shader: support V_CMPX_EQ_U16 (#816)

* shader: support V_CMPX_EQ_U16

Decode gfx11 VOPC 0xba and lower it as an unsigned 16-bit equality
compare with CMPX EXEC-mask semantics.

🤖 Generated with Codebuff 🤖
Co-Authored-By: Codebuff <noreply@codebuff.com>

* shader: correct CMPX EQ U16 runtime expectations

Remove the added CFG tests and correct two EXEC expectations in the compute regression case.

---------

Co-authored-by: Matthew Hurley <171001439+theantipopau@users.noreply.github.com>
Co-authored-by: Codebuff <noreply@codebuff.com>
Co-authored-by: nmzik <Nmzik@mail.ru>

CMakeLists.txt

@@ -677,7 +677,10 @@ if(KYTY_CLANG_CL)
 	)
 	add_custom_command(TARGET kyty_emulator POST_BUILD COMMAND ${CMAKE_COMMAND} -E copy_if_different "${KYTY_THIRD_PARTY_DIR}/winpthread/bin/libwinpthread-1.dll" $<TARGET_FILE_DIR:kyty_emulator>/libwinpthread-1.dll)
 elseif(WIN32 OR LINUX)
-	set_target_properties(kyty_emulator PROPERTIES LINK_FLAGS "${KYTY_LD_OPTIONS} -Wl,-Map=${KYTY_EMULATOR_MAP_LINK_PATH}")
+	set_target_properties(kyty_emulator PROPERTIES LINK_FLAGS "${KYTY_LD_OPTIONS}")
+	# LINK_FLAGS is a space-delimited string: a -Map path containing spaces would
+	# be split into bogus arguments, so pass it as a discrete link option instead.
+	target_link_options(kyty_emulator PRIVATE "-Wl,-Map=${KYTY_EMULATOR_MAP_LINK_PATH}")
 endif()
 
 if(LINUX)

README.md

@@ -4,13 +4,13 @@
 [![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)
 [![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-early%20development-orange.svg)](#current-status)
+[![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)
 
 KytyPS5 is a free and open-source PlayStation 5 emulator written in C++ for Windows and Linux,
 with experimental macOS support. It is based on a heavily modified version of
-[Kyty](https://github.com/InoriRus/Kyty). The project is in an early stage of development, so
-compatibility is limited and behavior may change significantly between builds.
+[Kyty](https://github.com/InoriRus/Kyty). The project is in active development, and behavior
+can change significantly between builds.
 
 > [!IMPORTANT]
 > KytyPS5 is not affiliated with Sony Interactive Entertainment or PlayStation. The project does
@@ -20,12 +20,12 @@ compatibility is limited and behavior may change significantly between builds.
 ## Current Status
 
 KytyPS5 can boot 2D games and a selection of 3D games, including titles built with Unreal Engine
-4/5, Unity, and custom engines. No external low-level emulation modules are currently required.
+4/5, Unity, and custom engines. External low-level emulation modules are neither required nor
+planned.
 
-Development is focused on compatibility and boot reliability.
+Development is currently focused on expanding game compatibility and improving boot reliability.
 
-Windows is the primary platform and receives the most testing. Linux builds and runs; see
-[Building on Linux](#building-on-linux).
+Windows and Linux are the primary platforms and receive the most testing.
 
 macOS support is experimental. The emulator is built for x86-64 and runs on Apple Silicon under
 Rosetta 2, with Vulkan provided by MoltenVK. A small number of titles have been verified in-game
@@ -36,16 +36,16 @@ Community game test results are available in the
 
 ## Bugs and Issues
 
-The project is in an early stage, so please be mindful when opening new issues. Expect crashes,
-graphical glitches, low compatibility, and poor performance.
+Compatibility, stability, and performance can vary between versions. You may encounter crashes
+or graphical glitches, so please include the version you tested when reporting an issue.
 
 ## Screenshots
 
 <table align="center">
   <tr>
     <td align="center">
-      <strong>Disgaea 6</strong><br>
-      <img src="docs/screenshots/ps5-01.png" width="300" alt="Disgaea 6 running in KytyPS5">
+      <strong>Astro Bot</strong><br>
+      <img src="docs/screenshots/ps5-01.png" width="300" alt="Astro Bot running in KytyPS5">
     </td>
     <td align="center">
       <strong>Dreaming Sarah</strong><br>
@@ -64,12 +64,12 @@ graphical glitches, low compatibility, and poor performance.
   </tr>
   <tr>
     <td align="center">
-      <strong>Hellboy</strong><br>
-      <img src="docs/screenshots/ps5-02.png" width="300" alt="Hellboy running in KytyPS5">
+      <strong>Demon's Souls</strong><br>
+      <img src="docs/screenshots/ps5-02.png" width="300" alt="Demon's Souls running in KytyPS5">
     </td>
     <td align="center">
-      <strong>Paleo Pines</strong><br>
-      <img src="docs/screenshots/ps5-06.png" width="300" alt="Paleo Pines running in KytyPS5">
+      <strong>Hellboy</strong><br>
+      <img src="docs/screenshots/ps5-06.png" width="300" alt="Hellboy running in KytyPS5">
     </td>
   </tr>
 </table>
@@ -370,5 +370,6 @@ licenses included with those components.
 
 - [InoriRus/Kyty](https://github.com/InoriRus/Kyty) — KytyPS5 is based on a heavily modified version
   of the original Kyty project.
-- [shadps4-emu/shadPS4](https://github.com/shadps4-emu/shadPS4) — reference for memory-model
-  understanding and the AVPlayer implementation.
+- [shadps4-emu/shadPS4](https://github.com/shadps4-emu/shadPS4) — reference for understanding PS4
+  memory behavior, GPU resource aliasing and cache coherency,
+  and the AVPlayer implementation.

src/common/timer.cpp

@@ -38,16 +38,7 @@ bool Timer::IsPaused() const {
 
 // return time in milliseconds
 double Timer::GetTimeMs() const {
-	if (m_is_paused) {
-		return 1000.0 * (static_cast<double>(m_PauseTime - m_StartTime)) /
-		       static_cast<double>(m_Frequency);
-	}
-
-	uint64_t current_time = 0;
-	SysQueryPerformanceCounter(&current_time);
-
-	return 1000.0 * (static_cast<double>(current_time - m_StartTime)) /
-	       static_cast<double>(m_Frequency);
+	return GetTimeS() * 1000.0;
 }
 
 // return time in seconds

src/graphics/guest_gpu/graphicsRun.cpp

@@ -976,6 +976,16 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 	const auto args_size = indexed ? sizeof(DrawIndexedIndirectArgs) : sizeof(DrawIndirectArgs);
 	EXIT_NOT_IMPLEMENTED(stride_in_bytes < args_size);
 
+	uint64_t index_size = 0;
+	if (indexed) {
+		switch (m_index_type_and_size) {
+			case 0: index_size = 2; break;
+			case 1: index_size = 4; break;
+			case 2: index_size = 1; break;
+			default: EXIT("unknown index_type_and_size: %u\n", m_index_type_and_size);
+		}
+	}
+
 	for (uint32_t i = 0; i < draw_count; i++) {
 		const auto args_addr = m_draw_indirect_args_base_addr + data_offset +
 		                       static_cast<uint64_t>(i) * stride_in_bytes;
@@ -993,14 +1003,6 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 
 		auto* args = reinterpret_cast<const DrawIndexedIndirectArgs*>(args_addr);
 
-		uint64_t index_size = 0;
-		switch (m_index_type_and_size) {
-			case 0: index_size = 2; break;
-			case 1: index_size = 4; break;
-			case 2: index_size = 1; break;
-			default: EXIT("unknown index_type_and_size: %u\n", m_index_type_and_size);
-		}
-
 		auto* index_addr = reinterpret_cast<const void*>(
 		    m_index_base_addr + static_cast<uint64_t>(args->start_index_location) * index_size);
 

src/graphics/host_gpu/renderer/depthRenderTarget.h

@@ -41,13 +41,42 @@ struct RenderDepthInfo {
 	[[nodiscard]] vk::ImageAspectFlags AttachmentWriteAspects() const;
 };
 
+inline vk::ImageAspectFlags DepthFeedbackAspects(vk::ImageAspectFlags draw_writes,
+                                                 const ImageViewInfo& target,
+                                                 const ImageViewInfo& sampled) {
+	if (!ImageRangeOverlaps(target.base_level, target.level_count, sampled.base_level,
+	                        sampled.level_count) ||
+	    !ImageRangeOverlaps(target.base_layer, target.layer_count, sampled.base_layer,
+	                        sampled.layer_count)) {
+		return {};
+	}
+	return draw_writes & sampled.aspect;
+}
+
+inline vk::ImageAspectFlags DepthReadableAspects(vk::ImageLayout layout) {
+	switch (layout) {
+		case vk::ImageLayout::eDepthReadOnlyOptimal:
+		case vk::ImageLayout::eDepthReadOnlyStencilAttachmentOptimal:
+			return vk::ImageAspectFlagBits::eDepth;
+		case vk::ImageLayout::eStencilReadOnlyOptimal:
+		case vk::ImageLayout::eDepthAttachmentStencilReadOnlyOptimal:
+			return vk::ImageAspectFlagBits::eStencil;
+		case vk::ImageLayout::eDepthStencilReadOnlyOptimal:
+		case vk::ImageLayout::eAttachmentFeedbackLoopOptimalEXT:
+		case vk::ImageLayout::eGeneral:
+			return vk::ImageAspectFlagBits::eDepth | vk::ImageAspectFlagBits::eStencil;
+		default:
+			return {};
+	}
+}
+
 inline vk::ImageLayout depth_attachment_layout(const RenderDepthInfo& depth) {
 	const auto available     = ImageViewOps::DepthAspectMask(depth.desc.view_info.format);
 	const auto writes        = depth.AttachmentWriteAspects();
 	const bool has_depth     = static_cast<bool>(available & vk::ImageAspectFlagBits::eDepth);
 	const bool has_stencil   = static_cast<bool>(available & vk::ImageAspectFlagBits::eStencil);
-	// LoadOp clears are separate from guest depth writes.
-	const bool depth_write   = depth.depth_write_enable;
+	// The attachment layout must permit load clears as well as draw writes.
+	const bool depth_write   = static_cast<bool>(writes & vk::ImageAspectFlagBits::eDepth);
 	const bool stencil_write = static_cast<bool>(writes & vk::ImageAspectFlagBits::eStencil);
 	if (!has_stencil) {
 		return depth_write ? vk::ImageLayout::eDepthAttachmentOptimal

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

@@ -329,7 +329,9 @@ struct VideoOutFormatPolicy {
 	VideoOutPixelFormatInfo info;
 };
 
-inline constexpr std::array<VideoOutFormatPolicy, 6> VIDEO_OUT_FORMAT_POLICIES {{
+inline constexpr uint64_t VIDEO_OUT_PIXEL_FORMAT_R10_G10_B10_A2_BT2100_PQ = 0x8100070422000000ull;
+
+inline constexpr std::array<VideoOutFormatPolicy, 7> VIDEO_OUT_FORMAT_POLICIES {{
     {0x8000000022000000ull,
      {vk::Format::eR8G8B8A8Srgb, Prospero::BufferFormat::k8_8_8_8Srgb, 4, false}},
     {0x8000000000000000ull,
@@ -338,6 +340,8 @@ inline constexpr std::array<VideoOutFormatPolicy, 6> VIDEO_OUT_FORMAT_POLICIES {
      {vk::Format::eA2B10G10R10UnormPack32, Prospero::BufferFormat::k10_10_10_2UNorm, 4, false}},
     {0x8100000000000000ull,
      {vk::Format::eA2R10G10B10UnormPack32, Prospero::BufferFormat::k10_10_10_2UNorm, 4, false}},
+    {VIDEO_OUT_PIXEL_FORMAT_R10_G10_B10_A2_BT2100_PQ,
+     {vk::Format::eA2B10G10R10UnormPack32, Prospero::BufferFormat::k10_10_10_2UNorm, 4, false}},
     {0xc001000622000000ull,
      {vk::Format::eR16G16B16A16Sfloat, Prospero::BufferFormat::k16_16_16_16Float, 8, false}},
     {0xc001000600000000ull,

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

@@ -17,6 +17,7 @@
 #include "graphics/host_gpu/hostMemory.h"
 #include "graphics/host_gpu/renderer/colorRenderTarget.h"
 #include "graphics/host_gpu/renderer/debug.h"
+#include "graphics/host_gpu/renderer/depthRenderTarget.h"
 #include "graphics/host_gpu/renderer/image/imageView.h"
 #include "graphics/host_gpu/renderer/image/textureCommon.h"
 #include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
@@ -749,18 +750,27 @@ void RenderExecutor::ResetBindings() {
 	m_bound_images.clear();
 }
 
-PreparedBindings RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime) {
+void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
+                                     PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(!runtime);
 	const auto& program  = *runtime.program;
 	const auto& snapshot = *runtime.resources;
-	PreparedBindings prepared;
 	prepared.runtime = &runtime;
-	prepared.images.reserve(program.info.images.size());
+	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();
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		auto binding = ResolveTexture(program.info.images[i], snapshot.images[i]);
 		BindImage(binding.image_id, binding.desc.type == TextureCache::BindingType::Storage);
-		prepared.images.push_back(std::move(binding));
+		binding.mip_views.swap(prepared.images[i].mip_views);
+		binding.mip_views.clear();
+		prepared.images[i] = std::move(binding);
 	}
 	prepared.samplers.reserve(program.info.samplers.size());
 	for (uint32_t i = 0; i < program.info.samplers.size(); i++) {
@@ -775,7 +785,6 @@ PreparedBindings RenderExecutor::PrepareBindings(const ShaderStageRuntime& runti
 	        program.bindings, ShaderRecompiler::IR::DescriptorBindingKind::Gds) != nullptr) {
 		prepared.gds.buffer = m_context.GetBufferCache().GetGdsBuffer()->Handle();
 	}
-	return prepared;
 }
 
 void RenderExecutor::FindBuffers(PreparedBindings& prepared) {
@@ -989,19 +998,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 					    std::ranges::find(image.views, binding.image_view, &CachedImageView::view);
 					EXIT_IF(storage || host_view == image.views.end());
 					const auto aspect = host_view->info.aspect;
-					const bool depth_feedback =
-					    layout == vk::ImageLayout::eAttachmentFeedbackLoopOptimalEXT &&
-					    program.stage == ShaderType::Pixel;
-					const bool depth_read =
-					    depth_feedback || layout == vk::ImageLayout::eDepthReadOnlyOptimal ||
-					    layout == vk::ImageLayout::eDepthStencilReadOnlyOptimal ||
-					    layout == vk::ImageLayout::eDepthReadOnlyStencilAttachmentOptimal;
-					const bool stencil_read =
-					    layout == vk::ImageLayout::eStencilReadOnlyOptimal ||
-					    layout == vk::ImageLayout::eDepthStencilReadOnlyOptimal ||
-					    layout == vk::ImageLayout::eDepthAttachmentStencilReadOnlyOptimal;
-					if ((aspect & vk::ImageAspectFlagBits::eDepth && !depth_read) ||
-					    (aspect & vk::ImageAspectFlagBits::eStencil && !stencil_read)) {
+					if (aspect & ~DepthReadableAspects(layout)) {
 						EXIT("sampling a writable depth/stencil attachment aspect\n");
 					}
 				}

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

@@ -204,6 +204,7 @@ struct PipelineCache::ProgramCache {
 		ShaderRecompiler::IR::ResourceSnapshot       resources;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		std::vector<Permutation>                    permutations;
+		bool                                        skip_dispatch = false;
 	};
 
 	struct ProgramKeyHash {
@@ -283,6 +284,9 @@ struct PipelineCache::ProgramCache {
 		lookup_key.code_size       = static_cast<uint32_t>(params.code.size());
 		BuildStageStaticKey(input_info, lookup_key.static_state);
 		auto                                         entry = programs.find(lookup_key);
+		if (entry != programs.end() && entry->second.skip_dispatch) {
+			return {};
+		}
 		const ShaderRecompiler::IR::SrtRuntime       runtime {
 		    .user_data                  = user_data,
 		    .shader_base                = params.Base(),
@@ -348,6 +352,11 @@ struct PipelineCache::ProgramCache {
 			options.wave_size = input_info.wave_size;
 		}
 		auto translated = ShaderRecompiler::TranslateProgram(params.code, options);
+		if (translated.skip_dispatch) {
+			entry = programs.try_emplace(lookup_key, ShaderRecompiler::IR::ResourcePlan {}).first;
+			entry->second.skip_dispatch = true;
+			return {};
+		}
 		if (entry == programs.end()) {
 			entry = programs.try_emplace(lookup_key,
 			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program)).first;

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

@@ -368,6 +368,8 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	// MoltenVK lacks VK_EXT_depth_clip_enable; omit the depth-clip struct on macOS and accept
 	// Vulkan's default depth clipping (enabled) instead of the PS5's clamp behavior.
 #if !defined(__APPLE__)
+	// The DB clamps depth to the viewport range after polygon offset is applied.
+	rasterizer.depthClampEnable = VK_TRUE;
 	rasterizer.pNext = &clip_ext;
 #endif
 	vk::PipelineRasterizationProvokingVertexStateCreateInfoEXT provoking_vertex {};

src/graphics/host_gpu/renderer/render.h

@@ -160,7 +160,7 @@ public:
 	void DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer, uint64_t args_addr,
 	                      uint32_t mode);
 
-	[[nodiscard]] PreparedBindings PrepareBindings(const ShaderStageRuntime& runtime);
+	void PrepareBindings(const ShaderStageRuntime& runtime, PreparedBindings& prepared);
 	void                           FindBuffers(PreparedBindings& bindings);
 	void                           RebindBuffers(PreparedBindings& bindings);
 	void                           RebindImages(PreparedBindings& bindings);
@@ -196,7 +196,8 @@ private:
 	                         bool primitive_restart_enable);
 	[[nodiscard]] RenderState AcquireRenderTargets(CommandBuffer& buffer, RenderColorInfo* colors,
 	                                               uint32_t color_count, RenderDepthInfo& depth,
-	                                               const std::optional<PreparedBindings>& pixel = std::nullopt);
+	                                               vk::ImageAspectFlags& feedback_aspects,
+	                                               std::span<PreparedBindings* const> stages = {});
 	[[nodiscard]] bool        ResolveColorTargets(CommandBuffer& buffer,
 	                                              uint32_t render_target_slice_offset);
 	void                      BindImage(ImageId id, bool storage);
@@ -209,6 +210,8 @@ private:
 	                                              uint32_t group_y, uint32_t group_z, uint32_t mode);
 
 	RenderContext&                        m_context;
+	GraphicsBindings                     m_graphics_bindings;
+	PreparedBindings                     m_compute_bindings;
 	std::vector<ImageId>                  m_bound_images;
 	std::vector<vk::DescriptorBufferInfo> m_descriptor_buffers;
 	std::vector<vk::DescriptorImageInfo>  m_descriptor_images;

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -256,6 +256,10 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.dispatch_thread_dimensions = use_thread_dimensions;
 	const auto compute_program =
 	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
+	if (!compute_program) {
+		// Temporary until RT is implemented.
+		return;
+	}
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
@@ -362,7 +366,8 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	buffer.EndRendering();
 	auto& pipeline =
 	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
-	auto bindings = PrepareBindings(input_info.stage);
+	auto& bindings = m_compute_bindings;
+	PrepareBindings(input_info.stage, bindings);
 	FindBuffers(bindings);
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
@@ -412,9 +417,14 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	ShaderComputeInputInfo input_info {};
 	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
 	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
+	if (!compute_program) {
+		// Temporary until RT is implemented.
+		return;
+	}
 	buffer.EndRendering();
 	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
-	auto bindings = PrepareBindings(input_info.stage);
+	auto& bindings = m_compute_bindings;
+	PrepareBindings(input_info.stage, bindings);
 	FindBuffers(bindings);
 	const auto& program = *input_info.stage.program;
 	if (program.info.uses_dma) {

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -453,8 +453,10 @@ struct DrawCallInfo {
 
 RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderColorInfo* colors,
                                                  uint32_t color_count, RenderDepthInfo& depth,
-                                                 const std::optional<PreparedBindings>& pixel) {
+                                                 vk::ImageAspectFlags& feedback_aspects,
+                                                 std::span<PreparedBindings* const> stages) {
 	EXIT_IF(colors == nullptr || color_count > RENDER_COLOR_ATTACHMENTS_MAX);
+	feedback_aspects = {};
 	auto&       cache = m_context.GetTextureCache();
 	RenderState state {};
 	state.width                 = std::numeric_limits<uint32_t>::max();
@@ -529,28 +531,29 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		                     image.info.data.address, depth.desc.view_info.base_layer,
 		                     depth.desc.view_info.layer_count);
 		EXIT_IF(image_view == nullptr || image.backing.samples != depth.desc.info.samples);
-		const bool feedback = depth.depth_write_enable && pixel &&
-		    std::ranges::any_of(pixel->images, [&](const TextureBinding& binding) {
-			    if (binding.image_id != depth.image_id ||
-			        binding.desc.type != TextureCache::BindingType::Texture) {
-				    return false;
-			    }
-			    const auto native =
-			        std::ranges::find(image.views, binding.image_view, &CachedImageView::view);
-			    EXIT_IF(native == image.views.end());
-			    const auto& sampled = native->info;
-			    const auto& target = depth.desc.view_info;
-			    return (sampled.aspect & vk::ImageAspectFlagBits::eDepth) &&
-			           ImageRangeOverlaps(sampled.base_level, sampled.level_count,
-			                              target.base_level, target.level_count) &&
-			           ImageRangeOverlaps(sampled.base_layer, sampled.layer_count,
-			                              target.base_layer, target.layer_count);
-		    });
-		if (feedback && !m_context.GetGraphics().attachment_feedback_loop_enabled) {
+		const auto draw_writes = depth.AttachmentWriteAspects();
+		vk::ImageAspectFlags sampled_aspects;
+		for (const auto* stage: stages) {
+			for (const auto& binding: stage->images) {
+				if (binding.image_id != depth.image_id ||
+				    binding.desc.type != TextureCache::BindingType::Texture) continue;
+				const auto native =
+				    std::ranges::find(image.views, binding.image_view, &CachedImageView::view);
+				EXIT_IF(native == image.views.end());
+				sampled_aspects |= native->info.aspect;
+				feedback_aspects |= DepthFeedbackAspects(draw_writes, depth.desc.view_info,
+				                                         native->info);
+			}
+		}
+		if (feedback_aspects && !m_context.GetGraphics().attachment_feedback_loop_enabled) {
 			EXIT("depth attachment feedback loop is not supported by the host\n");
 		}
-		const auto layout = feedback ? vk::ImageLayout::eAttachmentFeedbackLoopOptimalEXT
-		                             : depth_attachment_layout(depth);
+		auto layout = depth_attachment_layout(depth);
+		if (sampled_aspects & ~DepthReadableAspects(layout)) {
+			layout = m_context.GetGraphics().attachment_feedback_loop_enabled
+			             ? vk::ImageLayout::eAttachmentFeedbackLoopOptimalEXT
+			             : vk::ImageLayout::eGeneral;
+		}
 		// The attachment store writes even when guest depth/stencil tests do not.
 		const auto access = vk::AccessFlagBits2::eDepthStencilAttachmentRead |
 		                    vk::AccessFlagBits2::eDepthStencilAttachmentWrite;
@@ -683,11 +686,13 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 	        vs_input_info.buffers_num > ShaderVertexInputInfo::RES_MAX);
 
 	// Collect the non-empty guest vertex ranges.
+	std::array<uint64_t, ShaderVertexInputInfo::RES_MAX>          sizes {};
 	std::array<VertexBufferRange, ShaderVertexInputInfo::RES_MAX> ranges {};
 	uint32_t                                                      range_count = 0;
 	for (int i = 0; i < vs_input_info.buffers_num; i++) {
 		const auto& vertex = vs_input_info.buffers[i];
 		const auto  size   = VertexBufferDescriptorSize(vertex, vs_input_info);
+		sizes[i]           = size;
 		if (size == 0) {
 			continue;
 		}
@@ -736,7 +741,7 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 	vk::Buffer null_buffer = nullptr;
 	for (int i = 0; i < vs_input_info.buffers_num; i++) {
 		const auto& vertex = vs_input_info.buffers[i];
-		const auto  size   = VertexBufferDescriptorSize(vertex, vs_input_info);
+		const auto  size   = sizes[i];
 		if (size == 0) {
 			if (null_buffer == nullptr) {
 				null_buffer = cache.GetBuffer(NULL_BUFFER_ID).Handle();
@@ -1073,15 +1078,16 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		                              index_source.guest_element_size);
 	}
 	LogDrawPhase(draw.Name(), "PrepareBindings");
-	GraphicsBindings                 bindings;
+	auto&                            bindings = m_graphics_bindings;
 	std::array<PreparedBindings*, 4> descriptor_stages {};
 	uint32_t                         stage_count = 0;
 	for (uint32_t i = 0; i < vertex_stages.size(); i++) {
-		bindings.vertex[i]               = PrepareBindings(state.vertex_info[i].stage);
+		PrepareBindings(state.vertex_info[i].stage, bindings.vertex[i]);
 		descriptor_stages[stage_count++] = &bindings.vertex[i];
 	}
 	if (state.ps_active) {
-		bindings.pixel.emplace(PrepareBindings(state.ps_input_info.stage));
+		if (!bindings.pixel) bindings.pixel.emplace();
+		PrepareBindings(state.ps_input_info.stage, *bindings.pixel);
 		descriptor_stages[stage_count++] = &*bindings.pixel;
 	}
 	const auto stages = std::span {descriptor_stages.data(), stage_count};
@@ -1100,9 +1106,10 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	    std::span {state.color_info, state.color_count}, state.depth_info, vertex_stages, buffer,
 	    state.ps_active ? &state.ps_input_info : nullptr, topology, primitive_restart_enable,
 	    state.programs);
+	vk::ImageAspectFlags feedback_aspects;
 	const auto rendering =
 	    AcquireRenderTargets(buffer, state.color_info, state.color_count, state.depth_info,
-	                         bindings.pixel);
+	                         feedback_aspects, stages);
 
 	// Resource preparation above may synchronously finish and restart the scheduler. From this
 	// point onward, every operation targets the current command buffer and cannot touch guest
@@ -1112,7 +1119,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (!mesh_active) {
 		CommitVertexBuffers(vk_buffer, vertex_bindings);
 	}
-	if (bindings.pixel && !draw.IsIndexed()) {
+	if (state.ps_active && !draw.IsIndexed()) {
 		SetDrawDebugPhase(buffer, submit_id, draw, 0x300u);
 	}
 	CommitBindings(buffer, vk::PipelineBindPoint::eGraphics, pipeline, stages);
@@ -1134,11 +1141,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 
 	SetGraphicsDynamicParams(buffer, vk_buffer, vertex_stages.back(), state.depth_info, rendering);
 	if (m_context.GetGraphics().attachment_feedback_loop_enabled) {
-		vk_buffer.setAttachmentFeedbackLoopEnableEXT(
-		    rendering.depth_stencil_attachment.image_layout ==
-		            vk::ImageLayout::eAttachmentFeedbackLoopOptimalEXT
-		        ? vk::ImageAspectFlags {vk::ImageAspectFlagBits::eDepth}
-		        : vk::ImageAspectFlags {});
+		vk_buffer.setAttachmentFeedbackLoopEnableEXT(feedback_aspects);
 	}
 
 	LogDrawPhase(draw.Name(), "BeginRendering");

src/graphics/presentation/videoOut.cpp

@@ -24,6 +24,8 @@
 
 #include <algorithm>
 #include <array>
+#include <atomic>
+#include <fmt/format.h>
 #include <list>
 #include <thread>
 #include <vector>
@@ -541,6 +543,14 @@ Graphics::ImageInfo BufferAttributeGroup::ImageInfo(const VideoOutBuffer& buffer
 	if (!Graphics::DecodeVideoOutPixelFormat(attribute.pixel_format, pixel_format)) {
 		EXIT("unsupported video-out pixel format: 0x%016" PRIx64 "\n", attribute.pixel_format);
 	}
+	if (attribute.pixel_format == Graphics::VIDEO_OUT_PIXEL_FORMAT_R10_G10_B10_A2_BT2100_PQ) {
+		static std::atomic_flag warned = ATOMIC_FLAG_INIT;
+		if (!warned.test_and_set(std::memory_order_relaxed)) {
+			Log::WriteToConsoleAndLog(fmt::format(
+			    "Warning: HDR-to-SDR conversion is not implemented; displaying PQ video-out "
+			    "pixels unchanged (format=0x{:016x}).\n", attribute.pixel_format));
+		}
+	}
 	const auto tile_mode = Graphics::Prospero::TileMode::kRenderTarget;
 	const auto pitch =
 	    Graphics::TileGetTexturePitch(pixel_format.guest_format, attribute.width, tile_mode);
@@ -1546,7 +1556,9 @@ KYTY_SYSV_ABI int VideoOutIsFlipPending(int handle) {
 	VideoOutFlipStatus status {};
 	DriverState().GetFlipQueue().GetFlipStatus(*ctx, status);
 
-	LOGF("\t flipPendingNum = %d\n", status.flipPendingNum);
+	if (Config::GraphicsDebugDumpEnabled()) {
+		LOGF("\t flipPendingNum = %d\n", status.flipPendingNum);
+	}
 
 	return status.flipPendingNum;
 }
@@ -1680,21 +1692,6 @@ KYTY_SYSV_ABI int VideoOutGetOutputStatus(int handle, VideoOutOutputStatus* stat
 	return OK;
 }
 
-KYTY_SYSV_ABI int VideoOutGetVrrStatus(int handle, int32_t* status) {
-	PRINT_NAME();
-
-	if (status == nullptr) {
-		return VIDEO_OUT_ERROR_INVALID_ADDRESS;
-	}
-	if (DriverState().Get(handle) == nullptr) {
-		return VIDEO_OUT_ERROR_INVALID_HANDLE;
-	}
-
-	// Kyty currently presents at a fixed refresh rate and does not negotiate VRR.
-	*status = 0;
-	return OK;
-}
-
 static int ValidateOutputConfig(int handle, uint64_t mode, const VideoOutOutputOptions* options,
                                 void* reserved_ptr, uint64_t reserved) {
 	if (!DriverState().IsOpened(handle)) {

src/graphics/presentation/videoOut.h

@@ -87,7 +87,6 @@ KYTY_SYSV_ABI int VideoOutGetEventData(const LibKernel::EventQueue::KernelEvent*
 KYTY_SYSV_ABI int VideoOutGetEventCount(const LibKernel::EventQueue::KernelEvent* ev);
 KYTY_SYSV_ABI int VideoOutWaitVblank(int handle);
 KYTY_SYSV_ABI int VideoOutGetOutputStatus(int handle, VideoOutOutputStatus* status);
-KYTY_SYSV_ABI int VideoOutGetVrrStatus(int handle, int32_t* status);
 KYTY_SYSV_ABI int VideoOutInitializeOutputOptions(VideoOutOutputOptions* options);
 KYTY_SYSV_ABI int VideoOutIsOutputSupported(int handle, uint64_t mode,
                                             const VideoOutOutputOptions* options,

src/graphics/presentation/window/vulkanWindow.cpp

@@ -238,6 +238,10 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 #endif
 		}
 #if !defined(__APPLE__)
+		if (device_features2.features.depthClamp != VK_TRUE) {
+			LOGF("depthClamp is not supported\n");
+			skip_device = true;
+		}
 		if (fragment_barycentric.fragmentShaderBarycentric != VK_TRUE) {
 			LOGF("fragmentShaderBarycentric is not supported\n");
 			skip_device = true;
@@ -607,6 +611,7 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	device_features.robustBufferAccess       = VK_TRUE;
 #if !defined(__APPLE__)
 	device_features.depthBounds = VK_TRUE; // unsupported by MoltenVK
+	device_features.depthClamp  = VK_TRUE;
 #endif
 	device_features.shaderStorageImageWriteWithoutFormat = VK_TRUE;
 	device_features.shaderImageGatherExtended            = VK_TRUE;

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -20,6 +20,7 @@
 
 #include <algorithm>
 #include <array>
+#include <atomic>
 #include <chrono>
 #include <fmt/format.h>
 #include <map>
@@ -523,6 +524,20 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
 	     static_cast<uint64_t>(decoded.instructions.size()), phase_ms());
 
+	// Temporary workaround for games that compile ray-tracing shaders before
+	// the player can select a mode without ray tracing.
+	if (options.stage == ShaderType::Compute && decoded.has_bvh) {
+		static std::atomic_flag warned = ATOMIC_FLAG_INIT;
+		if (!warned.test_and_set(std::memory_order_relaxed)) {
+			const auto& bvh = decoded.instructions.back();
+			Log::WriteToConsoleAndLog(fmt::format(
+			    "Warning: ray tracing is not implemented; skipping compute dispatches containing "
+			    "BVH intersection instructions (shader=0x{:016x}, pc=0x{:08x}, opcode=0x{:02x}).\n",
+			    options.shader_hash, bvh.pc, bvh.opcode_id));
+		}
+		return {.skip_dispatch = true};
+	}
+
 	std::string decoded_dump;
 	if (options.dump_ir) {
 		decoded_dump = Decoder::ProgramToString(decoded);
@@ -615,6 +630,7 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 CompileResult CompileProgram(TranslateResult translated, const CompileOptions& options,
                              const IR::ResourceSpecialization& specialization,
                              uint32_t push_data_start_dword) {
+	EXIT_IF(translated.skip_dispatch);
 	const auto emit_begin = std::chrono::steady_clock::now();
 	auto& ir = translated.program;
 	IR::ApplyResourceSpecialization(ir, specialization);

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -28,6 +28,7 @@ struct TranslateResult {
 	IR::Program program;
 	std::string decoded_dump;
 	std::string cfg_dump;
+	bool        skip_dispatch = false;
 };
 
 struct CompileResult {

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

@@ -227,6 +227,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					requirements.subgroup_local_invocation_id = true;
 				}
 				if (memory.kind == IR::ResourceKind::Buffer) {
+					requirements.coherent_buffers |= memory.coherent;
 					if (memory.resource >= program.info.buffers.size()) {
 						Fail(program, "buffer operation has invalid resource metadata");
 					}

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

@@ -784,11 +784,11 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		                         ? &state.program.descriptor_sources[image.source]
 		                         : nullptr;
 		if (handle == nullptr || source == nullptr || !source->indirect_image.has_value() ||
-		    source->indirect_image->key_arg >= handle->NumArgs()) {
+		    handle->NumArgs() == 0u) {
 			ctx.Fail(inst, "has invalid indirect image key provenance");
 			return;
 		}
-		const auto key = ctx.Def(handle->Arg(source->indirect_image->key_arg));
+		const auto key = ctx.Def(handle->Arg(0));
 		if (state.flattened_srt_variable == 0 || image.indirect_search_iterations == 0u ||
 		    image.indirect_resources.size() < 2u) {
 			ctx.Fail(inst, "has no indirect image runtime mapping");

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

@@ -67,6 +67,7 @@ struct SpirvRequirements {
 	bool function_scratch             = false;
 	bool pixel_valid_mask             = false;
 	bool buffer_int64_atomics         = false;
+	bool coherent_buffers             = false;
 };
 
 SpirvRequirements AnalyzeProgramRequirements(const IR::Program& program);
@@ -371,12 +372,13 @@ void EmitMemoryOffsets(EmitterState& state);
 uint32_t LdsDwordCount(const EmitterState& state);
 
 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;
+	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;
+	spv::MemoryAccessMask memory_access    = spv::MemoryAccessMaskNone;
 };
 
 MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::MemoryInfo& mem);

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

@@ -313,7 +313,8 @@ uint32_t LoadWordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess& res
                           uint32_t index) {
 	const auto value   = ctx.state.builder.AllocateId();
 	const auto pointer = EmitMemoryElementPointer(ctx.state, resource, index);
-	ctx.state.builder.AddFunction(spv::OpLoad, TypeU32(ctx.state), value, pointer);
+	ctx.state.builder.AddFunction(spv::OpLoad, TypeU32(ctx.state), value, pointer,
+	                              resource.memory_access);
 	return value;
 }
 
@@ -498,21 +499,21 @@ void StoreSubword(ValueEmitContext& ctx, const IR::Inst& inst, IR::MemoryInfo me
 	});
 }
 
+void StoreWordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess& resource, uint32_t index,
+                       uint32_t data) {
+	ctx.state.builder.AddFunction(spv::OpStore,
+	                              EmitMemoryElementPointer(ctx.state, resource, index), data,
+	                              resource.memory_access);
+}
+
 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));
 	EmitIfCondition(ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index), [&]() {
-		ctx.state.builder.AddFunction(spv::OpStore,
-		                              EmitMemoryElementPointer(ctx.state, resource, index), data);
+		StoreWordInBounds(ctx, resource, index, data);
 	});
 }
 
-void StoreWordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess& resource, uint32_t index,
-                       uint32_t data) {
-	ctx.state.builder.AddFunction(spv::OpStore,
-	                              EmitMemoryElementPointer(ctx.state, 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);
@@ -710,10 +711,14 @@ void StoreFormattedInBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
                             uint32_t data) {
 	if (component >= plan.info.component_count) return;
 	const auto bits = plan.info.component_bits[component];
-	if (plan.info.type == Format::ComponentType::Snorm && bits == 16u) {
+	if (bits == 16u && (plan.info.type == Format::ComponentType::Snorm ||
+	                    plan.info.type == Format::ComponentType::Float)) {
 		const auto value = EmitBitCastF32U32(ctx.state, data);
-		data = EmitPackSnorm2x16(
-		    ctx.state, EmitCompositeConstructF32x2(ctx.state, value, ConstantF32Value(ctx.state, 0.0f)));
+		const auto pair = EmitCompositeConstructF32x2(ctx.state, value,
+		                                               ConstantF32Value(ctx.state, 0.0f));
+		data = plan.info.type == Format::ComponentType::Float
+		           ? EmitPackHalf2x16(ctx.state, pair)
+		           : EmitPackSnorm2x16(ctx.state, pair);
 	}
 	if (bits == 8u || bits == 16u) {
 		StoreSubwordInBounds(ctx, mem, plan.resource, plan.addresses[component],

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

@@ -86,7 +86,9 @@ MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState& state,
 	}
 	const auto array_index =
 	    ResourceForDescriptor(state, IR::DescriptorBindingKind::Buffers, mem.resource);
-	MemoryResourceAccess access {.kind = mem.kind};
+	MemoryResourceAccess access {
+	    .kind = mem.kind,
+	    .memory_access = mem.coherent ? spv::MemoryAccessVolatileMask : spv::MemoryAccessMaskNone};
 	access.object_pointer = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpAccessChain, pointer_type, access.object_pointer, variable,
 	                          ConstantU32(state, array_index));

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

@@ -201,6 +201,16 @@ void DefineDescriptors(EmitterState& state) {
 					state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_u64_variable,
 					                            spv::DecorationAliased);
 				}
+				if (state.requirements.coherent_buffers) {
+					// 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);
+					}
+				}
 				break;
 			case IR::DescriptorBindingKind::BdaPagetable:
 				state.bda_pagetable_variable = Define(StorageBufferU64Type(state), "bda_pagetable");

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

@@ -21,6 +21,7 @@ constexpr MemoryOpcodeInfo SMEM_OPCODE_LIST[] = {
     {0x04u, Opcode::S_LOAD_DWORDX16, 16, 32},      {0x08u, Opcode::S_BUFFER_LOAD_DWORD, 1, 32},
     {0x09u, Opcode::S_BUFFER_LOAD_DWORDX2, 2, 32}, {0x0au, Opcode::S_BUFFER_LOAD_DWORDX4, 4, 32},
     {0x0bu, Opcode::S_BUFFER_LOAD_DWORDX8, 8, 32}, {0x0cu, Opcode::S_BUFFER_LOAD_DWORDX16, 16, 32},
+    {0x25u, Opcode::S_MEMREALTIME, 2, 32},
 };
 
 constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
@@ -86,6 +87,7 @@ constexpr MemoryOpcodeInfo FLAT_OPCODE_LIST[] = {
 
 constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x00u, Opcode::DS_ADD_U32, 1, 32},          {0x01u, Opcode::DS_SUB_U32, 1, 32},
+    {0x03u, Opcode::DS_INC_U32, 1, 32},          {0x04u, Opcode::DS_DEC_U32, 1, 32},
     {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},
@@ -107,6 +109,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x4du, Opcode::DS_WRITE_B64, 2, 32},        {0x4eu, Opcode::DS_WRITE2_B64, 4, 32},
     {0x4fu, Opcode::DS_WRITE2ST64_B64, 4, 32},   {0x76u, Opcode::DS_READ_B64, 2, 32},
     {0x77u, Opcode::DS_READ2_B64, 4, 32},        {0x78u, Opcode::DS_READ2ST64_B64, 4, 32},
+    {0xa0u, Opcode::DS_WRITE_B8_D16_HI, 1, 8},
     {0xa1u, Opcode::DS_WRITE_B16_D16_HI, 1, 16},
     {0xa6u, Opcode::DS_READ_U16_D16, 1, 16},
     {0xa7u, Opcode::DS_READ_U16_D16_HI, 1, 16},
@@ -147,6 +150,7 @@ bool IsDsWriteOpcode(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::DS_WRITE_B8:
 		case Opcode::DS_WRITE_B16:
+		case Opcode::DS_WRITE_B8_D16_HI:
 		case Opcode::DS_WRITE_B16_D16_HI:
 		case Opcode::DS_WRITE2_B32:
 		case Opcode::DS_WRITE2ST64_B32:
@@ -166,7 +170,9 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 		case Opcode::DS_ADD_RTN_U32:
 		case Opcode::DS_SUB_U32:
 		case Opcode::DS_SUB_RTN_U32:
+		case Opcode::DS_INC_U32:
 		case Opcode::DS_INC_RTN_U32:
+		case Opcode::DS_DEC_U32:
 		case Opcode::DS_DEC_RTN_U32:
 		case Opcode::DS_MIN_I32:
 		case Opcode::DS_MIN_RTN_I32:
@@ -240,6 +246,10 @@ void DecodeSmem(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	}
 
 	DecodeScalarDestination(sdst, pc, inst.dst);
+	if (inst.opcode == Opcode::S_MEMREALTIME) {
+		inst.src_count = 0;
+		return;
+	}
 	// SMEM encodes SBASE in SGPR pairs. Scalar-buffer loads still use the same
 	// pair index; their descriptor operand consumes four SGPRs from that base.
 	DecodeScalarSource(sbase * 2u, pc, inst.src0);
@@ -263,6 +273,7 @@ void DecodeMubuf(uint32_t pc, std::span<const uint32_t> code, uint32_t word_inde
 	inst.idxen       = ((word0 >> 13u) & 1u) != 0;
 	inst.offen       = ((word0 >> 12u) & 1u) != 0;
 	inst.glc         = ((word0 >> 14u) & 1u) != 0;
+	inst.dlc         = ((word0 >> 15u) & 1u) != 0;
 	inst.slc         = ((word1 >> 22u) & 1u) != 0;
 	inst.family      = Family::MUBUF;
 	inst.opcode_id   = opcode;
@@ -298,6 +309,7 @@ void DecodeMtbuf(uint32_t pc, std::span<const uint32_t> code, uint32_t word_inde
 	inst.idxen         = ((word0 >> 13u) & 1u) != 0;
 	inst.offen         = ((word0 >> 12u) & 1u) != 0;
 	inst.glc           = ((word0 >> 14u) & 1u) != 0;
+	inst.dlc           = ((word0 >> 15u) & 1u) != 0;
 	inst.slc           = ((word1 >> 22u) & 1u) != 0;
 	inst.family        = Family::MTBUF;
 	inst.opcode_id     = opcode;
@@ -429,7 +441,9 @@ void DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 	}
 	DecodeVectorGpr(addr, inst.src0);
 	DecodeVectorGpr(data0, inst.src1);
-	if (inst.opcode == Opcode::DS_WRITE_B16_D16_HI) {
+	if (inst.opcode == Opcode::DS_WRITE_B8_D16_HI) {
+		inst.src1.sdwa_sel = 2u;
+	} else if (inst.opcode == Opcode::DS_WRITE_B16_D16_HI) {
 		inst.src1.sdwa_sel = 5u;
 	}
 	DecodeVectorGpr(data1, inst.src2);

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

@@ -73,11 +73,11 @@ std::string FormatMemory(const Instruction& inst) {
 		text += OperandToString(*sources[i]);
 	}
 	text += fmt::format(" ; offset={} offset2={} dwords={} bits={} dfmt={} nfmt={} signed={} "
-	                    "typed={} formatted={} segment={} glc={} slc={} idxen={} offen={}",
+	                    "typed={} formatted={} segment={} glc={} dlc={} slc={} idxen={} offen={}",
 	                    inst.offset, inst.secondary_offset, inst.data_dwords, inst.data_bits,
 	                    inst.data_format, inst.number_format, inst.data_signed ? 1u : 0u,
 	                    inst.typed ? 1u : 0u, inst.formatted ? 1u : 0u, inst.memory_segment,
-	                    inst.glc ? 1u : 0u, inst.slc ? 1u : 0u, inst.idxen ? 1u : 0u,
+	                    inst.glc ? 1u : 0u, inst.dlc ? 1u : 0u, inst.slc ? 1u : 0u, inst.idxen ? 1u : 0u,
 	                    inst.offen ? 1u : 0u);
 	return text;
 }
@@ -403,6 +403,7 @@ void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 	program.instructions.clear();
 	program.instructions.reserve(code.size());
 	program.code = code;
+	program.has_bvh = false;
 
 	std::vector<bool> branch_targets;
 	for (uint32_t word_index = 0; word_index < code.size();) {
@@ -411,6 +412,10 @@ void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 
 		const auto& inst = program.instructions.back();
 		word_index += inst.word_count;
+		if (inst.family == Family::MIMG && (inst.opcode_id == 0xe6u || inst.opcode_id == 0xe7u)) {
+			program.has_bvh = true;
+			return;
+		}
 
 		if (IsDirectBranch(inst.opcode)) {
 			const auto target_index = inst.branch_target / sizeof(uint32_t);
@@ -659,7 +664,9 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::DS_ADD_RTN_U32:
 		case Opcode::DS_SUB_U32:
 		case Opcode::DS_SUB_RTN_U32:
+		case Opcode::DS_INC_U32:
 		case Opcode::DS_INC_RTN_U32:
+		case Opcode::DS_DEC_U32:
 		case Opcode::DS_DEC_RTN_U32:
 		case Opcode::DS_MIN_I32:
 		case Opcode::DS_MIN_RTN_I32:
@@ -694,6 +701,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::DS_READ_B128:
 		case Opcode::DS_WRITE_B8:
 		case Opcode::DS_WRITE_B16:
+		case Opcode::DS_WRITE_B8_D16_HI:
 		case Opcode::DS_WRITE_B16_D16_HI:
 		case Opcode::DS_WRITE2_B32:
 		case Opcode::DS_WRITE2ST64_B32:

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

@@ -404,6 +404,7 @@ enum class Opcode {
 	V_CMP_LE_U16,
 	V_CMP_GT_U16,
 	V_CMPX_LT_U16,
+	V_CMPX_EQ_U16,
 	V_CMPX_GT_U16,
 	V_CMP_NE_U16,
 	V_CMP_GE_U16,
@@ -439,6 +440,7 @@ enum class Opcode {
 	S_BUFFER_LOAD_DWORDX4,
 	S_BUFFER_LOAD_DWORDX8,
 	S_BUFFER_LOAD_DWORDX16,
+	S_MEMREALTIME,
 	BUFFER_LOAD_FORMAT_X,
 	BUFFER_LOAD_FORMAT_XY,
 	BUFFER_LOAD_FORMAT_XYZ,
@@ -502,7 +504,9 @@ enum class Opcode {
 	DS_ADD_RTN_U32,
 	DS_SUB_U32,
 	DS_SUB_RTN_U32,
+	DS_INC_U32,
 	DS_INC_RTN_U32,
+	DS_DEC_U32,
 	DS_DEC_RTN_U32,
 	DS_MIN_I32,
 	DS_MIN_RTN_I32,
@@ -541,6 +545,7 @@ enum class Opcode {
 	DS_READ_B128,
 	DS_WRITE_B8,
 	DS_WRITE_B16,
+	DS_WRITE_B8_D16_HI,
 	DS_WRITE_B16_D16_HI,
 	DS_WRITE2_B32,
 	DS_WRITE2ST64_B32,
@@ -708,6 +713,7 @@ struct Instruction {
 	bool           formatted                                    = false;
 	bool           gds                                          = false;
 	bool           glc                                          = false;
+	bool           dlc                                          = false;
 	bool           slc                                          = false;
 	bool           idxen                                        = false;
 	bool           offen                                        = false;
@@ -726,6 +732,7 @@ struct Instruction {
 struct Program {
 	std::span<const uint32_t> code;
 	std::vector<Instruction>  instructions;
+	bool                     has_bvh = false;
 };
 
 // Code spans are trusted to contain complete instructions, valid branch targets, and 32-bit PCs.

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

@@ -40,7 +40,7 @@ constexpr Vop2OpcodeInfo VOP2_OPCODE_LIST[] = {
     {0x0fu, Opcode::V_MIN_F32},
     {0x10u, Opcode::V_MAX_F32},
     {0x11u, Opcode::V_MIN_I32},
-    {0x12u, Opcode::V_MAX_I32},
+    {0x12u, Opcode::V_MAX_I32, Vop2SdwaProfile::IntegerFullDestination},
     {0x13u, Opcode::V_MIN_U32, Vop2SdwaProfile::IntegerPartialDestination},
     {0x14u, Opcode::V_MAX_U32, Vop2SdwaProfile::IntegerFullDestination},
     {0x15u, Opcode::V_LSHR_B32},
@@ -231,7 +231,7 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0xa9u, Opcode::V_CMP_LT_U16},         {0xaau, Opcode::V_CMP_EQ_U16},
     {0xabu, Opcode::V_CMP_LE_U16},         {0xacu, Opcode::V_CMP_GT_U16},
     {0xadu, Opcode::V_CMP_NE_U16},         {0xaeu, Opcode::V_CMP_GE_U16},
-    {0xb9u, Opcode::V_CMPX_LT_U16, false},
+    {0xb9u, Opcode::V_CMPX_LT_U16, false}, {0xbau, Opcode::V_CMPX_EQ_U16, false},
     {0xbcu, Opcode::V_CMPX_GT_U16},        {0xc0u, Opcode::V_CMP_F_U32},
     {0xc1u, Opcode::V_CMP_LT_U32},         {0xc2u, Opcode::V_CMP_EQ_U32},
     {0xc3u, Opcode::V_CMP_LE_U32},         {0xc4u, Opcode::V_CMP_GT_U32},
@@ -1454,6 +1454,7 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_NE_I64:
 		case Opcode::V_CMPX_NE_U64:
 		case Opcode::V_CMPX_LT_U16:
+		case Opcode::V_CMPX_EQ_U16:
 		case Opcode::V_CMPX_GT_U16:
 		case Opcode::V_CMPX_LT_F16:
 		case Opcode::V_CMPX_EQ_F16:

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

@@ -302,23 +302,95 @@ void Translator::S_MOV_B64(const Decoder::Instruction& inst) {
 }
 
 void Translator::S_WQM(const Decoder::Instruction& inst, bool wide) {
-	const auto source = wide ? ReadU64(inst.src0)
-	                         : ir.ConstructU64(ReadU32(inst.src0), IR::U32(IR::Value(0u)));
+	const bool source_is_mask = inst.src0.kind == Decoder::OperandKind::Sgpr ||
+	                            inst.src0.kind == Decoder::OperandKind::ExecLo ||
+	                            inst.src0.kind == Decoder::OperandKind::VccLo ||
+	                            (!wide && IsExecOrVcc(inst.src0));
+	const auto source_valid = ReadMaskValid(inst.src0);
+	auto source_live = IR::U1(IR::Value(false));
+	if (source_is_mask) {
+		if (inst.src0.kind == Decoder::OperandKind::VccLo ||
+		    inst.src0.kind == Decoder::OperandKind::VccHi) {
+			source_live = ir.GetVcc();
+		} else {
+			source_live = ReadMask(inst.src0);
+		}
+	}
+	const auto retained_live = ir.LogicalAnd(source_valid, source_live);
+
+	IR::U64 source;
+	if (wide) {
+		source = ReadU64(inst.src0);
+	} else {
+		source = ir.ConstructU64(ReadU32(inst.src0), IR::U32(IR::Value(0u)));
+	}
 	const auto result = IR::U64(ir.Emit(IR::ValueOpcode::WqmU64, {source}));
 	if (!wide) {
 		const auto low = ir.CompositeExtract(result, 0);
+		const bool destination_is_mask = IsExecOrVcc(inst.dst);
+		const bool destination_is_exec = inst.dst.kind == Decoder::OperandKind::ExecLo ||
+		                                 inst.dst.kind == Decoder::OperandKind::ExecHi;
+		auto old_live = IR::U1(IR::Value(false));
+		if (destination_is_mask) {
+			if (destination_is_exec) {
+				old_live = ir.GetExec();
+			} else {
+				old_live = ir.GetVcc();
+			}
+		}
+
 		// A 32-bit write preserves the other EXEC/VCC half and invalidates any
 		// overlapping SGPR-pair mask provenance through the ordinary scalar path.
 		WriteRawU32(inst.dst, low);
+		if (destination_is_mask) {
+			const auto is_high_half = [](const Decoder::Operand& operand) {
+				return operand.kind == Decoder::OperandKind::ExecHi ||
+				       operand.kind == Decoder::OperandKind::VccHi;
+			};
+			const bool destination_is_high = is_high_half(inst.dst);
+			auto affected_lanes = IR::U1(IR::Value(!destination_is_high));
+			if (program.wave_size == 64u) {
+				const auto lane = IR::U32(ir.Emit(IR::ValueOpcode::LaneId));
+				affected_lanes = ir.ULessThan(lane, IR::U32(IR::Value(32u)));
+				if (destination_is_high) {
+					affected_lanes = ir.LogicalNot(affected_lanes);
+				}
+			}
+
+			IR::U1 live;
+			if (destination_is_exec) {
+				live = ir.GetExec();
+			} else {
+				live = ir.GetVcc();
+			}
+			// A cross-half copy refers to different lanes, so only scalar bits apply.
+			if (is_high_half(inst.src0) == destination_is_high) {
+				live = ir.LogicalOr(retained_live, live);
+			}
+			live = IR::U1(ir.Emit(IR::ValueOpcode::SelectU1, {affected_lanes, live, old_live}));
+			if (destination_is_exec) {
+				ir.SetExec(live);
+			} else {
+				ir.SetVcc(live);
+			}
+		}
 		ir.SetScc(ir.INotEqual(low, IR::U32(IR::Value(0u))));
 		return;
 	}
-	const auto mask_valid = ReadMaskValid(inst.src0);
+
+	// WQM adds live lanes but never removes a source lane. Keep that fact separate
+	// from the scalar mask: reconstructing an entry-true predicate from a ballot
+	// otherwise introduces a spurious dependency into every vector write.
+	const auto live = ir.LogicalOr(retained_live, ThreadBit(ExtractU64(result)));
 	WriteOperand(DestinationOperand(inst), result);
-	if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
+	if (inst.dst.kind == Decoder::OperandKind::ExecLo) {
+		ir.SetExec(live);
+	} else if (inst.dst.kind == Decoder::OperandKind::VccLo) {
+		ir.SetVcc(live);
+	} else if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
 		const auto dst = static_cast<IR::ScalarReg>(inst.dst.reg);
-		ir.SetThreadBitScalarReg(dst, ThreadBit(ExtractU64(result)));
-		ir.SetScalarMaskTag(dst, mask_valid);
+		ir.SetThreadBitScalarReg(dst, live);
+		ir.SetScalarMaskTag(dst, source_valid);
 	}
 	ir.SetScc(IR::U1(ir.Emit(IR::ValueOpcode::INotEqual64, {result, IR::Value(uint64_t {0})})));
 }

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

@@ -1,8 +1,10 @@
+#include "common/logging/log.h"
 #include "graphics/shader/recompiler/frontend/translate/Translator.h"
 #include "graphics/shader/recompiler/frontend/decode/ImageOps.h"
 
 #include <algorithm>
 #include <array>
+#include <atomic>
 
 namespace Libs::Graphics::ShaderRecompiler::Frontend {
 
@@ -123,6 +125,11 @@ IR::MemoryInfo MemoryInfoFromDecoded(const Decoder::Instruction& decoded) {
 	memory.image_r128    = decoded.image_r128;
 	memory.idxen         = decoded.idxen;
 	memory.offen         = decoded.offen;
+	// Vector loads use GLC/DLC to bypass L0/GL1; atomics use GLC only to return data.
+	const bool buffer_atomic = decoded.opcode >= Decoder::Opcode::BUFFER_ATOMIC_SWAP &&
+	                           decoded.opcode <= Decoder::Opcode::BUFFER_ATOMIC_FMAX;
+	memory.coherent = memory.kind == ResourceKind::Buffer && !buffer_atomic &&
+	                  (decoded.glc || decoded.dlc);
 	memory.resource      = ResourceIndexFromOperand(decoded.src1);
 	memory.sampler       = ResourceIndexFromOperand(decoded.src2);
 	if (memory.kind == ResourceKind::ScalarBuffer) {
@@ -235,6 +242,7 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 				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:
 			case Decoder::Opcode::DS_WRITE_B16_D16_HI:
 			case Decoder::Opcode::DS_WRITE_B32:
 			case Decoder::Opcode::DS_WRITE_B64:
@@ -890,6 +898,17 @@ bool Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX4:
 		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX8:
 		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX16: return S_LOAD(inst, false);
+		case Decoder::Opcode::S_MEMREALTIME: {
+			static std::atomic_flag warned = ATOMIC_FLAG_INIT;
+			if (!warned.test_and_set(std::memory_order_relaxed)) {
+				Log::WriteToConsoleAndLog(
+				    "Warning: S_MEMREALTIME uses placeholder UINT64_MAX; real-time clock not implemented.\n");
+			}
+			for (uint32_t component = 0; component < 2; component++) {
+				WriteOperand(ScalarDestinationOperand(inst.dst, component), IR::Value(UINT32_MAX));
+			}
+			return true;
+		}
 
 		case Decoder::Opcode::BUFFER_LOAD_UBYTE:
 		case Decoder::Opcode::BUFFER_LOAD_SBYTE:
@@ -962,8 +981,12 @@ bool Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicISub32, false);
 		case Decoder::Opcode::DS_SUB_RTN_U32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicISub32, true);
+		case Decoder::Opcode::DS_INC_U32:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicInc32, false);
 		case Decoder::Opcode::DS_INC_RTN_U32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicInc32, true);
+		case Decoder::Opcode::DS_DEC_U32:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicDec32, false);
 		case Decoder::Opcode::DS_DEC_RTN_U32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicDec32, true);
 		case Decoder::Opcode::DS_MIN_I32:
@@ -1076,6 +1099,7 @@ bool Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::DS_WRITE2ST64_B64: return DS_WRITE2(inst);
 		case Decoder::Opcode::DS_WRITE_B8:
 		case Decoder::Opcode::DS_WRITE_B16:
+		case Decoder::Opcode::DS_WRITE_B8_D16_HI:
 		case Decoder::Opcode::DS_WRITE_B16_D16_HI:
 		case Decoder::Opcode::DS_WRITE_B32:
 		case Decoder::Opcode::DS_WRITE_B64:

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

@@ -141,6 +141,9 @@ bool Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMPX_LT_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::ULessThan32, false, true);
 			return true;
+		case O::V_CMPX_EQ_U16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, false, true);
+			return true;
 		case O::V_CMP_LE_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::ULessThanEqual32, false, false);
 			return true;

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

@@ -454,10 +454,8 @@ void ValidateProgram(const Program& program, bool require_ssa) {
 					                        ValueOpcodeName(inst.GetOpcode())));
 				}
 				if (memory.kind == ResourceKind::IndirectBuffer &&
-				    (memory.formatted || memory.typed ||
-				     (inst.GetOpcode() != ValueOpcode::LoadBufferU32x2 &&
-				      inst.GetOpcode() != ValueOpcode::LoadBufferU32x4))) {
-					return Fail("indirect buffer requires a raw DWORD x2/x4 load");
+				    !memory.SupportsIndirectBufferLoad(inst.GetOpcode())) {
+					return Fail("indirect buffer requires a raw DWORD x2/x3/x4 load");
 				}
 				if (buffer_components > 1u &&
 				    (!vector_buffer || memory.data_bits != 32u ||

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

@@ -67,8 +67,15 @@ struct MemoryInfo {
 	bool                    image_r128                                            = false;
 	bool                    idxen                                                 = false;
 	bool                    offen                                                 = false;
+	bool                    coherent                                              = false;
 	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);
+	}
+
 	bool operator==(const MemoryInfo& other) const = default;
 };
 
@@ -463,9 +470,9 @@ struct DescriptorSource {
 		uint32_t table_source    = 0;
 		uint32_t selector_stride = 0;
 		uint32_t selector_offset = 0;
-		uint32_t key_arg         = 0;
 		uint32_t table_offset    = 0;
-		uint32_t key_count       = 0;
+		Value    key_count;
+		Value    selector_mask;
 
 		bool operator==(const IndirectImage& other) const = default;
 	};
@@ -541,6 +548,7 @@ struct ResourcePlan {
 	std::vector<SrtRead>                srt_reads;
 	std::vector<uint8_t>                clean_flat_slots;
 	bool                                requires_specialization_memory = false;
+	bool                                has_address_writes = false;
 	bool                                srt_plan_complete          = false;
 	bool                                resource_tracking_complete = false;
 	ShaderInfo                          info;
@@ -553,6 +561,7 @@ struct ResourcePlan {
 	mutable std::vector<uint8_t>            visited_blocks;
 	mutable std::vector<uint32_t>           pending_blocks;
 	mutable std::vector<uint32_t>           material_keys;
+	mutable std::vector<std::pair<uint64_t, uint64_t>> specialization_reads;
 };
 
 struct Program: ResourcePlan {

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

@@ -200,6 +200,13 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 	switch (inst.GetOpcode()) {
 		case ValueOpcode::Phi: FoldPhi(inst); return;
 		case ValueOpcode::SelectU1:
+			if (!FoldSelect(inst) && IsImmediate(Arg(inst, 2), Type::U1) &&
+			    !Arg(inst, 2).U1()) {
+				auto result = block.PrependNewInst(instruction, ValueOpcode::LogicalAnd,
+				                                   {Arg(inst, 0), Arg(inst, 1)});
+				Replace(inst, Value(&*result));
+			}
+			return;
 		case ValueOpcode::SelectU32:
 		case ValueOpcode::SelectF32: FoldSelect(inst); return;
 		case ValueOpcode::BitFieldInsert: {
@@ -431,6 +438,10 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 			return;
 		case ValueOpcode::ShiftRightLogical32:
 			if (!FoldU32(inst, [](uint32_t a, uint32_t b) { return a >> (b & 31u); })) {
+				if (IsImmediate(Arg(inst, 0), Type::U32) && Arg(inst, 0).U32() == 0u) {
+					Replace(inst, Value(0u));
+					return;
+				}
 				const auto shift = Arg(inst, 1);
 				if (IsImmediate(shift, Type::U32) && (shift.U32() & 31u) == 0u) {
 					Replace(inst, Arg(inst, 0));
@@ -573,6 +584,23 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 			if (!FoldLogical(inst, [](bool a, bool b) { return a && b; })) {
 				const auto lhs = Arg(inst, 0);
 				const auto rhs = Arg(inst, 1);
+				const auto simplify = [&](Value assumption, Value expression) {
+					const auto* disjunction = expression.TryInstruction();
+					if (disjunction == nullptr || disjunction->GetOpcode() != ValueOpcode::LogicalOr) {
+						return false;
+					}
+					for (uint32_t i = 0; i < 2u; ++i) {
+						const auto* inverse = disjunction->Arg(i).Resolve().TryInstruction();
+						if (inverse != nullptr && inverse->GetOpcode() == ValueOpcode::LogicalNot &&
+						    inverse->Arg(0).Resolve() == assumption) {
+							inst.SetArg(assumption == lhs ? 1u : 0u,
+							            disjunction->Arg(i ^ 1u));
+							return true;
+						}
+					}
+					return false;
+				};
+				if (simplify(lhs, rhs) || simplify(rhs, lhs)) return;
 				if (IsImmediate(lhs, Type::U1)) {
 					Replace(inst, lhs.U1() ? rhs : lhs);
 				} else if (IsImmediate(rhs, Type::U1)) {

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

@@ -138,6 +138,50 @@ bool DecodeBufferDescriptor(const DescriptorValue& descriptor, ShaderBufferResou
 	return true;
 }
 
+struct ReadCapture {
+	SrtRuntime                                  source;
+	std::vector<std::pair<uint64_t, uint64_t>>& ranges;
+};
+
+bool CaptureStrictRead(void* userdata, uint64_t address, std::span<uint32_t> values) {
+	auto& capture = *static_cast<ReadCapture*>(userdata);
+	if (!capture.source.read_specialization_memory(capture.source.userdata, address, values)) {
+		return false;
+	}
+	capture.ranges.emplace_back(address, values.size_bytes());
+	return true;
+}
+
+bool CaptureOrdinaryRead(void* userdata, uint64_t address, std::span<uint32_t> values) {
+	auto& capture = *static_cast<ReadCapture*>(userdata);
+	if (!capture.source.read_memory(capture.source.userdata, address, values)) {
+		return false;
+	}
+	capture.ranges.emplace_back(address, values.size_bytes());
+	return true;
+}
+
+bool WrittenBuffersDisjoint(const ResourcePlan& program, const ResourceSnapshot& snapshot,
+                            std::span<const std::pair<uint64_t, uint64_t>> reads) {
+	for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
+		if (!program.info.buffers[i].written) continue;
+		ShaderBufferResource buffer;
+		if (!DecodeBufferDescriptor(snapshot.buffers[i], buffer)) return false;
+		const auto base = buffer.Base48();
+		const auto size = buffer.GetSize();
+		if (size == 0u) continue;
+		if (size - 1u > AddressMask - base) return false;
+		for (const auto [address, bytes]: reads) {
+			if (bytes == 0u) continue;
+			if (address > AddressMask || bytes - 1u > AddressMask - address ||
+			    (address <= base ? base - address < bytes : address - base < size)) {
+				return false;
+			}
+		}
+	}
+	return true;
+}
+
 const DescriptorSource* Source(const ResourcePlan& program, uint32_t source) {
 	if (source >= program.descriptor_sources.size()) {
 		return nullptr;
@@ -146,12 +190,15 @@ const DescriptorSource* Source(const ResourcePlan& program, uint32_t source) {
 }
 
 void MarkCleanFlatSlots(const ResourcePlan& program, const DescriptorSource* source,
-                        std::span<uint8_t> slots) {
-	if (source == nullptr) {
+                        std::span<uint8_t> slots, Value extra = {}) {
+	if (source == nullptr && extra.IsEmpty()) {
 		return;
 	}
-	std::vector<Value>       pending(source->dwords.begin(),
-	                                 source->dwords.begin() + source->dword_count);
+	std::vector<Value>       pending;
+	if (source != nullptr) {
+		pending.assign(source->dwords.begin(), source->dwords.begin() + source->dword_count);
+	}
+	if (!extra.IsEmpty()) pending.push_back(extra);
 	std::vector<const Inst*> visited;
 	while (!pending.empty()) {
 		auto value = pending.back().Resolve();
@@ -198,7 +245,8 @@ bool MaterializeIndirectImage(const ResourcePlan& program,
                               const DescriptorSource::IndirectImage& indirect,
                               const DescriptorValue& material_value,
                               const DescriptorValue& table_value, uint32_t image_index,
-                              const SrtRuntime& runtime, ResourceSnapshot& snapshot,
+                              const SrtRuntime& runtime, SrtWalker& clean,
+                              ResourceSnapshot& snapshot,
                               ResourceSpecialization& specialization) {
 	uint64_t table_base = 0;
 	uint64_t table_size = UINT64_MAX; // Scalar addresses have no buffer descriptor bounds.
@@ -214,12 +262,41 @@ bool MaterializeIndirectImage(const ResourcePlan& program,
 	auto& keys = program.material_keys;
 	keys.clear();
 	if (indirect.material_source == UINT32_MAX) {
-		if (table_value.dword_count != 2u || indirect.key_count == 0u ||
-		    indirect.key_count > MaxIndirectImageProbes) {
+		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) {
 			return false;
 		}
-		keys.resize(indirect.key_count);
+		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) {
+			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 ||
@@ -779,6 +856,7 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	plan.memory_info                = program.memory_info;
 	plan.srt_plan_complete          = program.srt_plan_complete;
 	plan.resource_tracking_complete = program.resource_tracking_complete;
+	plan.has_address_writes         = program.has_address_writes;
 
 	std::unordered_map<const Inst*, Inst*> cloned;
 	std::function<Value(Value)>            Clone = [&](Value value) -> Value {
@@ -816,6 +894,10 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		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);
+		}
 		for (uint32_t dword = 0; dword < source.dword_count; dword++) {
 			target.dwords[dword] = Clone(source.dwords[dword]);
 		}
@@ -840,7 +922,7 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		}
 		plan.requires_specialization_memory = true;
 		MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->material_source),
-		                   plan.clean_flat_slots);
+		                   plan.clean_flat_slots, source->indirect_image->selector_mask);
 		MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->table_source),
 		                   plan.clean_flat_slots);
 	}
@@ -853,8 +935,29 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	    (program.requires_specialization_memory && runtime.read_specialization_memory == nullptr)) {
 		return false;
 	}
-	SrtWalker clean(program, CleanRuntime(runtime));
-	SrtWalker walker(program, runtime, program.clean_flat_slots, &clean);
+	const bool masked_image = std::ranges::any_of(program.info.images, [&](const auto& image) {
+		const auto* source = Source(program, image.source);
+		return source != nullptr && source->indirect_image.has_value() &&
+		       !source->indirect_image->selector_mask.IsEmpty();
+	});
+	if (masked_image &&
+	    (program.has_address_writes ||
+	     std::ranges::any_of(program.info.images, &ImageResource::written))) {
+		return false;
+	}
+	const bool capture_reads = masked_image &&
+	    std::ranges::any_of(program.info.buffers, &BufferResource::written);
+	auto& reads = program.specialization_reads;
+	ReadCapture capture {runtime, reads};
+	SrtRuntime observed = runtime;
+	if (capture_reads) {
+		reads.clear();
+		observed.userdata = &capture;
+		observed.read_specialization_memory = CaptureStrictRead;
+		if (observed.read_memory != nullptr) observed.read_memory = CaptureOrdinaryRead;
+	}
+	SrtWalker clean(program, CleanRuntime(observed));
+	SrtWalker walker(program, observed, program.clean_flat_slots, &clean);
 	const auto active = clean.FindActiveSources();
 	if (!walker.RefreshFlatBuffer(snapshot.flattened_srt)) {
 		return false;
@@ -888,6 +991,15 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			return false;
 		}
 	}
+	if (capture_reads) {
+		for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
+			const auto& buffer = program.info.buffers[i];
+			if (!buffer.written || (!active.empty() && !active[buffer.source])) continue;
+			DescriptorValue strict;
+			if (!clean.EvaluateDescriptor(buffer.source, strict) ||
+			    strict != snapshot.buffers[i]) return false;
+		}
+	}
 	snapshot.images.resize(program.info.images.size());
 	specialization.images.clear();
 	specialization.images.reserve(program.info.images.size());
@@ -921,7 +1033,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			if ((indirect.material_source != UINT32_MAX &&
 			     !clean.EvaluateDescriptor(indirect.material_source, material)) ||
 			    !clean.EvaluateDescriptor(indirect.table_source, table) ||
-			    !MaterializeIndirectImage(program, indirect, material, table, i, runtime, snapshot,
+			    !MaterializeIndirectImage(program, indirect, material, table, i, observed, clean, snapshot,
 			                              specialization)) {
 				return false;
 			}
@@ -940,6 +1052,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			return false;
 		}
 	}
+	if (capture_reads && !WrittenBuffersDisjoint(program, snapshot, reads)) return false;
 	snapshot.user_data.assign(runtime.user_data.begin(), runtime.user_data.end());
 	return BuildResourceSpecialization(program, snapshot, specialization);
 }

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

@@ -313,9 +313,20 @@ private:
 		for (uint32_t candidate = 0; candidate < m_sources.size(); candidate++) {
 			const auto& current = m_sources[candidate];
 			if (current.dword_count != descriptor.dword_count ||
-			    current.indirect_image != descriptor.indirect_image) {
+			    current.indirect_image.has_value() != descriptor.indirect_image.has_value()) {
 				continue;
 			}
+			if (current.indirect_image.has_value()) {
+				const auto& a = *current.indirect_image;
+				const auto& b = *descriptor.indirect_image;
+				if (a.material_source != b.material_source || a.table_source != b.table_source ||
+				    a.selector_stride != b.selector_stride || a.selector_offset != b.selector_offset ||
+				    a.table_offset != b.table_offset ||
+				    !EquivalentValue(m_program, a.key_count, b.key_count) ||
+				    a.selector_mask.IsEmpty() != b.selector_mask.IsEmpty() ||
+				    (!a.selector_mask.IsEmpty() &&
+				     !EquivalentValue(m_program, a.selector_mask, b.selector_mask))) continue;
+			}
 			bool same = true;
 			for (uint32_t i = 0; i < descriptor.dword_count; i++) {
 				same = same && EquivalentValue(m_program, current.dwords[i], descriptor.dwords[i]);
@@ -504,6 +515,355 @@ private:
 		}
 	}
 
+	const Block* FindBlock(uint32_t id) const {
+		const auto info = std::ranges::find(m_program.block_info, id, &BlockInfo::id);
+		return info == m_program.block_info.end()
+		           ? nullptr
+		           : m_program.blocks[info - m_program.block_info.begin()];
+	}
+
+	bool CanReach(const Block* start, const Block* target, const Block* avoid) const {
+		std::vector<const Block*> pending {start};
+		std::vector<const Block*> visited;
+		while (!pending.empty()) {
+			const auto* block = pending.back();
+			pending.pop_back();
+			if (block == nullptr || block == avoid ||
+			    std::ranges::find(visited, block) != visited.end()) continue;
+			if (block == target) return true;
+			visited.push_back(block);
+			for (const auto* next: block->ImmSuccessors()) pending.push_back(next);
+		}
+		return false;
+	}
+
+	bool ConditionalEdge(const Block* from, const Block* to, Value& condition,
+	                     bool& positive) const {
+		const auto position = std::ranges::find(m_program.blocks, from);
+		const auto target = std::ranges::find(m_program.blocks, to);
+		if (position == m_program.blocks.end() || target == m_program.blocks.end()) return false;
+		const auto& info = m_program.block_info[position - m_program.blocks.begin()];
+		const auto& term = info.terminator;
+		const auto id = m_program.block_info[target - m_program.blocks.begin()].id;
+		if (term.kind != CFG::TerminatorKind::ConditionalBranch ||
+		    (term.true_block == id) == (term.false_block == id)) return false;
+		condition = info.condition;
+		positive = term.true_block == id;
+		while (const auto* inst = condition.Resolve().TryInstruction()) {
+			if (inst->GetOpcode() != ValueOpcode::LogicalNot) break;
+			condition = inst->Arg(0);
+			positive = !positive;
+		}
+		condition = condition.Resolve();
+		return true;
+	}
+
+	Value PositiveUseGuard(const Block* use) const {
+		if (use == nullptr || use->ImmPredecessors().size() != 1u) return {};
+		Value condition;
+		bool positive = false;
+		return ConditionalEdge(use->ImmPredecessors()[0], use, condition, positive) && positive
+		           ? condition : Value {};
+	}
+
+	Value SimplifyGuard(Value guard) const {
+		const auto invariant = ResolveInvariantPhi(m_program, guard);
+		return (invariant.IsEmpty() ? guard : invariant).Resolve();
+	}
+
+	bool Implies(Value guard, Value required) const {
+		guard = SimplifyGuard(guard);
+		required = required.Resolve();
+		if (EquivalentValue(m_program, guard, required)) return true;
+		const auto* inst = guard.TryInstruction();
+		return inst != nullptr && inst->GetOpcode() == ValueOpcode::LogicalAnd &&
+		       (Implies(inst->Arg(0), required) || Implies(inst->Arg(1), required));
+	}
+
+	Value EqualLocalKey(Value guard, Value key) const {
+		guard = SimplifyGuard(guard);
+		const auto* inst = guard.TryInstruction();
+		if (inst == nullptr) return {};
+		if (inst->GetOpcode() == ValueOpcode::LogicalAnd) {
+			const auto left = EqualLocalKey(inst->Arg(0), key);
+			return left.IsEmpty() ? EqualLocalKey(inst->Arg(1), key) : left;
+		}
+		if (inst->GetOpcode() != ValueOpcode::IEqual32 || inst->NumArgs() != 2u) return {};
+		if (EquivalentValue(m_program, inst->Arg(0), key)) return inst->Arg(1).Resolve();
+		if (EquivalentValue(m_program, inst->Arg(1), key)) return inst->Arg(0).Resolve();
+		return {};
+	}
+
+	struct AffineOffset {
+		Value    index;
+		uint64_t stride = 0;
+		uint64_t offset = 0;
+	};
+
+	bool MatchAffineOffset(Value value, Value guard, AffineOffset& out,
+	                       uint32_t depth = 0) const {
+		if (depth > 16u) return false;
+		value = value.Resolve();
+		uint32_t immediate = 0;
+		if (ImmediateU32(value, immediate)) {
+			out.offset = immediate;
+			return true;
+		}
+		const auto* inst = value.TryInstruction();
+		if (inst == nullptr) return false;
+		if (inst->GetOpcode() == ValueOpcode::SelectU32 && inst->NumArgs() == 3u &&
+		    Implies(guard, inst->Arg(0))) {
+			return MatchAffineOffset(inst->Arg(1), guard, out, depth + 1u);
+		}
+		if (inst->GetOpcode() == ValueOpcode::IAdd32 && inst->NumArgs() == 2u) {
+			AffineOffset left, right;
+			if (!MatchAffineOffset(inst->Arg(0), guard, left, depth + 1u) ||
+			    !MatchAffineOffset(inst->Arg(1), guard, right, depth + 1u) ||
+			    (!left.index.IsEmpty() && !right.index.IsEmpty() &&
+			     !EquivalentValue(m_program, left.index, right.index))) return false;
+			out.index = left.index.IsEmpty() ? right.index : left.index;
+			out.stride = left.stride + right.stride;
+			out.offset = left.offset + right.offset;
+			return out.stride <= UINT32_MAX && out.offset <= UINT32_MAX;
+		}
+		if (inst->GetOpcode() == ValueOpcode::ShiftLeftLogical32 && inst->NumArgs() == 2u &&
+		    ImmediateU32(inst->Arg(1), immediate) && immediate < 32u) {
+			if (!MatchAffineOffset(inst->Arg(0), guard, out, depth + 1u)) return false;
+			out.stride <<= immediate;
+			out.offset <<= immediate;
+			return out.stride <= UINT32_MAX && out.offset <= UINT32_MAX;
+		}
+		out.index = value;
+		out.stride = 1u;
+		return value.GetType() == Type::U32;
+	}
+
+	bool ImpliesNonzero(Value guard, Value value) const {
+		guard = SimplifyGuard(guard);
+		const auto* inst = guard.TryInstruction();
+		if (inst == nullptr) return false;
+		if (inst->GetOpcode() == ValueOpcode::LogicalAnd) {
+			return ImpliesNonzero(inst->Arg(0), value) ||
+			       ImpliesNonzero(inst->Arg(1), value);
+		}
+		if (inst->GetOpcode() != ValueOpcode::INotEqual32 || inst->NumArgs() != 2u)
+			return false;
+		uint32_t zero = 1u;
+		return (ImmediateU32(inst->Arg(0), zero) && zero == 0u &&
+		        EquivalentValue(m_program, inst->Arg(1), value)) ||
+		       (ImmediateU32(inst->Arg(1), zero) && zero == 0u &&
+		        EquivalentValue(m_program, inst->Arg(0), value));
+	}
+
+	bool ImpliesIndexBelow32(Value guard, Value index) const {
+		guard = SimplifyGuard(guard);
+		const auto* inst = guard.TryInstruction();
+		if (inst == nullptr) return false;
+		if (inst->GetOpcode() == ValueOpcode::LogicalAnd) {
+			return ImpliesIndexBelow32(inst->Arg(0), index) ||
+			       ImpliesIndexBelow32(inst->Arg(1), index);
+		}
+		if (inst->NumArgs() != 2u) return false;
+		uint32_t limit = 0;
+		return (inst->GetOpcode() == ValueOpcode::SLessThan32 &&
+		        EquivalentValue(m_program, inst->Arg(0), index) &&
+		        ImmediateU32(inst->Arg(1), limit) && limit == 32u) ||
+		       (inst->GetOpcode() == ValueOpcode::SGreaterThan32 &&
+		        ImmediateU32(inst->Arg(0), limit) && limit == 32u &&
+		        EquivalentValue(m_program, inst->Arg(1), index));
+	}
+
+	bool ClearsFirstSetBit(Value value, Value mask) const {
+		const auto* update = value.Resolve().TryInstruction();
+		if (update == nullptr || update->GetOpcode() != ValueOpcode::BitwiseXor32 ||
+		    update->NumArgs() != 2u) return false;
+		Value bit;
+		if (EquivalentValue(m_program, update->Arg(0), mask)) bit = update->Arg(1);
+		else if (EquivalentValue(m_program, update->Arg(1), mask)) bit = update->Arg(0);
+		else return false;
+		const auto* shift = bit.Resolve().TryInstruction();
+		uint32_t one = 0;
+		if (shift == nullptr || shift->GetOpcode() != ValueOpcode::ShiftLeftLogical32 ||
+		    shift->NumArgs() != 2u || !ImmediateU32(shift->Arg(0), one) || one != 1u)
+			return false;
+		Value position = shift->Arg(1).Resolve();
+		const auto* masked = position.TryInstruction();
+		uint32_t lane_mask = 0;
+		if (masked != nullptr && masked->GetOpcode() == ValueOpcode::BitwiseAnd32 &&
+		    masked->NumArgs() == 2u) {
+			if (ImmediateU32(masked->Arg(0), lane_mask) && lane_mask == 31u)
+				position = masked->Arg(1);
+			else if (ImmediateU32(masked->Arg(1), lane_mask) && lane_mask == 31u)
+				position = masked->Arg(0);
+		}
+		const auto* first = position.Resolve().TryInstruction();
+		return first != nullptr && first->GetOpcode() == ValueOpcode::FindILsb32 &&
+		       first->NumArgs() == 1u && EquivalentValue(m_program, first->Arg(0), mask);
+	}
+
+	Value InitialClearingMask(Value value, const Block* update_block) const {
+		const auto* phi = value.Resolve().TryInstruction();
+		if (phi == nullptr || phi->GetOpcode() != ValueOpcode::Phi ||
+		    phi->NumArgs() != 2u || phi->GetType() != Type::U32) return {};
+		for (uint32_t back = 0; back < 2u; ++back) {
+			if (phi->PhiBlock(back) != update_block ||
+			    !ClearsFirstSetBit(phi->Arg(back), value)) continue;
+			const auto initial = phi->Arg(back ^ 1u).Resolve();
+			return ValidateRuntimeValue(m_program, initial, RuntimeValueType::Integer)
+			           ? initial : Value {};
+		}
+		return {};
+	}
+
+	bool EdgeOn(const Block* from, const Block* to, Value predicate, bool positive) const {
+		Value condition;
+		bool taken = false;
+		return ConditionalEdge(from, to, condition, taken) && taken == positive &&
+		       EquivalentValue(m_program, condition, predicate);
+	}
+
+	Value BoundedSetBitMask(Value index, Value guard) const {
+		const auto* phi = index.Resolve().TryInstruction();
+		if (phi == nullptr || phi->GetOpcode() != ValueOpcode::Phi ||
+		    phi->NumArgs() != 3u || phi->GetType() != Type::U32 ||
+		    !ImpliesIndexBelow32(guard, index)) return {};
+		for (uint32_t bit_arm = 0; bit_arm < 3u; ++bit_arm) {
+			const auto bit_guard = PositiveUseGuard(phi->PhiBlock(bit_arm));
+			const auto* selected = phi->Arg(bit_arm).Resolve().TryInstruction();
+			if (selected == nullptr || selected->GetOpcode() != ValueOpcode::SelectU32 ||
+			    selected->NumArgs() != 3u ||
+			    !Implies(bit_guard, selected->Arg(0))) continue;
+			const auto* first = selected->Arg(1).Resolve().TryInstruction();
+			if (first == nullptr || first->GetOpcode() != ValueOpcode::FindILsb32 ||
+			    first->NumArgs() != 1u || !ImpliesNonzero(bit_guard, first->Arg(0)))
+				continue;
+			const auto initial_mask = InitialClearingMask(first->Arg(0), phi->PhiBlock(bit_arm));
+			if (initial_mask.IsEmpty()) continue;
+			for (uint32_t sentinel_arm = 0; sentinel_arm < 3u; ++sentinel_arm) {
+				if (sentinel_arm == bit_arm) continue;
+				const auto sentinel_guard = PositiveUseGuard(phi->PhiBlock(sentinel_arm));
+				const auto* sentinel = phi->Arg(sentinel_arm).Resolve().TryInstruction();
+				uint32_t bound = 0;
+				if (sentinel == nullptr || sentinel->GetOpcode() != ValueOpcode::SelectU32 ||
+				    sentinel->NumArgs() != 3u ||
+				    !Implies(sentinel_guard, sentinel->Arg(0)) ||
+				    !ImmediateU32(sentinel->Arg(1), bound) || bound != 32u) continue;
+				const auto loop_active = sentinel->Arg(0).Resolve();
+				const auto* active_phi = loop_active.TryInstruction();
+				if (active_phi == nullptr || active_phi->GetOpcode() != ValueOpcode::Phi ||
+				    active_phi->NumArgs() != 2u) continue;
+				bool invariant = false;
+				for (uint32_t initial = 0; initial < 2u; ++initial) {
+					invariant = Implies(guard, active_phi->Arg(initial)) &&
+					            EdgeOn(active_phi->PhiBlock(initial ^ 1u), active_phi->Parent(),
+					                   active_phi->Arg(initial ^ 1u), true);
+					if (invariant) break;
+				}
+				if (!invariant) continue;
+				const auto other_arm = 3u - bit_arm - sentinel_arm;
+				if (EdgeOn(phi->PhiBlock(other_arm), phi->Parent(), loop_active, false))
+					return initial_mask;
+			}
+		}
+		return {};
+	}
+
+	bool MatchUniformizedMaterialKey(Value key, const Inst& image,
+	                                DescriptorSource::IndirectImage& indirect,
+	                                DescriptorSource& material_source, uint32_t pc) {
+		const auto guard = PositiveUseGuard(image.Parent());
+		if (guard.IsEmpty()) return false;
+		const auto local = EqualLocalKey(guard, key);
+		const auto* selected = local.Resolve().TryInstruction();
+		if (selected == nullptr || selected->GetOpcode() != ValueOpcode::SelectU32 ||
+		    selected->NumArgs() != 3u || !Implies(guard, selected->Arg(0))) return false;
+		const auto active = selected->Arg(0).Resolve();
+		const auto* read = selected->Arg(1).Resolve().TryInstruction();
+		if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadAddressU32 ||
+		    read->NumArgs() != 4u || !EquivalentValue(m_program, read->Arg(3), active))
+			return false;
+		uint32_t high = 1u;
+		if (!ImmediateU32(read->Arg(2), high) || high != 0u) return false;
+		const auto memory_index = read->Flags<MemoryFlags>().index;
+		if (memory_index >= m_program.memory_info.size()) return false;
+		const auto& memory = m_program.memory_info[memory_index];
+		if (memory.kind != ResourceKind::Global || memory.data_bits != 32u ||
+		    memory.data_dwords != 1u || !MemoryIndexBelongsTo(memory_index, *read)) return false;
+		const auto* material_handle = read->Arg(0).Resolve().TryInstruction();
+		if (material_handle == nullptr ||
+		    material_handle->GetOpcode() != ValueOpcode::GetAddressResource ||
+		    !MakeRuntimeTableSource(*material_handle, pc, material_source)) return false;
+		AffineOffset offset;
+		if (!MatchAffineOffset(read->Arg(1), active, offset) || offset.index.IsEmpty() ||
+		    offset.stride == 0u || offset.offset + memory.offset > UINT32_MAX ||
+		    offset.offset + memory.offset + 31u * offset.stride + 4u >
+		        static_cast<uint64_t>(UINT32_MAX) + 1u) return false;
+		const auto mask = BoundedSetBitMask(offset.index, active);
+		if (mask.IsEmpty()) return false;
+		indirect.material_source = InternSource(material_source);
+		indirect.selector_stride = static_cast<uint32_t>(offset.stride);
+		indirect.selector_offset = static_cast<uint32_t>(offset.offset + memory.offset);
+		indirect.key_count = Value(32u);
+		indirect.selector_mask = mask;
+		return true;
+	}
+
+	Value BoundedLoopCount(Value key, const Block* use) const {
+		const auto* phi = key.Resolve().TryInstruction();
+		if (m_shader_writes || phi == nullptr || phi->GetOpcode() != ValueOpcode::Phi ||
+		    phi->GetType() != Type::U32 || phi->NumArgs() != 2u ||
+		    m_program.blocks.size() != m_program.block_info.size()) {
+			return {};
+		}
+		bool induction = false;
+		for (uint32_t initial = 0; initial < 2u; ++initial) {
+			const auto zero = phi->Arg(initial).Resolve();
+			const auto* step = phi->Arg(initial ^ 1u).Resolve().TryInstruction();
+			if (!zero.IsImmediate() || zero.GetType() != Type::U32 || zero.U32() != 0u ||
+			    step == nullptr || step->GetOpcode() != ValueOpcode::IAdd32 ||
+			    step->Parent() != phi->PhiBlock(initial ^ 1u)) {
+				continue;
+			}
+			uint32_t increment = 0;
+			induction = (step->Arg(0).Resolve() == key &&
+			             ImmediateU32(step->Arg(1), increment) && increment == 1u) ||
+			            (step->Arg(1).Resolve() == key &&
+			             ImmediateU32(step->Arg(0), increment) && increment == 1u);
+			if (induction) break;
+		}
+		if (!induction) return {};
+
+		const auto contains = [&](auto&& self, Value value, const Inst* comparison) -> bool {
+			value = value.Resolve();
+			if (value.TryInstruction() == comparison) return true;
+			const auto* inst = value.TryInstruction();
+			return inst != nullptr && inst->GetOpcode() == ValueOpcode::LogicalAnd &&
+			       (self(self, inst->Arg(0), comparison) || self(self, inst->Arg(1), comparison));
+		};
+		for (const auto& use_of_key: phi->Uses()) {
+			const auto* compare = use_of_key.user;
+			if (compare->GetOpcode() != ValueOpcode::SLessThan32 || use_of_key.operand != 0u ||
+			    !ValidateRuntimeValue(m_program, compare->Arg(1))) continue;
+			for (uint32_t i = 0; i < m_program.block_info.size(); ++i) {
+				const auto& info = m_program.block_info[i];
+				const auto* negated = info.condition.Resolve().TryInstruction();
+				if (info.terminator.kind != CFG::TerminatorKind::ConditionalBranch ||
+				    compare->Parent() != m_program.blocks[i] || negated == nullptr ||
+				    negated->GetOpcode() != ValueOpcode::LogicalNot ||
+				    !contains(contains, negated->Arg(0), compare) ||
+				    use == m_program.blocks[i] ||
+				    CanReach(m_program.blocks.front(), use, m_program.blocks[i]) ||
+				    CanReach(phi->Parent(), use, m_program.blocks[i]) ||
+				    CanReach(FindBlock(info.terminator.true_block), use, phi->Parent()) ||
+				    !CanReach(FindBlock(info.terminator.false_block), use, phi->Parent())) {
+					continue;
+				}
+				return compare->Arg(1);
+			}
+		}
+		return {};
+	}
+
 	bool TryMakeIndirectImage(Inst& handle, uint32_t pc, IndirectImagePlan& plan) {
 		if (handle.GetOpcode() != ValueOpcode::GetImageResource || handle.NumArgs() != 8u) {
 			return false;
@@ -562,13 +922,20 @@ private:
 		indirect.table_offset = table_offset;
 		if (table_source.dword_count == 2u) {
 			const auto* selector = key.Resolve().TryInstruction();
-			if (m_shader_writes || selector == nullptr ||
-			    selector->GetOpcode() != ValueOpcode::FindILsb32 || selector->NumArgs() != 1u ||
-			    !NonzeroOnEntry(selector->Arg(0), handle.Parent()) ||
-			    (table_offset & 3u) != 0u || table_offset > UINT32_MAX - (32u * 32u - 1u)) {
+			const bool bitscan = selector != nullptr && selector->GetOpcode() == ValueOpcode::FindILsb32 &&
+			    selector->NumArgs() == 1u && !m_shader_writes &&
+			    NonzeroOnEntry(selector->Arg(0), handle.Parent());
+			if (bitscan) {
+				indirect.key_count = Value(32u);
+			} else {
+				indirect.key_count = BoundedLoopCount(key, handle.Parent());
+			}
+			if (indirect.key_count.IsEmpty() &&
+			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source, pc)) {
 				return false;
 			}
-			indirect.key_count = 32u;
+			if ((table_offset & 3u) != 0u ||
+			    (bitscan && table_offset > UINT32_MAX - (32u * 32u - 1u))) return false;
 		} else {
 			auto* material_read = key.Resolve().TryInstruction();
 			uint32_t material_memory_index = 0;
@@ -839,11 +1206,10 @@ private:
 		if (buffer != BufferAccess::None) {
 			if (!GetHandle(inst.Arg(0), ValueOpcode::GetBufferResource, 4, flags.pc, handle,
 			               source)) {
-				if (memory.kind != ResourceKind::Buffer || memory.formatted || memory.typed ||
-				    (op != ValueOpcode::LoadBufferU32x2 && op != ValueOpcode::LoadBufferU32x4)) {
+				if (memory.kind != ResourceKind::Buffer || !memory.SupportsIndirectBufferLoad(op)) {
 					Fail(flags.pc,
 					     "buffer descriptor is not a valid runtime value; GPU-selected access "
-					     "requires a raw DWORD x2/x4 load");
+					     "requires a raw DWORD x2/x3/x4 load");
 				}
 				m_program.memory_info[flags.index].kind = ResourceKind::IndirectBuffer;
 				m_info.uses_dma                         = true;
@@ -873,6 +1239,9 @@ private:
 				return;
 			}
 			ValidateAddressHandle(inst.Arg(0), flags.pc);
+			if (address_info.access == AddressAccess::Write) {
+				m_program.has_address_writes = true;
+			}
 			m_info.uses_dma = true;
 			return;
 		}

src/graphics/shader/shader.cpp

@@ -388,6 +388,8 @@ static void ShaderApplyAttribSemantics(ShaderVertexInputInfo& info,
 
 	EXIT_IF(attrib == nullptr || buffer == nullptr);
 
+	const bool debug_dump = Config::GraphicsDebugDumpEnabled();
+
 	for (uint32_t i = 0; i < num_input_semantics; i++) {
 		const auto& in = input_semantics[i];
 
@@ -396,7 +398,7 @@ static void ShaderApplyAttribSemantics(ShaderVertexInputInfo& info,
 		uint32_t reg  = in.hardware_mapping;
 		uint32_t size = in.size_in_elements;
 
-		if (Config::GraphicsDebugDumpEnabled()) {
+		if (debug_dump) {
 			LOGF("reg = %u, size = %u, va[%u] = 0x%08" PRIx32 "\n", reg, size, i,
 			     attrib[in.semantic]);
 		}

src/kernel/memory.cpp

@@ -200,6 +200,10 @@ static bool IsPrivateCommittedRangeType(VirtualRangeType type) {
 	       type == VirtualRangeType::Runtime;
 }
 
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+static bool g_test_fail_next_range_replace = false;
+#endif
+
 class VirtualRanges {
 public:
 	struct Range {
@@ -288,36 +292,22 @@ public:
 		return true;
 	}
 
-	bool ConsumeReserved(uint64_t start, uint64_t size,
-	                     VirtualRangeType type = VirtualRangeType::Reserved) {
+	bool ReplaceSpan(uint64_t start, uint64_t size, VirtualRangeType expected_type,
+	                 uint64_t offset, int protection, int memory_type, VirtualRangeType type,
+	                 const char* name, bool disallow_merge = false) {
 		Common::LockGuard lock(m_mutex);
 
-		auto end = End(start, size);
-		for (const auto& r: m_ranges) {
-			if (r.type == type && start >= r.start && end <= End(r.start, r.size)) {
-				RemoveUnlocked(start, size);
-				MergeUnlocked();
-				return true;
-			}
-		}
-
-		return false;
-	}
-
-	bool ConsumeReservedSpan(uint64_t start, uint64_t size, Range* first_range = nullptr,
-	                         VirtualRangeType type = VirtualRangeType::Reserved) {
-		Common::LockGuard lock(m_mutex);
-
-		if (size == 0) {
+		if (start == 0 || size == 0 || size > UINT64_MAX - start) {
 			return false;
 		}
 
 		auto current = start;
-		auto end     = End(start, size);
+		const auto end = start + size;
 		while (current < end) {
 			const Range* candidate = nullptr;
 			for (const auto& r: m_ranges) {
-				if (r.type == type && current >= r.start && current < End(r.start, r.size)) {
+				if (r.type == expected_type && current >= r.start &&
+				    current < End(r.start, r.size)) {
 					candidate = &r;
 					break;
 				}
@@ -325,14 +315,29 @@ public:
 			if (candidate == nullptr) {
 				return false;
 			}
-			if (current == start && first_range != nullptr) {
-				*first_range = *candidate;
-			}
 			current = std::min(end, End(candidate->start, candidate->size));
 		}
 
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+		if (g_test_fail_next_range_replace) {
+			g_test_fail_next_range_replace = false;
+			return false;
+		}
+#endif
+		Range replacement {};
+		replacement.start          = start;
+		replacement.size           = size;
+		replacement.offset         = offset;
+		replacement.protection     = protection;
+		replacement.memory_type    = memory_type;
+		replacement.type           = type;
+		replacement.disallow_merge = disallow_merge;
+		CopyVirtualRangeName(replacement.name, name);
 		RemoveUnlocked(start, size);
-		MergeUnlocked();
+		auto position = LowerBound(start);
+		const auto index = static_cast<size_t>(position - m_ranges.begin());
+		m_ranges.insert(position, replacement);
+		MergeAroundUnlocked(index);
 		return true;
 	}
 
@@ -2250,10 +2255,9 @@ int32_t KYTY_SYSV_ABI KernelMapNamedFlexibleMemory(void** addr_in_out, size_t le
 		EXIT("unknown prot: %d\n", prot);
 	}
 
-	auto                 in_addr              = reinterpret_cast<uint64_t>(*addr_in_out);
-	uint64_t             out_addr             = 0;
-	bool                 consumed_reservation = false;
-	VirtualRanges::Range consumed_range {};
+	auto     in_addr         = reinterpret_cast<uint64_t>(*addr_in_out);
+	uint64_t out_addr        = 0;
+	bool     reserved_target = false;
 
 	if ((flags & GUEST_MAP_FIXED) != 0) {
 		if (in_addr == 0 || (in_addr & (PAGE_SIZE - 1)) != 0 ||
@@ -2269,16 +2273,12 @@ int32_t KYTY_SYSV_ABI KernelMapNamedFlexibleMemory(void** addr_in_out, size_t le
 			    return range.type == VirtualRangeType::Reserved;
 		    })) {
 			UnmapGpuRange(in_addr, len);
-			consumed_range = reserved_ranges.front();
-			if (g_virtual_ranges->ConsumeReservedSpan(in_addr, len)) {
-				consumed_reservation = true;
-				out_addr             = in_addr;
-			}
+			reserved_target = true;
+			out_addr        = in_addr;
 		}
-		if (!consumed_reservation && ReplaceFixedRangeWithReserved(in_addr, len) &&
-		    g_virtual_ranges->ConsumeReservedSpan(in_addr, len, &consumed_range)) {
-			consumed_reservation = true;
-			out_addr             = in_addr;
+		if (!reserved_target && ReplaceFixedRangeWithReserved(in_addr, len)) {
+			reserved_target = true;
+			out_addr        = in_addr;
 		}
 	} else {
 		const auto search_addr = (in_addr != 0 ? in_addr : DEFAULT_PS5_BASE);
@@ -2291,30 +2291,25 @@ int32_t KYTY_SYSV_ABI KernelMapNamedFlexibleMemory(void** addr_in_out, size_t le
 	*addr_in_out = reinterpret_cast<void*>(out_addr);
 
 	if (out_addr == 0) {
-		if (consumed_reservation) {
-			g_virtual_ranges->Add(in_addr, len, 0, 0, 0, VirtualRangeType::Reserved,
-			                      consumed_range.name);
-		}
 		return KERNEL_ERROR_ENOMEM;
 	}
 
 	if (!g_flexible_memory->Map(out_addr, len, prot, mode, gpu_mode, name)) {
 		LOGF_COLOR(Log::Color::Red, "\t [Fail]\n");
-		if (consumed_reservation) {
-			EXIT_IF(!g_virtual_ranges->Add(out_addr, len, 0, 0, 0, VirtualRangeType::Reserved,
-			                               consumed_range.name));
-		}
 		return KERNEL_ERROR_ENOMEM;
 	}
 
-	if (!g_virtual_ranges->Add(out_addr, len, 0, prot, 0, VirtualRangeType::Flexible, name,
-	                           (map_flags & GUEST_MAP_NO_COALESCE) != 0)) {
+	const bool published = reserved_target
+	                           ? g_virtual_ranges->ReplaceSpan(
+	                                 out_addr, len, VirtualRangeType::Reserved, 0, prot, 0,
+	                                 VirtualRangeType::Flexible, name,
+	                                 (map_flags & GUEST_MAP_NO_COALESCE) != 0)
+	                           : g_virtual_ranges->Add(out_addr, len, 0, prot, 0,
+	                                                   VirtualRangeType::Flexible, name,
+	                                                   (map_flags & GUEST_MAP_NO_COALESCE) != 0);
+	if (!published) {
 		GpuAccessMode rollback_gpu_mode = GpuAccessMode::NoAccess;
 		EXIT_IF(!g_flexible_memory->Unmap(out_addr, len, &rollback_gpu_mode));
-		if (consumed_reservation) {
-			EXIT_IF(!g_virtual_ranges->Add(out_addr, len, 0, 0, 0, VirtualRangeType::Reserved,
-			                               consumed_range.name));
-		}
 		return KERNEL_ERROR_EBUSY;
 	}
 
@@ -2946,25 +2941,17 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 		return KERNEL_ERROR_ENOMEM;
 	}
 
-	auto                 in_addr              = reinterpret_cast<uint64_t>(*addr);
-	uint64_t             out_addr             = 0;
-	bool                 shared_backing       = false;
-	bool                 consumed_reservation = false;
-	VirtualRanges::Range consumed_range {};
-	auto                 shared_failure = GuestBackingStore::FailureReason::None;
+	auto     in_addr         = reinterpret_cast<uint64_t>(*addr);
+	uint64_t out_addr        = 0;
+	bool     shared_backing  = false;
+	bool     reserved_target = false;
+	auto     shared_failure  = GuestBackingStore::FailureReason::None;
 	// Direct mappings must remain views of the single backing object. Anonymous fallbacks break
 	// aliasing and lose direct-memory contents when a range is unmapped and mapped again.
 	auto map_shared_fixed = [&](uint64_t target_addr) -> bool {
 		return g_guest_address_space->MapBacking(target_addr, len, direct_memory_start, mode,
 		                                         &shared_failure);
 	};
-	auto map_consumed_reserved_fixed = [&]() {
-		if (map_shared_fixed(in_addr)) {
-			out_addr       = in_addr;
-			shared_backing = true;
-		}
-	};
-
 	if (fixed) {
 		if (in_addr == 0 || (in_addr & (PAGE_SIZE - 1u)) != 0 ||
 		    (alignment != 0 && in_addr % alignment != 0)) {
@@ -2980,20 +2967,18 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 			    return range.type == VirtualRangeType::Reserved;
 		    })) {
 			UnmapGpuRange(in_addr, len);
-			consumed_range = reserved_ranges.front();
-			if (g_virtual_ranges->ConsumeReservedSpan(in_addr, len)) {
-				consumed_reservation = true;
-				map_consumed_reserved_fixed();
-			}
+			reserved_target = true;
 		}
-		if (!consumed_reservation && ReplaceFixedRangeWithReserved(in_addr, len) &&
-		    g_virtual_ranges->ConsumeReservedSpan(in_addr, len, &consumed_range)) {
-			consumed_reservation = true;
-			map_consumed_reserved_fixed();
+		if (!reserved_target && ReplaceFixedRangeWithReserved(in_addr, len)) {
+			reserved_target = true;
 		}
-		if (!consumed_reservation) {
+		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;
 		alignment                          = (alignment != 0 ? alignment : DEFAULT_ALIGNMENT);
@@ -3003,16 +2988,13 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 			    return range.type == VirtualRangeType::Reserved;
 		    })) {
 			UnmapGpuRange(in_addr, len);
-			consumed_range = reserved_ranges.front();
-			if (g_virtual_ranges->ConsumeReservedSpan(in_addr, len)) {
-				consumed_reservation = true;
-				if (map_shared_fixed(in_addr)) {
-					out_addr       = in_addr;
-					shared_backing = true;
-				}
+			reserved_target = true;
+			if (map_shared_fixed(in_addr)) {
+				out_addr       = in_addr;
+				shared_backing = true;
 			}
 		}
-		if (!consumed_reservation) {
+		if (!reserved_target) {
 			out_addr = FindGuestFreeRange(in_addr, len, alignment);
 			if (out_addr != 0) {
 				UnmapGpuRange(out_addr, len);
@@ -3046,34 +3028,28 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 	     magic_enum::enum_name(gpu_mode), shared_backing ? "yes" : "no", shared_reason);
 
 	if (out_addr == 0) {
-		if (consumed_reservation) {
-			g_virtual_ranges->Add(in_addr, len, 0, 0, 0, VirtualRangeType::Reserved,
-			                      consumed_range.name);
-		}
 		return KERNEL_ERROR_ENOMEM;
 	}
 
 	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 (consumed_reservation) {
-			EXIT_IF(!g_virtual_ranges->Add(in_addr, len, 0, 0, 0, VirtualRangeType::Reserved,
-			                               consumed_range.name));
-		}
 		return KERNEL_ERROR_EBUSY;
 	}
 
 	PhysicalMemory::AllocatedBlock mapped_block {};
 	g_physical_memory->Find(direct_memory_start, false, &mapped_block);
-	if (!g_virtual_ranges->Add(out_addr, len, direct_memory_start, prot, mapped_block.memory_type,
-	                           VirtualRangeType::Direct, "")) {
+	const bool published = reserved_target
+	                           ? g_virtual_ranges->ReplaceSpan(
+	                                 out_addr, len, VirtualRangeType::Reserved, 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,
+	                                                   VirtualRangeType::Direct, "");
+	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 (consumed_reservation) {
-			EXIT_IF(!g_virtual_ranges->Add(in_addr, len, 0, 0, 0, VirtualRangeType::Reserved,
-			                               consumed_range.name));
-		}
 		return KERNEL_ERROR_EBUSY;
 	}
 
@@ -3442,6 +3418,10 @@ void TestFailNextFixedReserveRangeRegistration() {
 	g_test_fail_next_fixed_reserve_range_add = true;
 }
 
+void TestFailNextVirtualRangeReplacement() {
+	g_test_fail_next_range_replace = true;
+}
+
 bool TestPlaceholderRangeIsFree(uint64_t vaddr, uint64_t size) {
 	return g_guest_address_space->TestContainsFree(vaddr, size);
 }
@@ -3965,12 +3945,8 @@ int KYTY_SYSV_ABI KernelMemoryPoolReserve(void* addr_in, size_t len, size_t alig
 		    !IsReservedRangeType(reserved_range.type)) {
 			return KERNEL_ERROR_EBUSY;
 		}
-		g_virtual_ranges->Remove(out_vaddr, len);
-		if (!g_virtual_ranges->Add(out_vaddr, len, 0, 0, 0, VirtualRangeType::PoolReserved,
-		                           reserved_range.name)) {
-			g_virtual_ranges->Add(out_vaddr, len, reserved_range.offset, reserved_range.protection,
-			                      reserved_range.memory_type, reserved_range.type,
-			                      reserved_range.name);
+		if (!g_virtual_ranges->ReplaceSpan(out_vaddr, len, VirtualRangeType::Reserved, 0, 0, 0,
+		                                   VirtualRangeType::PoolReserved, reserved_range.name)) {
 			return KERNEL_ERROR_EBUSY;
 		}
 		*addr_out = out_addr;
@@ -4011,17 +3987,13 @@ int KYTY_SYSV_ABI KernelMemoryPoolCommit(void* addr, size_t len, int type, int p
 
 	VirtualRanges::Range old_range {};
 	if (!g_virtual_ranges->Query(vaddr, 0, &old_range) ||
-	    old_range.type != VirtualRangeType::PoolReserved) {
-		return KERNEL_ERROR_EACCES;
-	}
-
-	if (!g_virtual_ranges->ConsumeReserved(vaddr, len, VirtualRangeType::PoolReserved)) {
+	    old_range.type != VirtualRangeType::PoolReserved ||
+	    len > old_range.size - (vaddr - old_range.start)) {
 		return KERNEL_ERROR_EACCES;
 	}
 
 	std::vector<PooledMemory::Mapping> mappings;
 	if (!g_pooled_memory->Allocate(vaddr, len, gpu_mode, &mappings)) {
-		g_virtual_ranges->Add(vaddr, len, 0, 0, 0, VirtualRangeType::PoolReserved, old_range.name);
 		return KERNEL_ERROR_ENOMEM;
 	}
 
@@ -4034,7 +4006,6 @@ int KYTY_SYSV_ABI KernelMemoryPoolCommit(void* addr, size_t len, int type, int p
 		if (!g_pooled_memory->Release(vaddr, len, &rollback_gpu_mode)) {
 			EXIT("failed to release pooled-memory rollback allocation\n");
 		}
-		g_virtual_ranges->Add(vaddr, len, 0, 0, 0, VirtualRangeType::PoolReserved, old_range.name);
 	};
 
 	for (const auto& mapping: mappings) {
@@ -4050,8 +4021,8 @@ int KYTY_SYSV_ABI KernelMemoryPoolCommit(void* addr, size_t len, int type, int p
 		mapped.push_back(mapping);
 	}
 
-	if (!g_virtual_ranges->Add(vaddr, len, 0, prot, type, VirtualRangeType::Pooled,
-	                           old_range.name)) {
+	if (!g_virtual_ranges->ReplaceSpan(vaddr, len, VirtualRangeType::PoolReserved, 0, prot, type,
+	                                   VirtualRangeType::Pooled, old_range.name)) {
 		rollback();
 		return KERNEL_ERROR_EBUSY;
 	}

src/kernel/memory.h

@@ -190,6 +190,7 @@ void     TestFailNextPhysicalMemoryUnmap();
 void     TestFailPhysicalMemoryUnmapAfter(uint32_t successful_unmaps);
 void     TestFailGuestBackingStoreUnmapAfter(uint32_t successful_unmaps);
 void     TestFailNextFixedReserveRangeRegistration();
+void     TestFailNextVirtualRangeReplacement();
 bool     TestPlaceholderRangeIsFree(uint64_t vaddr, uint64_t size);
 bool     TestGuestAddressRangeIsOwned(uint64_t vaddr, uint64_t size);
 bool     TestGuestBackingOutsideAddressSpace();

src/libs/agc.cpp

@@ -2110,7 +2110,7 @@ uint32_t KYTY_SYSV_ABI AgcCbQueueEndOfPipeActionGetSize() {
 uint32_t* KYTY_SYSV_ABI AgcAcbResetQueue(CommandBuffer* buf, uint32_t op) {
 	PRINT_NAME();
 
-	LOGF("\t op    = 0x%08" PRIx32 "\n", op);
+	AgcTrace("\t op    = 0x%08" PRIx32 "\n", op);
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED((op & ~0x1c2u) != 0);
@@ -2130,9 +2130,9 @@ uint32_t* KYTY_SYSV_ABI AgcAcbResetQueue(CommandBuffer* buf, uint32_t op) {
 uint32_t* KYTY_SYSV_ABI AgcDcbResetQueue(CommandBuffer* buf, uint32_t op, uint32_t state) {
 	PRINT_NAME();
 
-	LOGF("\t op    = 0x%08" PRIx32 "\n"
-	     "\t state = 0x%08" PRIx32 "\n",
-	     op, state);
+	AgcTrace("\t op    = 0x%08" PRIx32 "\n"
+	         "\t state = 0x%08" PRIx32 "\n",
+	         op, state);
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED((op & ~0xfffu) != 0);
@@ -2749,9 +2749,9 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetBaseIndirectArgs(CommandBuffer* buf, uint32_t s
                                                   const volatile void* indirect_base_addr) {
 	PRINT_NAME();
 
-	LOGF("\t shader_type        = %" PRIu32 "\n"
-	     "\t indirect_base_addr = 0x%016" PRIx64 "\n",
-	     shader_type, reinterpret_cast<uint64_t>(indirect_base_addr));
+	AgcTrace("\t shader_type        = %" PRIu32 "\n"
+	         "\t indirect_base_addr = 0x%016" PRIx64 "\n",
+	         shader_type, reinterpret_cast<uint64_t>(indirect_base_addr));
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
@@ -2941,9 +2941,9 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDispatchIndirect(CommandBuffer* buf, uint32_t data
                                                uint32_t flags) {
 	PRINT_NAME();
 
-	LOGF("\t data_offset = 0x%" PRIx32 "\n"
-	     "\t flags       = 0x%08" PRIx32 "\n",
-	     data_offset_in_bytes, flags);
+	AgcTrace("\t data_offset = 0x%" PRIx32 "\n"
+	         "\t flags       = 0x%08" PRIx32 "\n",
+	         data_offset_in_bytes, flags);
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
@@ -3294,9 +3294,9 @@ uint32_t* KYTY_SYSV_ABI AgcAcbDispatchIndirect(CommandBuffer*       buf,
                                                uint32_t             modifier) {
 	PRINT_NAME();
 
-	LOGF("\t indirect_args = 0x%016" PRIx64 "\n"
-	     "\t modifier      = 0x%08" PRIx32 "\n",
-	     reinterpret_cast<uint64_t>(indirect_args), modifier);
+	AgcTrace("\t indirect_args = 0x%016" PRIx64 "\n"
+	         "\t modifier      = 0x%08" PRIx32 "\n",
+	         reinterpret_cast<uint64_t>(indirect_args), modifier);
 
 	if (buf == nullptr) {
 		return nullptr;
@@ -3724,9 +3724,9 @@ int KYTY_SYSV_ABI AgcCondExecPatchSetCommandAddress(uint32_t*                cmd
                                                     const volatile uint32_t* command) {
 	PRINT_NAME();
 
-	LOGF("\t cmd     = 0x%016" PRIx64 "\n"
-	     "\t command = 0x%016" PRIx64 "\n",
-	     reinterpret_cast<uint64_t>(cmd), reinterpret_cast<uint64_t>(command));
+	AgcTrace("\t cmd     = 0x%016" PRIx64 "\n"
+	         "\t command = 0x%016" PRIx64 "\n",
+	         reinterpret_cast<uint64_t>(cmd), reinterpret_cast<uint64_t>(command));
 
 	if (cmd == nullptr || command == nullptr) {
 		return GRAPHICS5_ERROR_INVALID_PACKET;
@@ -3868,9 +3868,9 @@ static uint32_t* get_agc_wait_packet(uint32_t* cmd) {
 int KYTY_SYSV_ABI AgcWaitRegMemPatchAddress(uint32_t* cmd, const volatile void* address) {
 	PRINT_NAME();
 
-	LOGF("\t cmd     = 0x%016" PRIx64 "\n"
-	     "\t address = 0x%016" PRIx64 "\n",
-	     reinterpret_cast<uint64_t>(cmd), reinterpret_cast<uint64_t>(address));
+	AgcTrace("\t cmd     = 0x%016" PRIx64 "\n"
+	         "\t address = 0x%016" PRIx64 "\n",
+	         reinterpret_cast<uint64_t>(cmd), reinterpret_cast<uint64_t>(address));
 
 	auto* wait = get_agc_wait_packet(cmd);
 	if (wait == nullptr) {
@@ -3913,9 +3913,9 @@ int KYTY_SYSV_ABI AgcQueueEndOfPipeActionPatchAddress(uint32_t*             cmd,
 
 	// Not sure
 
-	LOGF("\t cmd     = 0x%016" PRIx64 "\n"
-	     "\t address = 0x%016" PRIx64 "\n",
-	     reinterpret_cast<uint64_t>(cmd), reinterpret_cast<uint64_t>(address));
+	AgcTrace("\t cmd     = 0x%016" PRIx64 "\n"
+	         "\t address = 0x%016" PRIx64 "\n",
+	         reinterpret_cast<uint64_t>(cmd), reinterpret_cast<uint64_t>(address));
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
 
@@ -4246,7 +4246,7 @@ static void submit_acb(uint32_t queue, uint32_t* acb, uint32_t size_in_dwords) {
 	}
 
 	for (uint32_t i = 0; i < std::min<uint32_t>(size_in_dwords, 8); i++) {
-		LOGF("\t acb[%u] = 0x%08" PRIx32 "\n", i, acb[i]);
+		AgcTrace("\t acb[%u] = 0x%08" PRIx32 "\n", i, acb[i]);
 	}
 
 	GraphicsDbgDumpDcb("a", size_in_dwords, acb);
@@ -4322,17 +4322,17 @@ int KYTY_SYSV_ABI AgcDriverSubmitMultiCommandBuffers(void*            queue_cont
 int KYTY_SYSV_ABI AgcDriverSubmitAcb(uint32_t queue, const Packet* packet) {
 	PRINT_NAME();
 
-	LOGF("\t queue  = 0x%08" PRIx32 "\n"
-	     "\t packet = 0x%016" PRIx64 "\n",
-	     queue, reinterpret_cast<uint64_t>(packet));
+	AgcTrace("\t queue  = 0x%08" PRIx32 "\n"
+	         "\t packet = 0x%016" PRIx64 "\n",
+	         queue, reinterpret_cast<uint64_t>(packet));
 
 	if (packet == nullptr) {
 		return OK;
 	}
-	LOGF("\t acb   = 0x%016" PRIx64 "\n"
-	     "\t size  = 0x%08" PRIx32 "\n"
-	     "\t flags = 0x%02" PRIx8 "\n",
-	     reinterpret_cast<uint64_t>(packet->addr), packet->dw_num, packet->flags);
+	AgcTrace("\t acb   = 0x%016" PRIx64 "\n"
+	         "\t size  = 0x%08" PRIx32 "\n"
+	         "\t flags = 0x%02" PRIx8 "\n",
+	         reinterpret_cast<uint64_t>(packet->addr), packet->dw_num, packet->flags);
 
 	submit_acb(queue, packet->addr, packet->dw_num);
 	return OK;

src/libs/ajm.cpp

@@ -682,16 +682,9 @@ static AjmDecodeResult AjmControlInstance(uint32_t instance, uint64_t flags,
 
 static AjmDecodeResult AjmSetGaplessDecode(uint32_t instance, const void* gapless_decode,
                                            int reset) {
-	std::scoped_lock lock(g_ajm_instances_mutex);
-	auto*            state = AjmFindInstanceLocked(instance);
-	if (state == nullptr || state->decoder == nullptr || gapless_decode == nullptr) {
-		AjmDecodeResult result {};
-		result.result = AJM_RESULT_INVALID_PARAMETER;
-		return result;
-	}
-
-	state->gapless.Set(*static_cast<const AjmSidebandGaplessDecode*>(gapless_decode), reset != 0);
-	return state->decoder->MakeResult();
+	const auto flags = AJM_FLAG_SIDEBAND_GAPLESS | (reset != 0 ? AJM_FLAG_CONTROL_RESET : 0);
+	return AjmControlInstance(instance, flags, gapless_decode, sizeof(AjmSidebandGaplessDecode),
+	                          nullptr, nullptr);
 }
 
 static AjmDecodeResult AjmGetGaplessDecode(uint32_t                  instance,
@@ -930,18 +923,8 @@ int KYTY_SYSV_ABI AjmBatchJobClearContext(AjmBatchInfo* info, uint32_t instance,
 	PRINT_NAME();
 	EXIT_NOT_IMPLEMENTED(info == nullptr);
 
-	AjmDecodeResult decode_result {};
-	{
-		std::scoped_lock lock(g_ajm_instances_mutex);
-		auto*            state = AjmFindInstanceLocked(instance);
-		if (state != nullptr && state->decoder != nullptr) {
-			state->gapless.Reset();
-			state->decoder->Reset();
-			decode_result = state->decoder->MakeResult();
-		} else {
-			decode_result.result = AJM_RESULT_INVALID_PARAMETER;
-		}
-	}
+	const auto decode_result =
+	    AjmControlInstance(instance, AJM_FLAG_CONTROL_RESET, nullptr, 0, nullptr, nullptr);
 
 	AjmWriteResult(result, AJM_SIDEBAND_RESULT_SIZE, decode_result);
 	LOGF("\t instance = 0x%08" PRIx32 "\n", instance);

src/libs/libVideoOut.cpp

@@ -1,5 +1,6 @@
 #include "common/abi.h"
 #include "graphics/presentation/videoOut.h"
+#include "libs/errno.h"
 #include "libs/libs.h"
 #include "loader/symbolDatabase.h"
 
@@ -51,8 +52,14 @@ namespace LibGen5::VrrStatus {
 
 LIB_VERSION("VideoOutVrrStatus", 1, "VideoOut", 1, 1);
 
+static KYTY_SYSV_ABI int VideoOutVrrStatus_kP2L8t3j_aM() {
+	// The observed guest call passes no arguments.
+	// Return success for Kyty's fixed-refresh path.
+	return OK;
+}
+
 LIB_DEFINE(InitVideoOutVrrStatus_1) {
-	LIB_FUNC("kP2L8t3j-aM", VideoOut::VideoOutGetVrrStatus);
+	LIB_FUNC("kP2L8t3j-aM", VideoOutVrrStatus_kP2L8t3j_aM);
 }
 
 } // namespace LibGen5::VrrStatus

src/libs/libs.h

@@ -7,6 +7,8 @@
 #include "common/threads.h"
 #include "loader/timer.h" // IWYU pragma: keep
 
+#include <fmt/format.h>
+
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 #define PRINT_NAME_ENABLED g_print_name
 
@@ -49,12 +51,7 @@
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 #define PRINT_NAME()                                                                               \
 	if (PRINT_NAME_ENABLED) {                                                                      \
-		if (Log::GetDirection() != Log::Direction::Silent) {                                       \
-			const auto print_name_time = Loader::Timer::GetTime().ToString("HH24:MI:SS.FFF");      \
-			LOGF_COLOR(Log::Color::Cyan, "[%d][%s] %s::%s::%s()\n",                                \
-			           Common::Thread::GetThreadIdUnique(), print_name_time.c_str(), g_library,    \
-			           g_module, __func__);                                                        \
-		}                                                                                          \
+		Libs::PrintName(g_library, g_module, __func__);                                              \
 	}
 
 namespace Loader {
@@ -63,6 +60,19 @@ class SymbolDatabase;
 
 namespace Libs {
 
+// Keep the formatting path from inflating fiber functions' stack frames under LTO.
+[[gnu::noinline]] inline void PrintName(const char* library, const char* module, const char* function) {
+	if (Log::GetDirection() != Log::Direction::Silent) {
+		const auto elapsed_ms      = static_cast<uint64_t>(Loader::Timer::GetTimeMs());
+		const auto print_name_time = fmt::format(
+		    "{:02}:{:02}:{:02}.{:03}", elapsed_ms / 3600000, (elapsed_ms / 60000) % 60,
+		    (elapsed_ms / 1000) % 60, elapsed_ms % 1000);
+		LOGF_COLOR(Log::Color::Cyan, "[%d][%s] %s::%s::%s()\n",
+		           Common::Thread::GetThreadIdUnique(), print_name_time.c_str(), library, module,
+		           function);
+	}
+}
+
 void InitAll(Loader::SymbolDatabase* s);
 
 } // namespace Libs

src/libs/ngs2.cpp

@@ -10,6 +10,7 @@
 #include <algorithm>
 #include <array>
 #include <atomic>
+#include <cmath>
 #include <cstring>
 #include <deque>
 #include <limits>
@@ -428,11 +429,19 @@ struct Ngs2VoiceInternal {
 	std::unique_ptr<Ajm::AjmAt9Decoder> decoder;
 	std::vector<float>                 decoded_frame;
 	bool                               accepts_blocks = true;
+	uint32_t                           sample_rate    = 0;
+	uint64_t                           sample_phase   = 0;
+	uint64_t                           sample_step    = 0;
 
 	void SetupSampler(const Ngs2WaveformFormat& format) {
-		EXIT_NOT_IMPLEMENTED(format.sample_rate != rack->ngs->option.sample_rate ||
-		                     format.frame_margin != 0 || format.frame_offset != 0);
+		EXIT_NOT_IMPLEMENTED(format.frame_margin != 0 || format.frame_offset != 0 ||
+		                     format.sample_rate == 0 ||
+		                     (format.waveform_type == NGS2_WAVEFORM_TYPE_ATRAC9 &&
+		                      format.sample_rate != rack->ngs->option.sample_rate));
 		channels = format.num_channels;
+		sample_rate = format.sample_rate;
+		sample_phase = 0;
+		sample_step = uint64_t(sample_rate) << 32u;
 		blocks.clear();
 		decoder.reset();
 		decoded_frame.clear();
@@ -563,6 +572,16 @@ struct Ngs2SamplerVoiceState {
 	const void*    waveform_data;
 };
 
+struct Ngs2CustomSamplerVoiceState {
+	Ngs2VoiceState voice_state;
+	const void*    waveform_data;
+	uint64_t       num_decoded_samples;
+	uint64_t       decoded_data_size;
+	uint64_t       user_data;
+	uint32_t       reserved;
+	uint32_t       reserved2;
+};
+
 static Ngs2Internal*        g_ngs_list     = nullptr;
 static Ngs2RackInternal*    g_racks_list   = nullptr;
 static std::atomic_uint32_t g_next_ngs_uid = 1;
@@ -578,6 +597,7 @@ static_assert(sizeof(Ngs2VoiceState) == 8);
 static_assert(sizeof(Ngs2SubmixerVoiceState) == 20);
 static_assert(sizeof(Ngs2CustomMasteringVoiceState) == 16);
 static_assert(sizeof(Ngs2SamplerVoiceState) == 56);
+static_assert(sizeof(Ngs2CustomSamplerVoiceState) == 48);
 static_assert(sizeof(Ngs2WaveformFormat) == 24);
 static_assert(sizeof(Ngs2WaveformBlock) == 40);
 static_assert(sizeof(Ngs2WaveformInfo) == 232);
@@ -1244,76 +1264,118 @@ static void Ngs2ApplyEvent(Ngs2VoiceInternal& voice) {
 	voice.event = Ngs2VoicePlayEvent::None;
 }
 
+static void Ngs2FinishBlock(Ngs2VoiceInternal& voice) {
+	const auto& block = voice.blocks.front();
+	struct CallbackInfo {
+		uintptr_t   data, voice;
+		uint32_t    flag, reserved;
+		uintptr_t   user;
+		const void* block_data;
+		size_t      size;
+		uint32_t    repeats, attributes;
+	} info {voice.callback_data,
+	        reinterpret_cast<uintptr_t>(&voice),
+	        1,
+	        0,
+	        block.info.user_data,
+	        block.data,
+	        block.info.data_size,
+	        0,
+	        0};
+	static_assert(sizeof(CallbackInfo) == 56);
+	voice.blocks.pop_front();
+	if (voice.blocks.empty() && !voice.accepts_blocks) {
+		voice.state = Ngs2VoicePlayState::Empty;
+	}
+	if (voice.callback != 0 && (voice.callback_flags & 1u) != 0) {
+		reinterpret_cast<void KYTY_SYSV_ABI (*)(const CallbackInfo*)>(voice.callback)(&info);
+	}
+}
+
+static void Ngs2AdvancePcm(Ngs2VoiceInternal& voice) {
+	const auto output_rate = uint64_t(voice.rack->ngs->option.sample_rate) << 32u;
+	while (!voice.blocks.empty() && voice.sample_phase >= output_rate) {
+		auto&      block     = voice.blocks.front();
+		const auto available = block.info.num_samples - block.cursor;
+		const auto advance   = std::min(voice.sample_phase / output_rate,
+		                                uint64_t(available));
+		block.cursor += static_cast<uint32_t>(advance);
+		voice.sample_phase -= advance * output_rate;
+		if (block.cursor == block.info.num_samples) {
+			Ngs2FinishBlock(voice);
+		}
+	}
+}
+
 static void Ngs2ConsumeSamples(Ngs2VoiceInternal& voice, uint32_t grain) {
 	uint32_t output = 0;
 	while (output < grain && !voice.blocks.empty()) {
+		if (voice.decoder == nullptr) {
+			Ngs2AdvancePcm(voice);
+			if (voice.blocks.empty()) {
+				break;
+			}
+			const auto& block = voice.blocks.front();
+			const auto* pcm = reinterpret_cast<const int16_t*>(block.data) +
+			                  (block.info.num_skip_samples + block.cursor) * voice.channels;
+			const auto* next = pcm;
+			if (block.cursor + 1 < block.info.num_samples) {
+				next += voice.channels;
+			} else if (voice.blocks.size() > 1) {
+				const auto& next_block = voice.blocks[1];
+				next = reinterpret_cast<const int16_t*>(next_block.data) +
+				       next_block.info.num_skip_samples * voice.channels;
+			}
+			const float fraction = static_cast<float>(voice.sample_phase) /
+			                       static_cast<float>(uint64_t(voice.rack->ngs->option.sample_rate)
+			                                          << 32u);
+			for (uint32_t c = 0; c < voice.channels; ++c) {
+				voice.samples[c * grain + output] =
+				    (static_cast<float>(pcm[c]) +
+				     (static_cast<float>(next[c]) - pcm[c]) * fraction) /
+				    32768.0f;
+			}
+			voice.sample_phase += voice.sample_step;
+			++output;
+			continue;
+		}
 		auto&          block   = voice.blocks.front();
 		auto           count   = std::min(grain - output, block.info.num_samples - block.cursor);
-		const int16_t* pcm     = nullptr;
-		const float*   decoded = nullptr;
-		if (voice.decoder != nullptr) {
-			const auto skip =
-			    std::min(block.skip_remaining,
-			             static_cast<uint32_t>(voice.decoded_frame.size() / voice.channels));
-			Ajm::AjmGaplessState gapless;
-			gapless.Set({block.info.num_samples - block.cursor, static_cast<uint16_t>(skip), 0},
-			            true);
-			const auto result = voice.decoder->Decode(
-			    block.data + block.data_cursor, block.info.data_size - block.data_cursor,
-			    voice.decoded_frame.data(), voice.decoded_frame.size() * sizeof(float), false,
-			    &gapless);
-			EXIT_NOT_IMPLEMENTED(result.result != OK || result.input_consumed == 0);
-			block.data_cursor += result.input_consumed;
-			block.skip_remaining -= skip - gapless.current.skip_samples;
-			const auto decoded_samples =
-			    static_cast<uint32_t>(result.output_written / (sizeof(float) * voice.channels));
-			if (decoded_samples == 0) {
-				continue;
-			}
-			EXIT_NOT_IMPLEMENTED(decoded_samples > count);
-			count   = decoded_samples;
-			decoded = voice.decoded_frame.data();
-		} else {
-			pcm = reinterpret_cast<const int16_t*>(block.data) +
-			      (block.info.num_skip_samples + block.cursor) * voice.channels;
+		const auto skip =
+		    std::min(block.skip_remaining,
+		             static_cast<uint32_t>(voice.decoded_frame.size() / voice.channels));
+		Ajm::AjmGaplessState gapless;
+		gapless.Set({block.info.num_samples - block.cursor, static_cast<uint16_t>(skip), 0},
+		            true);
+		const auto result = voice.decoder->Decode(
+		    block.data + block.data_cursor, block.info.data_size - block.data_cursor,
+		    voice.decoded_frame.data(), voice.decoded_frame.size() * sizeof(float), false,
+		    &gapless);
+		EXIT_NOT_IMPLEMENTED(result.result != OK || result.input_consumed == 0);
+		block.data_cursor += result.input_consumed;
+		block.skip_remaining -= skip - gapless.current.skip_samples;
+		const auto decoded_samples =
+		    static_cast<uint32_t>(result.output_written / (sizeof(float) * voice.channels));
+		if (decoded_samples == 0) {
+			continue;
 		}
+		EXIT_NOT_IMPLEMENTED(decoded_samples > count);
+		count = decoded_samples;
 		for (uint32_t c = 0; c < voice.channels; ++c) {
 			for (uint32_t i = 0; i < count; ++i) {
 				const auto index = i * voice.channels + c;
-				voice.samples[c * grain + output + i] =
-				    decoded != nullptr ? decoded[index] : pcm[index] / 32768.0f;
+				voice.samples[c * grain + output + i] = voice.decoded_frame[index];
 			}
 		}
 		output += count;
 		block.cursor += count;
 		if (block.cursor == block.info.num_samples) {
-			struct CallbackInfo {
-				uintptr_t   data, voice;
-				uint32_t    flag, reserved;
-				uintptr_t   user;
-				const void* block_data;
-				size_t      size;
-				uint32_t    repeats, attributes;
-			} info {voice.callback_data,
-			        reinterpret_cast<uintptr_t>(&voice),
-			        1,
-			        0,
-			        block.info.user_data,
-			        block.data,
-			        block.info.data_size,
-			        0,
-			        0};
-			static_assert(sizeof(CallbackInfo) == 56);
-			voice.blocks.pop_front();
-			if (voice.blocks.empty() && !voice.accepts_blocks) {
-				voice.state = Ngs2VoicePlayState::Empty;
-			}
-			if (voice.callback != 0 && (voice.callback_flags & 1u) != 0) {
-				reinterpret_cast<void KYTY_SYSV_ABI (*)(const CallbackInfo*)>(voice.callback)(
-				    &info);
-			}
+			Ngs2FinishBlock(voice);
 		}
 	}
+	if (voice.decoder == nullptr) {
+		Ngs2AdvancePcm(voice);
+	}
 	voice.has_samples = output != 0;
 }
 
@@ -1925,7 +1987,12 @@ int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamH
 						};
 						const auto& blocks = *reinterpret_cast<const BlocksParam*>(param);
 						EXIT_NOT_IMPLEMENTED(!voice->accepts_blocks ||
-						                     (blocks.flags != 0 && blocks.flags != 0x11));
+						                     (blocks.flags != 0 && blocks.flags != 0x11 &&
+						                      blocks.flags != 0x4));
+						if (blocks.flags == 0x4) {
+							voice->blocks.clear();
+							voice->sample_phase = 0;
+						}
 						voice->accepts_blocks = (blocks.flags & 1u) != 0;
 						for (uint32_t i = 0; i < blocks.count; ++i) {
 							const auto& block = blocks.blocks[i];
@@ -1939,9 +2006,14 @@ int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamH
 						}
 						break;
 					}
-					case 5:
-						EXIT_NOT_IMPLEMENTED(*reinterpret_cast<const float*>(param + 1) != 1.0f);
+					case 5: {
+						const float ratio = *reinterpret_cast<const float*>(param + 1);
+						EXIT_NOT_IMPLEMENTED(!std::isfinite(ratio) || ratio < 0.0f || ratio > 4.0f ||
+						                     (voice->decoder != nullptr && ratio != 1.0f));
+						voice->sample_step = static_cast<uint64_t>(std::llround(
+						    static_cast<double>(voice->sample_rate) * ratio * 4294967296.0));
 						break;
+					}
 					default: EXIT("unsupported sampler control: 0x%08" PRIx32 "\n", param->id);
 				}
 				break;
@@ -2041,8 +2113,17 @@ int KYTY_SYSV_ABI Ngs2VoiceGetState(uintptr_t voice_handle, Ngs2VoiceState* stat
 			mastering->voice_state.state_flags = Ngs2GetStateFlags(voice);
 			break;
 		}
-		case Ngs2RackType::Sampler:
 		case Ngs2RackType::CustomSampler: {
+			const auto configured_size =
+			    voice->rack->option.custom_sampler.custom_rack_option.state_size;
+			EXIT_NOT_IMPLEMENTED(configured_size < sizeof(Ngs2CustomSamplerVoiceState));
+			EXIT_NOT_IMPLEMENTED(state_size != configured_size);
+			std::memset(state, 0, state_size);
+			auto* sampler = reinterpret_cast<Ngs2CustomSamplerVoiceState*>(state);
+			sampler->voice_state.state_flags = Ngs2GetStateFlags(voice);
+			break;
+		}
+		case Ngs2RackType::Sampler: {
 			if (state_size != sizeof(Ngs2SamplerVoiceState)) {
 				LOGF("\t warning: sampler state_size = 0x%016" PRIx64 ", expected 0x%016" PRIx64
 				     "\n",

src/loader/timer.cpp

@@ -1,7 +1,5 @@
 #include "common/timer.h"
 
-#include "common/abi.h"
-#include "common/dateTime.h"
 #include "loader/timer.h"
 
 namespace Loader::Timer {
@@ -16,8 +14,4 @@ double GetTimeMs() {
 	return g_timer.GetTimeMs();
 }
 
-Common::Time GetTime() {
-	return Common::Time(static_cast<int>(GetTimeMs()));
-}
-
 } // namespace Loader::Timer

src/loader/timer.h

@@ -1,14 +1,10 @@
 #ifndef EMULATOR_INCLUDE_EMULATOR_LOADER_TIMER_H_
 #define EMULATOR_INCLUDE_EMULATOR_LOADER_TIMER_H_
 
-#include "common/abi.h"
-#include "common/common.h"
-#include "common/dateTime.h"
 namespace Loader::Timer {
 
-void         Start();
-double       GetTimeMs();
-Common::Time GetTime();
+void   Start();
+double GetTimeMs();
 
 struct Lifecycle {
 	static constexpr const char* name       = "Timer";

tests/ResourceTrackingTests.cpp

@@ -614,7 +614,8 @@ void TestGuardedDirectImageTable() {
     const auto &indirect = fixture.program.descriptor_sources[source].indirect_image;
     Check(indirect && indirect->material_source == UINT32_MAX &&
               indirect->selector_stride == 0u && indirect->table_offset == 344u &&
-              indirect->key_count == 32u &&
+              indirect->key_count.Resolve().IsImmediate() &&
+              indirect->key_count.Resolve().U32() == 32u &&
               fixture.program.descriptor_sources[indirect->table_source].dword_count == 2u,
           "guarded direct image table lost its pointer or proven selector range");
     return ExtractResourcePlan(fixture.program);
@@ -699,6 +700,390 @@ void TestGuardedDirectImageTable() {
         "batched descriptor read crossed the 48-bit endpoint");
 }
 
+void TestBoundedComputeImageLoop() {
+  namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
+  enum class Variant {
+    Bounded, WrongGuard, EntryBypass, ExitBypass, GuardBlock, WrongStep
+  };
+  const auto make_plan = [](Variant variant) {
+    Fixture fixture;
+    auto *entry = fixture.block;
+    auto *header = fixture.AddBlock();
+    auto *body = fixture.AddBlock();
+    auto *latch = fixture.AddBlock();
+    auto *exit = fixture.AddBlock();
+    entry->AddBranch(header);
+    header->AddBranch(exit);
+    header->AddBranch(body);
+    body->AddBranch(latch);
+    latch->AddBranch(header);
+    if (variant == Variant::EntryBypass) entry->AddBranch(body);
+    if (variant == Variant::ExitBypass) exit->AddBranch(body);
+    fixture.program.block_info[0].terminator = {
+        .kind = variant == Variant::EntryBypass ? CFG::TerminatorKind::ConditionalBranch
+                                               : CFG::TerminatorKind::Branch,
+        .true_block = 1u, .false_block = 2u};
+    fixture.program.block_info[1].terminator = {
+        .kind = CFG::TerminatorKind::ConditionalBranch,
+        .true_block = 4u, .false_block = 2u};
+    fixture.program.block_info[2].terminator = {
+        .kind = CFG::TerminatorKind::Branch, .true_block = 3u};
+    fixture.program.block_info[3].terminator = {
+        .kind = CFG::TerminatorKind::Branch, .true_block = 1u};
+    fixture.program.block_info[4].terminator = {
+        .kind = variant == Variant::ExitBypass ? CFG::TerminatorKind::Branch
+                                              : CFG::TerminatorKind::Return,
+        .true_block = 2u};
+
+    auto &phi = header->AppendNewInst(ValueOpcode::Phi, {},
+                                      static_cast<uint64_t>(Type::U32));
+    const auto key = Value(&phi);
+    const auto count = fixture.UserData(2);
+    const auto in_range = fixture.Emit(ValueOpcode::SLessThan32,
+                                       {variant == Variant::WrongGuard
+                                            ? Value(0u) : key,
+                                        count}, 0, header);
+    const auto allowed = fixture.Emit(ValueOpcode::LogicalAnd,
+                                      {in_range, Value(true)}, 0, header);
+    fixture.program.block_info[1].condition =
+        fixture.Emit(ValueOpcode::LogicalNot, {allowed}, 0, header);
+    const auto step = fixture.Emit(ValueOpcode::IAdd32,
+                                   {key, Value(variant == Variant::WrongStep ? 2u : 1u)},
+                                   0, latch);
+    phi.AddPhiOperand(entry, Value(0u));
+    phi.AddPhiOperand(latch, step);
+
+    fixture.block = variant == Variant::GuardBlock ? header : body;
+    const auto table = fixture.Address(fixture.UserData(0), fixture.UserData(1));
+    const auto offset = fixture.Emit(ValueOpcode::IAdd32,
+        {fixture.Emit(ValueOpcode::ShiftLeftLogical32, {key, Value(5u)}),
+         Value(0x6b0u)});
+    std::array<Value, 8> words;
+    for (uint32_t word = 0; word < words.size(); ++word) {
+      MemoryInfo memory;
+      memory.kind = ResourceKind::ScalarAddress;
+      memory.offset = word * sizeof(uint32_t);
+      words[word] = fixture.Emit(
+          ValueOpcode::LoadAddressU32,
+          {table, offset, Value(0u), Value(true)},
+          fixture.AddMemory(memory, 0x29c));
+    }
+    const auto image = fixture.Image(words, 0x29c);
+    const auto sampler = fixture.Sampler({Value(0u), Value(0u), Value(0u), Value(0u)});
+    MemoryInfo sample;
+    sample.kind = ResourceKind::Image;
+    sample.image_dimension = Decoder::ImageDimension::Dim2D;
+    fixture.Emit(ValueOpcode::ImageSampleRaw,
+                 {image, sampler, fixture.ImageAddress()},
+                 fixture.AddMemory(sample, 0x29c));
+    fixture.PlanAndTrack();
+    const auto source = fixture.program.info.images[0].source;
+    const auto &indirect = fixture.program.descriptor_sources[source].indirect_image;
+    Check(indirect && indirect->material_source == UINT32_MAX &&
+              indirect->table_offset == 0x6b0u &&
+              indirect->key_count.Resolve() == count.Resolve(),
+          "bounded compute loop lost its runtime image count");
+    return ExtractResourcePlan(fixture.program);
+  };
+
+  auto plan = make_plan(Variant::Bounded);
+  CheckFatal([&] { make_plan(Variant::WrongGuard); },
+             "not a valid runtime value",
+             "compute image loop accepted an unrelated guard");
+  CheckFatal([&] { make_plan(Variant::EntryBypass); },
+             "not a valid runtime value",
+             "compute image loop accepted an entry bypass");
+  CheckFatal([&] { make_plan(Variant::ExitBypass); },
+             "not a valid runtime value",
+             "compute image loop accepted an exit bypass");
+  CheckFatal([&] { make_plan(Variant::GuardBlock); },
+             "not a valid runtime value",
+             "compute image loop accepted a descriptor read before the guard");
+  CheckFatal([&] { make_plan(Variant::WrongStep); },
+             "not a valid runtime value",
+             "compute image loop accepted a two-step induction");
+
+  LinearTestMemory memory;
+  const auto table = 0x1800u + 0x6b0u;
+  for (uint32_t key = 0; key < 3u; ++key) {
+    const auto word = (table - memory.base) / 4u + key * 8u;
+    memory.words[word] = 0x100u + key;
+    memory.words[word + 1u] = static_cast<uint32_t>(
+        Libs::Graphics::Prospero::BufferFormat::k32_32_32_32Float) << 20u;
+    memory.words[word + 3u] = Libs::Graphics::DstSel(4, 5, 6, 7) |
+        (static_cast<uint32_t>(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
+  }
+  std::array<uint32_t, 3> user_data{0x1800u, 0u, 2u};
+  SrtRuntime runtime{.user_data = user_data,
+                     .read_memory = ReadLinearTestMemory,
+                     .userdata = &memory,
+                     .read_specialization_memory = ReadLinearTestMemory};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  for (const uint32_t count : {2u, 3u, 2u}) {
+    user_data[2] = count;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              snapshot.images.size() == count &&
+              specialization.images.size() == count &&
+              snapshot.flattened_srt[
+                  specialization.images[0].indirect_mapping_offset] == count &&
+              snapshot.images.back().dwords[0] == 0x100u + count - 1u,
+          "compute image table did not refresh for a changed loop bound");
+  }
+  for (const uint32_t count : {0u, UINT32_MAX}) {
+    user_data[2] = count;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              snapshot.images.size() == 1u &&
+              specialization.images.size() == 1u &&
+              std::ranges::all_of(snapshot.images[0].dwords,
+                                  [](uint32_t word) { return word == 0u; }) &&
+              specialization.images[0].indirect_root ==
+                  ImageResource::NoIndirectImage &&
+              specialization.images[0].indirect_search_iterations == 0u,
+          "empty compute loop bound retained unreachable image candidates");
+  }
+  user_data[2] = 65537u;
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
+        "oversized compute loop bound was accepted for image enumeration");
+}
+
+void TestUniformizedMaterialImageKeys() {
+  namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
+  const auto make_plan = [](bool wrong_update, bool wrong_equality) {
+    Fixture fixture;
+    auto *entry = fixture.block;
+    auto *header = fixture.AddBlock();
+    auto *inactive = fixture.AddBlock();
+    auto *sentinel = fixture.AddBlock();
+    auto *bit = fixture.AddBlock();
+    auto *merge = fixture.AddBlock();
+    auto *choose = fixture.AddBlock();
+    auto *sample = fixture.AddBlock();
+    auto *done = fixture.AddBlock();
+    entry->AddBranch(header);
+    header->AddBranch(inactive);
+    inactive->AddBranch(merge);
+    inactive->AddBranch(sentinel);
+    sentinel->AddBranch(merge);
+    sentinel->AddBranch(bit);
+    bit->AddBranch(header);
+    bit->AddBranch(merge);
+    merge->AddBranch(choose);
+    choose->AddBranch(sample);
+    choose->AddBranch(done);
+    sample->AddBranch(done);
+    fixture.program.block_info[0].terminator = {
+        .kind = CFG::TerminatorKind::Branch, .true_block = 1u};
+    fixture.program.block_info[1].terminator = {
+        .kind = CFG::TerminatorKind::Branch, .true_block = 2u};
+    const auto active_on_entry = fixture.Emit(
+        ValueOpcode::INotEqual32, {fixture.UserData(5u), Value(0u)}, 0, entry);
+    auto &active_phi = header->AppendNewInst(
+        ValueOpcode::Phi, {}, static_cast<uint64_t>(Type::U1));
+    auto &mask_phi = header->AppendNewInst(
+        ValueOpcode::Phi, {}, static_cast<uint64_t>(Type::U32));
+    const auto mask = Value(&mask_phi);
+    const auto active = Value(&active_phi);
+    fixture.program.block_info[2].condition = fixture.Emit(
+        ValueOpcode::LogicalNot, {active}, 0, inactive);
+    fixture.program.block_info[2].terminator = {
+        .kind = CFG::TerminatorKind::ConditionalBranch,
+        .true_block = 5u, .false_block = 3u};
+    const auto nonzero = fixture.Emit(
+        ValueOpcode::INotEqual32, {Value(0u), mask}, 0, sentinel);
+    const auto bit_guard = fixture.Emit(
+        ValueOpcode::LogicalAnd, {active, nonzero}, 0, sentinel);
+    fixture.program.block_info[3].condition = fixture.Emit(
+        ValueOpcode::LogicalNot, {bit_guard}, 0, sentinel);
+    fixture.program.block_info[3].terminator = {
+        .kind = CFG::TerminatorKind::ConditionalBranch,
+        .true_block = 5u, .false_block = 4u};
+    const auto first = fixture.Emit(ValueOpcode::FindILsb32, {mask}, 0, bit);
+    const auto position = fixture.Emit(
+        ValueOpcode::BitwiseAnd32, {first, Value(31u)}, 0, bit);
+    const auto one_bit = fixture.Emit(
+        ValueOpcode::ShiftLeftLogical32, {Value(1u), position}, 0, bit);
+    const auto cleared = fixture.Emit(
+        wrong_update ? ValueOpcode::BitwiseOr32 : ValueOpcode::BitwiseXor32,
+        {mask, one_bit}, 0, bit);
+    const auto continuation = fixture.Emit(
+        ValueOpcode::LogicalAnd, {bit_guard, active_on_entry}, 0, bit);
+    fixture.program.block_info[4].condition = continuation;
+    fixture.program.block_info[4].terminator = {
+        .kind = CFG::TerminatorKind::ConditionalBranch,
+        .true_block = 1u, .false_block = 5u};
+    active_phi.AddPhiOperand(entry, active_on_entry);
+    active_phi.AddPhiOperand(bit, continuation);
+    mask_phi.AddPhiOperand(entry, fixture.UserData(4u));
+    mask_phi.AddPhiOperand(bit, cleared);
+    const auto arbitrary = fixture.UserData(6u);
+    const auto sentinel_index = fixture.Emit(
+        ValueOpcode::SelectU32, {active, Value(32u), arbitrary}, 0, sentinel);
+    const auto bit_index = fixture.Emit(
+        ValueOpcode::SelectU32, {bit_guard, first, sentinel_index}, 0, bit);
+    auto &index_phi = merge->AppendNewInst(
+        ValueOpcode::Phi, {}, static_cast<uint64_t>(Type::U32));
+    index_phi.AddPhiOperand(inactive, arbitrary);
+    index_phi.AddPhiOperand(sentinel, sentinel_index);
+    index_phi.AddPhiOperand(bit, bit_index);
+    const auto index = Value(&index_phi);
+    const auto below = fixture.Emit(
+        ValueOpcode::SGreaterThan32, {Value(32u), index}, 0, merge);
+    const auto material_guard = fixture.Emit(
+        ValueOpcode::LogicalAnd, {active_on_entry, below}, 0, merge);
+    fixture.program.block_info[5].terminator = {
+        .kind = CFG::TerminatorKind::Branch, .true_block = 6u};
+    const auto scaled = fixture.Emit(
+        ValueOpcode::ShiftLeftLogical32, {index, Value(4u)}, 0, choose);
+    const auto selected_scale = fixture.Emit(
+        ValueOpcode::SelectU32, {material_guard, scaled, Value(0u)}, 0, choose);
+    const auto times_eight = fixture.Emit(
+        ValueOpcode::ShiftLeftLogical32, {selected_scale, Value(3u)}, 0, choose);
+    const auto times_nine = fixture.Emit(
+        ValueOpcode::IAdd32, {times_eight, selected_scale}, 0, choose);
+    const auto material_offset = fixture.Emit(
+        ValueOpcode::SelectU32,
+        {material_guard,
+         fixture.Emit(ValueOpcode::IAdd32,
+                      {times_nine, Value(0xc00u)}, 0, choose),
+         times_nine}, 0, choose);
+    const auto base = fixture.Address(fixture.UserData(0u), fixture.UserData(1u));
+    MemoryInfo material_memory;
+    material_memory.kind = ResourceKind::Global;
+    const auto loaded = fixture.Emit(
+        ValueOpcode::LoadAddressU32,
+        {base, material_offset, Value(0u), material_guard},
+        fixture.AddMemory(material_memory, 0x1a88u), choose);
+    const auto local = fixture.Emit(
+        ValueOpcode::SelectU32, {material_guard, loaded, arbitrary}, 0, choose);
+    const auto key = fixture.Emit(
+        ValueOpcode::ReadLane, {local, Value(0u)}, 0, choose);
+    const auto compared = fixture.Emit(
+        ValueOpcode::IEqual32,
+        {key, wrong_equality ? arbitrary : local}, 0, choose);
+    const auto sample_guard = fixture.Emit(
+        ValueOpcode::LogicalAnd, {material_guard, compared}, 0, choose);
+    fixture.program.block_info[6].condition = fixture.Emit(
+        ValueOpcode::LogicalNot, {sample_guard}, 0, choose);
+    fixture.program.block_info[6].terminator = {
+        .kind = CFG::TerminatorKind::ConditionalBranch,
+        .true_block = 8u, .false_block = 7u};
+    const auto table_offset = fixture.Emit(
+        ValueOpcode::IAdd32,
+        {fixture.Emit(ValueOpcode::ShiftLeftLogical32,
+                      {key, Value(5u)}, 0, sample), Value(0x20e0u)}, 0, sample);
+    std::array<Value, 8> words;
+    for (uint32_t word = 0; word < words.size(); ++word) {
+      MemoryInfo memory;
+      memory.kind = ResourceKind::ScalarAddress;
+      memory.offset = word * 4u;
+      words[word] = fixture.Emit(
+          ValueOpcode::LoadAddressU32,
+          {base, table_offset, Value(0u), Value(true)},
+          fixture.AddMemory(memory, 0x1accu), sample);
+    }
+    const auto image = fixture.Emit(
+        ValueOpcode::GetImageResource,
+        {words[0], words[1], words[2], words[3],
+         words[4], words[5], words[6], words[7]}, 0, sample);
+    const auto sampler = fixture.Emit(
+        ValueOpcode::GetSamplerResource,
+        {Value(0u), Value(0u), Value(0u), Value(0u)}, 0, sample);
+    const auto address = fixture.Emit(
+        ValueOpcode::MakeImageAddress,
+        {Value(0u), Value(0u), Value(0u), Value(0u), Value(0u), Value(0u),
+         Value(0u), Value(0u), Value(0u), Value(0u), Value(0u), Value(0u),
+         Value(0u)}, 0, sample);
+    MemoryInfo image_memory;
+    image_memory.kind = ResourceKind::Image;
+    image_memory.image_dimension = Decoder::ImageDimension::Dim2D;
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, address},
+                 fixture.AddMemory(image_memory, 0x1aecu), sample);
+    fixture.program.block_info[7].terminator = {
+        .kind = CFG::TerminatorKind::Branch, .true_block = 8u};
+    fixture.program.block_info[8].terminator.kind = CFG::TerminatorKind::Return;
+    const auto output = fixture.Emit(
+        ValueOpcode::GetBufferResource,
+        {fixture.UserData(8u), fixture.UserData(9u),
+         fixture.UserData(10u), fixture.UserData(11u)}, 0, done);
+    MemoryInfo output_memory;
+    output_memory.kind = ResourceKind::Buffer;
+    fixture.Emit(ValueOpcode::StoreBufferU32,
+                 {output, Value(0u), Value(0u), Value(0u),
+                  Value(1u), Value(true)},
+                 fixture.AddMemory(output_memory, 0x1b00u), done);
+    fixture.PlanAndTrack();
+    const auto source = fixture.program.info.images[0].source;
+    const auto &indirect = fixture.program.descriptor_sources[source].indirect_image;
+    Check(indirect && indirect->selector_stride == 0x90u &&
+              indirect->selector_offset == 0xc00u &&
+              indirect->table_offset == 0x20e0u &&
+              !indirect->selector_mask.IsEmpty() &&
+              indirect->key_count.Resolve().IsImmediate() &&
+              indirect->key_count.Resolve().U32() == 32u,
+          "uniformized image key lost its finite material range");
+    return ExtractResourcePlan(fixture.program);
+  };
+  auto plan = make_plan(false, false);
+  Check(plan.requires_specialization_memory &&
+            plan.descriptor_sources[plan.info.images[0].source]
+                .indirect_image->selector_mask.Resolve().TryInstruction() != nullptr,
+        "uniformized image mask did not survive extraction");
+  LinearTestMemory memory;
+  memory.words.resize(0x23000u / 4u);
+  constexpr uint64_t base = 0x1000u;
+  constexpr uint64_t first_material = base + 0xc00u + 2u * 0x90u;
+  constexpr uint64_t second_material = base + 0xc00u + 29u * 0x90u;
+  constexpr uint64_t first_table = base + 0x20e0u + 7u * 32u;
+  constexpr uint64_t second_table = base + 0x20e0u + 4096u * 32u;
+  memory.words[(first_material - base) / 4u] = 7u;
+  memory.words[(second_material - base) / 4u] = 4096u;
+  const auto set_descriptor = [&](uint64_t address, uint32_t color) {
+    const auto word = (address - base) / 4u;
+    memory.words[word] = color;
+    memory.words[word + 1u] = static_cast<uint32_t>(
+        Libs::Graphics::Prospero::BufferFormat::k32_32_32_32Float) << 20u;
+    memory.words[word + 3u] = Libs::Graphics::DstSel(4, 5, 6, 7) |
+        (static_cast<uint32_t>(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
+  };
+  set_descriptor(first_table, 0x200u);
+  set_descriptor(second_table, 0x400u);
+  std::array<uint32_t, 12> user_data{};
+  user_data[0] = base;
+  user_data[4] = (1u << 2u) | (1u << 29u);
+  user_data[5] = 1u;
+  user_data[8] = 0x400000u;
+  user_data[9] = 16u << 16u;
+  user_data[10] = 1u;
+  memory.fail_address = base + 0xc00u + 3u * 0x90u;
+  const SrtRuntime runtime{
+      .user_data = user_data, .read_memory = ReadLinearTestMemory,
+      .userdata = &memory,
+      .read_specialization_memory = ReadLinearTestMemory};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            snapshot.images.size() == 2u &&
+            memory.reads == 4u && memory.descriptor_reads == 2u &&
+            snapshot.flattened_srt[
+                specialization.images[0].indirect_mapping_offset] == 2u &&
+            snapshot.flattened_srt[
+                specialization.images[0].indirect_mapping_offset + 1u] == 7u &&
+            snapshot.flattened_srt[
+                specialization.images[0].indirect_mapping_offset + 3u] == 4096u,
+        "material mask did not limit sparse descriptor reads");
+  user_data[8] = first_material;
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
+        "written buffer alias with a material key was accepted");
+  user_data[8] = first_table;
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
+        "written buffer alias with an image record was accepted");
+  CheckFatal([&] { make_plan(true, false); }, "not a valid runtime value",
+             "non-clearing material mask was accepted");
+  CheckFatal([&] { make_plan(false, true); }, "not a valid runtime value",
+             "unrelated ReadLane key was accepted");
+}
+
 void TestImageDescriptorFields() {
   constexpr std::array<std::pair<uint32_t, uint32_t>, 5> reserved{
       {{1u, 0x20000000u}, {2u, 0x70003000u}, {4u, 0xe000e000u},
@@ -2506,6 +2891,8 @@ int main() {
     Run("dynamic storage mips", TestDynamicStorageMipTracking);
     Run("invariant indirect images", TestInvariantIndirectImageMaterialization);
     Run("guarded direct image table", TestGuardedDirectImageTable);
+    Run("bounded compute image loop", TestBoundedComputeImageLoop);
+    Run("uniformized material image keys", TestUniformizedMaterialImageKeys);
     Run("image descriptor fields", TestImageDescriptorFields);
     Run("draw-uniform scalar image", TestUniformScalarBufferImage);
     Run("SRT runtime", TestSrtFlatteningAndRuntimeMemoization);

tests/SaveDataMemoryTests.cpp

@@ -27,8 +27,8 @@ bool SystemContentParamSfoGetString(const char*, std::string* value) {
 }
 void SymbolDatabase::Add(const SymbolResolve&, uint64_t, const std::string&) {}
 namespace Timer {
-Common::Time GetTime() {
-	return Common::Time(0);
+double GetTimeMs() {
+	return 0.0;
 }
 } // namespace Timer
 } // namespace Loader

tests/ShaderRayTracingTests.inc

@@ -0,0 +1,87 @@
+void TestRayTracingDispatchDetection() {
+  namespace D = ShaderRecompiler::Decoder;
+  const auto check_decoded = [](std::span<const uint32_t> code, bool expected,
+                                const char *message) {
+    D::Program decoded;
+    D::DecodeProgram(code, decoded);
+    Check(decoded.has_bvh == expected, message);
+    if (expected) {
+      const auto &last = decoded.instructions.back();
+      Check(last.family == D::Family::MIMG &&
+                (last.opcode_id == 0xe6u || last.opcode_id == 0xe7u),
+            "decoder continued past the first BVH instruction");
+      const auto result = ShaderRecompiler::TranslateProgram(
+          code, MakeCompileOptions(ShaderType::Compute));
+      Check(result.skip_dispatch && result.program.blocks.empty() && result.cfg_dump.empty(),
+            "BVH compute shader reached IR translation");
+    }
+  };
+  const auto check = [&](std::initializer_list<uint32_t> code, bool expected,
+                         const char *message) {
+    check_decoded(std::span<const uint32_t>{code.begin(), code.size()}, expected, message);
+  };
+
+  check({0xf1989f01u, 0x00081014u, EncodeSopp(0x01)}, true,
+        "captured contiguous BVH instruction did not skip dispatch");
+  check({0xf1989f07u, 0x00071001u, 0x3d3e3c15u, 0x38095851u, 0x00003a39u,
+         EncodeSopp(0x01)}, true,
+        "captured NSA BVH instruction did not skip dispatch");
+  check({0xf19c9f01u, 0x00081014u, EncodeSopp(0x01)}, true,
+        "RDNA2 BVH64 instruction did not skip dispatch");
+  check({0xbe8e210eu, 0xf1989f01u, 0x00081014u, EncodeSopp(0x01)}, true,
+        "unsupported SWAPPC before BVH prevented dispatch detection");
+  check({0xf1989f01u, 0x00081014u, 0xc0000000u, EncodeSopp(0x01)}, true,
+        "unknown instruction after BVH prevented dispatch skipping");
+
+  for (uint32_t opcode : {0xe6u, 0xe7u}) {
+    for (uint32_t nsa = 0; nsa < 4; ++nsa) {
+      for (uint32_t slc = 0; slc < 2; ++slc) {
+        std::array<uint32_t, 6> code{
+            EncodeMimg0(opcode, 0xf, true) | (nsa << 1u) | (slc << 25u),
+            EncodeMimg1(0, 0, 0, 0)};
+        code[2 + nsa] = EncodeSopp(0x01);
+        check_decoded(std::span(code).first(3 + nsa), true,
+                      "BVH cache flags or NSA length prevented dispatch detection");
+      }
+    }
+  }
+  for (uint32_t opcode : {0x66u, 0x67u, 0xe4u, 0xe5u, 0xe8u, 0xe9u}) {
+    check({EncodeMimg0(opcode, 0xf), EncodeMimg1(0, 0, 0, 0), EncodeSopp(0x01)},
+          false, "non-BVH image opcode was incorrectly skipped");
+  }
+
+  check({EncodeSMovB32(0, 255), 0xf1989f01u, EncodeSopp(0x01)}, false,
+        "scalar literal was mistaken for a BVH instruction");
+  check({EncodeMimg0(0x20, 0xf) | 6u, EncodeMimg1(0, 0, 0, 0),
+         0xf1989f01u, 0xf19c9f01u, 0u, EncodeSopp(0x01)}, false,
+        "NSA address words were mistaken for BVH instructions");
+  check({EncodeMimg0(0x20, 0xf), 0xf1989f01u, EncodeSopp(0x01)}, false,
+        "MIMG operand word was mistaken for a BVH instruction");
+  check({EncodeSopp(0x01), 0xf1989f01u, 0x00081014u}, false,
+        "data after the shader end was mistaken for a BVH instruction");
+  check({EncodeMimg0(0x20, 0xf), EncodeMimg1(0, 0, 0, 0), EncodeSopp(0x01)},
+        false, "ordinary image shader was incorrectly skipped");
+
+  // An end instruction on one branch does not end another branch's code.
+  check({EncodeSopp(0x04, 1), EncodeSopp(0x01), 0xf1989f01u, 0x00081014u,
+         EncodeSopp(0x01)}, true,
+        "BVH at a branch target after an early end was not detected");
+
+  const std::array bvh_code{0xf1989f01u, 0x00081014u, 0xf19c9f01u, 0x00081014u,
+                            EncodeSopp(0x01)};
+  const std::array ordinary_code{EncodeSMovB32(0, 255), 0xf1989f01u, EncodeSopp(0x01)};
+  D::Program reused;
+  D::DecodeProgram(bvh_code, reused);
+  Check(reused.has_bvh && reused.instructions.size() == 1,
+        "decoder did not stop at the first BVH instruction");
+  D::DecodeProgram(ordinary_code, reused);
+  Check(!reused.has_bvh, "decoder retained stale BVH metadata");
+  for (const auto stage : {ShaderType::Compute, ShaderType::Vertex, ShaderType::Pixel}) {
+    const auto options = MakeCompileOptions(stage);
+    Check(!ShaderRecompiler::TranslateProgram(ordinary_code, options).skip_dispatch,
+          "BVH-shaped literal marked an ordinary shader for skipping");
+    const auto compiled = RecompileForTest(ordinary_code, options);
+    Check(!compiled.spirv.empty(), "ordinary shader did not compile");
+    CheckSpirvBinaryValidates(compiled.spirv);
+  }
+}

tests/VirtualMemoryAllocationTests.cpp

@@ -462,6 +462,89 @@ void TestPrtBackingReadPreservesSparseResidency() {
 	std::printf("[host]    %-48s ok\n", test);
 }
 
+void TestPrtReadDuringDirectCommit() {
+	const char*        test       = "PrtReadDuringDirectCommit";
+	constexpr uint64_t chunk_size = SceKernelMemoryPoolCommitLen;
+	constexpr uint64_t chunks     = 32;
+	constexpr uint64_t size       = chunk_size * chunks;
+	int64_t            physical   = -1;
+	CheckOk(test,
+	        Libs::LibKernel::Memory::KernelAllocateDirectMemory(
+	            0, Libs::LibKernel::Memory::KernelGetDirectMemorySize(), size, chunk_size,
+	            SceKernelMtypeC, &physical),
+	        "KernelAllocateDirectMemory");
+	void* arena = reinterpret_cast<void*>(0x1000000000ull);
+	CheckOk(test, Libs::LibKernel::Memory::KernelReserveVirtualRange(&arena, size, 0, size),
+	        "KernelReserveVirtualRange");
+	const auto base = reinterpret_cast<uint64_t>(arena);
+	CheckOk(test, Libs::LibKernel::Memory::KernelSetPrtAperture(2, arena, size),
+	        "KernelSetPrtAperture");
+	const auto map_chunk = [&](uint64_t index) {
+		void* address = reinterpret_cast<void*>(base + index * chunk_size);
+		return Libs::LibKernel::Memory::KernelMapNamedDirectMemory(
+		    &address, chunk_size, SceKernelProtCpuRw, SceKernelMapFixed,
+		    physical + static_cast<int64_t>(index * chunk_size), chunk_size, "prt_direct");
+	};
+	CheckOk(test, map_chunk(0), "KernelMapNamedDirectMemory(first)");
+	CheckOk(test, map_chunk(chunks - 1), "KernelMapNamedDirectMemory(last)");
+	std::memset(reinterpret_cast<void*>(base), 0x3c, chunk_size);
+	std::memset(reinterpret_cast<void*>(base + (chunks - 1) * chunk_size), 0xa7, chunk_size);
+
+	std::vector<uint8_t> bytes(size, 0x5a);
+	Check(test, Libs::LibKernel::Memory::TryReadPrtBacking(base, bytes.data(), size),
+	      "PRT read rejected a partly committed direct mapping");
+	Check(test, bytes.front() == 0x3c && bytes[chunk_size] == 0 && bytes.back() == 0xa7,
+	      "PRT read lost resident bytes or sparse zeros");
+	Libs::LibKernel::Memory::TestFailNextVirtualRangeReplacement();
+	CheckFailed(test, map_chunk(1), "KernelMapNamedDirectMemory(injected replacement failure)");
+	ExpectRange(test, Query(test, base + chunk_size), base + chunk_size,
+	            base + (chunks - 1) * chunk_size, 0, 0, 0, 0, 0, "anon");
+	Check(test, !Libs::LibKernel::Memory::TryReadBacking(base + chunk_size, bytes.data(),
+	                                                    chunk_size) &&
+	                Libs::LibKernel::Memory::TryReadPrtBacking(base, bytes.data(), size) &&
+	                bytes[chunk_size] == 0,
+	      "failed direct publication did not restore sparse reservation and backing");
+
+	std::atomic<bool>     started {false};
+	std::atomic<bool>     stop {false};
+	std::atomic<bool>     read_failed {false};
+	std::atomic<uint32_t> reads {0};
+	std::thread reader([&] {
+		std::vector<uint8_t> snapshot(size);
+		started.store(true, std::memory_order_release);
+		while (!stop.load(std::memory_order_acquire)) {
+			if (!Libs::LibKernel::Memory::TryReadPrtBacking(base, snapshot.data(), size) ||
+			    snapshot.front() != 0x3c || snapshot.back() != 0xa7) {
+				read_failed.store(true, std::memory_order_relaxed);
+			}
+			reads.fetch_add(1, std::memory_order_relaxed);
+		}
+	});
+	while (!started.load(std::memory_order_acquire)) {
+		std::this_thread::yield();
+	}
+	int map_result = OK;
+	for (uint64_t index = 1; index + 1 < chunks; ++index) {
+		map_result = map_chunk(index);
+		if (map_result != OK) {
+			break;
+		}
+	}
+	stop.store(true, std::memory_order_release);
+	reader.join();
+
+	CheckOk(test, Libs::LibKernel::Memory::KernelSetPrtAperture(2, nullptr, 0),
+	        "KernelSetPrtAperture(clear)");
+	CheckOk(test, Libs::LibKernel::Memory::KernelMunmap(base, size), "KernelMunmap");
+	CheckOk(test, Libs::LibKernel::Memory::KernelReleaseDirectMemory(physical, size),
+	        "KernelReleaseDirectMemory");
+	CheckOk(test, map_result, "KernelMapNamedDirectMemory(middle)");
+	Check(test, reads.load(std::memory_order_relaxed) != 0 &&
+	                !read_failed.load(std::memory_order_relaxed),
+	      "PRT read observed a gap while direct backing was committed");
+	std::printf("[host]    %-48s ok\n", test);
+}
+
 void TestGuestAddressSpaceHasNoFixedFallback() {
 	const char* test            = "GuestAddressSpaceHasNoFixedFallback";
 	const auto  unowned_address = reinterpret_cast<void*>(0x10000);
@@ -3031,6 +3114,7 @@ int main(int argc, char** argv) {
 	RunTest(TestProsperoArgumentAndInfoSizeContracts);
 	RunTest(TestGuestAddressSpaceOwnsReservationsBeforeBacking);
 	RunTest(TestPrtBackingReadPreservesSparseResidency);
+	RunTest(TestPrtReadDuringDirectCommit);
 	RunTest(TestGuestAddressSpaceHasNoFixedFallback);
 	RunTest(TestGuestFreeRangeSearchDoesNotUnderflow);
 	RunTest(TestFlexibleMemoryCapacityIsBootFixed);

tests/shaderCfgTests.cpp

@@ -4458,6 +4458,23 @@ void TestNewShaderDecoderArchitecture() {
             d16_hi_write.src1.sdwa_sel == 5u,
         "DS decoder rejected the captured high-half D16 write");
 
+  // Captured from GTA V (PPSA04264) compute shader, pc 0x1598.
+  const uint32_t d16_hi_byte_write_ds[] = {0xda800200u, 0x00001413u};
+  Instruction d16_hi_byte_write;
+  ShaderRecompiler::Decoder::DecodeInstruction(d16_hi_byte_write_ds, 0u,
+                                               d16_hi_byte_write);
+  Check(d16_hi_byte_write.family == Family::DS &&
+            d16_hi_byte_write.opcode == Opcode::DS_WRITE_B8_D16_HI &&
+            d16_hi_byte_write.word_count == 2u &&
+            d16_hi_byte_write.src_count == 2u &&
+            d16_hi_byte_write.data_dwords == 1u &&
+            d16_hi_byte_write.data_bits == 8u &&
+            d16_hi_byte_write.offset == 0x200u && !d16_hi_byte_write.gds &&
+            d16_hi_byte_write.src0.reg == 19u &&
+            d16_hi_byte_write.src1.reg == 20u &&
+            d16_hi_byte_write.src1.sdwa_sel == 2u,
+        "DS decoder rejected the captured high-half byte write");
+
   constexpr uint32_t packed_source_selectors[][2] = {
       {0xcc0e0000u, 0x0c0a0300u}, // Source 0: instruction bit 59.
       {0xcc0e0000u, 0x140a0300u}, // Source 1: instruction bit 60.
@@ -6694,9 +6711,9 @@ void TestNewShaderRecompilerFormattedStoreUsesRuntimeArrayLengthOnly() {
   Check(result.resources.buffers.size() == 1 &&
             result.resources.buffers[0].dwords[2] == 5u,
         "formatted store test did not preserve descriptor NumRecords");
-  Check((result.decoded_dump.find("buffer_store_format_x") != std::string::npos),
+  Check((result.decoded_dump.find("BUFFER_STORE_FORMAT_X") != std::string::npos),
         "formatted store regression did not decode buffer_store_format_x");
-  Check((result.ir_dump.find("typed=0 formatted=1") != std::string::npos),
+  Check((result.decoded_dump.find("typed=0 formatted=1") != std::string::npos),
         "formatted store regression did not preserve formatted metadata");
   CheckSpirvBinaryValidates(result.spirv);
 
@@ -6706,6 +6723,15 @@ void TestNewShaderRecompilerFormattedStoreUsesRuntimeArrayLengthOnly() {
   Check(
       !SpirvSourceHasInstructionUsing(source, "OpULessThan", "%uint_5"),
       "formatted store SPIR-V baked descriptor NumRecords into a store guard");
+
+  user_data[1] = 8u << 16u;
+  user_data[3] = static_cast<uint32_t>(Prospero::BufferFormat::k16_16_16_16Float) << 12u;
+  options.user_data = user_data;
+  result = RecompileForTest(shader, options);
+  Check((DisassembleSpirvBinary(result.spirv).find("PackHalf2x16") !=
+         std::string::npos),
+        "formatted half-float store did not convert F32 to F16");
+  CheckSpirvBinaryValidates(result.spirv);
 }
 
 void TestNewShaderRecompilerTypedBufferTranslation() {
@@ -11482,6 +11508,27 @@ void TestNewShaderRecompilerVertexExportUsesInvocationExecMask() {
 }
 
 void TestNewShaderRecompilerPerInvocationMasksWithoutMirrors() {
+  for (uint32_t wave_size : {32u, 64u}) {
+    for (uint32_t opcode : {0x09u, 0x0au}) {
+      for (uint32_t source : {126u, 8u}) {
+        const uint32_t pixel_shader[] = {
+            EncodeSop1(0x04, 8, 126), // Save the entry live mask.
+            EncodeSop1(opcode, 126, source),
+            EncodeVop1(0x01, 0, 242), // v_mov_b32 v0, 1.0
+            EncodeExp0(0x00, 0x1), EncodeExp1(0, 0, 0, 0),
+            EncodeSopp(0x01),
+        };
+        auto pixel_options = MakeCompileOptions(ShaderType::Pixel);
+        pixel_options.wave_size = wave_size;
+        const auto pixel_result = RecompileForTest(pixel_shader, pixel_options);
+        CheckSpirvBinaryValidates(pixel_result.spirv);
+        Check(DisassembleSpirvBinary(pixel_result.spirv).find("OpGroupNonUniformBallot") ==
+                  std::string::npos,
+              "entry WQM lost the known live predicate through a scalar ballot");
+      }
+    }
+  }
+
   const uint32_t local_shader[] = {
       EncodeVopc(0xc1, 5 + 256, 8),    // v_cmp_lt_u32 vcc, v5, v8
       EncodeSop2(0x0f, 2, 126, 106),   // s_and_b64 s[2:3], exec, vcc
@@ -13470,6 +13517,8 @@ void TestNewShaderRecompilerSpirvSizeBaselines() {
   CheckSpirvPhiParents(dispatcher_result.spirv);
 }
 
+#include "ShaderRayTracingTests.inc"
+
 } // namespace
 } // namespace Libs::Graphics
 
@@ -13477,6 +13526,7 @@ int main() {
   using namespace Libs::Graphics;
 
   EnsureConfigInitialized();
+  TestRayTracingDispatchDetection();
   TestResourceDescriptorClassification();
   TestShaderBufferResourceSize();
   TestNativeShaderResourceDependencies();
@@ -13498,6 +13548,7 @@ int main() {
   TestNewShaderRecompilerNativeWideBufferIr();
   TestNewShaderRecompilerScalarB64LaneTranslation();
   TestNewShaderRecompilerMubufFormatTranslation();
+  TestNewShaderRecompilerFormattedStoreUsesRuntimeArrayLengthOnly();
   TestNewShaderRecompilerTypedBufferTranslation();
   TestNewShaderRecompilerDsReadWrite2Translation();
   TestNewShaderRecompilerDsWideAndAtomicTranslation();