriptor)) {
+			return SpecializationFail(fmt::format("buffer descriptor {} has invalid width", i));
+		}
+		if (descriptor.Type() != 0) {
+			descriptor_value.dwords.fill(0);
+			descriptor = {};
+		}
+		auto       packed_stride = descriptor.PackedStride();
+		const auto stride        = packed_stride & 0x3fffu;
+		const bool swizzle       = stride != 0u && ((packed_stride >> 14u) & 1u) != 0u;
+		if (stride == 0u) {
+			packed_stride &= ~((1u << 14u) | (3u << 16u));
+		} else if (!swizzle) {
+			packed_stride &= ~(3u << 16u);
+		}
+		auto& buffer         = specialization.buffers[i];
+		buffer.packed_stride = packed_stride;
+		buffer.descriptor_format =
+		    base.formatted ? descriptor.Format() : Prospero::BufferFormat::kInvalid;
+		buffer.descriptor_swizzle = base.formatted ? descriptor.DstSelXYZW() : DstSel(4, 5, 6, 7);
+		buffer.zero_stride_oob    = descriptor.OutOfBounds() == 0u && stride == 0u;
 	}
 	snapshot.images.resize(program.info.images.size());
 	specialization.images.resize(program.info.images.size());
@@ -1033,19 +1166,19 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		if (source == nullptr) {
 			return false;
 		}
-		if (source->indirect_image.has_value()) {
+		if (source->indirect_descriptor.has_value()) {
 			snapshot.images[i] = {.dword_count = 8u};
 			if (!active.empty() && !active[image.source]) {
 				continue;
 			}
-			const auto& indirect = *source->indirect_image;
-			DescriptorValue material;
-			DescriptorValue table;
-			if ((indirect.material_source != UINT32_MAX &&
-			     !clean.EvaluateDescriptor(indirect.material_source, material)) ||
-			    !clean.EvaluateDescriptor(indirect.table_source, table) ||
-			    !MaterializeIndirectImage(program, indirect, material, table, i, observed, clean, snapshot,
-			                              specialization)) {
+			const auto& indirect = *source->indirect_descriptor;
+			if (!MaterializeIndirectDescriptor(
+			        program, indirect, i, 8u, observed, clean, snapshot, snapshot.images,
+			        specialization.images, UINT32_MAX, [&](DescriptorValue& value) {
+				        if (NullImageDescriptor(value) || !ValidImageDescriptor(value, image.r128))
+					        value.dwords.fill(0u);
+				        return true;
+			        })) {
 				return false;
 			}
 		} else {
@@ -1062,6 +1195,16 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		if (!evaluate(program.info.samplers[i].source, snapshot.samplers[i])) {
 			return false;
 		}
+		if (program.info.samplers[i].gather_lod) {
+			const auto control = snapshot.samplers[i].dwords[2];
+			const auto filter = (control >> 26u) & 3u;
+			// MipNone always selects the base level. Explicit point gathers currently require
+			// encoded-zero primary and secondary bias; linear primary-mip selection is unsupported.
+			if (filter > 1u || (filter == 1u && (control & 0xfffffu) != 0u)) {
+				return SpecializationFail(
+				    "explicit-LOD gather requires mip filtering None or Point with zero LOD biases");
+			}
+		}
 	}
 	snapshot.user_data.assign(runtime.user_data.begin(), runtime.user_data.end());
 	return BuildResourceSpecialization(program, snapshot, specialization);
@@ -1070,22 +1213,42 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 void ApplyResourceSpecialization(Program& program, const ResourceSpecialization& specialization) {
 	EXIT_IF(!program.resource_tracking_complete || program.shader_info_complete ||
 	        program.binding_layout_complete);
-	EXIT_IF(program.info.buffers.size() != specialization.buffers.size() ||
+	EXIT_IF(program.info.buffers.size() > specialization.buffers.size() ||
 	        program.info.images.size() > specialization.images.size());
 
-	auto buffers = program.info.buffers;
-	for (size_t index = 0; index < buffers.size(); index++) {
-		buffers[index].packed_stride      = specialization.buffers[index].packed_stride;
-		buffers[index].descriptor_format  = specialization.buffers[index].descriptor_format;
-		buffers[index].descriptor_swizzle = specialization.buffers[index].descriptor_swizzle;
+	auto& buffers = program.info.buffers;
+	const auto original_buffer_count = buffers.size();
+	buffers.reserve(specialization.buffers.size());
+	for (uint32_t index = 0; index < specialization.buffers.size(); ++index) {
+		const auto& source = specialization.buffers[index];
+		if (index >= buffers.size()) {
+			EXIT_IF(source.indirect_root >= original_buffer_count);
+			buffers.push_back(buffers[source.indirect_root]);
+		}
+		auto& buffer                      = buffers[index];
+		buffer.packed_stride              = source.packed_stride;
+		buffer.descriptor_format          = source.descriptor_format;
+		buffer.descriptor_swizzle         = source.descriptor_swizzle;
+		buffer.indirect_root              = source.indirect_root;
+		buffer.indirect_mapping_offset    = source.indirect_mapping_offset;
+		buffer.indirect_search_iterations = source.indirect_search_iterations;
+		buffer.indirect_resources.clear();
 	}
-	auto images = program.info.images;
+	for (uint32_t index = 0; index < buffers.size(); ++index) {
+		const auto root = buffers[index].indirect_root;
+		if (root != BufferResource::NoIndirectBuffer) {
+			EXIT_IF(root >= buffers.size());
+			buffers[root].indirect_resources.push_back(index);
+		}
+	}
+	auto& images = program.info.images;
+	const auto original_image_count = images.size();
 	images.reserve(specialization.images.size());
 	for (uint32_t index = 0; index < specialization.images.size(); index++) {
 		const auto& source = specialization.images[index];
 		if (index >= images.size()) {
-			EXIT_IF(source.indirect_root >= program.info.images.size());
-			images.push_back(program.info.images[source.indirect_root]);
+			EXIT_IF(source.indirect_root >= original_image_count);
+			images.push_back(images[source.indirect_root]);
 		}
 		auto& image                      = images[index];
 		image.numeric_class              = source.numeric_class;
@@ -1108,15 +1271,17 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 	}
 
 	SamplerPlan sampler_plan;
-	EXIT_IF(!BuildSamplerPlan(program.info, images, sampler_plan));
-	auto samplers      = program.info.samplers;
-	auto sampled_pairs = program.info.sampled_pairs;
+	EXIT_IF(!BuildSamplerPlan(program.info, sampler_plan));
+	auto& samplers      = program.info.samplers;
+	auto& sampled_pairs = program.info.sampled_pairs;
+	const auto original_sampler_count = samplers.size();
 	samplers.reserve(sampler_plan.sampler_count);
 	for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {
 		const auto& binding = sampler_plan.bindings[index];
-		if (index >= program.info.samplers.size()) {
-			samplers.push_back(program.info.samplers[binding.source]);
+		if (index >= original_sampler_count) {
+			samplers.push_back(samplers[binding.source]);
 		}
+		samplers[index].snapshot_index = binding.source;
 		samplers[index].force_point_filtering = binding.type == SamplerClass::PointInteger;
 		samplers[index].integer_border        = binding.type != SamplerClass::Float;
 	}
@@ -1127,11 +1292,15 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 		samplers[pair.sampler].depth_compare |= images[pair.image].depth_compare;
 	}
 
-	auto memory_info = program.memory_info;
+	auto& memory_info = program.memory_info;
 	const ImageRemap image_remap(specialization);
 	for (auto* block: program.blocks) {
 		for (auto it = block->begin(); it != block->end(); ++it) {
 			auto& inst = *it;
+			if (inst.GetOpcode() == ValueOpcode::GetImageResource) {
+				inst.SetFlags(image_remap[inst.Flags<uint32_t>()]);
+				continue;
+			}
 			if (BufferAccessOf(inst.GetOpcode()) == BufferAccess::Read) {
 				const auto& memory = memory_info[inst.Flags<MemoryFlags>().index];
 				if (memory.kind == ResourceKind::Buffer &&
@@ -1202,7 +1371,7 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 				continue;
 			}
 			if (image_opcode.needs_sampler &&
-			    memory.sampler < program.info.samplers.size()) {
+			    memory.sampler < original_sampler_count) {
 				const auto type = static_cast<uint32_t>(ClassifySampler(image));
 				memory.sampler = sampler_plan.mapping[memory.sampler][type];
 				EXIT_IF(memory.sampler == UINT32_MAX);
@@ -1211,13 +1380,6 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 			        inst.GetOpcode() != ValueOpcode::ImageSampleRaw);
 		}
 	}
-	for (auto* block: program.blocks) {
-		for (auto& inst: *block) {
-			if (inst.GetOpcode() == ValueOpcode::GetImageResource) {
-				inst.SetFlags(image_remap[inst.Flags<uint32_t>()]);
-			}
-		}
-	}
 	for (auto& memory: memory_info) {
 		if (memory.kind == ResourceKind::Image && !memory.planning_only) {
 			memory.resource = image_remap[memory.resource];
@@ -1240,11 +1402,6 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 		}
 	}
 	image_remap.Apply(images);
-	program.info.buffers       = std::move(buffers);
-	program.info.images        = std::move(images);
-	program.info.samplers      = std::move(samplers);
-	program.info.sampled_pairs = std::move(sampled_pairs);
-	program.memory_info        = std::move(memory_info);
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -13,6 +13,9 @@ struct ResourceSpecialization {
 		Prospero::BufferFormat descriptor_format               = Prospero::BufferFormat::kInvalid;
 		uint32_t               descriptor_swizzle              = DstSel(4, 5, 6, 7);
 		bool                   zero_stride_oob                 = false;
+		uint32_t               indirect_root                   = BufferResource::NoIndirectBuffer;
+		uint32_t               indirect_mapping_offset         = 0;
+		uint32_t               indirect_search_iterations      = 0;
 		bool                   operator==(const Buffer&) const = default;
 	};
 

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

@@ -412,7 +412,7 @@ void CollectShaderInfo(Program& program, ShaderStageInputInfo input_info) {
 	ValidateOptions(program, input_info);
 	ValidateValueReferences(program, input_info);
 
-	auto next = program.info;
+	auto& next = program.info;
 	next.inputs.clear();
 	next.outputs.clear();
 	next.has_bitwise_xor =
@@ -433,7 +433,6 @@ void CollectShaderInfo(Program& program, ShaderStageInputInfo input_info) {
 	}
 	CollectBuiltinInputs(program, next);
 	CollectOutputs(program, input_info, next);
-	program.info                 = std::move(next);
 	program.shader_info_complete = true;
 }
 

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

@@ -23,9 +23,6 @@ namespace {
 constexpr uint64_t AddressMask = 0x0000ffffffffffffull;
 
 bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
-	if (base > AddressMask) {
-		return false;
-	}
 	if (offset < 0) {
 		const auto magnitude = uint64_t {0} - static_cast<uint64_t>(offset);
 		if (magnitude > base) {
@@ -35,7 +32,7 @@ bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
 		return true;
 	}
 	const auto magnitude = static_cast<uint64_t>(offset);
-	if (magnitude > AddressMask - base) {
+	if (magnitude > UINT64_MAX - base) {
 		return false;
 	}
 	result = base + magnitude;
@@ -44,16 +41,29 @@ bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
 
 bool IsRawRead(const ResourcePlan& values, const Inst& inst) {
 	const auto op = inst.GetOpcode();
-	if (op != ValueOpcode::LoadAddressU32 && op != ValueOpcode::ReadConstBuffer) {
+	if (op != ValueOpcode::LoadAddressU32 && op != ValueOpcode::ReadConstBuffer &&
+	    op != ValueOpcode::LoadBufferU32) {
 		return false;
 	}
 	const auto index = inst.Flags<MemoryFlags>().index;
 	if (index >= values.memory_info.size()) {
 		return false;
 	}
-	const auto kind = values.memory_info[index].kind;
-	return (op == ValueOpcode::LoadAddressU32 && kind == ResourceKind::ScalarAddress) ||
-	       (op == ValueOpcode::ReadConstBuffer && kind == ResourceKind::ScalarBuffer);
+	const auto& memory = values.memory_info[index];
+	if (op != ValueOpcode::LoadBufferU32) {
+		return (op == ValueOpcode::LoadAddressU32 && memory.kind == ResourceKind::ScalarAddress) ||
+		       (op == ValueOpcode::ReadConstBuffer && memory.kind == ResourceKind::ScalarBuffer);
+	}
+	if (inst.NumArgs() != 5u || memory.kind != ResourceKind::Buffer || memory.typed ||
+	    memory.formatted || memory.coherent || memory.data_bits != 32u ||
+	    memory.data_dwords != 1u || memory.idxen || memory.offen || memory.offset != 0u ||
+	    inst.Arg(4).GetType() != Type::U1) {
+		return false;
+	}
+	for (uint32_t arg = 1; arg <= 3; ++arg) {
+		if (inst.Arg(arg).Resolve() != Value(0u)) return false;
+	}
+	return true;
 }
 
 bool IsDescriptorHandle(ValueOpcode opcode) {
@@ -92,6 +102,7 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::ISub64:
 		case ValueOpcode::IMul32:
 		case ValueOpcode::IMul64:
+		case ValueOpcode::UMulHi:
 		case ValueOpcode::UMin32:
 		case ValueOpcode::ShiftLeftLogical32:
 		case ValueOpcode::ShiftLeftLogical64:
@@ -108,6 +119,7 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::SelectU32:
 		case ValueOpcode::SelectF32:
 		case ValueOpcode::ULessThan32:
+		case ValueOpcode::ULessThanEqual32:
 		case ValueOpcode::IEqual32:
 		case ValueOpcode::UGreaterThan32:
 		case ValueOpcode::SGreaterThanEqual32:
@@ -120,6 +132,7 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::FPOrdGreaterThanEqual32:
 		case ValueOpcode::FPIsNan32:
 		case ValueOpcode::FPMul32:
+		case ValueOpcode::FPRecipIFlag32:
 		case ValueOpcode::FPTrunc32: return true;
 		default: return false;
 	}
@@ -160,6 +173,7 @@ private:
 				default: return false;
 			}
 		}
+		if (require_uniform && !m_active_mask.IsEmpty() && value == m_active_mask) return true;
 		// Integer-only dependency checks do not depend on the active EXEC mask.
 		if (!require_uniform && m_validated_dependencies.contains(inst)) return true;
 		if (!m_visiting.insert(inst).second) {
@@ -249,7 +263,8 @@ private:
 			}
 			return finish(true);
 		}
-		if (op == ValueOpcode::LoadAddressU32 || op == ValueOpcode::ReadConstBuffer) {
+		if (op == ValueOpcode::LoadAddressU32 || op == ValueOpcode::ReadConstBuffer ||
+		    op == ValueOpcode::LoadBufferU32) {
 			const auto  expected = op == ValueOpcode::LoadAddressU32
 			                           ? ValueOpcode::GetAddressResource
 			                           : ValueOpcode::GetBufferResource;
@@ -258,6 +273,11 @@ private:
 			    handle->GetOpcode() != expected) {
 				return finish(false);
 			}
+			if (op == ValueOpcode::LoadBufferU32) {
+				const auto guard = inst->Arg(4).Resolve();
+				return finish(Validate(inst->Arg(0)) &&
+				              ((!m_active_mask.IsEmpty() && guard == m_active_mask) || Validate(guard)));
+			}
 		} else if (op == ValueOpcode::CompositeExtractU64) {
 			const auto index = inst->NumArgs() == 2 ? inst->Arg(1).Resolve() : Value {};
 			if (!index.IsImmediate() || index.GetType() != Type::U32 || index.U32() >= 2u) {
@@ -344,6 +364,10 @@ bool SrtWalker::EvaluateWide(Value value, uint64_t& result) {
 			default: return false;
 		}
 	}
+	if (!m_active_mask.IsEmpty() && value == m_active_mask) {
+		result = 1u;
+		return true;
+	}
 	auto* inst = value.TryInstruction();
 	if (inst == nullptr) {
 		return false;
@@ -433,6 +457,16 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 		return false;
 	}
 	const auto& mem    = m_program.memory_info[flags.index];
+	const bool vector = inst.GetOpcode() == ValueOpcode::LoadBufferU32;
+	const auto guard = vector ? inst.Arg(4).Resolve() : Value {};
+	if (vector && (guard.IsImmediate() || guard != m_active_mask)) {
+		uint64_t enabled = 0;
+		if (!Arg(inst, 4, enabled)) return false;
+		if (enabled == 0u) {
+			result = 0u;
+			return true;
+		}
+	}
 	const auto* handle = inst.Arg(0).ResolveInstruction();
 	if (handle == nullptr) {
 		return false;
@@ -443,10 +477,10 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 	if (!Arg(*handle, 0, low) || !Arg(*handle, 1, high) || !Arg(inst, 1, offset)) {
 		return false;
 	}
-	const auto base      = ((high << 32u) | static_cast<uint32_t>(low)) & AddressMask;
+	const auto base      = (high << 32u) | static_cast<uint32_t>(low);
 	const auto immediate = static_cast<int64_t>(static_cast<int32_t>(mem.offset));
 	uint64_t   address   = 0;
-	if (inst.GetOpcode() == ValueOpcode::ReadConstBuffer) {
+	if (inst.GetOpcode() == ValueOpcode::ReadConstBuffer || vector) {
 		uint64_t records = 0;
 		uint64_t word3   = 0;
 		if (handle->NumArgs() != 4u || !Arg(*handle, 2, records) || !Arg(*handle, 3, word3)) {
@@ -458,13 +492,22 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 		const auto byte_offset =
 		    (static_cast<uint64_t>(immediate) & ~uint64_t {3}) + (static_cast<uint32_t>(offset) & ~3u);
 		const auto stride  = (static_cast<uint32_t>(high) >> 16u) & 0x3fffu;
+		if (vector) {
+			// Only the uniform, unswizzled structured DWORD address is evaluated on the host.
+			if ((high & (1u << 31u)) != 0u || (word3 & ((1u << 23u) | 0xf0000000u)) != 0u)
+				return false;
+			if (stride == 0u || records == 0u || ((word3 >> 12u) & 0x7fu) == 0u) {
+				result = 0u;
+				return true;
+			}
+		}
 		const auto size = stride == 0u
 		                      ? static_cast<uint64_t>(static_cast<uint32_t>(records))
 		                      : static_cast<uint64_t>(stride) * static_cast<uint32_t>(records);
 		if (byte_offset > size || size - byte_offset < sizeof(uint32_t)) {
 			return false;
 		}
-		address = (base & ~uint64_t {3}) + byte_offset;
+		address = (base & AddressMask & ~uint64_t {3}) + byte_offset;
 	} else {
 		const auto relative = (immediate & ~int64_t {3}) +
 		                      static_cast<int64_t>(static_cast<uint32_t>(offset) & ~3u);
@@ -473,11 +516,13 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 		}
 	}
 	uint32_t word = 0;
-	if (m_runtime.read_memory != nullptr) {
-		if (!m_runtime.read_memory(m_runtime.userdata, address, {&word, 1})) {
+	const auto reader = vector ? m_runtime.read_specialization_memory : m_runtime.read_memory;
+	if (reader != nullptr) {
+		if (!reader(m_runtime.userdata, address, {&word, 1})) {
 			return false;
 		}
 	} else {
+		if (vector) return false;
 		std::memcpy(&word, reinterpret_cast<const void*>(address), sizeof(word));
 	}
 	result = word;
@@ -505,6 +550,9 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 		case ValueOpcode::GetShaderBase: result = m_runtime.shader_base; return true;
 		case ValueOpcode::Phi: return EvaluatePhi(inst, result);
 		case ValueOpcode::ReadFirstLane: {
+			if (!m_active_mask.IsEmpty() && inst.Arg(1).Resolve() == m_active_mask) {
+				return EvaluateWide(inst.Arg(0), result);
+			}
 			const auto clean_runtime = CleanRuntime(m_runtime);
 			SrtWalker  clean_active(m_program, clean_runtime, {}, nullptr, inst.Arg(1));
 			SrtWalker  active(m_program, m_runtime, m_clean_flat_slots, &clean_active,
@@ -537,7 +585,12 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 		}
 		case ValueOpcode::LoadAddressU32:
 		case ValueOpcode::ReadConstBuffer:
+		case ValueOpcode::LoadBufferU32:
 			if (IsRawRead(m_program, inst)) {
+				if (inst.GetOpcode() == ValueOpcode::LoadBufferU32 && m_clean_evaluator != nullptr &&
+				    m_clean_evaluator->m_active_mask == m_active_mask) {
+					return m_clean_evaluator->EvaluateWide(Value(const_cast<Inst*>(&inst)), result);
+				}
 				return EvaluateRawRead(inst, result);
 			}
 			break;
@@ -577,6 +630,12 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::UMulHi:
+			if (binary()) {
+				result = (static_cast<uint64_t>(static_cast<uint32_t>(a)) * static_cast<uint32_t>(b)) >> 32u;
+				return true;
+			}
+			return false;
 		case ValueOpcode::UMin32:
 			if (binary()) {
 				result = std::min(static_cast<uint32_t>(a), static_cast<uint32_t>(b));
@@ -612,6 +671,15 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::FPRecipIFlag32:
+			if (Arg(inst, 0, a)) {
+				const auto exponent = (a >> 23u) & 0xffu;
+				// Normal positive powers of two have exact normal reciprocals in every FP mode.
+				if ((a & 0x807fffffu) != 0u || exponent == 0u || exponent >= 254u) return false;
+				result = (254u - exponent) << 23u;
+				return true;
+			}
+			return false;
 		case ValueOpcode::FPIsNan32:
 			if (Arg(inst, 0, a)) {
 				result = std::isnan(Float32(a));
@@ -782,6 +850,12 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::ULessThanEqual32:
+			if (binary()) {
+				result = static_cast<uint32_t>(a) <= static_cast<uint32_t>(b);
+				return true;
+			}
+			return false;
 		case ValueOpcode::UGreaterThan32:
 			if (binary()) {
 				result = static_cast<uint32_t>(a) > static_cast<uint32_t>(b);

src/kernel/eventQueue.cpp

@@ -138,7 +138,13 @@ int KernelEqueuePrivate::GetTriggeredEvents(KernelEvent* ev, int num) {
 
 void KernelEqueuePrivate::TriggerExpiredTimers(uint64_t now_ns) {
 	for (auto& event: m_events) {
-		if (!event.triggered && event.deadline_ns != 0 && event.deadline_ns <= now_ns) {
+		if (event.deadline_ns != 0 && event.deadline_ns <= now_ns) {
+			if (event.event.filter == KERNEL_EVFILT_TIMER) {
+				const auto count = event.interval_ns == 0 ? !event.triggered :
+				    1 + (now_ns - event.deadline_ns) / event.interval_ns;
+				event.event.data += static_cast<intptr_t>(count);
+				event.deadline_ns += count * event.interval_ns;
+			}
 			event.triggered = true;
 		}
 	}
@@ -209,7 +215,9 @@ int KernelEqueuePrivate::AddEvent(const KernelEqueueEvent& event) {
 		                       return e.event.ident == ident && e.event.filter == filter;
 	                       });
 	if (it != m_events.end()) {
+		TriggerExpiredTimers(MonotonicTimeNs());
 		it->deadline_ns = event.deadline_ns;
+		it->interval_ns = event.interval_ns;
 		it->event.udata = event.event.udata;
 		for (auto& pending: it->pending_events) {
 			pending.udata = event.event.udata;
@@ -511,6 +519,22 @@ int KYTY_SYSV_ABI KernelDeleteUserEvent(KernelEqueue eq, int id) {
 	return KernelDeleteEvent(eq, static_cast<uintptr_t>(id), KERNEL_EVFILT_USER);
 }
 
+static int AddTimerEvent(KernelEqueue eq, int id, uint64_t delay_ns, bool periodic, void* udata) {
+	const auto now_ns = MonotonicTimeNs();
+	KernelEqueueEvent event {};
+	event.deadline_ns  = delay_ns <= UINT64_MAX - now_ns ? now_ns + delay_ns : UINT64_MAX;
+	event.interval_ns  = periodic ? delay_ns : 0;
+	event.event.ident  = static_cast<uintptr_t>(id);
+	event.event.filter = periodic ? KERNEL_EVFILT_TIMER : KERNEL_EVFILT_HRTIMER;
+	event.event.flags  = EV_ADD | (periodic ? EV_CLEAR : EV_ONESHOT);
+	event.event.udata  = udata;
+	return KernelAddEvent(eq, event);
+}
+
+int KYTY_SYSV_ABI KernelAddTimerEvent(KernelEqueue eq, int id, KernelUseconds usec, void* udata) {
+	return AddTimerEvent(eq, id, static_cast<uint64_t>(usec) * 1000, true, udata);
+}
+
 int KYTY_SYSV_ABI KernelAddHRTimerEvent(KernelEqueue eq, int id, const KernelTimespec* ts,
                                         void* udata) {
 	if (ts == nullptr) {
@@ -524,17 +548,7 @@ int KYTY_SYSV_ABI KernelAddHRTimerEvent(KernelEqueue eq, int id, const KernelTim
 
 	const auto delay_ns =
 	    static_cast<uint64_t>(ts->tv_sec) * 1000000000ull + static_cast<uint64_t>(ts->tv_nsec);
-	const auto now_ns = MonotonicTimeNs();
-
-	KernelEqueueEvent event {};
-	event.deadline_ns  = delay_ns <= UINT64_MAX - now_ns ? now_ns + delay_ns : UINT64_MAX;
-	event.event.ident  = static_cast<uintptr_t>(id);
-	event.event.filter = KERNEL_EVFILT_HRTIMER;
-	event.event.flags  = EV_ADD | EV_ONESHOT;
-	event.event.fflags = 0;
-	event.event.data   = 0;
-	event.event.udata  = udata;
-	return KernelAddEvent(eq, event);
+	return AddTimerEvent(eq, id, delay_ns, false, udata);
 }
 
 int KYTY_SYSV_ABI KernelDeleteHRTimerEvent(KernelEqueue eq, int id) {

src/kernel/eventQueue.h

@@ -51,6 +51,7 @@ struct KernelFilter {
 struct KernelEqueueEvent {
 	bool                    triggered   = false;
 	uint64_t                deadline_ns = 0;
+	uint64_t                interval_ns = 0;
 	KernelEvent             event;
 	KernelFilter            filter;
 	std::deque<KernelEvent> pending_events;
@@ -72,6 +73,8 @@ int KYTY_SYSV_ABI KernelAddUserEventEdge(KernelEqueue eq, int id);
 int KYTY_SYSV_ABI KernelTriggerUserEvent(KernelEqueue eq, int id, void* udata);
 int KYTY_SYSV_ABI KernelTriggerUserEventForAll(int id, void* udata);
 int KYTY_SYSV_ABI KernelDeleteUserEvent(KernelEqueue eq, int id);
+int KYTY_SYSV_ABI KernelAddTimerEvent(KernelEqueue eq, int id, KernelUseconds usec,
+                                      void* udata);
 int KYTY_SYSV_ABI KernelAddHRTimerEvent(KernelEqueue eq, int id, const KernelTimespec* ts,
                                         void* udata);
 int KYTY_SYSV_ABI KernelDeleteHRTimerEvent(KernelEqueue eq, int id);

src/kernel/memory.cpp

@@ -166,7 +166,8 @@ static bool VirtualRangesOverlap(uint64_t left_start, uint64_t left_size, uint64
 }
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
-static uint32_t g_test_backing_store_unmaps_before_failure = UINT32_MAX;
+static uint32_t        g_test_backing_store_unmaps_before_failure = UINT32_MAX;
+static callback_func_t g_test_before_backing_map                   = nullptr;
 #endif
 
 #include "memoryAddressSpace.inc"
@@ -880,13 +881,15 @@ bool TryReadBacking(uint64_t vaddr, void* data, uint64_t size) {
 	       g_guest_address_space->TryReadBacking(vaddr, data, size);
 }
 
-bool TryReadGpuCleanBacking(uint64_t vaddr, void* data, uint64_t size) {
+bool TryReadBufferBacking(uint64_t vaddr, void* data, uint64_t size) {
 	if (g_gpu_resources != nullptr && IsGpuAddressRange(vaddr, size)) {
-		if (!Graphics::GuestGpu::IsGpuThread() ||
-		    GetGpuResources().GetBufferCache().HasGpuDirtyBytes(vaddr, size) ||
-		    GetGpuResources().GetTextureCache().IsRegionGpuModified(vaddr, size)) {
+		if (!Graphics::GuestGpu::IsGpuThread()) {
 			return false;
 		}
+		auto& buffers = GetGpuResources().GetBufferCache();
+		if (buffers.HasGpuDirtyBytes(vaddr, size)) {
+			buffers.ReadMemory(vaddr, size);
+		}
 	}
 	return TryReadBacking(vaddr, data, size);
 }
@@ -3024,6 +3027,22 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 		return g_guest_address_space->MapBacking(target_addr, len, direct_memory_start, mode,
 		                                         &shared_failure);
 	};
+	std::vector<PhysicalMemory::AllocatedBlock> replaced_mappings;
+	auto restore_mappings = [&](size_t count, bool restore_views) {
+		for (size_t index = 0; index < count; ++index) {
+			const auto& old = replaced_mappings[index];
+			if (restore_views) {
+				EXIT_IF(!g_guest_address_space->MapBacking(old.map_vaddr, old.map_size,
+				                                           old.start_addr, old.mode));
+			}
+			EXIT_IF(!g_physical_memory->Map(old.map_vaddr, old.start_addr, old.map_size,
+			                               old.prot, old.mode, old.gpu_mode));
+			g_physical_memory->SetVirtualRangeName(old.map_vaddr, old.map_size, old.name);
+			g_physical_memory->SetVirtualRangeMemoryType(old.map_vaddr, old.map_size,
+			                                          old.memory_type);
+		}
+		MapGpuRange(in_addr, len);
+	};
 	if (fixed) {
 		if (in_addr == 0 || (in_addr & (PAGE_SIZE - 1u)) != 0 ||
 		    (alignment != 0 && in_addr % alignment != 0)) {
@@ -3034,22 +3053,68 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 		}
 
 		std::vector<VirtualRanges::Range> reserved_ranges;
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
 		if (g_virtual_ranges->QuerySpan(in_addr, len, &reserved_ranges) &&
 		    std::all_of(reserved_ranges.begin(), reserved_ranges.end(), [](const auto& range) {
-			    return range.type == VirtualRangeType::Reserved;
+			    return range.type == VirtualRangeType::Direct;
 		    })) {
+			uint64_t covered = 0;
+			{
+				Common::LockGuard lock(g_physical_memory->GetMutex());
+				for (const auto& mapping: g_physical_memory->GetMappings()) {
+					const auto start = std::max(in_addr, mapping.map_vaddr);
+					const auto end = std::min(in_addr + len, mapping.map_vaddr + mapping.map_size);
+					if (start >= end) {
+						continue;
+					}
+					auto old = mapping;
+					old.start_addr += start - old.map_vaddr;
+					old.map_vaddr = start;
+					old.size = old.map_size = end - start;
+					covered += old.map_size;
+					replaced_mappings.push_back(old);
+				}
+			}
+			if (covered != len) {
+				return KERNEL_ERROR_EBUSY;
+			}
 			UnmapGpuRange(in_addr, len);
-			reserved_target = true;
-		}
-		if (!reserved_target && ReplaceFixedRangeWithReserved(in_addr, len)) {
-			reserved_target = true;
-		}
-		if (!reserved_target) {
-			return KERNEL_ERROR_ENOMEM;
-		}
-		if (map_shared_fixed(in_addr)) {
+			for (size_t index = 0; index < replaced_mappings.size(); ++index) {
+				const auto& old = replaced_mappings[index];
+				GpuAccessMode old_gpu_mode = GpuAccessMode::NoAccess;
+				if (!g_physical_memory->Unmap(old.map_vaddr, old.map_size, &old_gpu_mode)) {
+					restore_mappings(index, false);
+					return KERNEL_ERROR_EBUSY;
+				}
+			}
+			// Keep the old view accessible until mmap atomically replaces it. In particular,
+			// do not publish a temporary reserved/PROT_NONE range between two direct views.
+			if (!map_shared_fixed(in_addr)) {
+				restore_mappings(replaced_mappings.size(), true);
+				return KERNEL_ERROR_ENOMEM;
+			}
 			out_addr       = in_addr;
 			shared_backing = true;
+		} else
+#endif
+		{
+			if (g_virtual_ranges->QuerySpan(in_addr, len, &reserved_ranges) &&
+			    std::all_of(reserved_ranges.begin(), reserved_ranges.end(), [](const auto& range) {
+				    return range.type == VirtualRangeType::Reserved;
+			    })) {
+				UnmapGpuRange(in_addr, len);
+				reserved_target = true;
+			}
+			if (!reserved_target && ReplaceFixedRangeWithReserved(in_addr, len)) {
+				reserved_target = true;
+			}
+			if (!reserved_target) {
+				return KERNEL_ERROR_ENOMEM;
+			}
+			if (map_shared_fixed(in_addr)) {
+				out_addr       = in_addr;
+				shared_backing = true;
+			}
 		}
 	} else {
 		constexpr size_t DEFAULT_ALIGNMENT = 0x4000;
@@ -3105,15 +3170,22 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 
 	if (!g_physical_memory->Map(out_addr, direct_memory_start, len, prot, mode, gpu_mode)) {
 		LOGF_COLOR(Log::Color::Red, "\t [Fail]\n");
-		EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
+		if (replaced_mappings.empty()) {
+			EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
+		} else {
+			restore_mappings(replaced_mappings.size(), true);
+		}
 		return KERNEL_ERROR_EBUSY;
 	}
 
 	PhysicalMemory::AllocatedBlock mapped_block {};
 	g_physical_memory->Find(direct_memory_start, false, &mapped_block);
-	const bool published = reserved_target
+	const bool published = !replaced_mappings.empty() || reserved_target
 	                           ? g_virtual_ranges->ReplaceSpan(
-	                                 out_addr, len, VirtualRangeType::Reserved, direct_memory_start,
+	                                 out_addr, len, replaced_mappings.empty()
+	                                                    ? VirtualRangeType::Reserved
+	                                                    : VirtualRangeType::Direct,
+	                                 direct_memory_start,
 	                                 prot, mapped_block.memory_type, VirtualRangeType::Direct, "")
 	                           : g_virtual_ranges->Add(out_addr, len, direct_memory_start, prot,
 	                                                   mapped_block.memory_type,
@@ -3121,7 +3193,11 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 	if (!published) {
 		GpuAccessMode rollback_gpu_mode = GpuAccessMode::NoAccess;
 		EXIT_IF(!g_physical_memory->Unmap(out_addr, len, &rollback_gpu_mode));
-		EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
+		if (replaced_mappings.empty()) {
+			EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
+		} else {
+			restore_mappings(replaced_mappings.size(), true);
+		}
 		return KERNEL_ERROR_EBUSY;
 	}
 
@@ -3504,6 +3580,10 @@ int KYTY_SYSV_ABI KernelReserveVirtualRange(void** addr, size_t len, int flags,
 }
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+void TestBeforeNextBackingMap(callback_func_t callback) {
+	g_test_before_backing_map = callback;
+}
+
 void TestFailNextPhysicalMemoryUnmap() {
 	TestFailPhysicalMemoryUnmapAfter(0);
 }

src/kernel/memory.h

@@ -114,7 +114,7 @@ int AllocateDirectMemory(int64_t search_start, int64_t search_end, size_t size,
 int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot);
 bool                   TryWriteBacking(uint64_t vaddr, const void* data, uint64_t size);
 bool                   TryReadBacking(uint64_t vaddr, void* data, uint64_t size);
-bool                   TryReadGpuCleanBacking(uint64_t vaddr, void* data, uint64_t size);
+bool                   TryReadBufferBacking(uint64_t vaddr, void* data, uint64_t size);
 bool                   TryReadSparseBacking(uint64_t vaddr, void* data, uint64_t size);
 [[nodiscard]] uint64_t ClampRangeSize(uint64_t vaddr, uint64_t size);
 void                   WriteBacking(uint64_t vaddr, const void* data, uint64_t size) noexcept;
@@ -191,6 +191,7 @@ bool ProtectGuestHostMemory(uint64_t vaddr, uint64_t size, Common::VirtualMemory
 bool FreeGuestMemory(uint64_t vaddr, uint64_t size);
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+void     TestBeforeNextBackingMap(callback_func_t callback);
 void     TestFailNextPhysicalMemoryUnmap();
 void     TestFailPhysicalMemoryUnmapAfter(uint32_t successful_unmaps);
 void     TestFailGuestBackingStoreUnmapAfter(uint32_t successful_unmaps);

src/kernel/memoryAddressSpace.inc

@@ -231,9 +231,8 @@ public:
 		return true;
 	}
 
-	bool MapExistingPlaceholderFixed(uint64_t vaddr, uint64_t size, uint64_t backing_offset,
-	                                 VirtualMemory::Mode mode,
-	                                 FailureReason*      failure_reason = nullptr) {
+	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)) {
 			SetFailure(failure_reason, !IsAvailable() ? FailureReason::BackingUnavailable
 			                                          : FailureReason::PhysicalRangeOutOfBounds);
@@ -257,14 +256,13 @@ public:
 		SetFailure(failure_reason, FailureReason::None);
 		return true;
 #else
+		std::lock_guard lock(m_mutex);
 		void* mapped = MapView(vaddr, size, backing_offset, mode, failure_reason);
 		if (mapped == nullptr || reinterpret_cast<uint64_t>(mapped) != vaddr) {
 			return false;
 		}
-		{
-			std::lock_guard lock(m_mutex);
-			EXIT_IF(!m_maps.emplace(vaddr, MapEntry {vaddr, size, backing_offset, mode}).second);
-		}
+		RemoveMappingsUnlocked(vaddr, size);
+		EXIT_IF(!m_maps.emplace(vaddr, MapEntry {vaddr, size, backing_offset, mode}).second);
 		SetFailure(failure_reason, FailureReason::None);
 		return true;
 #endif
@@ -302,8 +300,7 @@ public:
 			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(!MapExistingPlaceholderFixed(it->vaddr, it->size, it->backing_offset,
-					                                     it->mode));
+					EXIT_IF(!MapFixed(it->vaddr, it->size, it->backing_offset, it->mode));
 				}
 				if (placeholder_preserved != nullptr) {
 					*placeholder_preserved = false;
@@ -320,8 +317,7 @@ public:
 		// 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(!MapExistingPlaceholderFixed(it->vaddr, it->size, it->backing_offset,
-				                                     it->mode));
+				EXIT_IF(!MapFixed(it->vaddr, it->size, it->backing_offset, it->mode));
 			}
 			return false;
 		}
@@ -363,12 +359,10 @@ public:
 			return false;
 		}
 
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		const uint64_t left_size  = vaddr - old.vaddr;
 		const uint64_t right_addr = vaddr + size;
-		const uint64_t old_end    = old.vaddr + old.size;
-		const uint64_t right_size = old_end - right_addr;
-
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+		const uint64_t right_size = old.vaddr + old.size - right_addr;
 		const bool exact_unmap = (vaddr == old.vaddr && size == old.size);
 		if (exact_unmap) {
 			const bool unmapped =
@@ -393,14 +387,11 @@ public:
 				ok = SplitPlaceholder(vaddr, size) && ok;
 			}
 			if (left_size != 0) {
-				ok = MapExistingPlaceholderFixed(old.vaddr, left_size, old.backing_offset,
-				                                 old.mode) &&
-				     ok;
+				ok = MapFixed(old.vaddr, left_size, old.backing_offset, old.mode) && ok;
 			}
 			if (right_size != 0) {
-				ok = MapExistingPlaceholderFixed(right_addr, right_size,
-				                                 old.backing_offset + (right_addr - old.vaddr),
-				                                 old.mode) &&
+				ok = MapFixed(right_addr, right_size,
+				              old.backing_offset + (right_addr - old.vaddr), old.mode) &&
 				     ok;
 			}
 			if (ok) {
@@ -418,8 +409,7 @@ public:
 			}
 			EXIT_IF(VirtualFree(reinterpret_cast<void*>(old.vaddr), old.size,
 			                    MEM_RELEASE | MEM_COALESCE_PLACEHOLDERS) == 0);
-			EXIT_IF(
-			    !MapExistingPlaceholderFixed(old.vaddr, old.size, old.backing_offset, old.mode));
+			EXIT_IF(!MapFixed(old.vaddr, old.size, old.backing_offset, old.mode));
 			return false;
 		}
 		std::lock_guard lock(m_mutex);
@@ -433,16 +423,7 @@ public:
 		}
 
 		std::lock_guard lock(m_mutex);
-		m_maps.erase(old.vaddr);
-		if (left_size != 0) {
-			m_maps.emplace(old.vaddr,
-			               MapEntry {old.vaddr, left_size, old.backing_offset, old.mode});
-		}
-		if (right_size != 0) {
-			m_maps.emplace(right_addr,
-			               MapEntry {right_addr, right_size,
-			                         old.backing_offset + right_addr - old.vaddr, old.mode});
-		}
+		RemoveMappingsUnlocked(vaddr, size);
 
 		if (placeholder_preserved != nullptr) {
 			*placeholder_preserved = true;
@@ -516,6 +497,33 @@ private:
 		VirtualMemory::Mode mode           = VirtualMemory::Mode::NoAccess;
 	};
 
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
+	void RemoveMappingsUnlocked(uint64_t vaddr, uint64_t size) {
+		const auto end = vaddr + size;
+		auto       it = m_maps.upper_bound(vaddr);
+		if (it != m_maps.begin()) {
+			--it;
+		}
+		while (it != m_maps.end() && it->first < end) {
+			const auto old     = it->second;
+			const auto old_end = old.vaddr + old.size;
+			if (old_end <= vaddr) {
+				++it;
+				continue;
+			}
+			it = m_maps.erase(it);
+			if (old.vaddr < vaddr) {
+				m_maps.emplace(old.vaddr,
+				               MapEntry {old.vaddr, vaddr - old.vaddr, old.backing_offset, old.mode});
+			}
+			if (end < old_end) {
+				m_maps.emplace(end, MapEntry {end, old_end - end,
+				                             old.backing_offset + end - old.vaddr, old.mode});
+			}
+		}
+	}
+#endif
+
 	bool FindContainingUnlocked(uint64_t vaddr, uint64_t size, MapEntry* out) const {
 		EXIT_IF(out == nullptr);
 
@@ -621,6 +629,12 @@ private:
 
 	void* MapView(uint64_t vaddr, uint64_t size, uint64_t backing_offset, VirtualMemory::Mode mode,
 	              FailureReason* failure_reason) {
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+		if (const auto callback = g_test_before_backing_map) {
+			g_test_before_backing_map = nullptr;
+			callback(vaddr, size);
+		}
+#endif
 		void* ptr = mmap(reinterpret_cast<void*>(vaddr), size, GetProtect(mode),
 		                 MAP_SHARED | MAP_FIXED, m_fd, static_cast<off_t>(backing_offset));
 		if (ptr == MAP_FAILED) {
@@ -778,7 +792,12 @@ public:
 	                VirtualMemory::Mode               mode,
 	                GuestBackingStore::FailureReason* failure_reason = nullptr) {
 		std::lock_guard lock(m_mutex);
-		if (!ConsumeUnlocked(vaddr, size)) {
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
+		const bool replacing = ContainsMappedUnlocked(vaddr, size, MappingType::Backing);
+#else
+		constexpr bool replacing = false;
+#endif
+		if (!replacing && !ConsumeUnlocked(vaddr, size)) {
 			if (failure_reason != nullptr) {
 				*failure_reason = GuestBackingStore::FailureReason::AddressRangeUnavailable;
 			}
@@ -787,12 +806,15 @@ public:
 		const auto host_mode = VirtualMemory::IsExecute(mode)
 		                           ? VirtualMemory::Mode::ExecuteReadWrite
 		                           : VirtualMemory::Mode::ReadWrite;
-		if (m_backing->MapExistingPlaceholderFixed(vaddr, size, backing_offset, host_mode,
-		                                           failure_reason)) {
-			EXIT_IF(!AddMappedUnlocked(vaddr, size, MappingType::Backing));
+		if (m_backing->MapFixed(vaddr, size, backing_offset, host_mode, failure_reason)) {
+			if (!replacing) {
+				EXIT_IF(!AddMappedUnlocked(vaddr, size, MappingType::Backing));
+			}
 			return true;
 		}
-		AddFreeUnlocked(vaddr, size);
+		if (!replacing) {
+			AddFreeUnlocked(vaddr, size);
+		}
 		return false;
 	}
 

src/kernel/pthread.cpp

@@ -4208,6 +4208,10 @@ int KYTY_SYSV_ABI pthread_attr_setschedpolicy(LibKernel::PthreadAttr* attr, int
 	return POSIX_PTHREAD_CALL(LibKernel::PthreadAttrSetschedpolicy(attr, policy));
 }
 
+int KYTY_SYSV_ABI pthread_attr_setstack(LibKernel::PthreadAttr* attr, void* addr, size_t size) {
+	return POSIX_PTHREAD_CALL(LibKernel::PthreadAttrSetstack(attr, addr, size));
+}
+
 int KYTY_SYSV_ABI pthread_attr_setstacksize(LibKernel::PthreadAttr* attr, size_t stack_size) {
 	// PRINT_NAME();
 

src/kernel/pthread.h

@@ -249,6 +249,7 @@ int KYTY_SYSV_ABI pthread_attr_setinheritsched(LibKernel::PthreadAttr* attr, int
 int KYTY_SYSV_ABI pthread_attr_setschedparam(LibKernel::PthreadAttr*            attr,
                                              const LibKernel::KernelSchedParam* param);
 int KYTY_SYSV_ABI pthread_attr_setschedpolicy(LibKernel::PthreadAttr* attr, int policy);
+int KYTY_SYSV_ABI pthread_attr_setstack(LibKernel::PthreadAttr* attr, void* addr, size_t size);
 int KYTY_SYSV_ABI pthread_attr_setstacksize(LibKernel::PthreadAttr* attr, size_t stack_size);
 int KYTY_SYSV_ABI pthread_cond_broadcast(LibKernel::PthreadCond* cond);
 int KYTY_SYSV_ABI pthread_cond_signal(LibKernel::PthreadCond* cond);

src/launcher/CMakeLists.txt

@@ -21,6 +21,10 @@ add_executable(launcher WIN32
 
 
 target_link_libraries(launcher Qt6::Concurrent Qt6::Network Qt6::Widgets Vulkan::Headers SDL3::SDL3-static kyty_git_version common)
+if(LINUX)
+	find_package(Qt6 COMPONENTS DBus REQUIRED)
+	target_link_libraries(launcher Qt6::DBus)
+endif()
 if(WIN32)
 target_link_libraries(launcher setupapi)
 endif()

src/launcher/forms/configuration_edit_dialog.ui

@@ -23,7 +23,7 @@
    <property name="verticalSpacing">
     <number>8</number>
    </property>
-   <item row="0" column="0">
+   <item row="0" column="0" colspan="2">
     <layout class="QGridLayout" name="sectionsLayout" columnstretch="1,1">
      <property name="horizontalSpacing">
       <number>8</number>
@@ -72,27 +72,6 @@
           <property name="toolTip"><string>Language</string></property>
          </widget>
         </item>
-        <item row="3" column="0">
-         <widget class="QLabel" name="label_controller_color">
-          <property name="text"><string>DualSense lightbar:</string></property>
-         </widget>
-        </item>
-        <item row="3" column="1">
-         <layout class="QHBoxLayout" name="controllerColorLayout">
-          <item>
-           <widget class="QPushButton" name="button_controller_color">
-            <property name="toolTip"><string>Choose and preview a fixed LED color. Bluetooth DualSense may take several seconds to update.</string></property>
-            <property name="text"><string>Custom</string></property>
-           </widget>
-          </item>
-          <item>
-           <widget class="QPushButton" name="button_controller_color_reset">
-            <property name="toolTip"><string>Let the game control the lightbar</string></property>
-            <property name="text"><string>Default</string></property>
-           </widget>
-          </item>
-         </layout>
-        </item>
        </layout>
       </widget>
      </item>
@@ -194,6 +173,12 @@
           <property name="text"><string>Enable tessellation support (experimental)</string></property>
          </widget>
         </item>
+        <item row="8" column="1">
+         <widget class="QCheckBox" name="checkBox_hide_cursor">
+          <property name="toolTip"><string>Hide the host cursor after 2 seconds without mouse activity</string></property>
+          <property name="text"><string>Auto-hide cursor</string></property>
+         </widget>
+        </item>
        </layout>
       </widget>
      </item>
@@ -316,7 +301,82 @@
      </item>
     </layout>
    </item>
-   <item row="2" column="0">
+   <item row="1" column="1">
+    <widget class="QGroupBox" name="controller_group">
+     <property name="title"><string>Controller</string></property>
+     <layout class="QFormLayout" name="controllerLayout">
+      <property name="fieldGrowthPolicy"><enum>QFormLayout::AllNonFixedFieldsGrow</enum></property>
+      <item row="0" column="0">
+       <widget class="QLabel" name="label_controller_color">
+        <property name="text"><string>DualSense lightbar:</string></property>
+       </widget>
+      </item>
+      <item row="0" column="1">
+       <layout class="QHBoxLayout" name="controllerColorLayout">
+        <item>
+         <widget class="QPushButton" name="button_controller_color">
+          <property name="toolTip"><string>Choose and preview a fixed LED color. Bluetooth DualSense may take several seconds to update.</string></property>
+          <property name="text"><string>Custom</string></property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QPushButton" name="button_controller_color_reset">
+          <property name="toolTip"><string>Let the game control the lightbar</string></property>
+          <property name="text"><string>Default</string></property>
+         </widget>
+        </item>
+       </layout>
+      </item>
+      <item row="1" column="0">
+       <widget class="QLabel" name="label_controller_vibration">
+        <property name="text"><string>Vibration intensity:</string></property>
+       </widget>
+      </item>
+      <item row="1" column="1">
+       <layout class="QHBoxLayout" name="controllerVibrationLayout">
+        <item>
+         <widget class="QSlider" name="slider_controller_vibration">
+          <property name="toolTip"><string>Scale DualSense vibration over USB and Bluetooth</string></property>
+          <property name="maximum"><number>100</number></property>
+          <property name="value"><number>100</number></property>
+          <property name="orientation"><enum>Qt::Horizontal</enum></property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QLabel" name="label_controller_vibration_value">
+          <property name="minimumSize"><size><width>40</width><height>0</height></size></property>
+          <property name="text"><string>100%</string></property>
+         </widget>
+        </item>
+       </layout>
+      </item>
+      <item row="2" column="0">
+       <widget class="QLabel" name="label_controller_volume">
+        <property name="text"><string>Speaker volume:</string></property>
+       </widget>
+      </item>
+      <item row="2" column="1">
+       <layout class="QHBoxLayout" name="controllerVolumeLayout">
+        <item>
+         <widget class="QSlider" name="slider_controller_volume">
+          <property name="toolTip"><string>Scale the DualSense speaker over USB and Bluetooth. 50% is the previous default; 100% adds 6 dB of gain.</string></property>
+          <property name="maximum"><number>100</number></property>
+          <property name="value"><number>50</number></property>
+          <property name="orientation"><enum>Qt::Horizontal</enum></property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QLabel" name="label_controller_volume_value">
+          <property name="minimumSize"><size><width>40</width><height>0</height></size></property>
+          <property name="text"><string>50%</string></property>
+         </widget>
+        </item>
+       </layout>
+      </item>
+     </layout>
+    </widget>
+   </item>
+   <item row="2" column="0" colspan="2">
     <layout class="QHBoxLayout" name="buttonLayout">
      <item>
      <widget class="QPushButton" name="clear_button">

src/launcher/include/configuration.h

@@ -46,6 +46,30 @@ QString EnumToText(T value) {
 	                                                                     : key);
 }
 
+struct ControllerSettings {
+	QString color;
+	int     speaker_volume      = 50;
+	int     vibration_intensity = 100;
+
+	void WriteSettings(QSettings* s) const {
+		s->setValue("controller_color", color);
+		s->setValue("controller_speaker_volume", speaker_volume);
+		s->setValue("controller_vibration_intensity", vibration_intensity);
+	}
+
+	void ReadSettings(QSettings* s) {
+		const QColor saved_color(s->value("controller_color").toString());
+		color = saved_color.isValid() ? saved_color.name(QColor::HexRgb) : QString {};
+		const auto read_percent = [s](const char* key, int fallback) {
+			bool      ok    = false;
+			const int value = s->value(key, fallback).toInt(&ok);
+			return ok ? qBound(0, value, 100) : fallback;
+		};
+		speaker_volume      = read_percent("controller_speaker_volume", 50);
+		vibration_intensity = read_percent("controller_vibration_intensity", 100);
+	}
+};
+
 class Configuration: public QObject {
 	Q_OBJECT
 
@@ -85,14 +109,17 @@ public:
 	GameStatus game_status     = GameStatus::Unknown;
 	QString    game_comment;
 
+	// Controller preferences always come from the global configuration.
+	ControllerSettings controller;
+
 	Resolution             screen_resolution           = Resolution::R1280X720;
 	QString                user_name                   = "Kyty";
 	int                    user_id                     = Config::DEFAULT_USER_ID;
 	QString                audio_input_device;
-	QString                controller_color;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int                    gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
+	bool                   hide_cursor_enabled         = false;
 	bool                   readback_linear_images      = false;
 	bool                   tessellation_enabled        = false;
 	int                    vblank_frequency            = 60;
@@ -121,10 +148,10 @@ public:
 		user_name                   = other.user_name;
 		user_id                     = other.user_id;
 		audio_input_device          = other.audio_input_device;
-		controller_color            = other.controller_color;
 		present_mode                = other.present_mode;
 		gpu_index                   = other.gpu_index;
 		fullscreen_enabled          = other.fullscreen_enabled;
+		hide_cursor_enabled         = other.hide_cursor_enabled;
 		readback_linear_images      = other.readback_linear_images;
 		tessellation_enabled        = other.tessellation_enabled;
 		vblank_frequency            = other.vblank_frequency;
@@ -168,10 +195,10 @@ public:
 		KYTY_CFG_SET(user_name);
 		KYTY_CFG_SET(user_id);
 		KYTY_CFG_SET(audio_input_device);
-		KYTY_CFG_SET(controller_color);
 		KYTY_CFG_SET(present_mode);
 		KYTY_CFG_SET(gpu_index);
 		KYTY_CFG_SET(fullscreen_enabled);
+		KYTY_CFG_SET(hide_cursor_enabled);
 		KYTY_CFG_SET(readback_linear_images);
 		KYTY_CFG_SET(tessellation_enabled);
 		KYTY_CFG_SET(vblank_frequency);
@@ -208,14 +235,13 @@ public:
 		                         ? saved_user_id
 		                         : Config::DEFAULT_USER_ID;
 		audio_input_device = s->value("audio_input_device", audio_input_device).toString();
-		const QColor color(s->value("controller_color", controller_color).toString());
-		controller_color = color.isValid() ? color.name(QColor::HexRgb) : QString {};
 		KYTY_CFG_GET(present_mode);
 		gpu_index = s->value("gpu_index", -1).toInt();
 		if (EnumToText(present_mode).isEmpty()) {
 			present_mode = PresentMode::Mailbox;
 		}
 		KYTY_CFG_GET(fullscreen_enabled);
+		KYTY_CFG_GET(hide_cursor_enabled);
 		KYTY_CFG_GET(readback_linear_images);
 		KYTY_CFG_GET(tessellation_enabled);
 		vblank_frequency = s->value("vblank_frequency", vblank_frequency).toInt();

src/launcher/include/configurationEditDialog.h

@@ -28,7 +28,7 @@ public:
 	static void WriteSettings(QSettings& s);
 	static void ReadSettings(QSettings& s);
 
-	void                      SetGameDirectories(const QStringList& dirs);
+	void                      SetGlobalSettings(const QStringList& dirs);
 	[[nodiscard]] QStringList GetGameDirectories() const;
 
 signals:
@@ -40,7 +40,7 @@ private:
 	QGroupBox*                   m_game_dirs_group        = nullptr;
 	QListWidget*                 m_game_dirs_list         = nullptr;
 	QToolButton*                 m_remove_game_dir_button = nullptr;
-	bool                         m_show_game_dirs         = false;
+	bool                         m_global_settings        = false;
 
 protected:
 	void Init(const Configuration& info);

src/launcher/include/configurationListWidget.h

@@ -34,7 +34,7 @@ public:
 
 	[[nodiscard]] const QString& GetSettingsFile() const { return m_settings_file; }
 	[[nodiscard]] const QString& GetGlobalControllerColor() const {
-		return m_global_info.controller_color;
+		return m_global_info.controller.color;
 	}
 	[[nodiscard]] std::unique_ptr<Configuration>
 	CreateConfiguration(const ConfigurationItem& item) const;

src/launcher/src/configurationEditDialog.cpp

@@ -23,6 +23,7 @@
 #include <QPushButton>
 #include <QResizeEvent>
 #include <QSettings>
+#include <QSlider>
 #include <QSpinBox>
 #include <QStyle>
 #include <QToolButton>
@@ -39,7 +40,7 @@
 
 constexpr char SETTINGS_CFG_DIALOG[]               = "ConfigurationEditDialog";
 constexpr char SETTINGS_CFG_LAST_GEOMETRY[]        = "geometry";
-constexpr int  GLOBAL_SETTINGS_GAME_DIRS_MIN_WIDTH = 560;
+constexpr int  GLOBAL_SETTINGS_GAME_DIRS_MIN_WIDTH = 400;
 
 static void UpdateControllerColorButton(QPushButton* button, const QString& hex) {
 	const QColor color(hex);
@@ -123,6 +124,7 @@ ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* p
       m_info(info) {
 	m_ui->setupUi(this);
 	InitGameDirectories();
+	m_ui->controller_group->setVisible(false);
 
 	connect(m_ui->ok_button, &QPushButton::clicked, this, &ConfigurationEditDialog::save);
 	connect(m_ui->cancel_button, &QPushButton::clicked, this, &QDialog::reject);
@@ -142,6 +144,12 @@ ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* p
 		UpdateControllerColorButton(m_ui->button_controller_color, {});
 		emit PreviewControllerColor({});
 	});
+	connect(m_ui->slider_controller_vibration, &QSlider::valueChanged, this, [this](int value) {
+		m_ui->label_controller_vibration_value->setText(tr("%1%").arg(value));
+	});
+	connect(m_ui->slider_controller_volume, &QSlider::valueChanged, this, [this](int value) {
+		m_ui->label_controller_volume_value->setText(tr("%1%").arg(value));
+	});
 	connect(m_ui->comboBox_shader_log_direction, &QComboBox::currentTextChanged, this,
 	        [this](const QString& text) {
 		        auto log = TextToEnum<Configuration::LogDirection>(text);
@@ -201,7 +209,9 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->lineEdit_user_name->setMaxLength(static_cast<int>(Config::MAX_USER_NAME_LENGTH));
 	m_ui->lineEdit_user_name->setText(info.user_name);
 	m_ui->spinBox_user_id->setValue(info.user_id);
-	UpdateControllerColorButton(m_ui->button_controller_color, info.controller_color);
+	UpdateControllerColorButton(m_ui->button_controller_color, info.controller.color);
+	m_ui->slider_controller_vibration->setValue(info.controller.vibration_intensity);
+	m_ui->slider_controller_volume->setValue(info.controller.speaker_volume);
 	auto* microphone = m_ui->comboBox_audio_input_device;
 	microphone->clear();
 	microphone->addItem(tr("None"), QString {});
@@ -263,6 +273,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	    info.gpu_index >= 0 && info.gpu_index < m_ui->comboBox_gpu->count() - 1 ? info.gpu_index + 1
 	                                                                            : 0);
 	m_ui->checkBox_fullscreen->setChecked(info.fullscreen_enabled);
+	m_ui->checkBox_hide_cursor->setChecked(info.hide_cursor_enabled);
 	m_ui->checkBox_readback->setChecked(info.readback_linear_images);
 	m_ui->checkBox_tessellation->setChecked(info.tessellation_enabled);
 	m_ui->spinBox_vblank_frequency->setValue(info.vblank_frequency);
@@ -342,8 +353,8 @@ void ConfigurationEditDialog::InitGameDirectories() {
 	update_game_directory_buttons();
 }
 
-void ConfigurationEditDialog::SetGameDirectories(const QStringList& dirs) {
-	m_show_game_dirs = true;
+void ConfigurationEditDialog::SetGlobalSettings(const QStringList& dirs) {
+	m_global_settings = true;
 	m_game_dirs_list->clear();
 	m_game_dirs_group->setMinimumWidth(GLOBAL_SETTINGS_GAME_DIRS_MIN_WIDTH);
 
@@ -352,6 +363,7 @@ void ConfigurationEditDialog::SetGameDirectories(const QStringList& dirs) {
 	}
 
 	m_game_dirs_group->setVisible(true);
+	m_ui->controller_group->setVisible(true);
 	update_game_directory_buttons();
 	layout()->activate();
 	resize(size().expandedTo(minimumSizeHint()));
@@ -359,7 +371,7 @@ void ConfigurationEditDialog::SetGameDirectories(const QStringList& dirs) {
 
 QStringList ConfigurationEditDialog::GetGameDirectories() const {
 	QStringList dirs;
-	if (!m_show_game_dirs) {
+	if (!m_global_settings) {
 		return dirs;
 	}
 
@@ -399,17 +411,22 @@ void ConfigurationEditDialog::resizeEvent(QResizeEvent* event) {
 	g_last_geometry = saveGeometry();
 }
 
-static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui) {
+static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, bool global_settings) {
 	info.user_name = ui.lineEdit_user_name->text().trimmed();
 	info.user_id   = ui.spinBox_user_id->value();
 	info.audio_input_device = ui.comboBox_audio_input_device->currentData().toString();
-	info.controller_color   = ui.button_controller_color->property("controllerColor").toString();
+	if (global_settings) {
+		info.controller.color = ui.button_controller_color->property("controllerColor").toString();
+		info.controller.vibration_intensity = ui.slider_controller_vibration->value();
+		info.controller.speaker_volume      = ui.slider_controller_volume->value();
+	}
 	info.screen_resolution =
 	    TextToEnum<Configuration::Resolution>(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.hide_cursor_enabled       = ui.checkBox_hide_cursor->isChecked();
 	info.readback_linear_images    = ui.checkBox_readback->isChecked();
 	info.tessellation_enabled      = ui.checkBox_tessellation->isChecked();
 	info.vblank_frequency          = ui.spinBox_vblank_frequency->value();
@@ -453,7 +470,7 @@ void ConfigurationEditDialog::save() {
 		return;
 	}
 
-	UpdateInfo(m_info, *m_ui);
+	UpdateInfo(m_info, *m_ui, m_global_settings);
 
 	emit accept();
 }
@@ -461,9 +478,9 @@ void ConfigurationEditDialog::save() {
 void ConfigurationEditDialog::clear() {
 	Configuration default_info;
 	Init(default_info);
-	emit PreviewControllerColor({});
 
-	if (m_show_game_dirs) {
+	if (m_global_settings) {
+		emit PreviewControllerColor({});
 		m_game_dirs_list->clear();
 		update_game_directory_buttons();
 	}

src/launcher/src/configurationListWidget.cpp

@@ -36,6 +36,7 @@
 #include <QPainter>
 #include <QPalette>
 #include <QPointer>
+#include <QProcess>
 #include <QPushButton>
 #include <QRegularExpression>
 #include <QSet>
@@ -47,6 +48,12 @@
 #include <QUrl>
 #include <QtCore>
 
+#ifdef __linux__
+#include <QDBusConnection>
+#include <QDBusMessage>
+#include <QDBusPendingCallWatcher>
+#endif
+
 #include <memory>
 
 #include "ui_configuration_list_widget.h"
@@ -314,6 +321,7 @@ void ConfigurationListWidget::WriteSettings() {
 	s->remove(CONF_GLOBAL);
 	s->beginGroup(CONF_GLOBAL);
 	m_global_info.WriteSettings(s.get());
+	m_global_info.controller.WriteSettings(s.get());
 	s->endGroup();
 
 	s->remove(CONF_SECTION_NAME);
@@ -357,6 +365,7 @@ void ConfigurationListWidget::ReadSettings() {
 	if (!s->childKeys().isEmpty()) {
 		m_global_info.ReadSettings(s.get());
 	}
+	m_global_info.controller.ReadSettings(s.get());
 	s->endGroup();
 
 	qDeleteAll(m_custom_infos);
@@ -409,6 +418,7 @@ ConfigurationListWidget::CreateConfiguration(const ConfigurationItem& item) cons
 	const auto* custom = m_custom_infos.value(item.GetInfo().game_path);
 	info->CopyGameInfoFrom(item.GetInfo());
 	info->CopyEmulatorSettingsFrom(custom != nullptr ? *custom : m_global_info);
+	info->controller = m_global_info.controller;
 	if (custom != nullptr && !custom->elf.isEmpty()) {
 		info->elf = custom->elf;
 	}
@@ -787,16 +797,18 @@ void ConfigurationListWidget::delete_configuartion() {
 void ConfigurationListWidget::edit_global_settings() {
 	Configuration info;
 	info.CopyEmulatorSettingsFrom(m_global_info);
+	info.controller = m_global_info.controller;
 	info.name = tr("Global settings");
 
 	ConfigurationEditDialog dlg(info, this);
 	dlg.setWindowTitle(tr("Global settings"));
-	dlg.SetGameDirectories(m_game_dirs);
+	dlg.SetGlobalSettings(m_game_dirs);
 	connect(&dlg, &ConfigurationEditDialog::PreviewControllerColor, this,
 	        &ConfigurationListWidget::PreviewControllerColor);
 
 	if (dlg.exec() == QDialog::Accepted) {
 		m_global_info.CopyEmulatorSettingsFrom(info);
+		m_global_info.controller     = info.controller;
 		const auto game_dirs         = NormalizeGameDirectories(dlg.GetGameDirectories());
 		const bool game_dirs_changed = game_dirs != m_game_dirs;
 		m_game_dirs                  = game_dirs;
@@ -832,16 +844,52 @@ void ConfigurationListWidget::open_game_folder() {
 		return;
 	}
 
-	const auto base = item->GetInfo().basedir;
-	const QDir game_dir(GameContent::IsArchive(base) ? QFileInfo(base).absolutePath() : base);
+	const auto base       = item->GetInfo().basedir;
+	const bool is_archive = GameContent::IsArchive(base);
+	const QDir game_dir(is_archive ? QFileInfo(base).absolutePath() : base);
 	if (!game_dir.exists()) {
 		QMessageBox::warning(this, tr("Open game folder"), tr("Game folder does not exist."));
 		return;
 	}
 
-	if (!QDesktopServices::openUrl(QUrl::fromLocalFile(game_dir.absolutePath()))) {
-		QMessageBox::warning(this, tr("Open game folder"), tr("Could not open game folder."));
+	const auto open_directory = [this, game_dir] {
+		if (!QDesktopServices::openUrl(QUrl::fromLocalFile(game_dir.absolutePath()))) {
+			QMessageBox::warning(this, tr("Open game folder"), tr("Could not open game folder."));
+		}
+	};
+	if (is_archive) {
+		const auto path = QFileInfo(base).absoluteFilePath();
+#if defined(_WIN32)
+		QProcess explorer;
+		explorer.setProgram("explorer.exe");
+		explorer.setNativeArguments(
+		    QStringLiteral("/select,\"%1\"").arg(QDir::toNativeSeparators(path)));
+		if (explorer.startDetached()) {
+			return;
+		}
+#elif defined(__APPLE__)
+		if (QProcess::startDetached("/usr/bin/open", {"-R", path})) {
+			return;
+		}
+#elif defined(__linux__)
+		auto request = QDBusMessage::createMethodCall("org.freedesktop.FileManager1",
+		                                              "/org/freedesktop/FileManager1",
+		                                              "org.freedesktop.FileManager1", "ShowItems");
+		request << QStringList {QUrl::fromLocalFile(path).toString(QUrl::FullyEncoded)}
+		        << QString();
+		auto* watcher = new QDBusPendingCallWatcher(
+		    QDBusConnection::sessionBus().asyncCall(request, 5000), this);
+		connect(watcher, &QDBusPendingCallWatcher::finished, this,
+		        [open_directory](QDBusPendingCallWatcher* call) {
+			        call->deleteLater();
+			        if (call->isError()) {
+				        open_directory();
+			        }
+		        });
+		return;
+#endif
 	}
+	open_directory();
 }
 
 void ConfigurationListWidget::remove_save_data() {

src/launcher/src/mainDialog.cpp

@@ -228,9 +228,11 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	if (!info.audio_input_device.isEmpty()) {
 		args << "--mic" << info.audio_input_device;
 	}
-	if (!info.controller_color.isEmpty()) {
-		args << "--controller-color" << info.controller_color;
+	if (!info.controller.color.isEmpty()) {
+		args << "--controller-color" << info.controller.color;
 	}
+	args << "--controller-volume" << QString::number(info.controller.speaker_volume);
+	args << "--controller-vibration" << QString::number(info.controller.vibration_intensity);
 	args << "--present-mode" << EnumToText(info.present_mode);
 	if (info.gpu_index >= 0) {
 		args << "--gpu" << QString::number(info.gpu_index);
@@ -238,6 +240,9 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	if (info.fullscreen_enabled) {
 		args << "--fullscreen";
 	}
+	if (info.hide_cursor_enabled) {
+		args << "--hide-cursor";
+	}
 	args << "--readback-linear-images" << BoolArg(info.readback_linear_images);
 	if (info.tessellation_enabled) {
 		args << "--tessellation";
@@ -529,8 +534,7 @@ void MainDialogPrivate::Update() {
 		m_lightbar.Stop();
 		return;
 	}
-	m_lightbar.SetColor(item != nullptr ? m_ui->widget->CreateConfiguration(*item)->controller_color
-	                                    : m_ui->widget->GetGlobalControllerColor());
+	m_lightbar.SetColor(m_ui->widget->GetGlobalControllerColor());
 }
 
 #include "mainDialog.moc"

src/libs/controller.cpp

@@ -223,7 +223,15 @@ void GameController::CycleSetting(Setting setting) {
 
 float GameController::GetSettingScale(Setting setting) const {
 	const auto step = m_setting_steps[static_cast<size_t>(setting)].load(std::memory_order_relaxed);
-	return setting == Setting::SpeakerVolume ? SPEAKER_VOLUME[step] : INTENSITY[step];
+	if (setting == Setting::SpeakerVolume) {
+		// The old speaker level matches the PS5 slider's midpoint. Give the upper
+		// half of the global slider up to 6 dB of extra software gain.
+		return SPEAKER_VOLUME[step] * (Config::GetControllerSpeakerVolume() / 50.0f);
+	}
+	if (setting == Setting::VibrationIntensity) {
+		return INTENSITY[step] * (Config::GetControllerVibrationIntensity() / 100.0f);
+	}
+	return INTENSITY[step];
 }
 
 static uint8_t Scale(uint8_t value, float scale) {

src/libs/dialog.cpp

@@ -14,6 +14,81 @@
 // NOLINTNEXTLINE(modernize-concat-nested-namespaces)
 namespace Libs::Dialog {
 
+namespace {
+
+constexpr int DIALOG_STATUS_RUNNING  = 2;
+constexpr int DIALOG_STATUS_FINISHED = 3;
+
+struct DialogState {
+	int      status     = 0;
+	uint64_t generation = 0;
+};
+
+std::mutex  g_dialog_mutex;
+DialogState g_error;
+DialogState g_signin;
+uint64_t    g_dialog_generation        = 0;
+uint64_t    g_dialog_revision          = 0;
+int32_t     g_error_code               = 0;
+void (*g_dialog_visibility_callback)() = nullptr;
+
+void SetDialogStatus(DialogState& dialog, int status, std::unique_lock<std::mutex>& lock) {
+	const bool visibility_changed =
+	    (dialog.status == DIALOG_STATUS_RUNNING) != (status == DIALOG_STATUS_RUNNING);
+	dialog.status = status;
+	if (visibility_changed) {
+		g_dialog_revision++;
+		if (status == DIALOG_STATUS_RUNNING) {
+			dialog.generation = ++g_dialog_generation;
+		}
+	}
+	const auto callback = visibility_changed ? g_dialog_visibility_callback : nullptr;
+	lock.unlock();
+	if (callback != nullptr) {
+		callback();
+	}
+}
+
+} // namespace
+
+namespace SystemDialog {
+
+bool GetHostSnapshot(HostSnapshot* snapshot) {
+	std::scoped_lock lock(g_dialog_mutex);
+	if (g_signin.status == DIALOG_STATUS_RUNNING) {
+		*snapshot = {Kind::Signin, g_signin.generation, 0};
+		return true;
+	}
+	if (g_error.status == DIALOG_STATUS_RUNNING) {
+		*snapshot = {Kind::Error, g_error.generation, g_error_code};
+		return true;
+	}
+	return false;
+}
+
+VisualState GetVisualState() noexcept {
+	std::scoped_lock lock(g_dialog_mutex);
+	const bool       signin = g_signin.status == DIALOG_STATUS_RUNNING;
+	return {g_error.status == DIALOG_STATUS_RUNNING || signin, signin, g_dialog_revision};
+}
+
+void SetVisibilityCallback(void (*callback)()) {
+	std::scoped_lock lock(g_dialog_mutex);
+	g_dialog_visibility_callback = callback;
+}
+
+bool HostClose(uint64_t generation) {
+	std::unique_lock lock(g_dialog_mutex);
+	auto&            dialog = g_signin.status == DIALOG_STATUS_RUNNING ? g_signin : g_error;
+	if (dialog.status != DIALOG_STATUS_RUNNING || generation != dialog.generation) {
+		return false;
+	}
+	SetDialogStatus(dialog, DIALOG_STATUS_FINISHED, lock);
+	return true;
+}
+
+} // namespace SystemDialog
+
 namespace CommonDialog {
 
 LIB_NAME("CommonDialog", "CommonDialog");
@@ -207,6 +282,7 @@ constexpr int SIGNIN_ERROR_NOT_INITIALIZED     = static_cast<int>(0x81350001u);
 constexpr int SIGNIN_ERROR_ALREADY_INITIALIZED = static_cast<int>(0x81350002u);
 constexpr int SIGNIN_ERROR_PARAM_INVALID       = static_cast<int>(0x81350003u);
 constexpr int SIGNIN_ERROR_INVALID_STATE       = static_cast<int>(0x81350005u);
+constexpr int SIGNIN_ERROR_INVALID_USER_ID     = static_cast<int>(0x81350007u);
 
 struct SigninDialogResult {
 	int32_t result;
@@ -222,70 +298,78 @@ struct SigninDialogParam {
 static_assert(sizeof(SigninDialogResult) == 16);
 static_assert(sizeof(SigninDialogParam) == 16);
 
-static int g_signin_status = SIGNIN_STATUS_NONE;
-
 int KYTY_SYSV_ABI SigninDialogInitialize() {
 	PRINT_NAME();
-	if (g_signin_status != SIGNIN_STATUS_NONE) {
+	std::scoped_lock lock(g_dialog_mutex);
+	if (g_signin.status != SIGNIN_STATUS_NONE) {
 		return SIGNIN_ERROR_ALREADY_INITIALIZED;
 	}
-	g_signin_status = SIGNIN_STATUS_INITIALIZED;
+	g_signin.status = SIGNIN_STATUS_INITIALIZED;
 	return OK;
 }
 
 int KYTY_SYSV_ABI SigninDialogTerminate() {
 	PRINT_NAME();
-	if (g_signin_status == SIGNIN_STATUS_NONE) {
+	std::unique_lock lock(g_dialog_mutex);
+	if (g_signin.status == SIGNIN_STATUS_NONE) {
 		return SIGNIN_ERROR_NOT_INITIALIZED;
 	}
-	g_signin_status = SIGNIN_STATUS_NONE;
+	SetDialogStatus(g_signin, SIGNIN_STATUS_NONE, lock);
 	return OK;
 }
 
 int KYTY_SYSV_ABI SigninDialogOpen(const void* param) {
 	PRINT_NAME();
-	if (g_signin_status != SIGNIN_STATUS_INITIALIZED && g_signin_status != SIGNIN_STATUS_FINISHED) {
+	std::unique_lock lock(g_dialog_mutex);
+	if (g_signin.status == SIGNIN_STATUS_NONE) {
+		return SIGNIN_ERROR_NOT_INITIALIZED;
+	}
+	if (g_signin.status != SIGNIN_STATUS_INITIALIZED && g_signin.status != SIGNIN_STATUS_FINISHED) {
 		return SIGNIN_ERROR_INVALID_STATE;
 	}
-	if (param == nullptr ||
-	    static_cast<const SigninDialogParam*>(param)->size != sizeof(SigninDialogParam)) {
+	if (param == nullptr) {
 		return SIGNIN_ERROR_PARAM_INVALID;
 	}
-	g_signin_status = SIGNIN_STATUS_RUNNING;
+	const auto* p = static_cast<const SigninDialogParam*>(param);
+	if (p->size != sizeof(SigninDialogParam) || p->reserved[0] != 0 || p->reserved[1] != 0) {
+		return SIGNIN_ERROR_PARAM_INVALID;
+	}
+	if (p->user_id != Config::GetUserId()) {
+		return SIGNIN_ERROR_INVALID_USER_ID;
+	}
+	SetDialogStatus(g_signin, SIGNIN_STATUS_RUNNING, lock);
 	return OK;
 }
 
 int KYTY_SYSV_ABI SigninDialogClose() {
 	PRINT_NAME();
-	if (g_signin_status == SIGNIN_STATUS_NONE) {
+	std::unique_lock lock(g_dialog_mutex);
+	if (g_signin.status == SIGNIN_STATUS_NONE) {
 		return SIGNIN_ERROR_NOT_INITIALIZED;
 	}
-	if (g_signin_status != SIGNIN_STATUS_RUNNING && g_signin_status != SIGNIN_STATUS_FINISHED) {
-		return SIGNIN_ERROR_INVALID_STATE;
-	}
-	g_signin_status = SIGNIN_STATUS_FINISHED;
+	SetDialogStatus(g_signin, SIGNIN_STATUS_FINISHED, lock);
 	return OK;
 }
 
 int KYTY_SYSV_ABI SigninDialogUpdateStatus() {
 	PRINT_NAME();
-	if (g_signin_status == SIGNIN_STATUS_RUNNING) {
-		g_signin_status = SIGNIN_STATUS_FINISHED;
-	}
-	return g_signin_status;
+	std::scoped_lock lock(g_dialog_mutex);
+	return g_signin.status;
 }
 
 int KYTY_SYSV_ABI SigninDialogGetStatus() {
 	PRINT_NAME();
-	return g_signin_status;
+	std::scoped_lock lock(g_dialog_mutex);
+	return g_signin.status;
 }
 
 int KYTY_SYSV_ABI SigninDialogGetResult(void* result) {
 	PRINT_NAME();
-	if (g_signin_status == SIGNIN_STATUS_NONE) {
+	std::scoped_lock lock(g_dialog_mutex);
+	if (g_signin.status == SIGNIN_STATUS_NONE) {
 		return SIGNIN_ERROR_NOT_INITIALIZED;
 	}
-	if (g_signin_status != SIGNIN_STATUS_FINISHED) {
+	if (g_signin.status != SIGNIN_STATUS_FINISHED) {
 		return SIGNIN_ERROR_INVALID_STATE;
 	}
 	if (result == nullptr) {
@@ -661,64 +745,15 @@ struct ErrorDialogParam {
 
 static_assert(sizeof(ErrorDialogParam) == 16);
 
-static std::mutex g_error_mutex;
-static int        g_error_status     = STATUS_NONE;
-static uint64_t   g_error_generation = 0;
-static uint64_t   g_error_revision   = 0;
-static int32_t    g_error_code       = 0;
-static void (*g_error_visibility_callback)() = nullptr;
-
-static void SetStatus(int status, std::unique_lock<std::mutex>& lock) {
-	const bool visibility_changed =
-	    (g_error_status == STATUS_RUNNING) != (status == STATUS_RUNNING);
-	g_error_status = status;
-	if (visibility_changed) {
-		g_error_revision++;
-	}
-	const auto callback = visibility_changed ? g_error_visibility_callback : nullptr;
-	lock.unlock();
-	if (callback != nullptr) {
-		callback();
-	}
-}
-
-bool GetHostSnapshot(HostSnapshot* snapshot) {
-	std::scoped_lock lock(g_error_mutex);
-	if (g_error_status != STATUS_RUNNING) {
-		return false;
-	}
-	*snapshot = {g_error_generation, g_error_code};
-	return true;
-}
-
-VisualState GetVisualState() noexcept {
-	std::scoped_lock lock(g_error_mutex);
-	return {g_error_status == STATUS_RUNNING, g_error_revision};
-}
-
-void SetVisibilityCallback(void (*callback)()) {
-	std::scoped_lock lock(g_error_mutex);
-	g_error_visibility_callback = callback;
-}
-
-bool HostAccept(uint64_t generation) {
-	std::unique_lock lock(g_error_mutex);
-	if (g_error_status != STATUS_RUNNING || generation != g_error_generation) {
-		return false;
-	}
-	SetStatus(STATUS_FINISHED, lock);
-	return true;
-}
-
 int KYTY_SYSV_ABI ErrorDialogInitialize() {
 	PRINT_NAME();
 
-	std::scoped_lock lock(g_error_mutex);
-	if (g_error_status != STATUS_NONE) {
+	std::scoped_lock lock(g_dialog_mutex);
+	if (g_error.status != STATUS_NONE) {
 		return ERROR_ALREADY_INITIALIZED;
 	}
 
-	g_error_status = STATUS_INITIALIZED;
+	g_error.status = STATUS_INITIALIZED;
 
 	return OK;
 }
@@ -726,8 +761,8 @@ int KYTY_SYSV_ABI ErrorDialogInitialize() {
 int KYTY_SYSV_ABI ErrorDialogOpen(const void* param) {
 	PRINT_NAME();
 
-	std::unique_lock lock(g_error_mutex);
-	if (g_error_status != STATUS_INITIALIZED && g_error_status != STATUS_FINISHED) {
+	std::unique_lock lock(g_dialog_mutex);
+	if (g_error.status != STATUS_INITIALIZED && g_error.status != STATUS_FINISHED) {
 		return ERROR_INVALID_STATE;
 	}
 	if (param == nullptr) {
@@ -745,8 +780,7 @@ int KYTY_SYSV_ABI ErrorDialogOpen(const void* param) {
 	     p->size, static_cast<uint32_t>(p->error_code), p->user_id);
 
 	g_error_code = p->error_code;
-	g_error_generation++;
-	SetStatus(STATUS_RUNNING, lock);
+	SetDialogStatus(g_error, STATUS_RUNNING, lock);
 
 	return OK;
 }
@@ -754,12 +788,12 @@ int KYTY_SYSV_ABI ErrorDialogOpen(const void* param) {
 int KYTY_SYSV_ABI ErrorDialogClose() {
 	PRINT_NAME();
 
-	std::unique_lock lock(g_error_mutex);
-	if (g_error_status != STATUS_RUNNING) {
+	std::unique_lock lock(g_dialog_mutex);
+	if (g_error.status != STATUS_RUNNING) {
 		return ERROR_INVALID_STATE;
 	}
 
-	SetStatus(STATUS_FINISHED, lock);
+	SetDialogStatus(g_error, STATUS_FINISHED, lock);
 
 	return OK;
 }
@@ -767,12 +801,12 @@ int KYTY_SYSV_ABI ErrorDialogClose() {
 int KYTY_SYSV_ABI ErrorDialogTerminate() {
 	PRINT_NAME();
 
-	std::unique_lock lock(g_error_mutex);
-	if (g_error_status == STATUS_NONE) {
+	std::unique_lock lock(g_dialog_mutex);
+	if (g_error.status == STATUS_NONE) {
 		return ERROR_NOT_INITIALIZED;
 	}
 
-	SetStatus(STATUS_NONE, lock);
+	SetDialogStatus(g_error, STATUS_NONE, lock);
 
 	return OK;
 }
@@ -780,15 +814,15 @@ int KYTY_SYSV_ABI ErrorDialogTerminate() {
 int KYTY_SYSV_ABI ErrorDialogUpdateStatus() {
 	PRINT_NAME();
 
-	std::scoped_lock lock(g_error_mutex);
-	return g_error_status;
+	std::scoped_lock lock(g_dialog_mutex);
+	return g_error.status;
 }
 
 int KYTY_SYSV_ABI ErrorDialogGetStatus() {
 	PRINT_NAME();
 
-	std::scoped_lock lock(g_error_mutex);
-	return g_error_status;
+	std::scoped_lock lock(g_dialog_mutex);
+	return g_error.status;
 }
 
 } // namespace ErrorDialog

src/libs/dialog.h

@@ -68,22 +68,30 @@ int KYTY_SYSV_ABI MsgDialogProgressBarSetMsg(int target, const char* msg);
 
 } // namespace MsgDialog
 
-namespace ErrorDialog {
+namespace SystemDialog {
+
+enum class Kind { Error, Signin };
 
 struct HostSnapshot {
+	Kind     kind;
 	uint64_t generation;
 	int32_t  error_code;
 };
 
 struct VisualState {
 	bool     active;
+	bool     background;
 	uint64_t revision;
 };
 
 bool        GetHostSnapshot(HostSnapshot* snapshot);
 VisualState GetVisualState() noexcept;
 void        SetVisibilityCallback(void (*callback)());
-bool        HostAccept(uint64_t generation);
+bool        HostClose(uint64_t generation);
+
+} // namespace SystemDialog
+
+namespace ErrorDialog {
 
 int KYTY_SYSV_ABI ErrorDialogInitialize();
 int KYTY_SYSV_ABI ErrorDialogOpen(const void* param);

src/libs/dualSenseHaptics.cpp

@@ -178,7 +178,11 @@ SDL_AudioDeviceID FindDevice() {
 	for (int i = 0; i < count; i++) {
 		const char*   name = SDL_GetAudioDeviceName(devices[i]);
 		SDL_AudioSpec spec {};
-		if (name != nullptr && SDL_strcasestr(name, "DualSense") != nullptr &&
+		const bool is_dualsense_name =
+		    name != nullptr &&
+		    (SDL_strcasestr(name, "DualSense") != nullptr ||
+		     SDL_strcasestr(name, "Wireless Controller") != nullptr);
+		if (is_dualsense_name &&
 		    SDL_GetAudioDeviceFormat(devices[i], &spec, nullptr) && spec.channels == 4) {
 			if (device != 0) {
 				device = 0; // Multiple matching endpoints cannot be associated reliably either.

src/libs/libKernel.cpp

@@ -2189,6 +2189,7 @@ LIB_DEFINE(InitLibKernel_1_Posix) {
 	LIB_FUNC("wtkt-teR1so", Posix::pthread_attr_init);
 	LIB_FUNC("zHchY8ft5pk", Posix::pthread_attr_destroy);
 	LIB_FUNC("vQm4fDEsWi8", Posix::pthread_attr_getstack);
+	LIB_FUNC("-SrbXpGR1f0", Posix::pthread_attr_setstack);
 	LIB_FUNC("2Q0z6rnBrTE", Posix::pthread_attr_setstacksize);
 	LIB_FUNC("Ucsu-OK+els", Posix::pthread_attr_get_np);
 	LIB_FUNC("RtLRV-pBTTY", Posix::pthread_attr_getschedpolicy);
@@ -2827,6 +2828,7 @@ constexpr int32_t KERNEL_AIO_STATE_SUBMITTED  = 1;
 constexpr int32_t KERNEL_AIO_STATE_PROCESSING = 2;
 constexpr int32_t KERNEL_AIO_STATE_COMPLETED  = 3;
 constexpr int32_t KERNEL_AIO_STATE_ABORTED    = 4;
+constexpr int32_t KERNEL_AIO_STATE_NOTIFIED   = 0x10000;
 constexpr int32_t KERNEL_AIO_MAX_QUEUE        = 512;
 constexpr int32_t KERNEL_AIO_MAX_REQUESTS     = 128;
 
@@ -2939,10 +2941,31 @@ int KYTY_SYSV_ABI KernelAioPollRequest(int32_t id, int32_t* state) {
 	}
 
 	if (!kernel_aio_is_valid_id(id)) {
-		return LibKernel::KERNEL_ERROR_EINVAL;
+		*state = LibKernel::KERNEL_ERROR_ESRCH;
+		return OK;
 	}
 
-	*state = g_kernel_aio_state[id].load(std::memory_order_acquire);
+	auto current = g_kernel_aio_state[id].load(std::memory_order_acquire);
+	while (current == KERNEL_AIO_STATE_COMPLETED || current == KERNEL_AIO_STATE_ABORTED) {
+		if (g_kernel_aio_state[id].compare_exchange_weak(current, current | KERNEL_AIO_STATE_NOTIFIED,
+		                                                std::memory_order_acq_rel)) {
+			break;
+		}
+	}
+	*state = current == 0 ? LibKernel::KERNEL_ERROR_ESRCH : current;
+	return OK;
+}
+
+int KYTY_SYSV_ABI KernelAioPollRequests(int32_t* ids, int32_t num, int32_t* states) {
+	if (ids == nullptr || states == nullptr) {
+		return LibKernel::KERNEL_ERROR_EFAULT;
+	}
+	if (num <= 0 || num > KERNEL_AIO_MAX_REQUESTS) {
+		return LibKernel::KERNEL_ERROR_EINVAL;
+	}
+	for (int32_t i = 0; i < num; i++) {
+		KernelAioPollRequest(ids[i], &states[i]);
+	}
 	return OK;
 }
 
@@ -2954,7 +2977,7 @@ int KYTY_SYSV_ABI KernelAioWaitRequest(int32_t id, int32_t* state, uint32_t* use
 	}
 
 	if (!kernel_aio_is_valid_id(id)) {
-		return LibKernel::KERNEL_ERROR_EINVAL;
+		return KernelAioPollRequest(id, state);
 	}
 
 	uint32_t waited  = 0;
@@ -2970,8 +2993,7 @@ int KYTY_SYSV_ABI KernelAioWaitRequest(int32_t id, int32_t* state, uint32_t* use
 		current = g_kernel_aio_state[id].load(std::memory_order_acquire);
 	}
 
-	*state = current;
-	return OK;
+	return KernelAioPollRequest(id, state);
 }
 
 int KYTY_SYSV_ABI KernelAioDeleteRequest(int32_t id, int32_t* ret) {
@@ -2982,12 +3004,35 @@ int KYTY_SYSV_ABI KernelAioDeleteRequest(int32_t id, int32_t* ret) {
 	}
 
 	if (!kernel_aio_is_valid_id(id)) {
-		return LibKernel::KERNEL_ERROR_EINVAL;
+		*ret = LibKernel::KERNEL_ERROR_ESRCH;
+		return OK;
 	}
 
-	g_kernel_aio_state[id].store(KERNEL_AIO_STATE_ABORTED, std::memory_order_release);
-	*ret = OK;
+	auto current = g_kernel_aio_state[id].load(std::memory_order_acquire);
+	for (;;) {
+		const auto state = current & ~KERNEL_AIO_STATE_NOTIFIED;
+		if (state != KERNEL_AIO_STATE_COMPLETED && state != KERNEL_AIO_STATE_ABORTED) {
+			*ret = current == 0 ? LibKernel::KERNEL_ERROR_ESRCH : LibKernel::KERNEL_ERROR_EBUSY;
+			break;
+		}
+		if (g_kernel_aio_state[id].compare_exchange_weak(current, 0, std::memory_order_acq_rel)) {
+			*ret = OK;
+			break;
+		}
+	}
+	return OK;
+}
 
+int KYTY_SYSV_ABI KernelAioDeleteRequests(int32_t* ids, int32_t num, int32_t* rets) {
+	if (ids == nullptr || rets == nullptr) {
+		return LibKernel::KERNEL_ERROR_EFAULT;
+	}
+	if (num <= 0 || num > KERNEL_AIO_MAX_REQUESTS) {
+		return LibKernel::KERNEL_ERROR_EINVAL;
+	}
+	for (int32_t i = 0; i < num; i++) {
+		KernelAioDeleteRequest(ids[i], &rets[i]);
+	}
 	return OK;
 }
 
@@ -3103,6 +3148,7 @@ LIB_DEFINE(InitLibKernel_1_Equeue) {
 	LIB_FUNC("WDszmSbWuDk", EventQueue::KernelAddUserEventEdge);
 	LIB_FUNC("F6e0kwo4cnk", EventQueue::KernelTriggerUserEvent);
 	LIB_FUNC("LJDwdSNTnDg", EventQueue::KernelDeleteUserEvent);
+	LIB_FUNC("57ZK+ODEXWY", EventQueue::KernelAddTimerEvent);
 	LIB_FUNC("R74tt43xP6k", EventQueue::KernelAddHRTimerEvent);
 	LIB_FUNC("J+LF6LwObXU", EventQueue::KernelDeleteHRTimerEvent);
 	LIB_FUNC("bBfz7kMF2Ho", EventQueue::KernelAddAmprEvent);
@@ -3325,8 +3371,10 @@ LIB_DEFINE(InitLibKernel_1) {
 	LIB_FUNC("kbw4UHHSYy0", LibKernel::pthread_cxa_finalize);
 	LIB_FUNC("lLMT9vJAck0", LibKernel::clock_gettime);
 	LIB_FUNC("5TgME6AYty4", KernelAioDeleteRequest);
+	LIB_FUNC("Ft3EtsZzAoY", KernelAioDeleteRequests);
 	LIB_FUNC("HgX7+AORI58", KernelAioSubmitReadCommands);
 	LIB_FUNC("2pOuoWoCxdk", KernelAioPollRequest);
+	LIB_FUNC("o7O4z3jwKzo", KernelAioPollRequests);
 	LIB_FUNC("KOF-oJbQVvc", KernelAioWaitRequest);
 	LIB_FUNC("XQ8C8y+de+E", KernelAioSubmitWriteCommands);
 	LIB_FUNC("nu4a0-arQis", KernelAioInitializeParam);

src/libs/libNet.cpp

@@ -111,6 +111,10 @@ int KYTY_SYSV_ABI NetResolverDestroy(int rid) {
 	return FinishNetCall(Net::NetResolverDestroy(rid));
 }
 
+int KYTY_SYSV_ABI NetResolverAbort(int rid, int flags) {
+	return FinishNetCall(Net::NetResolverAbort(rid, flags));
+}
+
 int KYTY_SYSV_ABI NetResolverStartNtoa(int rid, const char* hostname, void* addr, int timeout,
                                        int retry, int flags) {
 	return FinishNetCall(Net::NetResolverStartNtoa(rid, hostname, addr, timeout, retry, flags));
@@ -203,6 +207,7 @@ LIB_DEFINE(InitNet_1_Net) {
 	LIB_FUNC("K7RlrTkI-mw", LibNet::NetPoolDestroy);
 	LIB_FUNC("C4UgDHHPvdw", LibNet::NetResolverCreate);
 	LIB_FUNC("kJlYH5uMAWI", LibNet::NetResolverDestroy);
+	LIB_FUNC("AzqoBha7js4", LibNet::NetResolverAbort);
 	LIB_FUNC("Nd91WaWmG2w", LibNet::NetResolverStartNtoa);
 	LIB_FUNC("8Kcp5d-q1Uo", LibNet::NetInetPton);
 	LIB_FUNC("9vA2aW+CHuA", LibNet::NetInetNtop);
@@ -3133,9 +3138,21 @@ namespace LibNpWebApi2 {
 
 LIB_VERSION("NpWebApi2", 1, "NpWebApi2", 1, 1);
 
-constexpr int NP_WEBAPI2_ERROR_INVALID_ARGUMENT  = -2141899774; /* 0x80553402 */
-constexpr int NP_WEBAPI2_ERROR_REQUEST_NOT_FOUND = -2141899770; /* 0x80553406 */
-constexpr int NP_WEBAPI2_ERROR_NOT_SIGNED_IN     = -2141899769; /* 0x80553407 */
+constexpr int NP_WEBAPI2_ERROR_INVALID_ARGUMENT       = -2141899774; /* 0x80553402 */
+constexpr int NP_WEBAPI2_ERROR_INVALID_LIB_CONTEXT_ID = -2141899773; /* 0x80553403 */
+constexpr int NP_WEBAPI2_ERROR_LIB_CONTEXT_NOT_FOUND  = -2141899772; /* 0x80553404 */
+constexpr int NP_WEBAPI2_ERROR_REQUEST_NOT_FOUND      = -2141899770; /* 0x80553406 */
+constexpr int NP_WEBAPI2_ERROR_NOT_SIGNED_IN          = -2141899769; /* 0x80553407 */
+
+struct NpWebApi2MemoryPoolStats {
+	size_t  pool_size;
+	size_t  max_inuse_size;
+	size_t  current_inuse_size;
+	int32_t reserved;
+};
+
+static std::mutex            g_np_webapi2_context_mutex;
+static std::map<int, size_t> g_np_webapi2_contexts;
 
 struct NpWebApi2ResponseInformationOption {
 	int32_t http_status;
@@ -3166,7 +3183,30 @@ static int KYTY_SYSV_ABI NpWebApi2Initialize(int lib_http_ctx_id, size_t pool_si
 
 	static int id = 0;
 
-	return ++id;
+	if (pool_size > std::numeric_limits<size_t>::max() - 0x3fff) {
+		return NP_WEBAPI2_ERROR_INVALID_ARGUMENT;
+	}
+	std::lock_guard lock(g_np_webapi2_context_mutex);
+	g_np_webapi2_contexts[++id] = (pool_size + 0x3fff) & ~size_t {0x3fff};
+	return id;
+}
+
+static int KYTY_SYSV_ABI NpWebApi2GetMemoryPoolStats(int lib_ctx_id,
+                                                   NpWebApi2MemoryPoolStats* stats) {
+	if (stats == nullptr) {
+		return NP_WEBAPI2_ERROR_INVALID_ARGUMENT;
+	}
+	if (lib_ctx_id <= 0) {
+		return NP_WEBAPI2_ERROR_INVALID_LIB_CONTEXT_ID;
+	}
+	std::lock_guard lock(g_np_webapi2_context_mutex);
+	const auto context = g_np_webapi2_contexts.find(lib_ctx_id);
+	if (context == g_np_webapi2_contexts.end()) {
+		return NP_WEBAPI2_ERROR_LIB_CONTEXT_NOT_FOUND;
+	}
+	// Requests currently use host storage, not allocations from the library pool.
+	*stats = {context->second, 0, 0, 0};
+	return OK;
 }
 
 static int KYTY_SYSV_ABI NpWebApi2PushEventCreateHandle(int lib_ctx_id) {
@@ -3433,11 +3473,16 @@ static int KYTY_SYSV_ABI NpWebApi2Terminate(int lib_ctx_id) {
 
 	LOGF("\t lib_ctx_id = %d\n", lib_ctx_id);
 
-	return 0;
+	if (lib_ctx_id <= 0) {
+		return NP_WEBAPI2_ERROR_INVALID_LIB_CONTEXT_ID;
+	}
+	std::lock_guard lock(g_np_webapi2_context_mutex);
+	return g_np_webapi2_contexts.erase(lib_ctx_id) != 0 ? OK : NP_WEBAPI2_ERROR_LIB_CONTEXT_NOT_FOUND;
 }
 
 LIB_DEFINE(InitNet_1_NpWebApi2) {
 	LIB_FUNC("+o9816YQhqQ", LibNpWebApi2::NpWebApi2Initialize);
+	LIB_FUNC("Xweb+naPZ8Y", LibNpWebApi2::NpWebApi2GetMemoryPoolStats);
 	LIB_FUNC("WV1GwM32NgY", LibNpWebApi2::NpWebApi2PushEventCreateHandle);
 	LIB_FUNC("sk54bi6FtYM", LibNpWebApi2::NpWebApi2CreateUserContext);
 	LIB_FUNC("9X9+cneTGUU", LibNpWebApi2::NpWebApi2DeleteUserContext);

src/libs/libSystemService.cpp

@@ -163,8 +163,10 @@ static int KYTY_SYSV_ABI SystemServiceGetStatus(SystemServiceStatus* status) {
 		return SYSTEM_SERVICE_ERROR_PARAMETER;
 	}
 
-	*status                      = SystemServiceStatus();
-	status->is_system_ui_overlaid = Dialog::ErrorDialog::GetVisualState().active;
+	const auto dialog                  = Dialog::SystemDialog::GetVisualState();
+	*status                            = SystemServiceStatus();
+	status->is_system_ui_overlaid      = dialog.active;
+	status->is_in_background_execution = dialog.background;
 
 	return OK;
 }

src/libs/network.cpp

@@ -849,20 +849,38 @@ using SocketIoLength                                = size_t;
 #endif
 
 struct SocketTransport {
-	explicit SocketTransport(NativeSocket value): socket(value) {}
+	SocketTransport(NativeSocket value, int socket_type): socket(value), type(socket_type) {}
 	~SocketTransport() { Close(); }
+	bool HasPendingReceive() const {
+#if defined(_WIN32)
+		return buffered_bytes.load(std::memory_order_relaxed) != 0 ||
+		       receive_shutdown.load(std::memory_order_relaxed);
+#else
+		return false;
+#endif
+	}
 	int Close() {
-		if (socket == INVALID_NATIVE_SOCKET) {
+#if defined(_WIN32)
+		receive_shutdown.store(true, std::memory_order_relaxed);
+#endif
+		const auto value = socket.exchange(INVALID_NATIVE_SOCKET);
+		if (value == INVALID_NATIVE_SOCKET) {
 			return 0;
 		}
-		const auto value = std::exchange(socket, INVALID_NATIVE_SOCKET);
 #if defined(_WIN32)
 		return closesocket(value);
 #else
 		return ::close(value);
 #endif
 	}
-	NativeSocket socket;
+	std::atomic<NativeSocket> socket;
+	const int                type;
+#if defined(_WIN32)
+	std::mutex         receive_mutex;
+	std::vector<char>  receive_buffer;
+	std::atomic_size_t buffered_bytes   = 0;
+	std::atomic_bool   receive_shutdown = false;
+#endif
 };
 
 struct P2pEndpoint {
@@ -1164,7 +1182,8 @@ static int ConvertHostSockaddr(const sockaddr_storage* addr, SocketLength addrle
 	return 0;
 }
 
-static bool GetSocketBackend(int guest_fd, NativeSocket* out, SocketSlot* state = nullptr) {
+static bool GetSocketBackend(int guest_fd, NativeSocket* out, SocketSlot* state = nullptr,
+                             std::shared_ptr<SocketTransport>* transport = nullptr) {
 	EXIT_IF(out == nullptr);
 
 	if (guest_fd < 0 || guest_fd >= SOCKET_FD_MAX) {
@@ -1188,6 +1207,9 @@ static bool GetSocketBackend(int guest_fd, NativeSocket* out, SocketSlot* state
 	if (state != nullptr) {
 		*state = slot;
 	}
+	if (transport != nullptr) {
+		*transport = slot.transport;
+	}
 	return true;
 }
 
@@ -1234,6 +1256,63 @@ static int AllocSocketFd(const std::shared_ptr<SocketTransport>& transport, bool
 	return -1;
 }
 
+static int SelectSockets(int nfds, fd_set* readfds, fd_set* writefds, fd_set* exceptfds,
+                         timeval* timeout,
+                         const std::vector<std::shared_ptr<SocketTransport>>& stream_reads) {
+#if defined(_WIN32)
+	if (!stream_reads.empty()) {
+		const auto requested_read   = *readfds;
+		const auto requested_write  = writefds != nullptr ? *writefds : fd_set {};
+		const auto requested_except = exceptfds != nullptr ? *exceptfds : fd_set {};
+		std::optional<std::chrono::steady_clock::time_point> deadline;
+		if (timeout != nullptr) {
+			deadline = std::chrono::steady_clock::now() + std::chrono::seconds(timeout->tv_sec) +
+			           std::chrono::microseconds(timeout->tv_usec);
+		}
+		for (;;) {
+			*readfds = requested_read;
+			if (writefds != nullptr) {
+				*writefds = requested_write;
+			}
+			if (exceptfds != nullptr) {
+				*exceptfds = requested_except;
+			}
+			const bool ready = std::any_of(stream_reads.begin(), stream_reads.end(),
+			    [](const auto& transport) { return transport->HasPendingReceive(); });
+			// A concurrent peek can move native bytes into the guest buffer after this check.
+			// Bound native waits so both readiness sources are rechecked without holding receive locks.
+			auto wait = std::chrono::microseconds(ready ? 0 : 10'000);
+			if (deadline) {
+				const auto remaining = std::chrono::duration_cast<std::chrono::microseconds>(
+				    *deadline - std::chrono::steady_clock::now());
+				wait = std::min(wait, std::max(remaining, std::chrono::microseconds::zero()));
+			}
+			timeval slice {};
+			slice.tv_usec = static_cast<long>(wait.count());
+			int result = ::select(nfds, readfds, writefds, exceptfds, &slice);
+			if (result < 0) {
+				return result;
+			}
+			for (const auto& transport: stream_reads) {
+				const auto socket = transport->socket.load();
+				if (socket == INVALID_NATIVE_SOCKET) {
+					WSASetLastError(WSAENOTSOCK);
+					return SOCKET_ERROR;
+				}
+				if (transport->HasPendingReceive() && !FD_ISSET(socket, readfds)) {
+					FD_SET(socket, readfds);
+					result++;
+				}
+			}
+			if (result != 0 || (deadline && std::chrono::steady_clock::now() >= *deadline)) {
+				return result;
+			}
+		}
+	}
+#endif
+	return ::select(nfds, readfds, writefds, exceptfds, timeout);
+}
+
 static bool GuestFdIsSet(const void* fds, int fd) {
 	if (fds == nullptr || fd < 0 || fd >= SOCKET_FD_MAX) {
 		return false;
@@ -1343,6 +1422,12 @@ int KYTY_SYSV_ABI NetResolverDestroy(int rid) {
 	return OK;
 }
 
+int KYTY_SYSV_ABI NetResolverAbort(int rid, int flags) {
+	PRINT_NAME();
+	LOGF("\t rid = %d\n\t flags = 0x%08x\n", rid, flags);
+	return OK;
+}
+
 int KYTY_SYSV_ABI NetResolverStartNtoa(int rid, const char* hostname, void* addr, int timeout,
                                        int retry, int flags) {
 	PRINT_NAME();
@@ -1615,6 +1700,7 @@ int KYTY_SYSV_ABI EpollWait(int eid, NetEpollEvent* events, int maxevents, int t
 	FD_ZERO(&host_except);
 
 	std::vector<HostRegistration> host_registrations;
+	std::vector<std::shared_ptr<SocketTransport>> stream_reads;
 	host_registrations.reserve(registrations.size());
 	int host_nfds = 0;
 	for (const auto& registration: registrations) {
@@ -1635,6 +1721,11 @@ int KYTY_SYSV_ABI EpollWait(int eid, NetEpollEvent* events, int maxevents, int t
 #endif
 		if ((registration.event.events & EPOLL_IN) != 0) {
 			FD_SET(socket, &host_read);
+#if defined(_WIN32)
+			if (state.transport->type == SOCK_STREAM) {
+				stream_reads.push_back(state.transport);
+			}
+#endif
 		}
 		if ((registration.event.events & EPOLL_OUT) != 0) {
 			FD_SET(socket, &host_write);
@@ -1654,8 +1745,8 @@ int KYTY_SYSV_ABI EpollWait(int eid, NetEpollEvent* events, int maxevents, int t
 		host_timeout.tv_usec = timeout % 1000000;
 		host_timeout_ptr     = &host_timeout;
 	}
-
-	const int result = ::select(host_nfds, &host_read, &host_write, &host_except, host_timeout_ptr);
+	const int result = SelectSockets(host_nfds, &host_read, &host_write, &host_except,
+	                                 host_timeout_ptr, stream_reads);
 	if (result < 0) {
 		return SetHostSocketError();
 	}
@@ -1721,7 +1812,12 @@ int KYTY_SYSV_ABI SocketClose(int s) {
 	}
 	RemoveSocketFromEpolls(s);
 
-	if (slot.transport.use_count() == 1 && slot.transport->Close() != 0) {
+	bool close_transport = slot.transport.use_count() == 1;
+#if defined(_WIN32)
+	// Retained receive/readiness state must not defer native TCP close cancellation.
+	close_transport |= !slot.p2p && slot.transport->type == SOCK_STREAM;
+#endif
+	if (close_transport && slot.transport->Close() != 0) {
 		return NET_ERROR_EBADF;
 	}
 
@@ -1756,7 +1852,7 @@ int KYTY_SYSV_ABI Socket(int family, int type, int protocol) {
 		return SetHostSocketError();
 	}
 
-	auto transport = std::make_shared<SocketTransport>(socket);
+	auto transport = std::make_shared<SocketTransport>(socket, p2p ? SOCK_DGRAM : type);
 	if (p2p) {
 		// Logical sockets keep their own blocking mode; the shared transport never blocks.
 #if defined(_WIN32)
@@ -1939,7 +2035,7 @@ int KYTY_SYSV_ABI Accept(int s, void* addr, uint32_t* addrlen) {
 		return SetHostSocketError();
 	}
 
-	auto      transport = std::make_shared<SocketTransport>(accepted);
+	auto      transport = std::make_shared<SocketTransport>(accepted, SOCK_STREAM);
 	const int fd        = AllocSocketFd(transport);
 	if (fd < 0) {
 		*Posix::GetErrorAddr() = Posix::POSIX_EMFILE;
@@ -1968,8 +2064,9 @@ int KYTY_SYSV_ABI Shutdown(int s, int how) {
 	     "\t how = %d\n",
 	     s, how);
 
-	NativeSocket socket = INVALID_NATIVE_SOCKET;
-	if (!GetSocketBackend(s, &socket)) {
+	NativeSocket                    socket = INVALID_NATIVE_SOCKET;
+	std::shared_ptr<SocketTransport> transport;
+	if (!GetSocketBackend(s, &socket, nullptr, &transport)) {
 		return -1;
 	}
 
@@ -1982,6 +2079,9 @@ int KYTY_SYSV_ABI Shutdown(int s, int how) {
 	if (::shutdown(socket, how) == SOCKET_ERROR) {
 		return SetHostSocketError();
 	}
+	if (how == SD_RECEIVE || how == SD_BOTH) {
+		transport->receive_shutdown.store(true, std::memory_order_relaxed);
+	}
 
 	return 0;
 #else
@@ -2199,6 +2299,77 @@ int64_t KYTY_SYSV_ABI Sendto(int s, const void* buf, uint64_t len, int flags, co
 	return result;
 }
 
+#if defined(_WIN32)
+static int ReceiveStream(SocketTransport& transport, char* output, int length, int flags,
+                         bool dontwait) {
+	std::unique_lock lock(transport.receive_mutex, std::defer_lock);
+	if (dontwait) {
+		if (!lock.try_lock()) {
+			return SetGuestSocketError(Posix::POSIX_EWOULDBLOCK);
+		}
+	} else {
+		lock.lock();
+	}
+	if (transport.receive_shutdown.load(std::memory_order_relaxed)) {
+		return SetGuestSocketError(Posix::POSIX_ESHUTDOWN);
+	}
+	auto&           buffered = transport.receive_buffer;
+	const bool      peek     = (flags & MSG_PEEK) != 0;
+	const bool      waitall  = (flags & MSG_WAITALL) != 0;
+	int             copied   = static_cast<int>(std::min<size_t>(buffered.size(), length));
+	if (copied != 0) {
+		std::memcpy(output, buffered.data(), copied);
+		if (!peek) {
+			buffered.erase(buffered.begin(), buffered.begin() + copied);
+			transport.buffered_bytes.store(buffered.size(), std::memory_order_relaxed);
+		}
+	}
+	if (copied == length || (copied != 0 && !waitall)) {
+		return copied;
+	}
+
+	// Winsock rejects PEEK | WAITALL. Consume into transport-owned storage when peeking,
+	// and let ordinary recv report EOF, timeouts, or nonblocking partial progress.
+	flags &= ~(MSG_PEEK | MSG_WAITALL);
+	do {
+		if (transport.receive_shutdown.load(std::memory_order_relaxed)) {
+			return copied != 0 ? copied : SetGuestSocketError(Posix::POSIX_ESHUTDOWN);
+		}
+		const auto socket = transport.socket.load();
+		if (dontwait) {
+			fd_set readable {};
+			FD_SET(socket, &readable);
+			timeval immediate {};
+			const int ready = ::select(0, &readable, nullptr, nullptr, &immediate);
+			if (ready <= 0) {
+				return copied != 0 ? copied : (ready < 0 ? SetHostSocketError() :
+				    SetGuestSocketError(Posix::POSIX_EWOULDBLOCK));
+			}
+		}
+		const auto previous_size = buffered.size();
+		const int  requested     = peek ? std::min(length - copied, 64 * 1024) : length - copied;
+		char*      destination   = output + copied;
+		if (peek) {
+			buffered.resize(previous_size + requested);
+			destination = buffered.data() + previous_size;
+		}
+		const int received = ::recv(socket, destination, requested, flags);
+		if (peek) {
+			if (received > 0) {
+				std::memcpy(output + copied, destination, received);
+			}
+			buffered.resize(previous_size + std::max(received, 0));
+			transport.buffered_bytes.store(buffered.size(), std::memory_order_relaxed);
+		}
+		if (received <= 0) {
+			return copied != 0 || received == 0 ? copied : SetHostSocketError();
+		}
+		copied += received;
+	} while (waitall && copied < length);
+	return copied;
+}
+#endif
+
 int64_t KYTY_SYSV_ABI Recv(int s, void* buf, uint64_t len, int flags) {
 	return Recvfrom(s, buf, len, flags, nullptr, nullptr);
 }
@@ -2221,8 +2392,9 @@ int64_t KYTY_SYSV_ABI Recvfrom(int s, void* buf, uint64_t len, int flags, void*
 		return -1;
 	}
 
-	NativeSocket socket = INVALID_NATIVE_SOCKET;
-	if (!GetSocketBackend(s, &socket)) {
+	NativeSocket                    socket = INVALID_NATIVE_SOCKET;
+	std::shared_ptr<SocketTransport> transport;
+	if (!GetSocketBackend(s, &socket, nullptr, &transport)) {
 		return -1;
 	}
 
@@ -2236,11 +2408,27 @@ int64_t KYTY_SYSV_ABI Recvfrom(int s, void* buf, uint64_t len, int flags, void*
 	sockaddr_storage host_addr {};
 	SocketLength     host_addrlen = sizeof(host_addr);
 	int64_t          result       = 0;
-	if (addr == nullptr) {
-		result = ::recv(socket, static_cast<char*>(buf), host_len, host_flags);
-	} else {
-		result = ::recvfrom(socket, static_cast<char*>(buf), host_len, host_flags,
-		                    reinterpret_cast<sockaddr*>(&host_addr), &host_addrlen);
+#if defined(_WIN32)
+	if (transport->type == SOCK_STREAM) {
+		if (addr != nullptr &&
+		    ::getpeername(socket, reinterpret_cast<sockaddr*>(&host_addr), &host_addrlen) != 0) {
+			return SetHostSocketError();
+		}
+		result = ReceiveStream(*transport, static_cast<char*>(buf), host_len, host_flags,
+		                       (flags & 0x80) != 0);
+		if (result < 0) {
+			return result;
+		}
+	} else
+#endif
+	{
+		transport.reset();
+		if (addr == nullptr) {
+			result = ::recv(socket, static_cast<char*>(buf), host_len, host_flags);
+		} else {
+			result = ::recvfrom(socket, static_cast<char*>(buf), host_len, host_flags,
+			                    reinterpret_cast<sockaddr*>(&host_addr), &host_addrlen);
+		}
 	}
 #if defined(_WIN32)
 	if (result < 0 && WSAGetLastError() == WSAEMSGSIZE) {
@@ -2297,6 +2485,7 @@ int KYTY_SYSV_ABI Select(int nfds, void* readfds, void* writefds, void* exceptfd
 	FD_ZERO(&host_except);
 
 	std::vector<std::pair<int, NativeSocket>> descriptors;
+	std::vector<std::shared_ptr<SocketTransport>> stream_reads;
 	int host_nfds = 0;
 	for (int fd = 0; fd < nfds; fd++) {
 		const bool read = GuestFdIsSet(readfds, fd);
@@ -2305,13 +2494,14 @@ int KYTY_SYSV_ABI Select(int nfds, void* readfds, void* writefds, void* exceptfd
 		if (!read && !write && !except) {
 			continue;
 		}
-		NativeSocket socket = INVALID_NATIVE_SOCKET;
+		NativeSocket                    socket = INVALID_NATIVE_SOCKET;
+		std::shared_ptr<SocketTransport> transport;
 #if !defined(_WIN32)
 		if (fd < 3) {
 			socket = fd;
 		} else
 #endif
-		if (!GetSocketBackend(fd, &socket)) {
+		if (!GetSocketBackend(fd, &socket, nullptr, &transport)) {
 			return -1;
 		}
 #if defined(_WIN32)
@@ -2329,6 +2519,11 @@ int KYTY_SYSV_ABI Select(int nfds, void* readfds, void* writefds, void* exceptfd
 		descriptors.emplace_back(fd, socket);
 		if (read) {
 			FD_SET(socket, &host_read);
+#if defined(_WIN32)
+			if (transport->type == SOCK_STREAM) {
+				stream_reads.push_back(std::move(transport));
+			}
+#endif
 		}
 		if (write) {
 			FD_SET(socket, &host_write);
@@ -2359,9 +2554,10 @@ int KYTY_SYSV_ABI Select(int nfds, void* readfds, void* writefds, void* exceptfd
 	} else
 #endif
 	{
-		result = ::select(host_nfds, readfds != nullptr ? &host_read : nullptr,
-		                  writefds != nullptr ? &host_write : nullptr,
-		                  exceptfds != nullptr ? &host_except : nullptr, host_timeout_ptr);
+		result = SelectSockets(host_nfds, readfds != nullptr ? &host_read : nullptr,
+		                       writefds != nullptr ? &host_write : nullptr,
+		                       exceptfds != nullptr ? &host_except : nullptr, host_timeout_ptr,
+		                       stream_reads);
 		if (result < 0) {
 			return SetHostSocketError();
 		}
@@ -2370,15 +2566,19 @@ int KYTY_SYSV_ABI Select(int nfds, void* readfds, void* writefds, void* exceptfd
 	GuestFdZero(readfds, nfds);
 	GuestFdZero(writefds, nfds);
 	GuestFdZero(exceptfds, nfds);
+	result = 0;
 	for (const auto& [fd, socket]: descriptors) {
 		if (FD_ISSET(socket, &host_read)) {
 			GuestFdSet(readfds, fd);
+			result++;
 		}
 		if (FD_ISSET(socket, &host_write)) {
 			GuestFdSet(writefds, fd);
+			result++;
 		}
 		if (FD_ISSET(socket, &host_except)) {
 			GuestFdSet(exceptfds, fd);
+			result++;
 		}
 	}
 	return result;

src/libs/network.h

@@ -41,6 +41,7 @@ int KYTY_SYSV_ABI NetPoolCreate(const char* name, int size, int flags);
 int KYTY_SYSV_ABI NetPoolDestroy(int memid);
 int KYTY_SYSV_ABI NetResolverCreate(const char* name, int memid, int flags);
 int KYTY_SYSV_ABI NetResolverDestroy(int rid);
+int KYTY_SYSV_ABI NetResolverAbort(int rid, int flags);
 int KYTY_SYSV_ABI NetResolverStartNtoa(int rid, const char* hostname, void* addr, int timeout,
                                        int retry, int flags);
 int KYTY_SYSV_ABI NetInetPton(int af, const char* src, void* dst);

src/loader/guestInstructionPatcher.cpp

@@ -1,12 +1,11 @@
 // SPDX-FileCopyrightText: Copyright 2026 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
-#include "loader/redZonePatcher.h"
+#include "loader/guestInstructionPatcher.h"
 
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "common/virtualMemory.h"
-#include "loader/x64InstructionEmulator.h"
 
 #include <Zydis/Zydis.h>
 #include <algorithm>
@@ -22,7 +21,7 @@
 #include <set>
 #include <unordered_set>
 #include <vector>
-#if defined(_WIN32)
+#if !defined(__APPLE__)
 #include <xbyak/xbyak.h>
 #include <xbyak/xbyak_util.h>
 #endif
@@ -34,7 +33,7 @@
 #undef max
 #endif
 
-#if defined(_WIN32)
+#if !defined(__APPLE__)
 using namespace Xbyak::util;
 #endif
 
@@ -53,7 +52,7 @@ using s64 = int64_t;
 
 constexpr size_t NearJumpSize = 5;
 
-#if defined(_WIN32)
+#if !defined(__APPLE__)
 
 struct PatchModule {
 	std::mutex           mutex {};
@@ -62,7 +61,7 @@ struct PatchModule {
 	std::set<u8*>        patched;
 	Xbyak::CodeGenerator patch_gen;
 	Xbyak::CodeGenerator trampoline_gen;
-	bool                 trampoline_exhausted = false;
+	bool                 trampoline_exhaustion_reported = false;
 
 	PatchModule(u8* module_ptr, u64 module_size, u8* trampoline_ptr, u64 trampoline_size)
 	    : start(module_ptr), end(module_ptr + module_size), patch_gen(module_size, module_ptr),
@@ -81,14 +80,14 @@ static PatchModule* GetContainingModule(const void* ptr) {
 	return address >= module->start && address < module->end ? module : nullptr;
 }
 
-static bool HandleTrampolineError(PatchModule* module, const Xbyak::Error& error) {
-	if (static_cast<int>(error) != Xbyak::ERR_CODE_IS_TOO_BIG) {
+static bool HandleTrampolineError(PatchModule* module, int error) {
+	if (error != Xbyak::ERR_CODE_IS_TOO_BIG) {
 		return false;
 	}
-	if (!module->trampoline_exhausted) {
-		LOGF("Windows guest red-zone trampoline space exhausted for module %p\n",
+	if (!module->trampoline_exhaustion_reported) {
+		LOGF("Guest instruction trampoline space exhausted for module %p\n",
 		     static_cast<void*>(module->start));
-		module->trampoline_exhausted = true;
+		module->trampoline_exhaustion_reported = true;
 	}
 	return true;
 }
@@ -103,9 +102,6 @@ static ZydisDecoder& GetDecoder() {
 	return decoder;
 }
 
-#endif
-#if defined(_WIN32)
-
 namespace {
 
 constexpr size_t GuestRedZoneSize = 128;
@@ -136,11 +132,39 @@ struct DecodedFunction {
 	bool                                        requires_conservative_red_zone_tracking {};
 };
 
-struct InstructionRewrite {
-	bool protect_red_zone {};
-	bool protected_indirect_call {};
+enum class InstructionReplacement {
+	None,
+	ReciprocalSquareRoot,
+	ExtractQ,
+	InsertQ,
+	ReadProcessorId,
+	CacheLineWriteBack
 };
 
+struct InstructionRewrite {
+	bool                   protect_red_zone {};
+	bool                   protected_indirect_call {};
+	InstructionReplacement replacement {};
+};
+
+InstructionPatchCounts& ReplacementCounts(GuestInstructionPatchResult& result,
+                                          InstructionReplacement       replacement) {
+	ASSERT(replacement != InstructionReplacement::None);
+	switch (replacement) {
+		case InstructionReplacement::ReciprocalSquareRoot: return result.reciprocal_sqrt;
+		case InstructionReplacement::ExtractQ: return result.extrq;
+		case InstructionReplacement::InsertQ: return result.insertq;
+		case InstructionReplacement::ReadProcessorId: return result.rdpid;
+		default: return result.clwb;
+	}
+}
+
+bool UsesInstructionTrap(InstructionReplacement replacement, bool trap_replacements) {
+	return replacement != InstructionReplacement::None &&
+	       (trap_replacements || replacement == InstructionReplacement::InsertQ ||
+	        replacement == InstructionReplacement::ReadProcessorId);
+}
+
 bool IsStackPointerRegister(ZydisRegister reg) {
 	return reg != ZYDIS_REGISTER_NONE &&
 	       ZydisRegisterGetLargestEnclosing(ZYDIS_MACHINE_MODE_LONG_64, reg) == ZYDIS_REGISTER_RSP;
@@ -210,10 +234,11 @@ DecodedCodeInstruction DecodeCodeInstruction(uintptr_t address, uintptr_t end) {
 
 	if (decoded.changes_stack_pointer) {
 		const auto& operands = decoded.operands;
-		if (decoded.instruction.mnemonic == ZYDIS_MNEMONIC_MOV &&
-		    operands[0].type == ZYDIS_OPERAND_TYPE_REGISTER &&
+		if (operands[0].type == ZYDIS_OPERAND_TYPE_REGISTER &&
 		    IsStackPointerRegister(operands[0].reg.value) &&
-		    operands[1].type == ZYDIS_OPERAND_TYPE_MEMORY) {
+		    (decoded.instruction.mnemonic == ZYDIS_MNEMONIC_RDPID ||
+		     (decoded.instruction.mnemonic == ZYDIS_MNEMONIC_MOV &&
+		      operands[1].type == ZYDIS_OPERAND_TYPE_MEMORY))) {
 			decoded.replaces_stack_pointer = true;
 		} else if ((decoded.instruction.mnemonic == ZYDIS_MNEMONIC_ADD ||
 		            decoded.instruction.mnemonic == ZYDIS_MNEMONIC_SUB) &&
@@ -266,7 +291,7 @@ DecodedCodeInstruction DecodeCodeInstruction(uintptr_t address, uintptr_t end) {
 		const s64 access_start = operand.mem.disp.value;
 		const s64 access_size  = std::max<s64>(operand.size / 8, 1);
 		const s64 range_start  = std::max(access_start, -static_cast<s64>(GuestRedZoneSize));
-		const s64 range_end    = std::min(access_start + access_size, 0LL);
+		const s64 range_end    = std::min<s64>(access_start + access_size, 0);
 		for (s64 offset = range_start; offset < range_end; ++offset) {
 			const size_t bit = static_cast<size_t>(offset + static_cast<s64>(GuestRedZoneSize));
 			if ((operand.actions & ZYDIS_OPERAND_ACTION_MASK_READ) != 0) {
@@ -565,7 +590,10 @@ DecodedFunction DecodeFunction(uintptr_t function_start, uintptr_t function_end,
 	return function;
 }
 
-RedZoneMask TranslateRedZoneMask(const RedZoneMask& mask, s64 stack_pointer_delta) {
+std::optional<RedZoneMask> TranslateRedZoneMask(const RedZoneMask& mask, s64 stack_pointer_delta) {
+	if (stack_pointer_delta == 0) {
+		return mask;
+	}
 	RedZoneMask translated;
 	for (size_t bit = 0; bit < GuestRedZoneSize; ++bit) {
 		if (!mask.test(bit)) {
@@ -573,9 +601,12 @@ RedZoneMask TranslateRedZoneMask(const RedZoneMask& mask, s64 stack_pointer_delt
 		}
 		const s64 after_offset  = static_cast<s64>(bit) - static_cast<s64>(GuestRedZoneSize);
 		const s64 before_offset = after_offset + stack_pointer_delta;
-		if (before_offset >= -static_cast<s64>(GuestRedZoneSize) && before_offset < 0) {
-			translated.set(static_cast<size_t>(before_offset + static_cast<s64>(GuestRedZoneSize)));
+		if (before_offset < -static_cast<s64>(GuestRedZoneSize) || before_offset >= 0) {
+			// A later inverse RSP adjustment can bring this byte back into the red zone.
+			// Losing its stack-slot identity makes the bounded analysis inconclusive.
+			return std::nullopt;
 		}
+		translated.set(static_cast<size_t>(before_offset + static_cast<s64>(GuestRedZoneSize)));
 	}
 	return translated;
 }
@@ -585,13 +616,16 @@ void AnalyzeRedZoneLiveness(DecodedFunction& function) {
 		return;
 	}
 
+	const auto protect_function = [&] {
+		for (auto& [_, decoded]: function.instructions) {
+			decoded.red_zone_live.set();
+		}
+	};
 	if (function.has_indirect_branch || function.requires_conservative_red_zone_tracking ||
 	    std::ranges::any_of(function.branch_targets, [&function](uintptr_t target) {
 		    return !function.instructions.contains(target);
 	    })) {
-		for (auto& [_, decoded]: function.instructions) {
-			decoded.red_zone_live.set();
-		}
+		protect_function();
 		return;
 	}
 
@@ -628,12 +662,14 @@ void AnalyzeRedZoneLiveness(DecodedFunction& function) {
 				add_successor(next_address);
 			}
 
-			const RedZoneMask translated_live_out =
-			    decoded.stack_pointer_delta.has_value()
-			        ? TranslateRedZoneMask(live_out, *decoded.stack_pointer_delta)
-			        : live_out;
+			const auto translated_live_out =
+			    TranslateRedZoneMask(live_out, decoded.stack_pointer_delta.value_or(0));
+			if (!translated_live_out) {
+				protect_function();
+				return;
+			}
 			const RedZoneMask new_live_in =
-			    decoded.red_zone_use | (translated_live_out & ~decoded.red_zone_def);
+			    decoded.red_zone_use | (*translated_live_out & ~decoded.red_zone_def);
 			if (new_live_in != live_in[reverse_index]) {
 				live_in[reverse_index] = new_live_in;
 				changed                = true;
@@ -647,13 +683,17 @@ void AnalyzeRedZoneLiveness(DecodedFunction& function) {
 }
 
 bool EncodeRelocatedInstruction(const DecodedCodeInstruction& decoded,
-                                Xbyak::CodeGenerator&         generator) {
+                                Xbyak::CodeGenerator& generator, s64 stack_adjustment = 0,
+                                ZydisMnemonic replacement = ZYDIS_MNEMONIC_INVALID) {
 	ZydisEncoderRequest request;
 	if (!ZYAN_SUCCESS(ZydisEncoderDecodedInstructionToEncoderRequest(
 	        &decoded.instruction, decoded.operands.data(),
 	        decoded.instruction.operand_count_visible, &request))) {
 		return false;
 	}
+	if (replacement != ZYDIS_MNEMONIC_INVALID) {
+		request.mnemonic = replacement;
+	}
 
 	for (u8 index = 0; index < decoded.instruction.operand_count_visible; ++index) {
 		const auto& operand = decoded.operands[index];
@@ -673,6 +713,9 @@ bool EncodeRelocatedInstruction(const DecodedCodeInstruction& decoded,
 				return false;
 			}
 			request.operands[index].mem.displacement = static_cast<ZyanI64>(absolute_address);
+		} else if (operand.type == ZYDIS_OPERAND_TYPE_MEMORY &&
+		           IsStackPointerRegister(operand.mem.base)) {
+			request.operands[index].mem.displacement += stack_adjustment;
 		}
 	}
 
@@ -738,9 +781,9 @@ bool GenerateProtectedIndirectCall(const DecodedCodeInstruction& decoded,
 	return true;
 }
 
-void CollectRedZoneMemoryInstructions(const DecodedFunction& function,
+void CollectRedZoneMemoryInstructions(const DecodedFunction&                   function,
                                       std::map<uintptr_t, InstructionRewrite>& rewrite_sites,
-                                      RedZonePatchResult& result) {
+                                      GuestInstructionPatchResult&             result) {
 	if (!function.uses_red_zone) {
 		return;
 	}
@@ -777,98 +820,284 @@ void CollectRedZoneMemoryInstructions(const DecodedFunction& function,
 	}
 }
 
-struct ReciprocalSquareRootSite {
-	uintptr_t address;
-	u8        length;
-	bool      requires_red_zone_protection;
-};
+// Compute 1/sqrt in double precision, rounded to float, with denormals treated as
+// signed zero. Preserve guest MXCSR, RFLAGS and the scratch register's full YMM.
+void GenerateReciprocalSquareRoot(const DecodedCodeInstruction& decoded,
+                                  Xbyak::CodeGenerator&         generator) {
+	const int destination   = decoded.operands[0].reg.value - ZYDIS_REGISTER_XMM0;
+	const int source        = decoded.operands[1].reg.value - ZYDIS_REGISTER_XMM0;
+	int       scratch_index = 0;
+	while (scratch_index == destination || scratch_index == source) {
+		++scratch_index;
+	}
+	const Xbyak::Ymm scratch(scratch_index);
+	Xbyak::Label     one;
+	Xbyak::Label     done;
+	constexpr size_t SpillSize = 48;
 
-void CollectReciprocalSquareRoots(const DecodedFunction& function,
-                                 std::map<uintptr_t, InstructionRewrite>& rewrite_sites,
-                                 std::vector<ReciprocalSquareRootSite>& sites) {
+	// Guest code uses the SysV red zone on both hosts; put our spills below it.
+	// LEA/MOV/SIMD leave RFLAGS intact, including when the guest red zone is live.
+	generator.lea(rsp, ptr[rsp - GuestRedZoneSize - SpillSize]);
+	generator.vmovdqu(ptr[rsp], scratch);
+	generator.stmxcsr(ptr[rsp + 32]);
+	generator.mov(dword[rsp + 36], 0x1fc0); // round to nearest, exceptions masked, DAZ
+	generator.ldmxcsr(ptr[rsp + 36]);
+	generator.vcvtps2pd(scratch, Xbyak::Xmm(source));
+	generator.vsqrtpd(scratch, scratch);
+	generator.vbroadcastsd(Xbyak::Ymm(destination), ptr[rip + one]);
+	generator.vdivpd(scratch, Xbyak::Ymm(destination), scratch);
+	generator.vcvtpd2ps(Xbyak::Xmm(destination), scratch); // Clears upper YMM lanes.
+	generator.ldmxcsr(ptr[rsp + 32]);
+	generator.vmovdqu(scratch, ptr[rsp]);
+	generator.lea(rsp, ptr[rsp + GuestRedZoneSize + SpillSize]);
+	generator.jmp(done);
+	// A failed label reference can have an incomplete displacement. Do not bind it.
+	if (Xbyak::GetError() != 0) {
+		return;
+	}
+	generator.L(one);
+	generator.dq(0x3ff0000000000000ULL);
+	generator.L(done);
+}
+
+bool IsSupportedBitFieldInstruction(const DecodedCodeInstruction& decoded) {
+	const bool insert = decoded.instruction.mnemonic == ZYDIS_MNEMONIC_INSERTQ;
+	if (!insert && decoded.instruction.mnemonic != ZYDIS_MNEMONIC_EXTRQ) {
+		return false;
+	}
+	const auto is_xmm = [](const ZydisDecodedOperand& operand) {
+		return operand.type == ZYDIS_OPERAND_TYPE_REGISTER &&
+		       operand.reg.value >= ZYDIS_REGISTER_XMM0 &&
+		       operand.reg.value <= ZYDIS_REGISTER_XMM15;
+	};
+	const size_t immediate_index = insert ? 2 : 1;
+	const bool   immediate =
+	    decoded.instruction.operand_count_visible == immediate_index + 2 &&
+	    decoded.operands[immediate_index].type == ZYDIS_OPERAND_TYPE_IMMEDIATE &&
+	    decoded.operands[immediate_index + 1].type == ZYDIS_OPERAND_TYPE_IMMEDIATE;
+	if (!is_xmm(decoded.operands[0]) || (insert && !is_xmm(decoded.operands[1])) ||
+	    (!immediate &&
+	     !(decoded.instruction.operand_count_visible == 2 && is_xmm(decoded.operands[1])))) {
+		return false;
+	}
+	// Match the trap handler's encoding support: 66/F2 [REX] 0F 78/79 ModRM [imm8 imm8].
+	const auto*  code   = reinterpret_cast<const u8*>(decoded.address);
+	const size_t opcode = (code[1] & 0xf0u) == 0x40u ? 2 : 1;
+	return code[0] == (insert ? 0xf2 : 0x66) && code[opcode] == 0x0f &&
+	       code[opcode + 1] == (immediate ? 0x78 : 0x79) &&
+	       decoded.instruction.length == opcode + (immediate ? 5 : 3);
+}
+
+void GenerateExtractQ(const DecodedCodeInstruction& decoded, Xbyak::CodeGenerator& generator) {
+	const Xbyak::Xmm destination(decoded.operands[0].reg.value - ZYDIS_REGISTER_XMM0);
+	// Keep spills below the guest red zone on both hosts, and restore all modified flags/GPRs.
+	generator.lea(rsp, ptr[rsp - GuestRedZoneSize]);
+	generator.pushfq();
+	generator.push(rax);
+	generator.push(rcx);
+	generator.push(rdx);
+	if (decoded.operands[1].type == ZYDIS_OPERAND_TYPE_IMMEDIATE) {
+		const u32 length = decoded.operands[1].imm.value.u & 63u;
+		const u32 index  = decoded.operands[2].imm.value.u & 63u;
+		const u64 mask   = length == 0 ? UINT64_MAX : (u64 {1} << length) - 1;
+		generator.movq(rax, destination);
+		if (index != 0) {
+			generator.shr(rax, index);
+		}
+		generator.mov(rdx, mask);
+	} else {
+		const Xbyak::Xmm source(decoded.operands[1].reg.value - ZYDIS_REGISTER_XMM0);
+		generator.movq(rax, source);
+		generator.mov(ecx, eax);
+		generator.neg(ecx);
+		generator.mov(rdx, UINT64_MAX);
+		// CL shifts use six bits: -length gives 64-length, with zero retaining all 64 bits.
+		generator.shr(rdx, cl);
+		generator.shr(rax, 8);
+		generator.mov(ecx, eax);
+		generator.movq(rax, destination);
+		generator.shr(rax, cl);
+	}
+	generator.and_(rax, rdx);
+	// Match the emulator: clear the upper XMM half, preserving the upper YMM half.
+	generator.movq(destination, rax);
+	generator.pop(rdx);
+	generator.pop(rcx);
+	generator.pop(rax);
+	generator.popfq();
+	generator.lea(rsp, ptr[rsp + GuestRedZoneSize]);
+}
+
+void CollectAmdInstructions(const DecodedFunction&                   function,
+                            std::map<uintptr_t, InstructionRewrite>& rewrite_sites,
+                            GuestInstructionPatchResult&             result,
+                            GuestInstructionHostFeatures             host_features) {
 	for (const auto& [address, decoded]: function.instructions) {
-		if (!X64InstructionEmulator::IsReciprocalSquareRoot(decoded.instruction,
-		                                                    decoded.operands.data())) {
+		const auto&            instruction = decoded.instruction;
+		InstructionReplacement replacement {};
+		if (instruction.mnemonic == ZYDIS_MNEMONIC_VRSQRTPS &&
+		    instruction.encoding == ZYDIS_INSTRUCTION_ENCODING_VEX &&
+		    instruction.raw.vex.offset == 0 && decoded.operands[0].size == 128 &&
+		    decoded.operands[1].type == ZYDIS_OPERAND_TYPE_REGISTER) {
+			replacement = InstructionReplacement::ReciprocalSquareRoot;
+		} else if (!host_features.sse4a && IsSupportedBitFieldInstruction(decoded)) {
+			replacement = instruction.mnemonic == ZYDIS_MNEMONIC_EXTRQ
+			                  ? InstructionReplacement::ExtractQ
+			                  : InstructionReplacement::InsertQ;
+		} else if (!host_features.rdpid && instruction.mnemonic == ZYDIS_MNEMONIC_RDPID) {
+			replacement = InstructionReplacement::ReadProcessorId;
+		} else if (!host_features.clwb && instruction.mnemonic == ZYDIS_MNEMONIC_CLWB) {
+			replacement = InstructionReplacement::CacheLineWriteBack;
+		} else {
 			continue;
 		}
-		const bool protect_red_zone = decoded.red_zone_live.any();
-		sites.push_back({address, decoded.instruction.length, protect_red_zone});
-		if (protect_red_zone) {
-			rewrite_sites[address].protect_red_zone = true;
-		}
+		++ReplacementCounts(result, replacement).found;
+		rewrite_sites[address].replacement = replacement;
 	}
 }
 
-uint64_t ApplyReciprocalSquareRootPatches(const PatchModule& module,
-                                        std::span<const ReciprocalSquareRootSite> sites,
-                                        uint64_t trampoline_addr, uint64_t trampoline_size) {
-	// Validate every required relocation before introducing any traps.
-	for (const auto& site: sites) {
-		if (site.requires_red_zone_protection &&
-		    !module.patched.contains(reinterpret_cast<u8*>(site.address))) {
-			EXIT("Cannot preserve the guest red zone at emulated VRSQRTPS 0x%016" PRIx64 "\n",
-			     static_cast<u64>(site.address));
-		}
+void MarkInstructionTrap(u8* code, const ZydisDecodedInstruction& instruction,
+                         InstructionReplacement replacement) {
+	if (replacement == InstructionReplacement::ReciprocalSquareRoot) {
+		// Clear a reserved VEX.vvvv bit, retaining both register operands for the handler.
+		code[instruction.raw.vex.size - 1] &= ~0x08u;
 	}
-
-	uint64_t patched = 0;
-	for (const auto& site: sites) {
-		if (!module.patched.contains(reinterpret_cast<u8*>(site.address))) {
-			patched += X64InstructionEmulator::PatchReciprocalSquareRoots(site.address, site.length);
-		}
-	}
-	return patched +
-	       X64InstructionEmulator::PatchReciprocalSquareRoots(trampoline_addr, trampoline_size);
+	// Other replacements already raise #UD on hosts without their CPU feature.
 }
 
-void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& function,
-                                const std::map<uintptr_t, InstructionRewrite>& rewrite_sites,
-                                RedZonePatchResult& result) {
+bool GenerateInstructionTrap(const DecodedCodeInstruction& decoded,
+                             InstructionReplacement replacement, Xbyak::CodeGenerator& generator,
+                             s64 stack_adjustment) {
+	if (replacement == InstructionReplacement::CacheLineWriteBack) {
+		return EncodeRelocatedInstruction(decoded, generator, stack_adjustment);
+	}
+	std::array<u8, ZYDIS_MAX_INSTRUCTION_LENGTH> code {};
+	std::memcpy(code.data(), reinterpret_cast<const void*>(decoded.address),
+	            decoded.instruction.length);
+	MarkInstructionTrap(code.data(), decoded.instruction, replacement);
+	generator.db(code.data(), decoded.instruction.length);
+	return true;
+}
+
+bool NeedsTrapRedZoneProtection(const DecodedCodeInstruction& decoded) {
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+	return decoded.red_zone_live.any();
+#else
+	return false;
+#endif
+}
+
+void TrapUnrelocatedInstructions(const PatchModule& module, const DecodedFunction& function,
+                                 const std::map<uintptr_t, InstructionRewrite>& rewrite_sites,
+                                 GuestInstructionPatchResult&                   result) {
+	for (const auto& [address, rewrite]: rewrite_sites) {
+		if (rewrite.replacement == InstructionReplacement::None ||
+		    module.patched.contains(reinterpret_cast<u8*>(address))) {
+			continue;
+		}
+		const auto& decoded = function.instructions.at(address);
+		if (NeedsTrapRedZoneProtection(decoded)) {
+			if (rewrite.replacement != InstructionReplacement::ReciprocalSquareRoot) {
+				// Leaving an unsupported instruction unchanged still traps with a live red zone.
+				EXIT("AMD CPU compatibility: cannot safely trap %s at %p (guest red zone is "
+				     "live)\n",
+				     ZydisMnemonicGetString(decoded.instruction.mnemonic),
+				     reinterpret_cast<void*>(address));
+			}
+			continue;
+		}
+		MarkInstructionTrap(reinterpret_cast<u8*>(address), decoded.instruction,
+		                    rewrite.replacement);
+		++ReplacementCounts(result, rewrite.replacement).trapped;
+	}
+}
+
+void RelocateGuestInstructions(PatchModule* module, const DecodedFunction& function,
+                               const std::map<uintptr_t, InstructionRewrite>& rewrite_sites,
+                               GuestInstructionPatchResult&                   result) {
 	struct RelocationSpan {
 		std::vector<const DecodedCodeInstruction*> instructions;
 		uintptr_t                                  patch_start {};
 		uintptr_t                                  continuation {};
 		size_t                                     patch_size {};
+		bool                                       trap_replacements {};
 	};
 
-	const auto emit_span = [&](const RelocationSpan& span) -> std::optional<size_t> {
-		const size_t trampoline_offset = module->trampoline_gen.getSize();
-		if (module->trampoline_exhausted) {
-			return std::nullopt;
-		}
-		try {
+	const auto emit_span = [&](RelocationSpan& span) -> std::optional<size_t> {
+		auto&        generator         = module->trampoline_gen;
+		const size_t trampoline_offset = generator.getSize();
+		const bool   has_replacements =
+		    std::ranges::any_of(span.instructions, [&](const auto* decoded) {
+			    const auto rewrite = rewrite_sites.find(decoded->address);
+			    return rewrite != rewrite_sites.end() &&
+			           rewrite->second.replacement != InstructionReplacement::None;
+		    });
+		for (bool trap_replacements: {false, true}) {
+			span.trap_replacements = trap_replacements;
+			Xbyak::ClearError();
+			bool encoded = true;
 			for (const auto* decoded: span.instructions) {
-				const auto rewrite = rewrite_sites.find(decoded->address);
+				const auto rewrite     = rewrite_sites.find(decoded->address);
+				const auto replacement = rewrite != rewrite_sites.end()
+				                             ? rewrite->second.replacement
+				                             : InstructionReplacement::None;
+				const bool uses_trap   = UsesInstructionTrap(replacement, trap_replacements);
 				const bool protected_indirect_call =
 				    rewrite != rewrite_sites.end() && rewrite->second.protected_indirect_call;
-				const bool protect_red_zone = rewrite != rewrite_sites.end() &&
-				                              rewrite->second.protect_red_zone &&
-				                              !protected_indirect_call;
+				const bool protect_red_zone =
+				    (rewrite != rewrite_sites.end() && rewrite->second.protect_red_zone &&
+				     !protected_indirect_call) ||
+				    ((uses_trap || replacement == InstructionReplacement::CacheLineWriteBack) &&
+				     NeedsTrapRedZoneProtection(*decoded));
 				if (protect_red_zone) {
-					module->trampoline_gen.lea(rsp, ptr[rsp - GuestRedZoneSize]);
+					generator.lea(rsp, ptr[rsp - GuestRedZoneSize]);
 				}
-				if (protected_indirect_call) {
-					if (!GenerateProtectedIndirectCall(*decoded, module->trampoline_gen)) {
-						module->trampoline_gen.setSize(trampoline_offset);
-						return std::nullopt;
+				if (replacement != InstructionReplacement::None) {
+					if (uses_trap) {
+						encoded = GenerateInstructionTrap(*decoded, replacement, generator,
+						                                  protect_red_zone ? GuestRedZoneSize : 0);
+					} else if (replacement == InstructionReplacement::ReciprocalSquareRoot) {
+						GenerateReciprocalSquareRoot(*decoded, generator);
+					} else if (replacement == InstructionReplacement::ExtractQ) {
+						GenerateExtractQ(*decoded, generator);
+					} else {
+						// CLFLUSH writes back dirty data too; invalidation is a permitted stronger
+						// action.
+						encoded = EncodeRelocatedInstruction(
+						    *decoded, generator, protect_red_zone ? GuestRedZoneSize : 0,
+						    ZYDIS_MNEMONIC_CLFLUSH);
 					}
-				} else if (!EncodeRelocatedInstruction(*decoded, module->trampoline_gen)) {
-					module->trampoline_gen.setSize(trampoline_offset);
-					return std::nullopt;
+				} else if (protected_indirect_call) {
+					encoded = GenerateProtectedIndirectCall(*decoded, generator);
+				} else {
+					encoded = EncodeRelocatedInstruction(*decoded, generator);
 				}
 				if (protect_red_zone && !decoded->replaces_stack_pointer) {
-					module->trampoline_gen.lea(rsp, ptr[rsp + GuestRedZoneSize]);
+					generator.lea(rsp, ptr[rsp + GuestRedZoneSize]);
+				}
+				if (!encoded || Xbyak::GetError() != 0) {
+					break;
 				}
 			}
-			module->trampoline_gen.jmp(reinterpret_cast<void*>(span.continuation));
-		} catch (const Xbyak::Error& error) {
-			module->trampoline_gen.setSize(trampoline_offset);
-			if (HandleTrampolineError(module, error)) {
-				return std::nullopt;
+			if (encoded && Xbyak::GetError() == 0) {
+				generator.jmp(reinterpret_cast<void*>(span.continuation));
+				if (Xbyak::GetError() == 0) {
+					return trampoline_offset;
+				}
+			}
+			const int error = Xbyak::GetError();
+			// Roll back before retrying this uncommitted span with smaller trap replacements.
+			generator.reset();
+			generator.setSize(trampoline_offset);
+			if (error != 0 && !HandleTrampolineError(module, error)) {
+				EXIT("Guest instruction trampoline generation failed: %s\n",
+				     Xbyak::ConvertErrorToString(error));
+			}
+			if (error != Xbyak::ERR_CODE_IS_TOO_BIG || !has_replacements) {
+				break;
 			}
-			throw;
 		}
-		return trampoline_offset;
+		return std::nullopt;
 	};
 
 	const auto record_rewrites = [&](const RelocationSpan& span) {
@@ -881,6 +1110,14 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 			if (rewrite->second.protect_red_zone && decoded->accesses_memory) {
 				++result.patched_memory_instruction_count;
 			}
+			if (rewrite->second.replacement != InstructionReplacement::None) {
+				auto& counts = ReplacementCounts(result, rewrite->second.replacement);
+				if (UsesInstructionTrap(rewrite->second.replacement, span.trap_replacements)) {
+					++counts.trapped;
+				} else {
+					++counts.native;
+				}
+			}
 		}
 	};
 
@@ -983,6 +1220,7 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 		patch_gen.setSize(span.patch_start - reinterpret_cast<uintptr_t>(patch_gen.getCode()));
 		patch_gen.jmp(trampoline, Xbyak::CodeGenerator::LabelType::T_NEAR);
 		patch_gen.nop(span.patch_size - NearJumpSize);
+		EXIT_IF(Xbyak::GetError() != 0);
 
 		record_rewrites(span);
 		patched_spans.emplace_back(span.patch_start, span.continuation);
@@ -1021,7 +1259,7 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 			continue;
 		}
 
-		const RelocationSpan site_span {
+		RelocationSpan site_span {
 		    .instructions = {&site_instruction->second},
 		    .patch_start  = site,
 		    .continuation = site + site_instruction->second.instruction.length,
@@ -1049,12 +1287,14 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 			patch_gen.reset();
 			patch_gen.setSize(*relay_slot - reinterpret_cast<uintptr_t>(patch_gen.getCode()));
 			patch_gen.jmp(site_trampoline, Xbyak::CodeGenerator::LabelType::T_NEAR);
+			EXIT_IF(Xbyak::GetError() != 0);
 
 			patch_gen.reset();
 			patch_gen.setSize(site - reinterpret_cast<uintptr_t>(patch_gen.getCode()));
 			patch_gen.jmp(reinterpret_cast<void*>(*relay_slot),
 			              Xbyak::CodeGenerator::LabelType::T_SHORT);
 			patch_gen.nop(site_span.patch_size - ShortJumpSize);
+			EXIT_IF(Xbyak::GetError() != 0);
 
 			record_rewrites(site_span);
 			patched_spans.emplace_back(site_span.patch_start, site_span.continuation);
@@ -1187,6 +1427,7 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 			patch_gen.reset();
 			patch_gen.setSize(relay_address - reinterpret_cast<uintptr_t>(patch_gen.getCode()));
 			patch_gen.jmp(site_trampoline, Xbyak::CodeGenerator::LabelType::T_NEAR);
+			EXIT_IF(Xbyak::GetError() != 0);
 			std::erase(relay_slots, relay_address);
 			std::erase(short_relay_slots, relay_address);
 		} else {
@@ -1201,6 +1442,7 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 			patch_gen.jmp(host_trampoline, Xbyak::CodeGenerator::LabelType::T_NEAR);
 			patch_gen.jmp(site_trampoline, Xbyak::CodeGenerator::LabelType::T_NEAR);
 			patch_gen.nop(host_span->patch_size - NearJumpSize * 2);
+			EXIT_IF(Xbyak::GetError() != 0);
 		}
 
 		uintptr_t jump_target = relay_address;
@@ -1210,6 +1452,7 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 			                  reinterpret_cast<uintptr_t>(patch_gen.getCode()));
 			patch_gen.jmp(reinterpret_cast<void*>(jump_target),
 			              Xbyak::CodeGenerator::LabelType::T_SHORT);
+			EXIT_IF(Xbyak::GetError() != 0);
 			jump_target       = final_short_jump;
 			const auto parent = relay_parent.find(final_short_jump);
 			ASSERT(parent != relay_parent.end());
@@ -1223,6 +1466,7 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 		patch_gen.jmp(reinterpret_cast<void*>(jump_target),
 		              Xbyak::CodeGenerator::LabelType::T_SHORT);
 		patch_gen.nop(site_span.patch_size - ShortJumpSize);
+		EXIT_IF(Xbyak::GetError() != 0);
 
 		if (host_span) {
 			record_rewrites(*host_span);
@@ -1240,10 +1484,30 @@ void RelocateRedZoneInstructions(PatchModule* module, const DecodedFunction& fun
 }
 } // namespace
 
-RedZonePatchResult PatchGuestInstructions(u64 segment_addr, u64 segment_size,
-                                          std::span<const uintptr_t> function_starts,
-                                          bool protect_memory, bool emulate_rsqrt) {
-	RedZonePatchResult result {};
+GuestInstructionHostFeatures GetGuestInstructionHostFeatures() {
+	static const auto features = [] {
+		const Xbyak::util::Cpu cpu;
+		uint32_t               leaf[4] {};
+		Xbyak::util::Cpu::getCpuid(0, leaf);
+		bool rdpid = false;
+		if (leaf[0] >= 7) {
+			Xbyak::util::Cpu::getCpuidEx(7, 0, leaf);
+			rdpid = (leaf[2] & (1u << 22)) != 0;
+		}
+		return GuestInstructionHostFeatures {cpu.has(Xbyak::util::Cpu::tSSE4a), rdpid,
+		                                     cpu.has(Xbyak::util::Cpu::tCLWB)};
+	}();
+	return features;
+}
+
+GuestInstructionPatchResult PatchGuestInstructions(u64 segment_addr, u64 segment_size,
+                                                   std::span<const uintptr_t> function_starts,
+                                                   bool protect_memory, bool emulate_amd,
+                                                   GuestInstructionHostFeatures host_features) {
+	GuestInstructionPatchResult result {};
+	if (!protect_memory && !emulate_amd) {
+		return result;
+	}
 	auto*              module = GetContainingModule(reinterpret_cast<void*>(segment_addr));
 	if (module == nullptr || function_starts.empty()) {
 		return result;
@@ -1262,8 +1526,11 @@ RedZonePatchResult PatchGuestInstructions(u64 segment_addr, u64 segment_size,
 	starts.erase(unique_end, starts.end());
 
 	std::unique_lock lock {module->mutex};
-	const size_t trampoline_begin = module->trampoline_gen.getSize();
-	std::vector<ReciprocalSquareRootSite> reciprocal_sqrt_sites;
+	const size_t     trampoline_begin = module->trampoline_gen.getSize();
+	bool analyze_red_zone = protect_memory;
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+	analyze_red_zone |= emulate_amd;
+#endif
 	for (size_t function_index = 0; function_index < starts.size(); ++function_index) {
 		const uintptr_t function_start = starts[function_index];
 		const uintptr_t function_end =
@@ -1274,7 +1541,9 @@ RedZonePatchResult PatchGuestInstructions(u64 segment_addr, u64 segment_size,
 
 		++result.function_count;
 		auto function = DecodeFunction(function_start, function_end, segment_addr, segment_end);
-		AnalyzeRedZoneLiveness(function);
+		if (analyze_red_zone) {
+			AnalyzeRedZoneLiveness(function);
+		}
 		result.instruction_count += function.instructions.size();
 
 		std::map<uintptr_t, InstructionRewrite> rewrite_sites;
@@ -1286,22 +1555,17 @@ RedZonePatchResult PatchGuestInstructions(u64 segment_addr, u64 segment_size,
 		if (protect_memory) {
 			CollectRedZoneMemoryInstructions(function, rewrite_sites, result);
 		}
-		if (emulate_rsqrt) {
-			CollectReciprocalSquareRoots(function, rewrite_sites, reciprocal_sqrt_sites);
+		if (emulate_amd) {
+			CollectAmdInstructions(function, rewrite_sites, result, host_features);
 		}
 		if (!rewrite_sites.empty()) {
-			RelocateRedZoneInstructions(module, function, rewrite_sites, result);
+			RelocateGuestInstructions(module, function, rewrite_sites, result);
+			TrapUnrelocatedInstructions(*module, function, rewrite_sites, result);
 		}
 	}
-	// Preserve valid instruction encodings throughout CFG analysis and relocation.
-	// Only now turn reciprocal roots into traps, including the relocated copies.
 	const auto trampoline_addr =
 	    reinterpret_cast<u64>(module->trampoline_gen.getCode()) + trampoline_begin;
 	const auto trampoline_size = module->trampoline_gen.getSize() - trampoline_begin;
-	if (emulate_rsqrt) {
-		result.reciprocal_sqrt_instruction_count = ApplyReciprocalSquareRootPatches(
-		    *module, reciprocal_sqrt_sites, trampoline_addr, trampoline_size);
-	}
 	Common::VirtualMemory::FlushInstructionCache(segment_addr, segment_size);
 	if (trampoline_size != 0) {
 		Common::VirtualMemory::FlushInstructionCache(trampoline_addr, trampoline_size);
@@ -1311,7 +1575,12 @@ RedZonePatchResult PatchGuestInstructions(u64 segment_addr, u64 segment_size,
 
 #else
 
-RedZonePatchResult PatchGuestInstructions(u64, u64, std::span<const uintptr_t>, bool, bool) {
+GuestInstructionHostFeatures GetGuestInstructionHostFeatures() {
+	return {};
+}
+
+GuestInstructionPatchResult PatchGuestInstructions(u64, u64, std::span<const uintptr_t>, bool, bool,
+                                                   GuestInstructionHostFeatures) {
 	return {};
 }
 
@@ -1492,9 +1761,9 @@ bool DecodeEhFrameFunctionStarts(uint64_t eh_frame_header_addr, uint64_t eh_fram
 	return true;
 }
 
-void RegisterRedZonePatchModule(void* module_ptr, uint64_t module_size, void* trampoline_area_ptr,
-                                uint64_t trampoline_area_size) {
-#if defined(_WIN32)
+void RegisterGuestInstructionPatchModule(void* module_ptr, uint64_t module_size,
+                                         void* trampoline_area_ptr, uint64_t trampoline_area_size) {
+#if !defined(__APPLE__)
 	EXIT_IF(module_ptr == nullptr || module_size == 0 || trampoline_area_ptr == nullptr ||
 	        trampoline_area_size == 0);
 	const auto module_addr = reinterpret_cast<u64>(module_ptr);
@@ -1511,8 +1780,8 @@ void RegisterRedZonePatchModule(void* module_ptr, uint64_t module_size, void* tr
 #endif
 }
 
-void UnregisterRedZonePatchModule(void* module_ptr) {
-#if defined(_WIN32)
+void UnregisterGuestInstructionPatchModule(void* module_ptr) {
+#if !defined(__APPLE__)
 	g_patch_modules.erase(reinterpret_cast<u64>(module_ptr));
 #else
 	(void)module_ptr;

src/loader/guestInstructionPatcher.h

@@ -0,0 +1,70 @@
+// SPDX-FileCopyrightText: Copyright 2026 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#ifndef KYTY_LOADER_GUEST_INSTRUCTION_PATCHER_H_
+#define KYTY_LOADER_GUEST_INSTRUCTION_PATCHER_H_
+
+#include "common/common.h"
+
+#include <span>
+#include <vector>
+
+namespace Loader {
+
+struct GuestInstructionHostFeatures {
+	bool sse4a = false;
+	bool rdpid = false;
+	bool clwb  = false;
+};
+
+GuestInstructionHostFeatures GetGuestInstructionHostFeatures();
+
+struct InstructionPatchCounts {
+	uint64_t found   = 0;
+	uint64_t native  = 0;
+	uint64_t trapped = 0;
+
+	uint64_t Skipped() const { return found - native - trapped; }
+
+	InstructionPatchCounts& operator+=(const InstructionPatchCounts& other) {
+		found += other.found;
+		native += other.native;
+		trapped += other.trapped;
+		return *this;
+	}
+};
+
+struct GuestInstructionPatchResult {
+	uint64_t               function_count                           = 0;
+	uint64_t               instruction_count                        = 0;
+	uint64_t               red_zone_function_count                  = 0;
+	uint64_t               memory_instruction_count                 = 0;
+	uint64_t               short_memory_instruction_count           = 0;
+	uint64_t               patched_memory_instruction_count         = 0;
+	uint64_t               stack_dependent_memory_instruction_count = 0;
+	uint64_t               control_flow_memory_instruction_count    = 0;
+	uint64_t               unrelocatable_memory_instruction_count   = 0;
+	uint64_t               indirect_red_zone_function_count         = 0;
+	InstructionPatchCounts reciprocal_sqrt;
+	InstructionPatchCounts extrq;
+	InstructionPatchCounts insertq;
+	InstructionPatchCounts rdpid;
+	InstructionPatchCounts clwb;
+};
+
+void RegisterGuestInstructionPatchModule(void* module_ptr, uint64_t module_size,
+                                         void* trampoline_area_ptr, uint64_t trampoline_area_size);
+void UnregisterGuestInstructionPatchModule(void* module_ptr);
+
+// Apply enabled instruction fixes using native trampolines or safe trap fallbacks.
+GuestInstructionPatchResult PatchGuestInstructions(
+    uint64_t segment_addr, uint64_t segment_size, std::span<const uintptr_t> function_starts,
+    bool protect_memory, bool emulate_amd,
+    GuestInstructionHostFeatures host_features = GetGuestInstructionHostFeatures());
+
+bool DecodeEhFrameFunctionStarts(uint64_t eh_frame_header_addr, uint64_t eh_frame_header_size,
+                                 std::vector<uintptr_t>* function_starts);
+
+} // namespace Loader
+
+#endif /* KYTY_LOADER_GUEST_INSTRUCTION_PATCHER_H_ */

src/loader/redZonePatcher.h

@@ -1,42 +0,0 @@
-// SPDX-FileCopyrightText: Copyright 2026 shadPS4 Emulator Project
-// SPDX-License-Identifier: GPL-2.0-or-later
-
-#ifndef KYTY_LOADER_RED_ZONE_PATCHER_H_
-#define KYTY_LOADER_RED_ZONE_PATCHER_H_
-
-#include "common/common.h"
-
-#include <span>
-#include <vector>
-
-namespace Loader {
-
-struct RedZonePatchResult {
-	uint64_t function_count                           = 0;
-	uint64_t instruction_count                        = 0;
-	uint64_t red_zone_function_count                   = 0;
-	uint64_t memory_instruction_count                 = 0;
-	uint64_t short_memory_instruction_count           = 0;
-	uint64_t patched_memory_instruction_count         = 0;
-	uint64_t stack_dependent_memory_instruction_count = 0;
-	uint64_t control_flow_memory_instruction_count    = 0;
-	uint64_t unrelocatable_memory_instruction_count   = 0;
-	uint64_t indirect_red_zone_function_count         = 0;
-	uint64_t reciprocal_sqrt_instruction_count        = 0;
-};
-
-void RegisterRedZonePatchModule(void* module_ptr, uint64_t module_size, void* trampoline_area_ptr,
-                                uint64_t trampoline_area_size);
-void UnregisterRedZonePatchModule(void* module_ptr);
-
-// Analyze and protect fault sites before replacing reciprocal roots with traps.
-RedZonePatchResult PatchGuestInstructions(uint64_t segment_addr, uint64_t segment_size,
-                                          std::span<const uintptr_t> function_starts,
-                                          bool protect_memory, bool emulate_rsqrt);
-
-bool DecodeEhFrameFunctionStarts(uint64_t eh_frame_header_addr, uint64_t eh_frame_header_size,
-                                 std::vector<uintptr_t>* function_starts);
-
-} // namespace Loader
-
-#endif /* KYTY_LOADER_RED_ZONE_PATCHER_H_ */

src/loader/runtimeLinker.cpp

@@ -17,8 +17,8 @@
 #include "kernel/pthread.h"
 #include "loader/elf.h"
 #include "loader/gamePatch.h"
+#include "loader/guestInstructionPatcher.h"
 #include "loader/jit.h"
-#include "loader/redZonePatcher.h"
 #include "loader/symbolDatabase.h"
 #include "loader/x64InstructionEmulator.h"
 
@@ -1812,38 +1812,39 @@ void RuntimeLinker::LoadProgramToMemory(Program* program) {
 	uint64_t tls_handler_size = is_shared ? 0 : Jit::SafeCall::GetSize();
 	EXIT_IF(tls_handler_size > UINT64_MAX - program->base_size_aligned);
 	program->mapped_size = program->base_size_aligned + tls_handler_size;
-	const bool emulate_rsqrt = Config::AmdCpuEnabled();
-
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-	const bool         protect_memory_faults   = Config::RedZoneProtectionEnabled();
-	const bool         use_red_zone_protection = protect_memory_faults || emulate_rsqrt;
-	constexpr uint64_t RED_ZONE_TRAMPOLINE_SIZE = 8u * 1024u * 1024u;
-	if (use_red_zone_protection) {
-		EXIT_IF(RED_ZONE_TRAMPOLINE_SIZE > UINT64_MAX - program->mapped_size);
-		program->mapped_size += RED_ZONE_TRAMPOLINE_SIZE;
-	}
+#if !defined(__APPLE__)
+	const bool emulate_amd = Config::AmdCpuEnabled();
+#else
+	const bool emulate_amd = false;
 #endif
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+	const bool protect_memory_faults = Config::RedZoneProtectionEnabled();
+#else
+	const bool protect_memory_faults = false;
+#endif
+	const bool patch_guest_instructions = protect_memory_faults || emulate_amd;
+
+	constexpr uint64_t INSTRUCTION_TRAMPOLINE_SIZE = 8u * 1024u * 1024u;
+	if (patch_guest_instructions) {
+		EXIT_IF(INSTRUCTION_TRAMPOLINE_SIZE > UINT64_MAX - program->mapped_size);
+		program->mapped_size += INSTRUCTION_TRAMPOLINE_SIZE;
+	}
 
 	program->base_vaddr = Libs::LibKernel::Memory::AllocateProgramMemory(
 	    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);
 
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-	if (use_red_zone_protection) {
-		program->red_zone_trampoline_vaddr = program->base_vaddr + program->base_size_aligned;
-		program->red_zone_trampoline_size  = RED_ZONE_TRAMPOLINE_SIZE;
-		RegisterRedZonePatchModule(reinterpret_cast<void*>(program->base_vaddr),
-		                           program->base_size_aligned,
-		                           reinterpret_cast<void*>(program->red_zone_trampoline_vaddr),
-		                           program->red_zone_trampoline_size);
+	if (patch_guest_instructions) {
+		const auto trampoline_addr           = program->base_vaddr + program->base_size_aligned;
+		program->instruction_trampoline_size = INSTRUCTION_TRAMPOLINE_SIZE;
+		RegisterGuestInstructionPatchModule(
+		    reinterpret_cast<void*>(program->base_vaddr), program->base_size_aligned,
+		    reinterpret_cast<void*>(trampoline_addr), program->instruction_trampoline_size);
 	}
-#endif
 	if (!is_shared) {
-		program->tls.handler_vaddr = program->base_vaddr + program->base_size_aligned;
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-		program->tls.handler_vaddr += program->red_zone_trampoline_size;
-#endif
+		program->tls.handler_vaddr =
+		    program->base_vaddr + program->base_size_aligned + program->instruction_trampoline_size;
 	}
 
 	g_desired_base_addr += CODE_BASE_INCR * (1 + program->mapped_size / CODE_BASE_INCR);
@@ -1863,10 +1864,8 @@ void RuntimeLinker::LoadProgramToMemory(Program* program) {
 	}
 
 	std::vector<std::pair<uint64_t, uint64_t>> executable_segments;
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	uint64_t                                   eh_frame_header_addr = 0;
 	uint64_t                                   eh_frame_header_size = 0;
-#endif
 
 	for (Elf64_Half i = 0; i < ehdr->e_phnum; i++) {
 		if (phdr[i].p_memsz != 0 && (phdr[i].p_type == PT_LOAD || phdr[i].p_type == PT_OS_RELRO)) {
@@ -1925,52 +1924,68 @@ void RuntimeLinker::LoadProgramToMemory(Program* program) {
 			program->proc_param_vaddr = phdr[i].p_vaddr + program->base_vaddr;
 		}
 
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-		if (use_red_zone_protection && phdr[i].p_type == PT_GNU_EH_FRAME) {
+		if (patch_guest_instructions && phdr[i].p_type == PT_GNU_EH_FRAME) {
 			eh_frame_header_addr = phdr[i].p_vaddr + program->base_vaddr;
 			eh_frame_header_size = phdr[i].p_memsz;
 		}
-#endif
 	}
 
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-	std::vector<uintptr_t> function_starts;
-	if (use_red_zone_protection) {
-		if (!DecodeEhFrameFunctionStarts(eh_frame_header_addr, eh_frame_header_size,
-		                                 &function_starts)) {
-			LOGF("Windows guest red-zone patching could not decode function boundaries for %s\n",
-			     Common::PathToString(program->file_name).c_str());
+	if (patch_guest_instructions) {
+		std::vector<uintptr_t> function_starts;
+		const bool             have_function_starts =
+		    DecodeEhFrameFunctionStarts(eh_frame_header_addr, eh_frame_header_size,
+		                                &function_starts) &&
+		    !function_starts.empty();
+		const auto module_name = Common::PathToString(program->file_name.filename());
+		if (!have_function_starts) {
+			Log::WriteToConsoleAndLog(
+			    fmt::format("{}: {} not patched (function boundaries unavailable)\n",
+			                emulate_amd ? "AMD CPU compatibility" : "Guest red-zone protection",
+			                module_name));
 		}
-	}
-#endif
-	for (const auto& [segment_addr, segment_size]: executable_segments) {
-		uint64_t reciprocal_sqrt_count = 0;
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-		if (use_red_zone_protection) {
-			const auto result =
-			    PatchGuestInstructions(segment_addr, segment_size, function_starts,
-			                           protect_memory_faults, emulate_rsqrt);
-			LOGF("Windows guest red-zone patching: %s, functions=%" PRIu64 ", red_zone=%" PRIu64
-			     ", memory=%" PRIu64 ", patched=%" PRIu64 ", short=%" PRIu64 ", stack=%" PRIu64
-			     ", control=%" PRIu64 ", unrelocatable=%" PRIu64 "\n",
-			     Common::PathToString(program->file_name.filename()).c_str(), result.function_count,
-			     result.red_zone_function_count, result.memory_instruction_count,
-			     result.patched_memory_instruction_count, result.short_memory_instruction_count,
-			     result.stack_dependent_memory_instruction_count,
-			     result.control_flow_memory_instruction_count,
-			     result.unrelocatable_memory_instruction_count);
-			reciprocal_sqrt_count = result.reciprocal_sqrt_instruction_count;
+		GuestInstructionPatchResult totals {};
+		for (const auto& [segment_addr, segment_size]: executable_segments) {
+			const auto result = PatchGuestInstructions(segment_addr, segment_size, function_starts,
+			                                           protect_memory_faults, emulate_amd);
+			totals.reciprocal_sqrt += result.reciprocal_sqrt;
+			totals.extrq += result.extrq;
+			totals.insertq += result.insertq;
+			totals.rdpid += result.rdpid;
+			totals.clwb += result.clwb;
+			if (protect_memory_faults) {
+				LOGF("Windows guest red-zone patching: %s, functions=%" PRIu64 ", red_zone=%" PRIu64
+				     ", memory=%" PRIu64 ", patched=%" PRIu64 ", short=%" PRIu64 ", stack=%" PRIu64
+				     ", control=%" PRIu64 ", unrelocatable=%" PRIu64 "\n",
+				     module_name.c_str(),
+				     result.function_count, result.red_zone_function_count,
+				     result.memory_instruction_count, result.patched_memory_instruction_count,
+				     result.short_memory_instruction_count,
+				     result.stack_dependent_memory_instruction_count,
+				     result.control_flow_memory_instruction_count,
+				     result.unrelocatable_memory_instruction_count);
+			}
 		}
-#else
-		if (emulate_rsqrt) {
-			reciprocal_sqrt_count =
-			    X64InstructionEmulator::PatchReciprocalSquareRoots(segment_addr, segment_size);
-			Common::VirtualMemory::FlushInstructionCache(segment_addr, segment_size);
-		}
-#endif
-		if (reciprocal_sqrt_count != 0) {
-			LOGF("Guest VRSQRTPS emulation: %s, instructions=%" PRIu64 "\n",
-			     Common::PathToString(program->file_name.filename()).c_str(), reciprocal_sqrt_count);
+		if (emulate_amd && have_function_starts) {
+			InstructionPatchCounts combined {};
+			std::string            details;
+			for (const auto& [name, counts]: {std::pair {"VRSQRTPS", totals.reciprocal_sqrt},
+			                                  {"EXTRQ", totals.extrq},
+			                                  {"INSERTQ", totals.insertq},
+			                                  {"RDPID", totals.rdpid},
+			                                  {"CLWB", totals.clwb}}) {
+				combined += counts;
+				if (!details.empty()) details += "; ";
+				details += fmt::format("{}: native={}, trapped={}, skipped={}", name, counts.native,
+				                       counts.trapped, counts.Skipped());
+			}
+			const auto  found   = combined.found;
+			const auto  skipped = combined.Skipped();
+			const char* status  = found == 0         ? "no matching instructions"
+			                      : skipped == found ? "not patched"
+			                      : skipped != 0     ? "partially patched"
+			                                         : "patched";
+			Log::WriteToConsoleAndLog(
+			    fmt::format("AMD CPU compatibility: {} {} ({})\n", module_name, status, details));
 		}
 	}
 
@@ -1999,11 +2014,9 @@ void RuntimeLinker::DeleteProgram(Program* p) {
 
 	if (program->base_vaddr != 0 || program->mapped_size != 0) {
 		EXIT_IF(program->base_vaddr == 0 || program->mapped_size == 0);
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-		if (program->red_zone_trampoline_size != 0) {
-			UnregisterRedZonePatchModule(reinterpret_cast<void*>(program->base_vaddr));
+		if (program->instruction_trampoline_size != 0) {
+			UnregisterGuestInstructionPatchModule(reinterpret_cast<void*>(program->base_vaddr));
 		}
-#endif
 		EXIT_IF(
 		    !Libs::LibKernel::Memory::FreeGuestMemory(program->base_vaddr, program->mapped_size));
 	}

src/loader/runtimeLinker.h

@@ -123,10 +123,7 @@ struct Program {
 	uint64_t                     base_size         = 0;
 	uint64_t                     base_size_aligned = 0;
 	uint64_t                     mapped_size       = 0;
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-	uint64_t red_zone_trampoline_vaddr = 0;
-	uint64_t red_zone_trampoline_size  = 0;
-#endif
+	uint64_t instruction_trampoline_size = 0;
 	std::unique_ptr<SymbolDatabase> export_symbols;
 	std::unique_ptr<SymbolDatabase> import_symbols;
 	ThreadLocalStorage              tls;

src/loader/x64InstructionEmulator.cpp

@@ -5,10 +5,9 @@
 #include <Zydis/Zydis.h>
 #include <bit>
 #include <cstring>
-#if !defined(__APPLE__)
 #include <emmintrin.h>
 #include <xmmintrin.h>
-#endif
+#include <cpuid.h>
 
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 #include <windows.h> // IWYU pragma: keep
@@ -407,12 +406,12 @@ struct Context {
 	void                   Advance(size_t length) { native->Rip += length; }
 	[[nodiscard]] void*    Xmm(uint8_t index) const { return &native->Xmm0 + index; }
 
-	void LoadGprs(uint64_t (&gpr)[16]) const {
-		const uint64_t registers[] = {native->Rax, native->Rcx, native->Rdx, native->Rbx,
-		                              native->Rsp, native->Rbp, native->Rsi, native->Rdi,
-		                              native->R8,  native->R9,  native->R10, native->R11,
-		                              native->R12, native->R13, native->R14, native->R15};
-		std::memcpy(gpr, registers, sizeof(gpr));
+	[[nodiscard]] auto& Gpr(uint8_t index) const {
+		DWORD64* registers[] = {&native->Rax, &native->Rcx, &native->Rdx, &native->Rbx,
+		                        &native->Rsp, &native->Rbp, &native->Rsi, &native->Rdi,
+		                        &native->R8,  &native->R9,  &native->R10, &native->R11,
+		                        &native->R12, &native->R13, &native->R14, &native->R15};
+		return *registers[index];
 	}
 
 	void ClearUpperYmm(uint8_t index) const {
@@ -438,6 +437,14 @@ struct Context {
 	void Advance(size_t length) {
 		native->uc_mcontext->__ss.__rip += static_cast<uint64_t>(length);
 	}
+	[[nodiscard]] auto& Gpr(uint8_t index) const {
+		auto*     state       = &native->uc_mcontext->__ss;
+		uint64_t* registers[] = {&state->__rax, &state->__rcx, &state->__rdx, &state->__rbx,
+		                         &state->__rsp, &state->__rbp, &state->__rsi, &state->__rdi,
+		                         &state->__r8,  &state->__r9,  &state->__r10, &state->__r11,
+		                         &state->__r12, &state->__r13, &state->__r14, &state->__r15};
+		return *registers[index];
+	}
 	// Darwin names the XMM file __fpu_xmm0..__fpu_xmm15 instead of exposing an array.
 	[[nodiscard]] void* Xmm(uint8_t index) const {
 		auto* fs = &native->uc_mcontext->__fs;
@@ -477,13 +484,11 @@ struct Context {
 		return native->uc_mcontext.fpregs->_xmm[index].element;
 	}
 
-	void LoadGprs(uint64_t (&gpr)[16]) const {
+	[[nodiscard]] auto& Gpr(uint8_t index) const {
 		constexpr int registers[] = {REG_RAX, REG_RCX, REG_RDX, REG_RBX, REG_RSP, REG_RBP,
 		                             REG_RSI, REG_RDI, REG_R8,  REG_R9,  REG_R10, REG_R11,
 		                             REG_R12, REG_R13, REG_R14, REG_R15};
-		for (size_t i = 0; i < 16; ++i) {
-			gpr[i] = static_cast<uint64_t>(native->uc_mcontext.gregs[registers[i]]);
-		}
+		return native->uc_mcontext.gregs[registers[index]];
 	}
 
 	void ClearUpperYmm(uint8_t index) const {
@@ -507,6 +512,12 @@ struct Context {
 		}
 	}
 #endif
+
+	void LoadGprs(uint64_t (&gpr)[16]) const {
+		for (uint8_t i = 0; i < 16; ++i) {
+			gpr[i] = static_cast<uint64_t>(Gpr(i));
+		}
+	}
 };
 
 #if !defined(__APPLE__)
@@ -575,8 +586,8 @@ static bool TryEmulateSse4a(Context& context) {
 	if (rip[offset] != 0x0f) {
 		return false;
 	}
-	const bool register_extract = prefix == 0x66 && rip[offset + 1] == 0x79;
-	if (rip[offset + 1] != 0x78 && !register_extract) {
+	const bool register_form = rip[offset + 1] == 0x79;
+	if (rip[offset + 1] != 0x78 && !register_form) {
 		return false;
 	}
 
@@ -590,7 +601,7 @@ static bool TryEmulateSse4a(Context& context) {
 
 	// Immediate EXTRQ encodes its destination in r/m; the two-register form uses reg.
 	uint8_t dest_index = reg;
-	if (prefix == 0x66 && !register_extract) {
+	if (prefix == 0x66 && !register_form) {
 		dest_index = rm;
 	}
 	auto* dest_xmm = context.Xmm(dest_index);
@@ -599,15 +610,17 @@ static bool TryEmulateSse4a(Context& context) {
 		return false;
 	}
 	uint64_t dest[2] {};
-	uint64_t source = 0;
+	uint64_t source[2] {};
 	std::memcpy(dest, dest_xmm, sizeof(dest));
-	std::memcpy(&source, src_xmm, sizeof(source));
+	std::memcpy(source, src_xmm, sizeof(source));
 	uint8_t length             = 0;
 	uint8_t index              = 0;
 	size_t  instruction_length = offset + 3;
-	if (register_extract) {
-		length = static_cast<uint8_t>(source);
-		index  = static_cast<uint8_t>(source >> 8u);
+	if (register_form) {
+		// EXTRQ controls occupy the low quadword; INSERTQ uses the upper quadword.
+		const uint64_t controls = source[prefix == 0x66 ? 0 : 1];
+		length                  = static_cast<uint8_t>(controls);
+		index                   = static_cast<uint8_t>(controls >> 8u);
 	} else {
 		length = rip[offset + 3];
 		index  = rip[offset + 4];
@@ -617,13 +630,79 @@ static bool TryEmulateSse4a(Context& context) {
 		dest[0] = ExtractBitField(dest[0], length, index);
 		dest[1] = 0;
 	} else {
-		dest[0] = InsertBitField(dest[0], source, length, index);
+		dest[0] = InsertBitField(dest[0], source[0], length, index);
 	}
 	std::memcpy(dest_xmm, dest, sizeof(dest));
 	context.Advance(instruction_length);
 	return true;
 }
 
+static bool ReadTscAux(uint32_t& auxiliary) {
+	uint32_t eax = 0;
+	uint32_t ebx = 0;
+	uint32_t ecx = 0;
+	uint32_t edx = 0;
+	if (!__get_cpuid(0x80000001u, &eax, &ebx, &ecx, &edx) || (edx & (1u << 27u)) == 0) {
+		return false;
+	}
+	asm volatile("rdtscp" : "=a"(eax), "=d"(edx), "=c"(auxiliary));
+	return true;
+}
+
+static void FlushCacheLine(uint64_t address, ZydisRegister segment) {
+	// Preserve FS/GS addressing rather than treating a segment-relative offset as a pointer.
+	if (segment == ZYDIS_REGISTER_FS) {
+		asm volatile("clflush %%fs:(%0)" : : "r"(address) : "memory");
+	} else if (segment == ZYDIS_REGISTER_GS) {
+		asm volatile("clflush %%gs:(%0)" : : "r"(address) : "memory");
+	} else {
+		_mm_clflush(reinterpret_cast<const void*>(address));
+	}
+}
+
+static bool TryEmulateCpuExtensions(Context& context) {
+	ZydisDecoder            decoder {};
+	ZydisDecodedInstruction instruction {};
+	ZydisDecodedOperand     operands[ZYDIS_MAX_OPERAND_COUNT] {};
+	if (!ZYAN_SUCCESS(
+	        ZydisDecoderInit(&decoder, ZYDIS_MACHINE_MODE_LONG_64, ZYDIS_STACK_WIDTH_64)) ||
+	    !ZYAN_SUCCESS(ZydisDecoderDecodeFull(&decoder, reinterpret_cast<const void*>(context.Rip()),
+	                                         ZYDIS_MAX_INSTRUCTION_LENGTH, &instruction,
+	                                         operands))) {
+		return false;
+	}
+	if (instruction.mnemonic == ZYDIS_MNEMONIC_RDPID &&
+	    operands[0].type == ZYDIS_OPERAND_TYPE_REGISTER) {
+		const auto reg_class = ZydisRegisterGetClass(operands[0].reg.value);
+		const auto index     = ZydisRegisterGetId(operands[0].reg.value);
+		uint32_t   auxiliary = 0;
+		if ((reg_class != ZYDIS_REGCLASS_GPR32 && reg_class != ZYDIS_REGCLASS_GPR64) || index < 0 ||
+		    index >= 16 || !ReadTscAux(auxiliary)) {
+			return false;
+		}
+		// RDPID and guest RDTSCP observe the same host TSC_AUX; only RDPID's destination changes.
+		context.Gpr(static_cast<uint8_t>(index)) = auxiliary;
+	} else if (instruction.mnemonic == ZYDIS_MNEMONIC_CLWB &&
+	           operands[0].type == ZYDIS_OPERAND_TYPE_MEMORY) {
+		ZydisRegisterContext registers {};
+		for (uint8_t i = 0; i < 16; ++i) {
+			const auto value                         = static_cast<uint64_t>(context.Gpr(i));
+			registers.values[ZYDIS_REGISTER_RAX + i] = value;
+			registers.values[ZYDIS_REGISTER_EAX + i] = static_cast<uint32_t>(value);
+		}
+		uint64_t address = 0;
+		if (!ZYAN_SUCCESS(ZydisCalcAbsoluteAddressEx(&instruction, &operands[0], context.Rip(),
+		                                             &registers, &address))) {
+			return false;
+		}
+		FlushCacheLine(address, operands[0].mem.segment);
+	} else {
+		return false;
+	}
+	context.Advance(instruction.length);
+	return true;
+}
+
 #if !defined(__APPLE__)
 
 static bool TryEmulateMonitorxMwaitx(Context& context) {
@@ -713,46 +792,6 @@ static bool TryEmulateReciprocalSquareRoot(Context& context) {
 
 #endif
 
-bool IsReciprocalSquareRoot(const ZydisDecodedInstruction& instruction,
-                            const ZydisDecodedOperand* operands) {
-	return instruction.mnemonic == ZYDIS_MNEMONIC_VRSQRTPS &&
-	       instruction.encoding == ZYDIS_INSTRUCTION_ENCODING_VEX &&
-	       instruction.raw.vex.offset == 0 && operands[0].size == 128 &&
-	       operands[1].type == ZYDIS_OPERAND_TYPE_REGISTER;
-}
-
-uint64_t PatchReciprocalSquareRoots(uint64_t address, uint64_t size) {
-	uint64_t patched = 0;
-#if !defined(__APPLE__)
-	ZydisDecoder decoder {};
-	if (!ZYAN_SUCCESS(
-	        ZydisDecoderInit(&decoder, ZYDIS_MACHINE_MODE_LONG_64, ZYDIS_STACK_WIDTH_64))) {
-		return 0;
-	}
-	for (uint64_t offset = 0; offset < size;) {
-		auto*                   code = reinterpret_cast<uint8_t*>(address + offset);
-		ZydisDecodedInstruction instruction {};
-		ZydisDecodedOperand     operands[ZYDIS_MAX_OPERAND_COUNT] {};
-		if (!ZYAN_SUCCESS(
-		        ZydisDecoderDecodeFull(&decoder, code, size - offset, &instruction, operands))) {
-			++offset;
-			continue;
-		}
-		if (IsReciprocalSquareRoot(instruction, operands)) {
-			// vvvv is reserved (must be 1111b). Clear one bit to route this
-			// otherwise intact instruction through the illegal-instruction emulator.
-			code[instruction.raw.vex.size - 1] &= ~0x08u;
-			++patched;
-		}
-		offset += instruction.length;
-	}
-#else
-	(void)address;
-	(void)size;
-#endif
-	return patched;
-}
-
 bool TryEmulate(void* native_context) {
 	if (native_context == nullptr) {
 		return false;
@@ -769,13 +808,10 @@ bool TryEmulate(void* native_context) {
 	Context context {static_cast<ucontext_t*>(native_context)};
 #endif
 #if !defined(__APPLE__)
-	if (TryEmulateReciprocalSquareRoot(context)) {
-		return true;
-	}
-	return TryEmulateMonitorxMwaitx(context) || TryEmulateSse4a(context) ||
-	       TryEmulateShaNi(context);
+	return TryEmulateReciprocalSquareRoot(context) || TryEmulateMonitorxMwaitx(context) ||
+	       TryEmulateSse4a(context) || TryEmulateShaNi(context) || TryEmulateCpuExtensions(context);
 #else
-	return TryEmulateSse4a(context);
+	return TryEmulateSse4a(context) || TryEmulateCpuExtensions(context);
 #endif
 }
 

src/loader/x64InstructionEmulator.h

@@ -1,14 +1,8 @@
 #ifndef KYTY_LOADER_X64_INSTRUCTION_EMULATOR_H_
 #define KYTY_LOADER_X64_INSTRUCTION_EMULATOR_H_
 
-#include <Zydis/DecoderTypes.h>
-#include <cstdint>
-
 namespace Loader::X64InstructionEmulator {
 
-[[nodiscard]] bool IsReciprocalSquareRoot(const ZydisDecodedInstruction& instruction,
-                                         const ZydisDecodedOperand* operands);
-uint64_t           PatchReciprocalSquareRoots(uint64_t address, uint64_t size);
 [[nodiscard]] bool TryEmulate(void* native_context);
 
 } // namespace Loader::X64InstructionEmulator

src/main.cpp

@@ -54,11 +54,15 @@ static void PrintUsage() {
 	         Config::DEFAULT_USER_ID);
 	::printf("  --mic <name>                        Capture from this microphone; omit for silence.\n");
 	::printf("  --controller-color <#RRGGBB>        Override the controller lightbar color.\n");
+	::printf("  --controller-volume <0-100>         DualSense speaker volume. Default: 50.\n");
+	::printf("  --controller-vibration <0-100>      DualSense vibration intensity. Default: 100.\n");
 	::printf(
 	    "  --present-mode <value>               Fifo, Mailbox, or Immediate. Default: Mailbox.\n");
 	::printf(
 	    "  --gpu <index>                        Vulkan physical device index. Default: auto.\n");
 	::printf("  --fullscreen                         Run in borderless desktop fullscreen.\n");
+	::printf(
+	    "  --hide-cursor                        Hide the cursor after 2 s idle. Default: off.\n");
 	::printf("  --vr                                 Enable the virtual VR headset.\n");
 	::printf("  --amd-cpu                            Apply AMD CPU instruction patches.\n");
 	::printf("  --vblank-frequency <num>             Virtual vblank frequency. Default: 60.\n");
@@ -203,6 +207,11 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 			continue;
 		}
 
+		if (arg == "--hide-cursor") {
+			options.config.hide_cursor_enabled = true;
+			continue;
+		}
+
 		if (arg == "--vr") {
 			options.config.vr_enabled = true;
 			continue;
@@ -324,6 +333,18 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 				return false;
 			}
 			options.config.controller_color = color;
+		} else if (arg == "--controller-volume") {
+			if (!ParseUint32(value, options.config.controller_speaker_volume) ||
+			    options.config.controller_speaker_volume > 100) {
+				::printf("invalid controller volume: %s\n", value.c_str());
+				return false;
+			}
+		} else if (arg == "--controller-vibration") {
+			if (!ParseUint32(value, options.config.controller_vibration_intensity) ||
+			    options.config.controller_vibration_intensity > 100) {
+				::printf("invalid controller vibration intensity: %s\n", value.c_str());
+				return false;
+			}
 		} else if (arg == "--present-mode") {
 			if (!ParseEnum(value, options.config.present_mode)) {
 				::printf("invalid present mode: %s\n", value.c_str());