update screenshot

Co-authored-by: Behzad <behzadkhan1736@gmail.com>

README.md

@@ -52,8 +52,8 @@ or graphical glitches, so please include the version you tested when reporting a
       <img src="docs/screenshots/ps5-01.png" width="300" alt="Astro Bot running in KytyPS5">
     </td>
     <td align="center">
-      <strong>Dreaming Sarah</strong><br>
-      <img src="docs/screenshots/ps5-03.png" width="300" alt="Dreaming Sarah running in KytyPS5">
+      <strong>Saros</strong><br>
+      <img src="docs/screenshots/ps5-03.png" width="300" alt="Saros running in KytyPS5">
     </td>
   </tr>
   <tr>

src/graphics/guest_gpu/graphicsRun.cpp

@@ -1003,7 +1003,7 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 		}
 
 		m_renderer.GetRenderExecutor().DispatchDirect(m_submit_id, CurrentBuffer(), thread_group_x,
-		                                              thread_group_y, thread_group_z, mode);
+		                                              thread_group_y, thread_group_z, mode, IsAsyncComputeQueue());
 	}
 }
 
@@ -1015,7 +1015,8 @@ void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {
 		return;
 	}
 	m_sh_ctx.SetCsWaveSize(Pm4::ComputeWaveSize(mode));
-	m_renderer.GetRenderExecutor().DispatchIndirect(m_submit_id, CurrentBuffer(), args_addr, mode);
+	m_renderer.GetRenderExecutor().DispatchIndirect(m_submit_id, CurrentBuffer(), args_addr, mode,
+	                                                IsAsyncComputeQueue());
 }
 
 void CommandProcessor::DrawIndexAuto(DrawAutoArgs args) {

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

@@ -217,17 +217,53 @@ struct PipelineCache::ProgramCache {
 	};
 
 	struct Permutation {
+		std::vector<uint32_t>                       function_code;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		ShaderRecompiler::IR::CompiledShaderInfo     program;
 		ShaderProgram                                handle;
 	};
 
 	struct SourceEntry {
-		explicit SourceEntry(ShaderRecompiler::IR::ResourcePlan plan)
-		    : resource_plan(std::move(plan)) {
+		SourceEntry(ShaderRecompiler::IR::ResourcePlan plan,
+		            std::unique_ptr<ShaderRecompiler::ShaderSource> call_source)
+		    : call_source(std::move(call_source)), resource_plan(std::move(plan)) {
 			permutations.reserve(8);
 		}
 
+		std::span<const uint32_t> FunctionCode() const {
+			return call_source
+			    ? std::span<const uint32_t>(call_source->linked->code).subspan(call_source->code.size())
+			    : std::span<const uint32_t>{};
+		}
+
+		bool RefreshCallSource(const ShaderRecompiler::IR::SrtRuntime& runtime) {
+			const auto revision = call_source->revision;
+			ShaderRecompiler::RefreshShaderSource(*call_source, runtime);
+			resource_plan.source_reads = call_source->reads;
+			return revision != call_source->revision;
+		}
+
+		template <bool HasCalls>
+		Permutation* FindPermutation(const ShaderRecompiler::IR::SrtRuntime& runtime,
+		                             uint32_t push_data_cursor) {
+			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
+			    resource_plan, runtime, resources, specialization));
+			for (auto& candidate: permutations) {
+				const auto& layout = candidate.program.bindings;
+				if (layout.push_data_start_dword != ShaderRecompiler::IR::PushData::StartFor(
+				        push_data_cursor, layout.ShaderDataDwords()) ||
+				    candidate.specialization != specialization) {
+					continue;
+				}
+				if constexpr (HasCalls) {
+					if (!std::ranges::equal(candidate.function_code, FunctionCode())) continue;
+				}
+				return &candidate;
+			}
+			return nullptr;
+		}
+
+		std::unique_ptr<ShaderRecompiler::ShaderSource> call_source;
 		ShaderRecompiler::IR::ResourcePlan           resource_plan;
 		ShaderRecompiler::IR::ResourceSnapshot       resources;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
@@ -256,7 +292,8 @@ struct PipelineCache::ProgramCache {
 	                               const ShaderRecompiler::CompileOptions&      options,
 	                               ShaderRecompiler::TranslateResult            translated,
 	                               ShaderRecompiler::IR::ResourceSpecialization specialization,
-	                               uint32_t push_data_start_dword) {
+	                               uint32_t push_data_start_dword,
+	                               std::span<const uint32_t> function_code) {
 		auto result = ShaderRecompiler::CompileProgram(std::move(translated), options,
 		                                               specialization, push_data_start_dword);
 		if (!ValidateShaderSpirv(options.dump_label, options.shader_hash, result.spirv)) {
@@ -273,6 +310,7 @@ struct PipelineCache::ProgramCache {
 			     static_cast<uint64_t>(result.spirv.size()), options.wave_size);
 		}
 		return {
+		    .function_code  = {function_code.begin(), function_code.end()},
 		    .specialization = std::move(specialization),
 		    .program        = std::move(result.program).TakeCompiledInfo(),
 		    .handle         = {.id = ++next_shader_id, .module = module},
@@ -307,21 +345,18 @@ struct PipelineCache::ProgramCache {
 		if constexpr (std::is_same_v<InputInfo, ShaderComputeInputInfo>) {
 			runtime.workgroup_counts = input_info.workgroup_counts;
 		}
+		bool source_changed = false;
 		if (entry != programs.end()) {
-			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
-			    entry->second.resource_plan, runtime, entry->second.resources,
-			    entry->second.specialization));
-			if (const auto permutation = std::ranges::find_if(
-			        entry->second.permutations, [&](const Permutation& candidate) {
-				        const auto& layout = candidate.program.bindings;
-				        return layout.push_data_start_dword ==
-				                   ShaderRecompiler::IR::PushData::StartFor(
-				                       push_data_cursor, layout.ShaderDataDwords()) &&
-				               candidate.specialization == entry->second.specialization;
-			        });
-			    permutation != entry->second.permutations.end()) {
-				input_info.stage = {.program   = &permutation->program,
-				                    .resources = &entry->second.resources};
+			auto& cached = entry->second;
+			Permutation* permutation = nullptr;
+			if (cached.call_source) {
+				source_changed = cached.RefreshCallSource(runtime);
+				if (!source_changed) permutation = cached.FindPermutation<true>(runtime, push_data_cursor);
+			} else {
+				permutation = cached.FindPermutation<false>(runtime, push_data_cursor);
+			}
+			if (permutation != nullptr) {
+				input_info.stage = {.program = &permutation->program, .resources = &entry->second.resources};
 				permutation->program.bindings.AdvancePushData(push_data_cursor);
 				return permutation->handle;
 			}
@@ -368,16 +403,39 @@ struct PipelineCache::ProgramCache {
 			options.wave_size = input_info.wave_size;
 		}
 		DumpShaderOriginal(stage_name, options.shader_hash, params.code);
-		auto translated = ShaderRecompiler::TranslateProgram(params.code, options);
+		ShaderRecompiler::ShaderSource source;
+		if (entry == programs.end()) source = ShaderRecompiler::PrepareShaderSource(params.code, options);
+		auto* current = entry == programs.end() ? &source : entry->second.call_source.get();
+		ShaderRecompiler::TranslateResult translated;
+		if (current != nullptr) {
+			if (current->call) {
+				if (entry == programs.end()) ShaderRecompiler::RefreshShaderSource(*current, runtime);
+				translated = ShaderRecompiler::TranslateProgram(current->linked->decoded, options);
+				translated.program.source_reads = current->reads;
+			} else {
+				translated = ShaderRecompiler::TranslateProgram(current->decoded, options);
+			}
+		} else {
+			translated = ShaderRecompiler::TranslateProgram(params.code, options);
+		}
 		if (entry == programs.end()) {
+			auto retained = source.call ? std::make_unique<ShaderRecompiler::ShaderSource>(std::move(source)) : nullptr;
 			entry = programs.try_emplace(lookup_key,
-			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program)).first;
+			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program), std::move(retained)).first;
 			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
 			    entry->second.resource_plan, runtime, entry->second.resources,
 			    entry->second.specialization));
+		} else if (source_changed) {
+			entry->second.resource_plan = ShaderRecompiler::IR::ExtractResourcePlan(translated.program);
+			if (const auto* permutation = entry->second.FindPermutation<true>(runtime, push_data_cursor)) {
+				input_info.stage = {.program = &permutation->program, .resources = &entry->second.resources};
+				permutation->program.bindings.AdvancePushData(push_data_cursor);
+				return permutation->handle;
+			}
 		}
 		entry->second.permutations.push_back(CompilePermutation(
-		    stage_name, options, std::move(translated), entry->second.specialization, push_data_cursor));
+		    stage_name, options, std::move(translated), entry->second.specialization, push_data_cursor,
+		    entry->second.FunctionCode()));
 		const auto& permutation = entry->second.permutations.back();
 		input_info.stage = {.program = &permutation.program, .resources = &entry->second.resources};
 		permutation.program.bindings.AdvancePushData(push_data_cursor);

src/graphics/host_gpu/renderer/render.h

@@ -156,9 +156,9 @@ public:
 	KYTY_CLASS_NO_COPY(RenderExecutor);
 
 	void DispatchDirect(uint64_t submit_id, CommandBuffer& buffer, uint32_t thread_group_x,
-	                    uint32_t thread_group_y, uint32_t thread_group_z, uint32_t mode);
+	                    uint32_t thread_group_y, uint32_t thread_group_z, uint32_t mode, bool async_compute);
 	void DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer, uint64_t args_addr,
-	                      uint32_t mode);
+	                      uint32_t mode, bool async_compute);
 
 	void PrepareBindings(const ShaderStageRuntime& runtime, PreparedBindings& prepared);
 	void                           FindBuffers(std::span<PreparedBindings* const> stages);

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -226,7 +226,7 @@ static void BindSharedMemory(RenderContext& context, ShaderComputeInputInfo& inp
 
 void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
                                     uint32_t thread_group_x, uint32_t thread_group_y,
-                                    uint32_t thread_group_z, uint32_t mode) {
+                                    uint32_t thread_group_z, uint32_t mode, bool async_compute) {
 	EXIT_IF(buffer.IsInvalid());
 	m_context.GetCommandScheduler().PopPendingOperations();
 	auto& ctx    = buffer.GetRegisters();
@@ -281,6 +281,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	const auto& sh_regs = ctx.GetShaderRegisters();
 
 	ShaderComputeInputInfo input_info {};
+	input_info.async_compute = async_compute;
 	const bool use_thread_dimensions = (mode & DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS) != 0;
 	input_info.dispatch_thread_dimensions = use_thread_dimensions;
 	input_info.workgroup_counts[0] = thread_group_x;
@@ -437,7 +438,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 }
 
 void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
-                                      uint64_t args_addr, uint32_t mode) {
+                                      uint64_t args_addr, uint32_t mode, bool async_compute) {
 	EXIT_IF(buffer.IsInvalid() || args_addr == 0 || (args_addr & 3u) != 0 ||
 	        (mode & Pm4::COMPUTE_DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS) != 0);
 	m_context.GetCommandScheduler().PopPendingOperations();
@@ -450,6 +451,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 		return;
 	}
 	ShaderComputeInputInfo input_info {};
+	input_info.async_compute = async_compute;
 	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
 	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
 	buffer.EndRendering();

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -7,6 +7,7 @@
 #include "graphics/shader/recompiler/frontend/cfg/ShaderCFG.h"
 #include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 #include "graphics/shader/recompiler/frontend/translate/Translate.h"
+#include "graphics/shader/recompiler/frontend/translate/Translator.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/recompiler/ir/passes/ConstantPropagation.h"
 #include "graphics/shader/recompiler/ir/passes/DeadCodeElimination.h"
@@ -20,6 +21,7 @@
 #include <array>
 #include <chrono>
 #include <fmt/format.h>
+#include <functional>
 #include <map>
 #include <span>
 #include <utility>
@@ -466,6 +468,7 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 	Decoder::DecodeInstruction(joined_code, front_words - 1u, result.instructions.back());
 	Decoder::Program back_program;
 	Decoder::DecodeProgram(back, back_program);
+	result.has_swap_pc |= back_program.has_swap_pc;
 	const auto back_pc = front_words * sizeof(uint32_t);
 	for (auto& inst: back_program.instructions) {
 		// A back-stage PC-relative data reference requires its guest code address.
@@ -478,11 +481,227 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 	return result;
 }
 
+void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
+	using namespace IR;
+	for (uint32_t i = 0; i < source.decoded.instructions.size(); ++i) {
+		if (source.decoded.instructions[i].opcode != Decoder::Opcode::S_SWAPPC_B64) continue;
+		EXIT_NOT_IMPLEMENTED(source.call.has_value() || options.stage != ShaderType::Compute);
+		source.call = ShaderSource::Call {.instruction = i};
+	}
+	EXIT_IF(options.input_info.compute == nullptr);
+	const auto& call = source.decoded.instructions[source.call->instruction];
+	EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
+	                     call.dst.kind != Decoder::OperandKind::Sgpr ||
+	                     call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
+	const auto graph = CFG::BuildGraph(source.decoded);
+	Program query;
+	query.stage = options.stage;
+	query.user_data_base = options.user_data_base;
+	query.user_data_count = static_cast<uint32_t>(options.user_data.size());
+	query.wave_size = options.wave_size;
+	auto& block = *query.block_storage.emplace_back(std::make_unique<Block>());
+	query.blocks.push_back(&block);
+	struct Producer { bool ready = false; std::map<uint32_t, Value> values; };
+	std::map<uint32_t, Producer> producers;
+	std::function<Value(uint32_t, uint32_t)> Register;
+	std::function<Value(uint32_t, Value)> Resolve = [&](uint32_t before, Value value) -> Value {
+		value = value.Resolve();
+		auto* inst = value.TryInstruction();
+		if (inst == nullptr) return value;
+		uint32_t code = UINT32_MAX;
+		switch (inst->GetOpcode()) {
+			case ValueOpcode::GetScalarRegister: code = static_cast<uint32_t>(inst->Arg(0).ScalarRegister()); break;
+			case ValueOpcode::GetVccLo: code = 106u; break;
+			case ValueOpcode::GetVccHi: code = 107u; break;
+			default: break;
+		}
+		if (code != UINT32_MAX) {
+			const auto result = Register(before, code);
+			inst->ReplaceUsesWith(result);
+			return result;
+		}
+		for (size_t i = 0; i < inst->NumArgs(); ++i) inst->SetArg(i, Resolve(before, inst->Arg(i)));
+		return value;
+	};
+	Register = [&](uint32_t before, uint32_t code) -> Value {
+		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code);
+		if (index == UINT32_MAX) {
+			EXIT_IF(code < query.user_data_base || code - query.user_data_base >= query.user_data_count);
+			return IREmitter(&block).GetUserData(static_cast<ScalarReg>(code));
+		}
+		auto [entry, inserted] = producers.try_emplace(index);
+		auto& producer = entry->second;
+		if (inserted) {
+			Block translated;
+			Frontend::Translator translator(query, &translated, 1u,
+			    (options.input_info.compute->float_mode & 0x10u) == 0u,
+			    !options.input_info.compute->async_compute);
+			translator.TranslateInstruction(source.decoded.instructions[index]);
+			for (const auto& inst: translated) {
+				uint32_t destination = UINT32_MAX;
+				Value value;
+				switch (inst.GetOpcode()) {
+					case ValueOpcode::SetScalarRegister:
+						destination = static_cast<uint32_t>(inst.Arg(0).ScalarRegister()); value = inst.Arg(1); break;
+					case ValueOpcode::SetVccLo: destination = 106u; value = inst.Arg(0); break;
+					case ValueOpcode::SetVccHi: destination = 107u; value = inst.Arg(0); break;
+					default: break;
+				}
+				if (destination != UINT32_MAX) producer.values.insert_or_assign(destination, value);
+			}
+			for (auto& [destination, value]: producer.values) value = Resolve(index, value);
+			for (auto& inst: translated) inst.SetParent(&block);
+			block.Instructions().splice(block.end(), translated.Instructions());
+			producer.ready = true;
+		}
+		EXIT_IF(!producer.ready || !producer.values.contains(code));
+		return producer.values.at(code).Resolve();
+	};
+	for (uint32_t i = 0; i < 2u; ++i) {
+		const auto value = Register(source.call->instruction, call.src0.reg + i);
+		EXIT_IF(!ValidateRuntimeValue(query, value, RuntimeValueType::Integer));
+		auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
+		root.SetArg(0, value);
+		source.call->target[i] = Value(&root);
+	}
+	RewriteToSsa(query.blocks);
+	ConstantPropagationPass(query.blocks, query.wave_size);
+	RemoveIdentities(query.blocks);
+	EliminateDeadCode(query.blocks);
+	for (auto& value: source.call->target) value = value.Instruction()->Arg(0).Resolve();
+	query.value_storage.clear();
+	for (auto& inst: block) {
+		inst.SetParent(nullptr);
+		if (inst.GetOpcode() == ValueOpcode::LoadAddressU32 || inst.GetOpcode() == ValueOpcode::ReadConstBuffer)
+			inst.SetFlags(SrtReadFlags {.index = inst.Flags<MemoryFlags>().index});
+	}
+	query.value_storage.splice(query.value_storage.end(), block.Instructions());
+	source.call_targets = std::move(static_cast<ResourcePlan&>(query));
+}
+
 } // namespace
 
+ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOptions& options) {
+	EXIT_IF(code.empty());
+	ShaderSource source;
+	if (!options.back_code.empty()) {
+		source.decoded = DecodeFusedProgram(code, options.back_code, source.code);
+	} else if (options.stage == ShaderType::Local) {
+		source.decoded = Decoder::DecodeFrontProgram(code);
+		auto& handoff = source.decoded.instructions.back();
+		handoff.opcode = Decoder::Opcode::S_ENDPGM;
+		handoff.src_count = 0;
+	} else {
+		Decoder::DecodeProgram(code, source.decoded);
+	}
+	if (source.decoded.has_swap_pc) {
+		PrepareCallTarget(source, options);
+		source.code.assign(code.begin(), code.end());
+		source.decoded.code = source.code;
+	}
+	return source;
+}
+
+const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime) {
+	auto& reads = source.reads;
+	reads.clear();
+	IR::SrtReadCapture capture(runtime, reads);
+	const auto clean = IR::CleanRuntime(capture.ObservedRuntime());
+	IR::SrtWalker walker(source.call_targets, clean);
+	uint32_t low = 0, high = 0;
+	const auto& call = source.decoded.instructions[source.call->instruction];
+	if (!walker.Evaluate(source.call->target[0], low) || !walker.Evaluate(source.call->target[1], high))
+		EXIT("shader call at pc 0x%08x has an unavailable scalar target", call.pc);
+	// SWAPPC reads the old aliased pair and ignores the target's low two bits.
+	const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
+	EXIT_IF(address == 0u);
+	const auto query_reads = reads.size();
+	if (source.linked) {
+		auto& linked = *source.linked;
+		const auto function = std::span(linked.code).subspan(source.code.size());
+		if (clean.read_memory(clean.userdata, address, linked.observed_function) &&
+		    std::ranges::equal(function, linked.observed_function)) {
+			return linked.decoded;
+		}
+		reads.resize(query_reads);
+	}
+	ShaderSource::Linked linked {.code = source.code,
+	                             .decoded = source.decoded};
+	const auto& last = source.decoded.instructions.back();
+	EXIT_NOT_IMPLEMENTED(last.opcode != Decoder::Opcode::S_ENDPGM);
+	// Keep the caller's data/footer bytes at their native offsets. The leaf's virtual CFG
+	// PCs follow decoded instructions; its physical code is independently read and owned.
+	const auto function_pc = last.pc + last.word_count * sizeof(uint32_t);
+	for (uint32_t index = 0;;) {
+		// The observed queue helper is a scalar leaf. Scalar encodings need at most one literal.
+		EXIT_IF(index >= 64u || address > UINT64_MAX - (uint64_t {index} + 2u) * sizeof(uint32_t));
+		std::array<uint32_t, 2> words {};
+		if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
+		                       std::span(words).first(1)))
+			EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
+		const auto family = Decoder::GetInstructionFamily(words[0]);
+		EXIT_NOT_IMPLEMENTED(family != Decoder::Family::SOP1 && family != Decoder::Family::SOP2 &&
+		                     family != Decoder::Family::SOPC && family != Decoder::Family::SOPK &&
+		                     family != Decoder::Family::SOPP);
+		Decoder::Instruction inst;
+		Decoder::DecodeInstruction(words, 0u, inst);
+		if (inst.word_count == 2u) {
+			if (!clean.read_memory(clean.userdata, address + uint64_t {index + 1u} * sizeof(uint32_t),
+			                       std::span(words).last(1)))
+				EXIT("shader call at pc 0x%08x cannot read leaf literal", call.pc);
+			inst = {};
+			Decoder::DecodeInstruction(words, 0u, inst);
+		}
+		inst.pc = index * sizeof(uint32_t);
+		if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
+		linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
+		index += inst.word_count;
+		EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
+		                     inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
+		                     inst.opcode == Decoder::Opcode::S_ENDPGM);
+		EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) || CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
+		const bool returns = inst.opcode == Decoder::Opcode::S_SETPC_B64;
+		if (returns) {
+			EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
+			inst.opcode = Decoder::Opcode::S_BRANCH;
+			inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
+		} else if (Decoder::IsDirectBranch(inst.opcode)) {
+			EXIT_IF(inst.branch_target > UINT32_MAX - function_pc);
+			inst.branch_target += function_pc;
+		}
+		inst.pc += function_pc;
+		linked.decoded.instructions.push_back(std::move(inst));
+		if (returns) break;
+	}
+	const auto function_end = linked.decoded.instructions.back().pc +
+	                          linked.decoded.instructions.back().word_count * sizeof(uint32_t);
+	for (size_t i = source.decoded.instructions.size(); i + 1u < linked.decoded.instructions.size(); ++i) {
+		const auto& inst = linked.decoded.instructions[i];
+		if (Decoder::IsDirectBranch(inst.opcode))
+			EXIT_IF(inst.branch_target < function_pc || inst.branch_target >= function_end);
+	}
+	reads.resize(query_reads);
+	reads.emplace_back(address, (linked.code.size() - source.code.size()) * sizeof(uint32_t));
+	linked.decoded.instructions[source.call->instruction].branch_target = function_pc;
+	linked.decoded.code = linked.code;
+	linked.observed_function.resize(linked.code.size() - source.code.size());
+	source.linked = std::move(linked);
+	++source.revision;
+	return source.linked->decoded;
+}
+
 TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOptions& options) {
-	if (code.empty()) {
-		EXIT("shader recompiler input is empty\n");
+	auto source = PrepareShaderSource(code, options);
+	return TranslateProgram(source.decoded, options);
+}
+
+TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileOptions& options) {
+	EXIT_IF(decoded.instructions.empty());
+	if (decoded.has_swap_pc) {
+		for (const auto& inst: decoded.instructions) {
+			if (inst.opcode == Decoder::Opcode::S_SWAPPC_B64 && inst.branch_target == UINT32_MAX)
+				EXIT("shader call at pc 0x%08x must be linked before translation", inst.pc);
+		}
 	}
 	if (options.stage != ShaderType::Compute && options.stage != ShaderType::Vertex &&
 	    options.stage != ShaderType::Pixel && options.stage != ShaderType::Mesh &&
@@ -493,34 +712,11 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	}
 
 	const auto compile_begin = std::chrono::steady_clock::now();
-	const auto phase_ms      = [&compile_begin]() {
+	const auto phase_ms = [&compile_begin]() {
 		return static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::milliseconds>(
-		                                 std::chrono::steady_clock::now() - compile_begin)
-		                                 .count());
+		    std::chrono::steady_clock::now() - compile_begin).count());
 	};
 
-	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " code_words=%" PRIu64 " decode\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
-	     static_cast<uint64_t>(code.size()));
-
-	Decoder::Program decoded;
-	std::vector<uint32_t> joined_code;
-	if (!options.back_code.empty()) {
-		decoded = DecodeFusedProgram(code, options.back_code, joined_code);
-	} else if (options.stage == ShaderType::Local) {
-		decoded = Decoder::DecodeFrontProgram(code);
-		// The separately compiled hull half runs in the next Vulkan stage.
-		auto& handoff     = decoded.instructions.back();
-		handoff.opcode    = Decoder::Opcode::S_ENDPGM;
-		handoff.src_count = 0;
-	} else {
-		Decoder::DecodeProgram(code, decoded);
-	}
-	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " decode instructions=%" PRIu64
-	     " elapsed_ms=%" PRIu64 "\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
-	     static_cast<uint64_t>(decoded.instructions.size()), phase_ms());
-
 	std::string decoded_dump;
 	if (options.dump_ir) {
 		decoded_dump = Decoder::ProgramToString(decoded);
@@ -606,6 +802,12 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 		IR::RemoveIdentities(ir.blocks);
 		IR::EliminateDeadCode(ir.blocks);
 	}
+	if (decoded.has_swap_pc) {
+		// Leaf returns are resolved CFG edges; native PC values must not escape into GPU data.
+		for (const auto* block: ir.blocks)
+			for (const auto& inst: *block)
+				EXIT_NOT_IMPLEMENTED(inst.GetOpcode() == IR::ValueOpcode::GetShaderBase);
+	}
 	LowerTessellationMemory(ir, options);
 	std::string cfg_dump;
 	if (options.dump_ir) {
@@ -656,7 +858,11 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 				const auto resource = inst.Flags<uint32_t>();
 				first = resource < ir.info.images.size() &&
 				                ir.info.images[resource].indirect_root == resource ? 1u : 0u;
-			} else if (op != IR::ValueOpcode::GetSamplerResource) {
+			} else if (op == IR::ValueOpcode::GetSamplerResource) {
+				const auto resource = inst.Flags<uint32_t>();
+				first = resource < ir.info.samplers.size() &&
+				                !ir.info.samplers[resource].indirect_resources.empty() ? 1u : 0u;
+			} else {
 				continue;
 			}
 			for (size_t index = first; index < inst.NumArgs(); index++) {

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -4,8 +4,11 @@
 #include "common/common.h"
 #include "common/stringUtils.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 #include "graphics/shader/shader.h"
 
+#include <array>
+#include <optional>
 #include <span>
 #include <vector>
 
@@ -37,6 +40,32 @@ struct CompileResult {
 	IR::Program            program;
 };
 
+// Decoded source and typed scalar call queries are retained once per cache source.
+struct ShaderSource {
+	struct Call {
+		uint32_t instruction;
+		std::array<IR::Value, 2> target;
+	};
+	struct Linked {
+		std::vector<uint32_t> code;
+		Decoder::Program decoded;
+		std::vector<uint32_t> observed_function;
+	};
+	std::vector<uint32_t> code;
+	Decoder::Program decoded;
+	IR::ResourcePlan call_targets;
+	std::optional<Call> call;
+	std::optional<Linked> linked;
+	uint64_t revision = 0;
+	std::vector<std::pair<uint64_t, uint64_t>> reads;
+};
+
+[[nodiscard]] ShaderSource PrepareShaderSource(std::span<const uint32_t> code,
+	                                           const CompileOptions& options);
+const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime);
+[[nodiscard]] TranslateResult TranslateProgram(const Decoder::Program& decoded,
+	                                           const CompileOptions& options);
+
 [[nodiscard]] TranslateResult TranslateProgram(std::span<const uint32_t> code,
                                                const CompileOptions& options);
 [[nodiscard]] CompileResult CompileProgram(TranslateResult translated,

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

@@ -137,6 +137,15 @@ uint32_t ImageViewSizeType(EmitterState& state, ImageDimension dimension) {
 	}
 }
 
+namespace {
+void DecorateNonUniformDescriptor(EmitterState& state, uint32_t id) {
+	state.builder.RequireExtension("SPV_EXT_descriptor_indexing");
+	state.builder.RequireCapability(spv::CapabilityShaderNonUniform);
+	state.builder.RequireCapability(spv::CapabilitySampledImageArrayNonUniformIndexing);
+	state.builder.AddAnnotation(spv::OpDecorate, id, spv::DecorationNonUniform);
+}
+} // namespace
+
 uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip,
                              uint32_t array_index) {
 	const auto pointer = ImageDescriptorPointer(state, resource, mip, array_index);
@@ -144,24 +153,29 @@ uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mi
 	const auto kind = *IR::DescriptorBindingForImage(state.program.info.images[resource]);
 	state.builder.AddFunction(spv::OpLoad, state.images[IR::ImageBindingIndex(kind)].type,
 	                          image, pointer);
+	if (array_index != 0u) {
+		DecorateNonUniformDescriptor(state, image);
+	}
 	return image;
 }
 
-uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler) {
+uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler, uint32_t array_index) {
 	if (sampler >= state.program.info.samplers.size()) {
 		ExitDescriptorBindingFailure(state, IR::DescriptorBindingKind::Samplers, sampler,
 		                             "sampler resource index is out of range");
 	}
 	const auto pointer = DescriptorElementPointer(
-	    state, state.sampler_pointer_type, state.sampler_variable, ConstantU32(state, sampler),
+	    state, state.sampler_pointer_type, state.sampler_variable,
+	    array_index == 0u ? ConstantU32(state, sampler) : array_index,
 	    IR::DescriptorBindingKind::Samplers, sampler, "sampler descriptor array was not emitted");
 	const auto sampler_id = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpLoad, state.sampler_type, sampler_id, pointer);
+	if (array_index != 0u) DecorateNonUniformDescriptor(state, sampler_id);
 	return sampler_id;
 }
 
 uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id, uint32_t mip,
-                          uint32_t array_index) {
+                          uint32_t array_index, bool sampler_dynamic) {
 	const auto& image_resource = state.program.info.images.at(resource);
 	EXIT_IF(image_resource.resource_class != IR::ImageResourceClass::Sampled);
 	const auto  image          = LoadImageDescriptor(state, resource, mip, array_index);
@@ -173,11 +187,8 @@ uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampl
 		    spv::OpTypeSampledImage, state.images[IR::ImageBindingIndex(kind)].type);
 	}
 	state.builder.AddFunction(spv::OpSampledImage, sampled_type, sampled_image, image, sampler_id);
-	if (array_index != 0u) {
-		state.builder.RequireExtension("SPV_EXT_descriptor_indexing");
-		state.builder.RequireCapability(spv::CapabilityShaderNonUniform);
-		state.builder.RequireCapability(spv::CapabilitySampledImageArrayNonUniformIndexing);
-		state.builder.AddAnnotation(spv::OpDecorate, sampled_image, spv::DecorationNonUniform);
+	if (array_index != 0u || sampler_dynamic) {
+		DecorateNonUniformDescriptor(state, sampled_image);
 	}
 	return sampled_image;
 }

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

@@ -615,6 +615,90 @@ uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t d
 	return PackImageTexel(ctx, mem, texel);
 }
 
+template <typename Emit>
+uint32_t EmitImageAccess(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t result_type,
+                         const Emit& emit) {
+	auto& state = ctx.state;
+	const auto& mem = ctx.Memory(inst);
+	const auto& image = state.program.info.images.at(mem.resource);
+	if (image.indirect_root != mem.resource) return emit(mem.resource, 0u);
+	const auto* handle = inst.Arg(0).ResolveInstruction();
+	const auto* source = image.source < state.program.descriptor_sources.size()
+	                         ? &state.program.descriptor_sources[image.source]
+	                         : nullptr;
+	if (handle == nullptr || source == nullptr || !source->indirect_descriptor.has_value() ||
+	    handle->NumArgs() == 0u) {
+		ctx.Fail(inst, "has invalid indirect image key provenance");
+	}
+	const auto key = ctx.Def(handle->Arg(0));
+	if (state.flattened_srt_variable == 0 ||
+	    image.indirect_resources.size() < 2u) {
+		ctx.Fail(inst, "has no indirect image runtime mapping");
+	}
+	const auto selected = EmitIndirectResourceIndex(
+	    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
+	struct ImageRun {
+		uint32_t first;
+		uint32_t count;
+		uint32_t resource;
+		uint32_t slot_bias;
+	};
+	std::vector<ImageRun> runs;
+	for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
+		const auto resource = image.indirect_resources[ordinal];
+		const auto& candidate = state.program.info.images[resource];
+		const auto kind = *IR::DescriptorBindingForImage(candidate);
+		if (!runs.empty()) {
+			auto& run = runs.back();
+			const auto& first = state.program.info.images[run.resource];
+			const auto next_slot = run.slot_bias + ordinal;
+			if (candidate.dimension == first.dimension && candidate.cube == first.cube &&
+			    IR::DescriptorBindingForImage(first) == kind &&
+			    candidate.mip_count == 1u && first.mip_count == 1u &&
+			    (ordinal == 1u || candidate.descriptor_index == next_slot)) {
+				// Native roots precede appended children; only the root may have a slot gap.
+				if (ordinal == 1u)
+					run.slot_bias = candidate.descriptor_index - ordinal;
+				++run.count;
+				continue;
+			}
+		}
+		runs.push_back({ordinal, 1u, resource,
+		                candidate.descriptor_index - ordinal});
+	}
+	const auto EmitRun = [&](uint32_t index) {
+		const auto& run = runs[index];
+		if (run.count == 1u) return emit(run.resource, 0u);
+		auto slot = Binary(state, spv::OpIAdd, TypeU32(state), selected,
+		                   ConstantU32(state, run.slot_bias));
+		if (run.first == 0u) {
+			const auto root_slot = image.descriptor_index;
+			if (root_slot != run.slot_bias) {
+				const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
+				                            ConstantU32(state, 0u));
+				const auto root_or_child = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpSelect, TypeU32(state), root_or_child,
+				                          is_root, ConstantU32(state, root_slot), slot);
+				slot = root_or_child;
+			}
+		}
+		return emit(run.resource, slot);
+	};
+	auto run_index = ConstantU32(state, 0u);
+	for (uint32_t index = 1; index < runs.size(); ++index) {
+		const auto at_or_after = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state),
+		                               selected, ConstantU32(state, runs[index].first));
+		const auto next = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpSelect, TypeU32(state), next, at_or_after,
+		                          ConstantU32(state, index), run_index);
+		run_index = next;
+	}
+	return runs.size() == 1u
+	           ? EmitRun(0u)
+	           : EmitIndexSwitch(state, run_index, static_cast<uint32_t>(runs.size()),
+	                             result_type, EmitRun);
+}
+
 } // namespace
 
 void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
@@ -656,29 +740,35 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		return;
 	}
 	if (op == IR::ValueOpcode::ImageRead) {
-		const auto  dimension      = image.dimension;
-		const auto& dimension_info = ImageDimensionInfoFor(dimension);
-		const auto  numeric_class  = image.numeric_class;
-		const auto  condition      = ctx.Arg(inst, 2);
+		const auto numeric_class = image.numeric_class;
+		const auto condition = ctx.Arg(inst, 2);
 		ctx.Define(
 		    inst,
 		    EmitValueOrDefaultIfCondition(
 		        state, condition, TypeU32Vector(state, 4), ConstantU32CompositeZero(state, 4),
 		        [&]() {
-			        const auto descriptor = LoadImageDescriptor(state, mem.resource);
-			        const auto color      = state.builder.AllocateId();
-			        const auto coord      = CoordU32(ctx, mem, *address, dimension);
-			        if (dimension_info.multisampled != 0u) {
-				        state.builder.AddFunction(
-				            spv::OpImageFetch, ImageVectorType(state, numeric_class, 4), color,
-				            descriptor, coord, spv::ImageOperandsSampleMask,
-				            AddressU32(ctx, mem, *address, dimension_info.coordinate_components));
-			        } else {
-				        state.builder.AddFunction(spv::OpImageFetch,
-				                                  ImageVectorType(state, numeric_class, 4), color,
-				                                  descriptor, coord, spv::ImageOperandsLodMask,
-				                                  LodU32(ctx, mem, *address, dimension));
-			        }
+			        const auto color = EmitImageAccess(
+			            ctx, inst, ImageVectorType(state, numeric_class, 4),
+			            [&](uint32_t resource, uint32_t array_index) {
+				        const auto& candidate = state.program.info.images[resource];
+				        const auto dimension = candidate.dimension;
+				        const auto& dimension_info = ImageDimensionInfoFor(dimension);
+				        const auto descriptor = LoadImageDescriptor(state, resource, 0u, array_index);
+				        const auto fetched = state.builder.AllocateId();
+				        const auto coord = CoordU32(ctx, mem, *address, dimension);
+				        if (dimension_info.multisampled != 0u) {
+					        state.builder.AddFunction(
+					            spv::OpImageFetch, ImageVectorType(state, candidate.numeric_class, 4), fetched,
+					            descriptor, coord, spv::ImageOperandsSampleMask,
+					            AddressU32(ctx, mem, *address, dimension_info.coordinate_components));
+				        } else {
+					        state.builder.AddFunction(
+					            spv::OpImageFetch, ImageVectorType(state, candidate.numeric_class, 4), fetched,
+					            descriptor, coord, spv::ImageOperandsLodMask,
+					            LodU32(ctx, mem, *address, mem.image_dimension));
+				        }
+				        return fetched;
+			            });
 			        return ResultVector(ctx, UnpackImageTexel(ctx, mem, color), numeric_class,
 			                            false, mem);
 		        }));
@@ -808,106 +898,32 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			operands[1] = AddressF32(ctx, mem, *address, layout.bias);
 			operand_count = 2;
 		}
-		const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
-		const auto EmitSample = [&](uint32_t resource, uint32_t array_index = 0u) {
+		const auto& samplers = state.program.info.samplers.at(mem.sampler).indirect_resources;
+		uint32_t sampler_index = 0;
+		if (!samplers.empty()) {
+			const auto* handle = inst.Arg(1).ResolveInstruction();
+			if (handle == nullptr || handle->GetOpcode() != IR::ValueOpcode::GetSamplerResource)
+				ctx.Fail(inst, "has invalid finite sampler key provenance");
+			sampler_index = EmitIndexSwitch(
+			    state, ctx.Def(handle->Arg(0)), static_cast<uint32_t>(samplers.size()),
+			    TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, samplers[ordinal]); });
+		}
+		const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler, sampler_index);
+		const auto EmitSample = [&](uint32_t resource, uint32_t array_index) {
 			const auto& candidate = state.program.info.images[resource];
 			const auto coord =
 			    CoordF32(ctx, mem, *address, layout.coord,
 			             ImageDimensionInfoFor(candidate.dimension).coordinate_components,
 			             candidate.cube);
-			const auto sampled = MakeSampledImage(state, resource, sampler_id, 0u, array_index);
+			const auto sampled = MakeSampledImage(state, resource, sampler_id, 0u, array_index,
+			                                      sampler_index != 0u);
 			const auto sample = state.builder.AllocateId();
 			state.builder.AddFunction(opcode, result_type, sample, sampled, coord,
 			                          std::span(&dref_value, dref ? 1u : 0u),
 			                          std::span<const uint32_t>(operands).first(operand_count));
 			return sample;
 		};
-		if (image.indirect_root != mem.resource) {
-			const auto sample = EmitSample(mem.resource);
-			auto       result = sample;
-			if (!dref) {
-				result = UnpackImageTexel(ctx, mem, sample);
-			}
-			ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
-			return;
-		}
-		const auto* handle = image_arg.ResolveInstruction();
-		const auto* source = image.source < state.program.descriptor_sources.size()
-		                         ? &state.program.descriptor_sources[image.source]
-		                         : nullptr;
-		if (handle == nullptr || source == nullptr || !source->indirect_descriptor.has_value() ||
-		    handle->NumArgs() == 0u) {
-			ctx.Fail(inst, "has invalid indirect image key provenance");
-			return;
-		}
-		const auto key = ctx.Def(handle->Arg(0));
-		if (state.flattened_srt_variable == 0 ||
-		    image.indirect_resources.size() < 2u) {
-			ctx.Fail(inst, "has no indirect image runtime mapping");
-			return;
-		}
-		const auto selected = EmitIndirectResourceIndex(
-		    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
-		struct SampleRun {
-			uint32_t first;
-			uint32_t count;
-			uint32_t resource;
-			uint32_t slot_bias;
-		};
-		std::vector<SampleRun> runs;
-		for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
-			const auto resource = image.indirect_resources[ordinal];
-			const auto& candidate = state.program.info.images[resource];
-			const auto kind = *IR::DescriptorBindingForImage(candidate);
-			if (!runs.empty()) {
-				auto& run = runs.back();
-				const auto& first = state.program.info.images[run.resource];
-				const auto next_slot = run.slot_bias + ordinal;
-				if (candidate.dimension == first.dimension && candidate.cube == first.cube &&
-				    IR::DescriptorBindingForImage(first) == kind &&
-				    candidate.mip_count == 1u && first.mip_count == 1u &&
-				    (ordinal == 1u || candidate.descriptor_index == next_slot)) {
-					// Native roots precede appended children; only the root may have a slot gap.
-					if (ordinal == 1u)
-						run.slot_bias = candidate.descriptor_index - ordinal;
-					++run.count;
-					continue;
-				}
-			}
-			runs.push_back({ordinal, 1u, resource,
-			                candidate.descriptor_index - ordinal});
-		}
-		const auto EmitRun = [&](uint32_t index) {
-			const auto& run = runs[index];
-			if (run.count == 1u) return EmitSample(run.resource);
-			auto slot = Binary(state, spv::OpIAdd, TypeU32(state), selected,
-			                   ConstantU32(state, run.slot_bias));
-			if (run.first == 0u) {
-				const auto root_slot = image.descriptor_index;
-				if (root_slot != run.slot_bias) {
-					const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
-					                            ConstantU32(state, 0u));
-					const auto root_or_child = state.builder.AllocateId();
-					state.builder.AddFunction(spv::OpSelect, TypeU32(state), root_or_child,
-					                          is_root, ConstantU32(state, root_slot), slot);
-					slot = root_or_child;
-				}
-			}
-			return EmitSample(run.resource, slot);
-		};
-		auto run_index = ConstantU32(state, 0u);
-		for (uint32_t index = 1; index < runs.size(); ++index) {
-			const auto at_or_after = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state),
-			                               selected, ConstantU32(state, runs[index].first));
-			const auto next = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), next, at_or_after,
-			                          ConstantU32(state, index), run_index);
-			run_index = next;
-		}
-		auto result = runs.size() == 1u
-		                  ? EmitRun(0u)
-		                  : EmitIndexSwitch(state, run_index, static_cast<uint32_t>(runs.size()),
-		                                    result_type, EmitRun);
+		auto result = EmitImageAccess(ctx, inst, result_type, EmitSample);
 		if (!dref) {
 			result = UnpackImageTexel(ctx, mem, result);
 		}

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

@@ -322,10 +322,11 @@ uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t
 uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip = 0,
                              uint32_t array_index = 0);
 
-uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler);
+uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler, uint32_t array_index = 0);
 
 uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id,
-                          uint32_t mip = 0, uint32_t array_index = 0);
+                          uint32_t mip = 0, uint32_t array_index = 0,
+                          bool sampler_dynamic = false);
 
 void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_lod, uint32_t coord,
                            uint32_t texel);

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

@@ -1468,7 +1468,8 @@ Graph BuildGraph(const Decoder::Program& program) {
 	for (uint32_t i = 0; i < program.instructions.size(); i++) {
 		const auto& inst    = program.instructions[i];
 		const auto  next_pc = InstructionEndPc(inst);
-		if (Decoder::IsDirectBranch(inst.opcode)) {
+		if (Decoder::IsDirectBranch(inst.opcode) ||
+		    (inst.opcode == Opcode::S_SWAPPC_B64 && inst.branch_target != UINT32_MAX)) {
 			if (!IsValidTarget(inst.branch_target, program.instructions, end_pc)) {
 				ExitBuildFailure(graph, FailureKind::InvalidBranchTarget, UINT32_MAX,
 				                 fmt::format("branch at pc 0x{:08x} targets invalid pc 0x{:08x}",
@@ -1501,7 +1502,7 @@ Graph BuildGraph(const Decoder::Program& program) {
 			if (next_pc <= end_pc) {
 				labels.push_back(next_pc);
 			}
-		} else if (inst.opcode == Opcode::S_ENDPGM) {
+		} else if (inst.opcode == Opcode::S_ENDPGM || inst.opcode == Opcode::S_SWAPPC_B64) {
 			labels.push_back(next_pc);
 		}
 	}
@@ -1550,6 +1551,9 @@ Graph BuildGraph(const Decoder::Program& program) {
 		const auto  next_pc = InstructionEndPc(last);
 		if (last.opcode == Opcode::S_ENDPGM) {
 			block.terminator.kind = TerminatorKind::Return;
+		} else if (last.opcode == Opcode::S_SWAPPC_B64 && last.branch_target != UINT32_MAX) {
+			block.terminator.kind = TerminatorKind::Branch;
+			block.terminator.true_block = block_at_pc(last.branch_target);
 		} else if (last.opcode == Opcode::S_SETPC_B64) {
 			const auto& target_info = setpc_targets.at(last.pc);
 			if (target_info.indirect) {
@@ -1676,6 +1680,36 @@ Graph Structurize(const Graph& graph) {
 	return GotoStructurizer(graph).Run();
 }
 
+bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code) {
+	uint32_t destination = UINT32_MAX;
+	// Scalar ALU pair widths are not all decoded. The typed producer must confirm this word.
+	return InstructionWritesScalarCode(inst, code) ||
+	       (ScalarOperandCode(inst.dst, destination) && code == destination + 1u);
+}
+
+uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
+                              uint32_t before, uint32_t code) {
+	EXIT_IF(before >= program.instructions.size());
+	const auto found = std::ranges::find_if(graph.blocks, [before](const auto& block) {
+		return block.inst_begin <= before && before < block.inst_end;
+	});
+	EXIT_IF(found == graph.blocks.end());
+	const auto* block = &*found;
+	for (size_t depth = 0; depth < graph.blocks.size(); ++depth) {
+		for (uint32_t i = before; i > block->inst_begin;) {
+			const auto& inst = program.instructions[--i];
+			EXIT_IF(inst.opcode == Opcode::S_SWAPPC_B64);
+			if (MayWriteScalarRegister(inst, code)) return i;
+		}
+		if (block->id == graph.entry_block) return UINT32_MAX;
+		EXIT_IF(block->predecessors.size() != 1u);
+		block = graph.FindBlock(block->predecessors[0]);
+		EXIT_IF(block == nullptr);
+		before = block->inst_end;
+	}
+	EXIT("scalar shader call source has cyclic reaching definitions");
+}
+
 std::string BranchConditionToString(BranchCondition condition) {
 	switch (condition) {
 		case BranchCondition::Always: return "always";

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

@@ -113,6 +113,9 @@ struct Graph {
 };
 
 Graph       BuildGraph(const Decoder::Program& program);
+bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code);
+uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
+                              uint32_t before, uint32_t code);
 // Returns structured control flow or failure diagnostics without changing the native graph.
 // On failure, failure_block is an original block ID or UINT32_MAX.
 Graph       Structurize(const Graph& graph);

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

@@ -61,6 +61,7 @@ constexpr OpcodeMap SOP1_OPCODE_LIST[] = {
     {0x1eu, Opcode::S_BITSET1_B64},
     {0x1fu, Opcode::S_GETPC_B64},
     {0x20u, Opcode::S_SETPC_B64},
+    {0x21u, Opcode::S_SWAPPC_B64},
     {0x24u, Opcode::S_AND_SAVEEXEC_B64},
     {0x28u, Opcode::S_ORN2_SAVEEXEC_B64},
     {0x2du, Opcode::S_QUADMASK_B64},
@@ -88,7 +89,7 @@ constexpr OpcodeMap SOPK_OPCODE_LIST[] = {
     {0x08u, Opcode::S_CMP_LE_I32}, {0x09u, Opcode::S_CMP_EQ_U32}, {0x0au, Opcode::S_CMP_LG_U32},
     {0x0bu, Opcode::S_CMP_GT_U32}, {0x0cu, Opcode::S_CMP_GE_U32}, {0x0du, Opcode::S_CMP_LT_U32},
     {0x0eu, Opcode::S_CMP_LE_U32}, {0x0fu, Opcode::S_ADD_I32},    {0x10u, Opcode::S_MULK_I32},
-    {0x13u, Opcode::S_SETREG_B32}, {0x17u, Opcode::S_WAITCNT_VSCNT}, {0x18u, Opcode::S_WAITCNT},
+    {0x12u, Opcode::S_GETREG_B32}, {0x13u, Opcode::S_SETREG_B32}, {0x17u, Opcode::S_WAITCNT_VSCNT}, {0x18u, Opcode::S_WAITCNT},
     {0x19u, Opcode::S_WAITCNT},    {0x1au, Opcode::S_WAITCNT},
     {0x1bu, Opcode::S_SUBVECTOR_LOOP_BEGIN}, {0x1cu, Opcode::S_SUBVECTOR_LOOP_END},
 };
@@ -155,6 +156,10 @@ void DecodeSop1(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	}
 
 	switch (inst.opcode) {
+		case Opcode::S_SWAPPC_B64:
+			inst.branch_target = UINT32_MAX;
+			inst.data_dwords = 2;
+			break;
 		case Opcode::S_GETPC_B64:
 			inst.src_count = 0;
 			DecodeScalarDestination(sdst, pc, inst.dst);
@@ -202,7 +207,7 @@ void DecodeSopk(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	const uint32_t word   = code[word_index];
 	const uint32_t opcode = (word >> 23u) & 0x1fu;
 	const uint32_t sdst   = (word >> 16u) & 0x7fu;
-	const auto     imm    = opcode >= 0x09u && opcode <= 0x0eu
+	const auto     imm    = (opcode >= 0x09u && opcode <= 0x0eu) || opcode == 0x12u
 	                           ? static_cast<int32_t>(word & 0xffffu)
 	                           : static_cast<int32_t>(static_cast<int16_t>(word & 0xffffu));
 
@@ -222,6 +227,7 @@ void DecodeSopk(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	}
 
 	switch (inst.opcode) {
+		case Opcode::S_GETREG_B32:
 		case Opcode::S_MOVK_I32: DecodeScalarDestination(sdst, pc, inst.dst); return;
 		case Opcode::S_SUBVECTOR_LOOP_BEGIN:
 		case Opcode::S_SUBVECTOR_LOOP_END:

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

@@ -389,6 +389,7 @@ Program DecodeFrontProgram(std::span<const uint32_t> front) {
 	while (front_words < front.size()) {
 		auto& inst = result.instructions.emplace_back();
 		DecodeInstruction(front, front_words, inst);
+		result.has_swap_pc |= inst.opcode == Opcode::S_SWAPPC_B64;
 		front_words += inst.word_count;
 		if (inst.opcode == Opcode::S_SETPC_B64) {
 			EXIT_NOT_IMPLEMENTED(inst.src0.kind != OperandKind::Sgpr || inst.src0.reg != 6u);
@@ -406,6 +407,7 @@ void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 	program.instructions.clear();
 	program.instructions.reserve(code.size());
 	program.code = code;
+	program.has_swap_pc = false;
 
 	std::vector<bool> branch_targets;
 	for (uint32_t word_index = 0; word_index < code.size();) {
@@ -413,6 +415,7 @@ void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 		DecodeInstruction(code, word_index, program.instructions.back());
 
 		const auto& inst = program.instructions.back();
+		program.has_swap_pc |= inst.opcode == Opcode::S_SWAPPC_B64;
 		word_index += inst.word_count;
 
 		if (IsDirectBranch(inst.opcode)) {

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

@@ -60,6 +60,7 @@ enum class Opcode {
 	S_QUADMASK_B64,
 	S_GETPC_B64,
 	S_SETPC_B64,
+	S_SWAPPC_B64,
 	S_SUBVECTOR_LOOP_BEGIN,
 	S_SUBVECTOR_LOOP_END,
 	S_AND_SAVEEXEC_B32,
@@ -678,6 +679,7 @@ enum class Opcode {
 	S_CBRANCH_CDBGSYS,
 	S_CBRANCH_CDBGSYS_OR_USER,
 	S_SENDMSG,
+	S_GETREG_B32,
 	S_SETREG_B32,
 	S_SLEEP,
 	S_TRAP,
@@ -816,6 +818,7 @@ struct Instruction {
 struct Program {
 	std::span<const uint32_t> code;
 	std::vector<Instruction>  instructions;
+	bool                     has_swap_pc = false;
 };
 
 // Code spans are trusted to contain complete instructions, valid branch targets, and 32-bit PCs.

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

@@ -10,13 +10,21 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_MOV_B64: S_MOV_B64(inst); return;
 		case O::S_WQM_B32: S_WQM(inst, false); return;
 		case O::S_WQM_B64: S_WQM(inst, true); return;
-		case O::S_GETPC_B64: S_GETPC_B64(inst); return;
+		case O::S_GETPC_B64:
+		case O::S_SWAPPC_B64: S_GETPC_B64(inst); return;
 		case O::S_SETPC_B64: return;
 		case O::S_SUBVECTOR_LOOP_BEGIN: S_SUBVECTOR_LOOP(inst, true); return;
 		case O::S_SUBVECTOR_LOOP_END: S_SUBVECTOR_LOOP(inst, false); return;
 		case O::S_CSELECT_B32: S_CSELECT_B32(inst); return;
 		case O::S_CSELECT_B64: ScalarSelect64(inst, inst.src1); return;
 		case O::S_CMOV_B64: ScalarSelect64(inst, inst.dst); return;
+		case O::S_GETREG_B32:
+			if (!graphics_compute || (inst.src0.value != 0x1818u && inst.src0.value != 0x0a18u))
+				EXIT("unsupported S_GETREG_B32 at pc 0x%08x: selector=0x%04x graphics_compute=%u",
+				     inst.pc, inst.src0.value, static_cast<unsigned>(graphics_compute));
+			// HW_ID2 identifies graphics compute as QUEUE=6, ME=0.
+			WriteOperand(inst.dst, IR::Value(inst.src0.value == 0x1818u ? 6u : 0u));
+			return;
 		case O::S_SETREG_B32: EmitControlNop(); return;
 		case O::S_WAITCNT_VSCNT: S_WAITCNT_VSCNT(inst); return;
 		case O::S_WAITCNT: return;

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

@@ -1356,7 +1356,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 	                              (options.input_info.compute->float_mode & 0x10u) == 0;
 	for (const auto& cfg_block: cfg.blocks) {
 		auto*      block = blocks_by_id.at(cfg_block.id);
-		Translator translator(result, block, vector_limit, flush_f32_inputs);
+		Translator translator(result, block, vector_limit, flush_f32_inputs,
+		    options.stage == ShaderType::Compute && !options.input_info.compute->async_compute);
 		for (uint32_t index = cfg_block.inst_begin; index < cfg_block.inst_end; index++) {
 			const auto& instruction = decoded.instructions[index];
 			if (IsCodeTableLoad(cfg, instruction.pc)) {

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

@@ -10,9 +10,9 @@ namespace Libs::Graphics::ShaderRecompiler::Frontend {
 class Translator {
 public:
 	Translator(IR::Program& program, IR::Block* block, uint32_t vector_limit,
-	           bool flush_f32_inputs)
+	           bool flush_f32_inputs, bool graphics_compute)
 	    : program(program), ir(block), current_vector_limit(vector_limit),
-	      flush_f32_inputs(flush_f32_inputs) {}
+	      flush_f32_inputs(flush_f32_inputs), graphics_compute(graphics_compute) {}
 
 	void TranslateInstruction(const Decoder::Instruction& inst);
 	void TranslateEmbeddedFetch(const Decoder::Instruction& inst, uint32_t attribute,
@@ -264,6 +264,7 @@ private:
 	uint32_t        current_pc           = 0;
 	uint32_t        current_vector_limit = 1;
 	bool            flush_f32_inputs;
+	bool            graphics_compute;
 };
 
 } // namespace Libs::Graphics::ShaderRecompiler::Frontend

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

@@ -161,6 +161,7 @@ struct SamplerResource {
 	uint32_t first_use_pc          = 0;
 	// Native filtering/border variants share the original sampler's runtime descriptor.
 	uint32_t snapshot_index        = 0;
+	std::vector<uint32_t> indirect_resources;
 	bool     force_point_filtering = false;
 	bool     depth_compare         = false;
 	bool     integer_border        = false;
@@ -590,6 +591,7 @@ struct ResourcePlan {
 	std::vector<ResourceBlock>          control_flow;
 	std::vector<SrtRead>                srt_reads;
 	bool                                requires_specialization_memory = false;
+	std::vector<std::pair<uint64_t, uint64_t>> source_reads;
 	bool                                capture_specialization_reads = false;
 	bool                                srt_plan_complete          = false;
 	bool                                resource_tracking_complete = false;

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

@@ -10,7 +10,6 @@
 #include <array>
 #include <bit>
 #include <cstdio>
-#include <cstring>
 #include <fmt/format.h>
 #include <functional>
 #include <numeric>
@@ -141,31 +140,6 @@ 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 != nullptr) {
-		if (!capture.source.read_memory(capture.source.userdata, address, values)) return false;
-	} else {
-		std::memcpy(values.data(), reinterpret_cast<const void*>(address), values.size_bytes());
-	}
-	capture.ranges.emplace_back(address, values.size_bytes());
-	return true;
-}
-
 const DescriptorSource* Source(const ResourcePlan& program, uint32_t source) {
 	if (source >= program.descriptor_sources.size()) {
 		return nullptr;
@@ -366,20 +340,22 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		} else if (!indirect.selector_first.IsEmpty()) {
 			ShaderBufferResource material;
 			uint32_t             first = 0, count = 0;
-			if (!DecodeBufferDescriptor(material_value, material) || material.Type() != 0u ||
-			    material.SwizzleEnabled() ||
-			    material.AddTid() || material.OutOfBounds() != 0u ||
-			    uint64_t {selector->offset} + 4u > material.Stride() ||
+			if (!DecodeBufferDescriptor(material_value, material) ||
 			    !clean.Evaluate(indirect.selector_first, first) ||
-			    !clean.Evaluate(indirect.key_count, count) || count > MaxIndirectDescriptorProbes ||
-			    uint64_t {first} + count > material.NumRecords())
+			    !clean.Evaluate(indirect.key_count, count) || count > MaxIndirectDescriptorProbes)
 				return false;
-			if (count != 0u && (uint64_t {first} + count - 1u) * material.Stride() +
-			                           selector->offset + 4u >
-			                       uint64_t {UINT32_MAX} + 1u)
+			// SMEM uses its instruction's byte stride; V# stride only supplies its bounds.
+			const auto step = selector->stride != 0u ? selector->stride : material.Stride();
+			if (selector->stride == 0u &&
+			    (material.Type() != 0u || material.SwizzleEnabled() || material.AddTid() ||
+			     material.OutOfBounds() != 0u || uint64_t {selector->offset} + 4u > step ||
+			     uint64_t {first} + count > material.NumRecords())) return false;
+			const auto end = count != 0u
+			    ? (uint64_t {first} + count - 1u) * step + selector->offset + 4u : 0u;
+			if (end > uint64_t {UINT32_MAX} + 1u || end > material.GetSize())
+				return false;
+			if (!read_keys(material, uint64_t {first} * step + selector->offset, step, count))
 				return false;
-			if (!read_keys(material, uint64_t {first} * material.Stride() + selector->offset,
-			               material.Stride(), count)) return false;
 		} else if (!indirect.selector_mask.IsEmpty()) {
 			uint32_t mask = 0;
 			uint32_t count = 0;
@@ -659,6 +635,10 @@ bool BuildSamplerPlan(const ShaderInfo& base, SamplerPlan& plan) {
 			return false;
 		}
 		usage[pair.sampler] |= 1u << static_cast<uint32_t>(ClassifySampler(base.images[pair.image]));
+		for (const auto candidate: base.samplers[pair.sampler].indirect_resources) {
+			if (candidate >= base.samplers.size()) return false;
+			usage[candidate] |= usage[pair.sampler];
+		}
 	}
 	for (uint32_t index = 0; index < base.samplers.size(); index++) {
 		auto& mapping = plan.mapping[index];
@@ -741,7 +721,31 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 		block.sources.erase(std::unique(block.sources.begin(), block.sources.end()),
 		                    block.sources.end());
 	}
-	for (auto& block: blocks) block.condition = predicate(block.condition);
+	// Only branches that can change future resource access belong in the host plan.
+	std::vector<uint8_t> resource_reachable(blocks.size());
+	std::vector<uint32_t> pending;
+	for (uint32_t i = 0; i < blocks.size(); ++i) {
+		if (!blocks[i].sources.empty() || !blocks[i].srt_reads.empty()) {
+			resource_reachable[i] = 1;
+			pending.push_back(i);
+		}
+	}
+	while (!pending.empty()) {
+		const auto index = pending.back();
+		pending.pop_back();
+		for (const auto* predecessor: program.blocks[index]->ImmPredecessors()) {
+			const auto found = indices.find(predecessor);
+			if (found == indices.end()) return {};
+			if (resource_reachable[found->second]) continue;
+			resource_reachable[found->second] = 1;
+			pending.push_back(found->second);
+		}
+	}
+	for (auto& block: blocks) {
+		block.condition = std::ranges::any_of(block.successors, [&](uint32_t successor) {
+			return resource_reachable[successor] != 0;
+		}) ? predicate(block.condition) : Value {};
+	}
 	if (std::ranges::none_of(
 	        blocks, [](const ResourceBlock& block) { return !block.condition.IsEmpty(); })) {
 		return {};
@@ -927,6 +931,7 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	plan.user_data_count            = program.user_data_count;
 	plan.info                       = program.info;
 	plan.memory_info                = program.memory_info;
+	plan.source_reads               = program.source_reads;
 	plan.srt_plan_complete          = program.srt_plan_complete;
 	plan.resource_tracking_complete = program.resource_tracking_complete;
 
@@ -1007,10 +1012,15 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	// A proven uniform factor can decide a branch even when its other lanes are unknown.
 	// Keep only that Boolean structure, never the varying shader dependency graph.
 	Value unknown;
+	bool capture_control_reads = false;
 	std::function<Value(Value)> ClonePredicate = [&](Value value) -> Value {
 		value = value.Resolve();
 		if (value.IsEmpty()) return {};
-		if (ValidateRuntimeValue(program, value, RuntimeValueType::Integer)) return Clone(value);
+		if (ValidateRuntimeValue(program, value, RuntimeValueType::Integer)) {
+			capture_control_reads |=
+			    !ValidateRuntimeValue(program, value, RuntimeValueType::ImmutableInteger);
+			return Clone(value);
+		}
 		const auto* inst = value.TryInstruction();
 		if (inst == nullptr) return {};
 		const auto op = inst->GetOpcode();
@@ -1040,7 +1050,8 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	for (uint32_t i = 0; i < plan.uniform_fill.fill.words; ++i) {
 		plan.uniform_fill.values[i] = Clone(plan.uniform_fill.values[i]);
 	}
-	plan.capture_specialization_reads |= capture_indirect_reads || !plan.control_flow.empty();
+	plan.capture_specialization_reads |=
+	    capture_indirect_reads || capture_control_reads || !plan.source_reads.empty();
 	if (plan.capture_specialization_reads) {
 		// Every former flat alias now targets its retained raw read. Mark those reads
 		// on the normalized graph, so discarded Phi/EXEC dependencies stay discarded.
@@ -1075,27 +1086,23 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 			if (image.written) MarkCleanReads(planned_reads, Source(plan, image.source));
 		}
 	}
-	if (capture_indirect_reads) plan.resource_tracking_complete &= !program.has_address_writes;
+	if (capture_indirect_reads || !plan.source_reads.empty())
+		plan.resource_tracking_complete &= !program.has_address_writes;
 	return plan;
 }
 
 bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime,
                           ResourceSnapshot& snapshot, ResourceSpecialization& specialization) {
 	if (!program.resource_tracking_complete ||
-	    (program.requires_specialization_memory && runtime.read_specialization_memory == nullptr)) {
+	    ((program.requires_specialization_memory || !program.source_reads.empty()) &&
+	     runtime.read_specialization_memory == nullptr)) {
 		return false;
 	}
-	const bool capture_reads = program.capture_specialization_reads;
+	const bool capture_reads = program.capture_specialization_reads || !program.source_reads.empty();
 	auto& reads = snapshot.specialization_reads;
-	reads.clear();
-	ReadCapture capture {runtime, reads};
-	SrtRuntime observed = runtime;
-	if (capture_reads) {
-		observed.userdata = &capture;
-		observed.read_specialization_memory = runtime.read_specialization_memory != nullptr
-		                                         ? CaptureStrictRead : nullptr;
-		observed.read_memory = CaptureOrdinaryRead;
-	}
+	reads = program.source_reads;
+	SrtReadCapture capture {runtime, reads};
+	const auto observed = capture_reads ? capture.ObservedRuntime() : runtime;
 	SrtWalker clean(program, CleanRuntime(observed));
 	SrtWalker walker(program, observed,
 	                 capture_reads || program.requires_specialization_memory ? &clean : nullptr);
@@ -1221,7 +1228,14 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	}
 	snapshot.samplers.resize(program.info.samplers.size());
 	for (uint32_t i = 0; i < program.info.samplers.size(); ++i) {
-		if (!evaluate(program.info.samplers[i].source, snapshot.samplers[i])) {
+		auto source = program.info.samplers[i].source;
+		if (source >= program.descriptor_sources.size()) return false;
+		const auto& descriptor = program.descriptor_sources[source];
+		if (descriptor.indirect_descriptor) {
+			if (descriptor.indirect_descriptor->sources.empty()) return false;
+			source = descriptor.indirect_descriptor->sources[0];
+		}
+		if (!evaluate(source, snapshot.samplers[i])) {
 			return false;
 		}
 		if (program.info.samplers[i].gather_lod) {
@@ -1313,11 +1327,20 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 		samplers[index].force_point_filtering = binding.type == SamplerClass::PointInteger;
 		samplers[index].integer_border        = binding.type != SamplerClass::Float;
 	}
+	for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {
+		const auto& binding = sampler_plan.bindings[index];
+		for (auto& candidate: samplers[index].indirect_resources) {
+			candidate = sampler_plan.mapping[candidate][static_cast<uint32_t>(binding.type)];
+			EXIT_IF(candidate == UINT32_MAX);
+		}
+	}
 	for (auto& pair: sampled_pairs) {
 		const auto type = static_cast<uint32_t>(ClassifySampler(images[pair.image]));
 		pair.sampler = sampler_plan.mapping[pair.sampler][type];
 		EXIT_IF(pair.sampler == UINT32_MAX);
 		samplers[pair.sampler].depth_compare |= images[pair.image].depth_compare;
+		for (const auto candidate: samplers[pair.sampler].indirect_resources)
+			samplers[candidate].depth_compare |= images[pair.image].depth_compare;
 	}
 
 	auto& memory_info = program.memory_info;
@@ -1405,7 +1428,8 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 				EXIT_IF(memory.sampler == UINT32_MAX);
 			}
 			EXIT_IF(image.indirect_root == memory.resource &&
-			        inst.GetOpcode() != ValueOpcode::ImageSampleRaw);
+			        inst.GetOpcode() != ValueOpcode::ImageSampleRaw &&
+			        inst.GetOpcode() != ValueOpcode::ImageRead);
 		}
 	}
 	for (auto& memory: memory_info) {

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

@@ -430,7 +430,7 @@ public:
 					plan.handle->SetArg(word, Value(0u));
 				continue;
 			}
-			if (plan.handle->GetOpcode() == ValueOpcode::GetBufferResource) {
+			if (plan.handle->NumArgs() == 4u) {
 				plan.handle->SetArg(0, plan.key);
 				for (uint32_t word = 1; word < 4u; ++word) plan.handle->SetArg(word, Value(0u));
 			} else {
@@ -609,23 +609,37 @@ private:
 		return selected.IsEmpty() ? value : selected;
 	}
 
-	Value LowerDescriptorPhi(Value value) {
-		value           = value.Resolve();
-		const auto* phi = value.TryInstruction();
-		if (m_shader_writes || phi == nullptr || phi->GetOpcode() != ValueOpcode::Phi ||
-		    phi->NumArgs() != 2u || phi->NumPhiBlocks() != 2u || phi->GetType() != Type::U32) {
-			return value;
-		}
-		const auto* merge = phi->Parent();
-		const auto* branch = phi->PhiBlock(0);
-		if (merge == nullptr || branch == nullptr || phi->PhiBlock(1) == nullptr ||
-		    branch == phi->PhiBlock(1)) {
-			return value;
-		}
-		for (const auto& [original, selected]: m_descriptor_selections) {
-			if (original == phi) {
-				return selected;
+	Value LowerDescriptorValue(Value value) {
+		value            = value.Resolve();
+		const auto* inst = value.TryInstruction();
+		if (inst == nullptr) return value;
+		// The scalar planner owns reads and lowers their address handles separately.
+		uint32_t memory_index = 0;
+		if (ScalarReadMemory(*inst, memory_index) != nullptr) return value;
+		const auto [cached, inserted] = m_descriptor_values.emplace(inst, value);
+		if (!inserted) return cached->second;
+		if (inst->GetOpcode() != ValueOpcode::Phi) {
+			Inst* projected = nullptr;
+			for (size_t arg = 0; arg < inst->NumArgs(); ++arg) {
+				const auto lowered = LowerDescriptorValue(inst->Arg(arg));
+				if (lowered == inst->Arg(arg).Resolve()) continue;
+				if (projected == nullptr) {
+					projected = &m_program.value_storage.emplace_back(inst->GetOpcode(), inst->Flags<uint64_t>());
+					for (size_t i = 0; i < inst->NumArgs(); ++i) projected->SetArg(i, inst->Arg(i));
+				}
+				projected->SetArg(arg, lowered);
 			}
+			if (projected != nullptr) cached->second = Value(projected);
+			return cached->second;
+		}
+		if (inst->NumArgs() != 2u || inst->NumPhiBlocks() != 2u || inst->GetType() != Type::U32) {
+			return value;
+		}
+		const auto* merge = inst->Parent();
+		const auto* branch = inst->PhiBlock(0);
+		if (merge == nullptr || branch == nullptr || inst->PhiBlock(1) == nullptr ||
+		    branch == inst->PhiBlock(1)) {
+			return value;
 		}
 		if (branch->ImmSuccessors().size() != 2u) {
 			if (branch->ImmPredecessors().size() != 1u) {
@@ -638,7 +652,7 @@ private:
 		}
 		std::array<const Block*, 2> targets;
 		for (uint32_t arm = 0; arm < 2; arm++) {
-			const auto* incoming = phi->PhiBlock(arm);
+			const auto* incoming = inst->PhiBlock(arm);
 			if (incoming == merge ||
 			    (incoming != branch &&
 			     (incoming->ImmPredecessors().size() != 1u ||
@@ -654,19 +668,20 @@ private:
 		if (term.kind != CFG::TerminatorKind::ConditionalBranch ||
 		    !((term.true_block == targets[0] && term.false_block == targets[1]) ||
 		      (term.false_block == targets[0] && term.true_block == targets[1])) ||
-		    !ValidateRuntimeValue(m_program, info.condition, RuntimeValueType::Integer) ||
-		    !ValidateRuntimeValue(m_program, phi->Arg(0)) ||
-		    !ValidateRuntimeValue(m_program, phi->Arg(1))) {
+		    !ValidateRuntimeValue(m_program, info.condition, m_shader_writes
+		        ? RuntimeValueType::ImmutableInteger : RuntimeValueType::Integer) ||
+		    !ValidateRuntimeValue(m_program, inst->Arg(0)) ||
+		    !ValidateRuntimeValue(m_program, inst->Arg(1))) {
 			return value;
 		}
 		// Retain a host expression; replacing the GPU Phi would break SSA dominance.
 		const auto true_arg = term.true_block == targets[0] ? 0u : 1u;
 		auto&      selected = m_program.value_storage.emplace_back(ValueOpcode::SelectU32);
 		selected.SetArg(0, info.condition);
-		selected.SetArg(1, phi->Arg(true_arg));
-		selected.SetArg(2, phi->Arg(true_arg ^ 1u));
-		m_descriptor_selections.emplace_back(phi, Value(&selected));
-		return Value(&selected);
+		selected.SetArg(1, inst->Arg(true_arg));
+		selected.SetArg(2, inst->Arg(true_arg ^ 1u));
+		cached->second = Value(&selected);
+		return cached->second;
 	}
 
 	void MakeSource(const Inst& handle, uint32_t width, bool sampler, bool sample_adjust,
@@ -686,7 +701,7 @@ private:
 		for (uint32_t i = 0; i < width; i++) {
 			const auto value = base_reg != UINT32_MAX
 			    ? NativeDescriptorSource(handle.Arg(i), base_reg + i, pc) : handle.Arg(i);
-			descriptor.dwords[i] = LowerDescriptorPhi(value);
+			descriptor.dwords[i] = LowerDescriptorValue(value);
 		}
 		if (sample_adjust) {
 			descriptor.dwords[3] = CanonicalizeSampleAdjustDword3(descriptor.dwords[3]);
@@ -792,15 +807,6 @@ private:
 				}
 			}
 		}
-		for (auto* block: m_program.blocks) {
-			for (auto& inst: *block) {
-				uint32_t index = 0;
-				if (inst.GetOpcode() == ValueOpcode::LoadAddressU32 &&
-				    ScalarReadMemory(inst, index) != nullptr && inst.Arg(1).Resolve().IsImmediate() &&
-				    ValidateRuntimeValue(m_program, Value(&inst)))
-					CollectScalarRead(Value(&inst), inst.Flags<MemoryFlags>().pc);
-			}
-		}
 		for (auto* read: m_scalar_reads) {
 			const auto flags = read->Flags<MemoryFlags>();
 			auto& memory = m_program.memory_info[flags.index];
@@ -1463,7 +1469,18 @@ private:
 			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;
+			    step->Parent() == nullptr) continue;
+			// Structurization can forward the native increment through a synthetic continue block.
+			const auto* incoming = phi->PhiBlock(initial ^ 1u);
+			const auto* next = phi->Parent();
+			for (size_t edges = 0; incoming != step->Parent() && edges < m_program.blocks.size(); ++edges) {
+				if (incoming == nullptr || incoming->terminator.kind != CFG::TerminatorKind::Branch ||
+				    incoming->terminator.true_block != next || incoming->ImmPredecessors().size() != 1u)
+					break;
+				next = incoming;
+				incoming = incoming->ImmPredecessors()[0];
+			}
+			if (incoming != step->Parent()) continue;
 			uint32_t increment = 0;
 			if ((step->Arg(0).Resolve() == key &&
 			     ImmediateU32(step->Arg(1), increment) && increment == 1u) ||
@@ -1501,6 +1518,30 @@ private:
 		});
 	}
 
+	bool MatchScalarMaterialKey(Value key, DescriptorSource::IndirectDescriptor& indirect,
+	                            DescriptorSource& material_source) {
+		const auto* read = key.Resolve().TryInstruction();
+		if (read == nullptr || read->GetOpcode() != ValueOpcode::ReadConstBuffer) return false;
+		uint32_t memory_index = 0;
+		const auto* memory = ScalarReadMemory(*read, memory_index);
+		Value index;
+		uint32_t offset = 0, stride = 0;
+		if (memory == nullptr || memory->kind != ResourceKind::ScalarBuffer ||
+		    memory->offset > INT32_MAX || !MemoryIndexBelongsTo(memory_index, *read) ||
+		    !MatchTableOffset(read->Arg(1), index, offset, stride) ||
+		    ((offset | stride | memory->offset) & 3u) != 0u ||
+		    uint64_t {offset} + memory->offset > UINT32_MAX) return false;
+		const auto* bound = BoundedLoop(index, read->Parent());
+		if (bound == nullptr || bound->GetOpcode() != ValueOpcode::ULessThan32 ||
+		    !MakeRuntimeTableSource(*read, material_source)) return false;
+		indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead {
+		    .source = InternSource(material_source), .stride = stride,
+		    .offset = offset + memory->offset});
+		indirect.selector_first = Value(0u);
+		indirect.key_count = bound->Arg(1);
+		return true;
+	}
+
 	void FoldBoundedLoopSelectors() {
 		for (auto* header: m_program.blocks) {
 			for (auto& induction: *header) {
@@ -1736,6 +1777,10 @@ private:
 			}
 			if ((table_offset & 3u) != 0u ||
 			    (bitscan && table_offset > UINT32_MAX - (32u * 32u - 1u))) return false;
+		} else if (!plan.table_indexed &&
+		           MatchScalarMaterialKey(key, indirect, material_source)) {
+			// The host enumerates the guarded scalar loop; its key and T# payload stay on the GPU.
+			plan.retain_reads = true;
 		} else {
 			// A bounded V# supplies the complete image table. Leave every GPU selector
 			// and descriptor read in the shader; the host only translates table bytes.
@@ -1899,17 +1944,21 @@ private:
 		return found;
 	}
 
-	bool TryMakeFiniteImage(Inst& handle, IndirectDescriptorPlan& plan) {
-		if (handle.GetOpcode() != ValueOpcode::GetImageResource || handle.NumArgs() != 8u)
+	bool TryMakeFiniteDescriptor(Inst& handle, IndirectDescriptorPlan& plan) {
+		const bool sampler = handle.GetOpcode() == ValueOpcode::GetSamplerResource;
+		if ((!sampler && handle.GetOpcode() != ValueOpcode::GetImageResource) ||
+		    handle.NumArgs() != (sampler ? 4u : 8u))
 			return false;
-		if (std::ranges::any_of(handle.Uses(), [](const Use& use) {
+		if (std::ranges::any_of(handle.Uses(), [&](const Use& use) {
 			const auto op = use.user->GetOpcode();
-			return op != ValueOpcode::ImageSampleRaw && op != ValueOpcode::ImageGatherRaw;
+			if (op != ValueOpcode::ImageSampleRaw) return sampler || op != ValueOpcode::ImageGatherRaw;
+			return sampler && (m_program.memory_info[use.user->Flags<MemoryFlags>().index]
+			                       .image_sample_flags & Decoder::ImageSampleFlagAdjust) != 0u;
 		})) return false;
 		bool has_phi = false;
 		for (size_t word = 0; word < handle.NumArgs(); ++word) has_phi |= handle.Arg(word).Resolve().IsPhi();
 		if (!has_phi) return false;
-		// Normalize the eight synchronized descriptor words to one GPU ordinal.
+		// Normalize synchronized descriptor words to one GPU ordinal.
 		// Build the complete graph before changing IR so an unsupported leaf is transactional.
 		struct Choice {
 			DescriptorSource descriptor;
@@ -1925,7 +1974,15 @@ private:
 			const auto index = static_cast<uint32_t>(choices.size());
 			choices.push_back({descriptor});
 			uint32_t bad = 0;
-			if (ValidateSource(descriptor, bad)) return index;
+			if (ValidateSource(descriptor, bad)) {
+				// Finite inline samplers have literal leaves; live descriptor tables use the
+				// existing direct sampler planning path.
+				if (sampler && !std::ranges::all_of(
+				    std::span(descriptor.dwords).first(descriptor.dword_count),
+				    [](Value word) { return word.IsImmediate() && word.GetType() == Type::U32; }))
+					return UINT32_MAX;
+				return index;
+			}
 			const auto* branch = descriptor.dwords[bad].TryInstruction();
 			if (branch == nullptr || branch->Parent() == nullptr ||
 			    branch->GetOpcode() != ValueOpcode::Phi ||
@@ -1962,14 +2019,14 @@ private:
 			return index;
 		};
 		DescriptorSource root;
-		root.dword_count = 8u;
+		root.dword_count = handle.NumArgs();
 		for (uint32_t word = 0; word < root.dword_count; ++word) root.dwords[word] = handle.Arg(word);
 		if (visit(visit, root) == UINT32_MAX || choices.front().branch == nullptr) return false;
-		DescriptorSource image_source;
-		image_source.dword_count = 8u;
-		image_source.dwords.fill(Value(0u));
-		image_source.indirect_descriptor.emplace(DescriptorSource::IndirectDescriptor{});
-		auto& sources = image_source.indirect_descriptor->sources;
+		DescriptorSource finite_source;
+		finite_source.dword_count = root.dword_count;
+		finite_source.dwords.fill(Value(0u));
+		finite_source.indirect_descriptor.emplace(DescriptorSource::IndirectDescriptor{});
+		auto& sources = finite_source.indirect_descriptor->sources;
 		for (auto& choice: choices) {
 			if (choice.branch != nullptr) continue;
 			const auto source = InternSource(choice.descriptor);
@@ -1998,9 +2055,9 @@ private:
 			}
 		}
 		plan.handle = &handle;
-		plan.source = InternSource(image_source);
+		plan.source = InternSource(finite_source);
 		plan.key = choices.front().key;
-		plan.roots = image_source.dwords;
+		plan.roots = finite_source.dwords;
 		plan.reads.fill(nullptr);
 		return true;
 	}
@@ -2031,27 +2088,33 @@ private:
 				    inst.NumArgs() == 0u) {
 					continue;
 				}
-				auto* handle = inst.Arg(0).Resolve().TryInstruction();
-				if (handle == nullptr || FindIndirectDescriptor(*handle) != nullptr) {
-					continue;
-				}
 				const auto flags = inst.Flags<MemoryFlags>();
 				const auto& memory = m_program.memory_info[flags.index];
-				DescriptorSource descriptor;
-				MakeSource(*handle, inst.GetOpcode() == ValueOpcode::StoreBufferU32 ? 4u : 8u,
-				           false, false, memory.resource * 4u, descriptor, flags.pc);
-				uint32_t bad_dword = 0;
-				if (ValidateSource(descriptor, bad_dword)) continue;
-				IndirectDescriptorPlan plan;
-				if (TryMakeIndirectImage(*handle, descriptor, plan) || TryMakeFiniteImage(*handle, plan) ||
-				    TryMakeIndirectBuffer(*handle, descriptor, plan)) {
-					for (const auto& previous: m_indirect_descriptors) {
-						if (plan.reads[0] != nullptr && previous.reads[0] == plan.reads[0] &&
-						    !EquivalentValue(m_program, previous.key, plan.key))
-							Fail(flags.pc, "shared descriptor read has incompatible keys");
+				const auto plan_handle = [&](Value value, uint32_t width, uint32_t base, bool sampler) {
+					auto* handle = value.Resolve().TryInstruction();
+					if (handle == nullptr || FindIndirectDescriptor(*handle) != nullptr) return;
+					DescriptorSource descriptor;
+					MakeSource(*handle, width, sampler, false, base, descriptor, flags.pc);
+					uint32_t bad_dword = 0;
+					if (ValidateSource(descriptor, bad_dword)) return;
+					IndirectDescriptorPlan plan;
+					if (sampler ? TryMakeFiniteDescriptor(*handle, plan)
+					             : (TryMakeIndirectImage(*handle, descriptor, plan) ||
+					                TryMakeFiniteDescriptor(*handle, plan) ||
+					                TryMakeIndirectBuffer(*handle, descriptor, plan))) {
+						for (const auto& previous: m_indirect_descriptors) {
+							if (plan.reads[0] != nullptr && previous.reads[0] == plan.reads[0] &&
+							    !EquivalentValue(m_program, previous.key, plan.key))
+								Fail(flags.pc, "shared descriptor read has incompatible keys");
+						}
+						m_indirect_descriptors.push_back(std::move(plan));
 					}
-					m_indirect_descriptors.push_back(std::move(plan));
-				}
+				};
+				plan_handle(inst.Arg(0), inst.GetOpcode() == ValueOpcode::StoreBufferU32 ? 4u : 8u,
+				            memory.resource * 4u, false);
+				if (inst.GetOpcode() == ValueOpcode::ImageSampleRaw &&
+				    (memory.image_sample_flags & Decoder::ImageSampleFlagAdjust) == 0u)
+					plan_handle(inst.Arg(1), 4u, memory.sampler * 4u, true);
 			}
 		}
 	}
@@ -2180,7 +2243,20 @@ private:
 			return UINT32_MAX;
 		}
 		m_info.samplers.push_back({source, pc});
-		return static_cast<uint32_t>(m_info.samplers.size() - 1);
+		const auto root = static_cast<uint32_t>(m_info.samplers.size() - 1);
+		const auto* descriptor = Source(source);
+		if (descriptor != nullptr && descriptor->indirect_descriptor) {
+			const auto& sources = descriptor->indirect_descriptor->sources;
+			if (sources.empty()) Fail(pc, "finite sampler has no candidates");
+			std::vector<uint32_t> resources {root};
+			for (size_t i = 1; i < sources.size(); ++i) {
+				const auto candidate = AddSampler(sources[i], pc);
+				if (candidate == UINT32_MAX) return UINT32_MAX;
+				resources.push_back(candidate);
+			}
+			m_info.samplers[root].indirect_resources = std::move(resources);
+		}
+		return root;
 	}
 
 	void AddSampledPair(uint32_t image, uint32_t sampler, uint32_t pc) {
@@ -2340,8 +2416,15 @@ private:
 			uint32_t   sampler_source = 0;
 			const bool sample_adjust =
 			    (memory.image_sample_flags & Decoder::ImageSampleFlagAdjust) != 0;
-			GetHandle(inst.Arg(1), ValueOpcode::GetSamplerResource, 4, flags.pc,
-			          memory.sampler * 4u, sampler_handle, sampler_source, true, sample_adjust);
+			sampler_handle = inst.Arg(1).Resolve().TryInstruction();
+			const auto* selected_sampler = sampler_handle != nullptr
+			    ? FindIndirectDescriptor(*sampler_handle) : nullptr;
+			if (selected_sampler != nullptr) {
+				sampler_source = selected_sampler->source;
+			} else {
+				GetHandle(inst.Arg(1), ValueOpcode::GetSamplerResource, 4, flags.pc,
+				          memory.sampler * 4u, sampler_handle, sampler_source, true, sample_adjust);
+			}
 			sampler = AddSampler(sampler_source, flags.pc);
 			if (sampler == UINT32_MAX) {
 				Fail(flags.pc, "sampler resource limit exceeded");
@@ -2397,7 +2480,7 @@ private:
 	std::vector<HandlePatch>                   m_handle_patches;
 	std::vector<MemoryPatch>                   m_memory_patches;
 	std::vector<IndirectDescriptorPlan>             m_indirect_descriptors;
-	std::vector<std::pair<const Inst*, Value>> m_descriptor_selections;
+	std::map<const Inst*, Value>              m_descriptor_values;
 	bool                                       m_shader_writes = false;
 };
 

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

@@ -11,6 +11,35 @@
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
+SrtRuntime SrtReadCapture::ObservedRuntime() {
+	auto runtime = m_source;
+	runtime.userdata = this;
+	runtime.read_specialization_memory = m_source.read_specialization_memory != nullptr
+	                                         ? ReadStrict : nullptr;
+	runtime.read_memory = ReadOrdinary;
+	return runtime;
+}
+
+bool SrtReadCapture::ReadStrict(void* userdata, uint64_t address, std::span<uint32_t> values) {
+	auto& capture = *static_cast<SrtReadCapture*>(userdata);
+	if (!capture.m_source.read_specialization_memory(capture.m_source.userdata, address, values)) {
+		return false;
+	}
+	capture.m_ranges.emplace_back(address, values.size_bytes());
+	return true;
+}
+
+bool SrtReadCapture::ReadOrdinary(void* userdata, uint64_t address, std::span<uint32_t> values) {
+	auto& capture = *static_cast<SrtReadCapture*>(userdata);
+	if (capture.m_source.read_memory != nullptr) {
+		if (!capture.m_source.read_memory(capture.m_source.userdata, address, values)) return false;
+	} else {
+		std::memcpy(values.data(), reinterpret_cast<const void*>(address), values.size_bytes());
+	}
+	capture.m_ranges.emplace_back(address, values.size_bytes());
+	return true;
+}
+
 SrtRuntime CleanRuntime(SrtRuntime runtime) {
 	runtime.read_memory = runtime.read_specialization_memory != nullptr
 	                          ? runtime.read_specialization_memory
@@ -159,7 +188,7 @@ private:
 	bool Validate(Value value, bool require_uniform = true) {
 		value = value.Resolve();
 		// Host floating-point evaluation does not model shader rounding/denormal modes.
-		if (m_type == RuntimeValueType::Integer &&
+		if (m_type != RuntimeValueType::Any &&
 		    TypesOverlap(value.GetType(), Type::F16 | Type::F32 | Type::F32x2)) {
 			return false;
 		}
@@ -188,6 +217,13 @@ private:
 			return valid;
 		};
 		const auto op = inst->GetOpcode();
+		if (m_type == RuntimeValueType::ImmutableInteger &&
+		    (op == ValueOpcode::ReadConst || BufferAccessOf(op) != BufferAccess::None ||
+		     AddressOpcodeInfoOf(op).access != AddressAccess::None ||
+		     ImageOpcodeInfoOf(op).access != ImageAccess::None ||
+		     SharedAccessOf(op) != SharedAccess::None)) {
+			return finish(false);
+		}
 		if (op == ValueOpcode::ReadConst) {
 			const auto slot = inst->NumArgs() == 2 ? inst->Arg(1).Resolve() : Value {};
 			if (inst->NumArgs() != 2 || inst->Arg(0).Resolve().TryInstruction() == nullptr ||
@@ -197,7 +233,7 @@ private:
 			    slot.U32() >= m_program.srt_reads.size()) {
 				return finish(false);
 			}
-			if (m_type == RuntimeValueType::Integer) {
+			if (m_type != RuntimeValueType::Any) {
 				const auto active_mask = m_active_mask;
 				m_active_mask          = {};
 				const bool valid       = Validate(m_program.srt_reads[slot.U32()].value);
@@ -209,7 +245,7 @@ private:
 		if (!m_active_mask.IsEmpty() && IsRuntimeSelect(op) && inst->NumArgs() == 3 &&
 		    inst->Arg(0).Resolve() == m_active_mask) {
 			// Empty EXEC reads lane zero, so ignored operands still require integer types.
-			if (m_type == RuntimeValueType::Integer && !Validate(inst->Arg(2), false)) {
+			if (m_type != RuntimeValueType::Any && !Validate(inst->Arg(2), false)) {
 				return finish(false);
 			}
 			return finish(Validate(inst->Arg(1)));
@@ -237,7 +273,7 @@ private:
 			return finish(true);
 		}
 		if (op == ValueOpcode::Phi) {
-			if (m_type == RuntimeValueType::Integer && !ValidateArguments(*inst, false)) {
+			if (m_type != RuntimeValueType::Any && !ValidateArguments(*inst, false)) {
 				return finish(false);
 			}
 			const auto invariant = ResolveInvariantPhi(m_program, value);
@@ -251,7 +287,7 @@ private:
 			    inst->Arg(1).GetType() != Type::U1) {
 				return finish(false);
 			}
-			if (m_type == RuntimeValueType::Integer && !Validate(inst->Arg(1), false)) {
+			if (m_type != RuntimeValueType::Any && !Validate(inst->Arg(1), false)) {
 				return finish(false);
 			}
 			const auto active_mask = m_active_mask;
@@ -959,7 +995,8 @@ bool SrtWalker::RefreshFlatBuffer(std::vector<uint32_t>& flat) {
 		}
 		uint32_t condition = 0;
 		auto& predicate = m_clean_evaluator != nullptr ? *m_clean_evaluator : *this;
-		if (!block.condition.IsEmpty() && m_runtime.read_specialization_memory != nullptr &&
+		if (!block.condition.IsEmpty() &&
+		    (m_runtime.read_specialization_memory != nullptr || !m_program.capture_specialization_reads) &&
 		    predicate.Evaluate(block.condition, condition)) {
 			pending.push_back(block.successors[condition != 0u ? 0u : 1u]);
 		} else {

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

@@ -20,6 +20,20 @@ struct SrtRuntime {
 	std::span<const uint32_t> workgroup_counts;
 };
 
+class SrtReadCapture {
+public:
+	SrtReadCapture(SrtRuntime source, std::vector<std::pair<uint64_t, uint64_t>>& ranges):
+	    m_source(source), m_ranges(ranges) {}
+	SrtRuntime ObservedRuntime();
+
+private:
+	static bool ReadStrict(void* userdata, uint64_t address, std::span<uint32_t> values);
+	static bool ReadOrdinary(void* userdata, uint64_t address, std::span<uint32_t> values);
+
+	SrtRuntime                                  m_source;
+	std::vector<std::pair<uint64_t, uint64_t>>& m_ranges;
+};
+
 // Retained reads no longer need the guest instruction PC. A clean read evaluates
 // its address and bounds through the strict reader as well as the final DWORD.
 struct SrtReadFlags {
@@ -27,7 +41,7 @@ struct SrtReadFlags {
 	uint32_t clean = 0;
 };
 
-enum class RuntimeValueType { Any, Integer };
+enum class RuntimeValueType { Any, Integer, ImmutableInteger };
 
 bool ValidateRuntimeValue(const ResourcePlan& program, Value value,
                           RuntimeValueType type = RuntimeValueType::Any);

src/graphics/shader/shader.cpp

@@ -574,10 +574,12 @@ static void ShaderGetStaticInputInfoPS(
 static void ShaderGetStaticInputInfoCS(const HW::ComputeShaderInfo& regs,
                                        const HW::ShaderRegisters& /*sh*/,
                                        const ShaderMappedData& data, ShaderComputeInputInfo& info) {
+	const bool async_compute             = info.async_compute;
 	const bool dispatch_thread_dimensions = info.dispatch_thread_dimensions;
 	const auto host_subgroup_size         = info.host_subgroup_size;
 	const auto workgroup_counts           = std::to_array(info.workgroup_counts);
 	info                                  = {};
+	info.async_compute                    = async_compute;
 	info.dispatch_thread_dimensions       = dispatch_thread_dimensions;
 	info.host_subgroup_size               = host_subgroup_size;
 	std::ranges::copy(workgroup_counts, info.workgroup_counts);
@@ -693,7 +695,8 @@ void BuildStageStaticKey(const ShaderComputeInputInfo& info, std::vector<uint32_
 	key.push_back(info.thread_ids_num);
 	key.push_back(info.lds_size_dwords);
 	key.push_back(info.scratch_size_dwords);
-	key.push_back(static_cast<uint32_t>(info.dispatch_thread_dimensions));
+	key.push_back(static_cast<uint32_t>(info.dispatch_thread_dimensions) |
+	              (static_cast<uint32_t>(info.async_compute) << 1u));
 	for (int i = 0; i < 3; i++) {
 		key.push_back(info.threads_num[i]);
 		key.push_back(static_cast<uint32_t>(info.group_id[i]));

src/graphics/shader/shader.h

@@ -146,6 +146,7 @@ struct ShaderVertexInputInfo {
 };
 
 struct ShaderComputeInputInfo: ShaderWorkgroupInputInfo {
+	bool               async_compute              = false;
 	uint8_t            float_mode                 = 0xc0;
 	uint32_t           dispatch_threads_num[3]    = {0, 0, 0};
 	uint32_t           workgroup_counts[3]        = {0, 0, 0};

src/launcher/forms/configuration_list_widget.ui

@@ -204,6 +204,72 @@
            </property>
           </spacer>
          </item>
+         <item>
+          <widget class="QSlider" name="grid_size_slider">
+           <property name="maximumSize">
+            <size><width>120</width><height>16777215</height></size>
+           </property>
+           <property name="toolTip">
+            <string>Game image size</string>
+           </property>
+           <property name="accessibleName">
+            <string>Game image size</string>
+           </property>
+           <property name="minimum">
+            <number>128</number>
+           </property>
+           <property name="maximum">
+            <number>320</number>
+           </property>
+           <property name="singleStep">
+            <number>16</number>
+           </property>
+           <property name="pageStep">
+            <number>32</number>
+           </property>
+           <property name="value">
+            <number>224</number>
+           </property>
+           <property name="orientation">
+            <enum>Qt::Horizontal</enum>
+           </property>
+          </widget>
+         </item>
+         <item>
+          <spacer name="gridSizeSpacer">
+           <property name="orientation">
+            <enum>Qt::Horizontal</enum>
+           </property>
+           <property name="sizeType">
+            <enum>QSizePolicy::Fixed</enum>
+           </property>
+           <property name="sizeHint" stdset="0">
+            <size><width>8</width><height>0</height></size>
+           </property>
+          </spacer>
+         </item>
+         <item>
+          <widget class="QToolButton" name="grid_view_button">
+           <property name="minimumSize">
+            <size><width>36</width><height>36</height></size>
+           </property>
+           <property name="text">
+            <string>Grid view</string>
+           </property>
+           <property name="toolTip">
+            <string>Toggle grid view</string>
+           </property>
+           <property name="iconSize">
+            <size><width>32</width><height>32</height></size>
+           </property>
+           <property name="checkable">
+            <bool>true</bool>
+           </property>
+           <property name="autoRaise">
+            <bool>true</bool>
+           </property>
+          </widget>
+         </item>
          <item>
           <widget class="QLineEdit" name="search_line_edit">
            <property name="sizePolicy">
@@ -238,50 +304,53 @@
         </layout>
        </item>
        <item>
-        <widget class="GameListTreeWidget" name="cfgs_list">
-         <property name="rootIsDecorated">
-          <bool>false</bool>
-         </property>
-         <column>
-          <property name="text">
-           <string>Game</string>
+        <widget class="QStackedWidget" name="game_views">
+         <widget class="GameListTreeWidget" name="cfgs_list">
+          <property name="rootIsDecorated">
+           <bool>false</bool>
           </property>
-         </column>
-         <column>
-          <property name="text">
-           <string>Serial</string>
-          </property>
-         </column>
-         <column>
-          <property name="text">
-           <string>Game Version</string>
-          </property>
-         </column>
-         <column>
-          <property name="text">
-           <string>Firmware Version</string>
-          </property>
-         </column>
-         <column>
-          <property name="text">
-           <string>Size</string>
-          </property>
-         </column>
-         <column>
-          <property name="text">
-           <string>Path</string>
-          </property>
-         </column>
-         <column>
-          <property name="text">
-           <string>Status</string>
-          </property>
-         </column>
-         <column>
-          <property name="text">
-           <string>Comments</string>
-          </property>
-         </column>
+          <column>
+           <property name="text">
+            <string>Game</string>
+           </property>
+          </column>
+          <column>
+           <property name="text">
+            <string>Serial</string>
+           </property>
+          </column>
+          <column>
+           <property name="text">
+            <string>Game Version</string>
+           </property>
+          </column>
+          <column>
+           <property name="text">
+            <string>Firmware Version</string>
+           </property>
+          </column>
+          <column>
+           <property name="text">
+            <string>Size</string>
+           </property>
+          </column>
+          <column>
+           <property name="text">
+            <string>Path</string>
+           </property>
+          </column>
+          <column>
+           <property name="text">
+            <string>Status</string>
+           </property>
+          </column>
+          <column>
+           <property name="text">
+            <string>Comments</string>
+           </property>
+          </column>
+         </widget>
+         <widget class="QListView" name="cfgs_grid"/>
         </widget>
        </item>
       </layout>

src/launcher/icons/grid-view.svg

@@ -0,0 +1,8 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32">
+  <g fill="none" stroke="#000000" stroke-width="2" stroke-linejoin="round">
+    <rect x="5" y="5" width="8" height="8" rx="1"/>
+    <rect x="19" y="5" width="8" height="8" rx="1"/>
+    <rect x="5" y="19" width="8" height="8" rx="1"/>
+    <rect x="19" y="19" width="8" height="8" rx="1"/>
+  </g>
+</svg>

src/launcher/include/configurationListWidget.h

@@ -12,9 +12,9 @@
 
 class ConfigurationItem;
 class CompatibilityDatabase;
+class ControllerPreview;
 class QEvent;
 class QTreeWidgetItem;
-class QPoint;
 
 namespace Ui {
 class ConfigurationListWidget;
@@ -29,6 +29,7 @@ public:
 
 	void SetRunEnabled(bool flag) { m_run_enabled = flag; }
 	void SetRuntimeDirectory(const QString& directory) { m_runtime_directory = directory; }
+	void SetControllerPreview(ControllerPreview* preview) { m_controller_preview = preview; }
 
 	[[nodiscard]] const ConfigurationItem* GetSelectedItem() const { return m_selected_item; }
 	ConfigurationItem*                     GetSelectedItem() { return m_selected_item; }
@@ -65,13 +66,13 @@ protected slots:
 	void edit_global_settings();
 	void edit_input_mapping();
 	void list_itemDoubleClicked(QTreeWidgetItem* witem, int column);
-	void show_context_menu(const QPoint& pos);
 	void open_game_folder();
 	void remove_save_data();
 	void filter_configurations(const QString& text);
 
 private:
 	void               SelectItem(QTreeWidgetItem* witem);
+	void               ShowContextMenu(QTreeWidgetItem* witem);
 	void               ApplyCompatibility();
 	void               UpdateToolbarIcons();
 	void               ImportGameSettings(QWidget* parent);
@@ -87,6 +88,7 @@ private:
 	Configuration                 m_global_info;
 	QMap<QString, Configuration*> m_custom_infos;
 	CompatibilityDatabase*        m_compatibility = nullptr;
+	ControllerPreview*            m_controller_preview = nullptr;
 };
 
 #endif // CONFIGURATION_LIST_WIDGET_H

src/launcher/include/controllerLightbar.h

@@ -1,29 +0,0 @@
-#ifndef CONTROLLER_LIGHTBAR_H
-#define CONTROLLER_LIGHTBAR_H
-
-#include <QColor>
-#include <QHash>
-#include <QObject>
-#include <QString>
-#include <QTimer>
-
-#include <SDL3/SDL_gamepad.h>
-
-class ControllerLightbar final: public QObject {
-public:
-	explicit ControllerLightbar(QObject* parent = nullptr);
-	~ControllerLightbar() override;
-
-	void SetColor(const QString& color);
-	void Stop();
-
-private:
-	void Refresh();
-	void ApplyColor(SDL_Gamepad* gamepad) const;
-
-	QTimer                              m_timer;
-	QHash<SDL_JoystickID, SDL_Gamepad*> m_gamepads;
-	QColor                              m_color;
-};
-
-#endif // CONTROLLER_LIGHTBAR_H

src/launcher/include/controllerPreview.h

@@ -0,0 +1,53 @@
+#ifndef CONTROLLER_PREVIEW_H
+#define CONTROLLER_PREVIEW_H
+
+#include <QColor>
+#include <QMap>
+#include <QObject>
+#include <QPointF>
+#include <QSet>
+#include <QString>
+#include <QTimer>
+
+#include <SDL3/SDL_gamepad.h>
+
+class ControllerPreview final: public QObject {
+	Q_OBJECT
+
+public:
+	explicit ControllerPreview(QObject* parent = nullptr);
+	~ControllerPreview() override;
+
+	void SetColor(const QString& color);
+	void SetPreviewEnabled(bool enabled);
+	void Stop();
+	void SetSuspended(bool suspended);
+
+	[[nodiscard]] const QSet<QString>& Pressed() const { return m_pressed; }
+	[[nodiscard]] QPointF              LeftStick() const { return m_left_stick; }
+	[[nodiscard]] QPointF              RightStick() const { return m_right_stick; }
+	[[nodiscard]] const QString&       DeviceName() const { return m_device_name; }
+	[[nodiscard]] const QString&       Error() const { return m_error; }
+
+signals:
+	void StateChanged();
+
+private:
+	void UpdateRunning();
+	void OpenGamepad(SDL_JoystickID id);
+	void Refresh();
+	void ApplyColor(SDL_Gamepad* gamepad) const;
+
+	QTimer                             m_timer;
+	QMap<SDL_JoystickID, SDL_Gamepad*> m_gamepads;
+	QColor                             m_color;
+	bool                               m_preview_enabled = false;
+	bool                               m_suspended       = false;
+	QSet<QString>                      m_pressed;
+	QPointF                            m_left_stick;
+	QPointF                            m_right_stick;
+	QString                            m_device_name;
+	QString                            m_error;
+};
+
+#endif // CONTROLLER_PREVIEW_H

src/launcher/include/dualsenseWidget.h

@@ -0,0 +1,40 @@
+#ifndef LAUNCHER_INCLUDE_DUALSENSE_WIDGET_H_
+#define LAUNCHER_INCLUDE_DUALSENSE_WIDGET_H_
+
+#include <QPointF>
+#include <QSet>
+#include <QString>
+#include <QTransform>
+#include <QWidget>
+
+class DualSenseWidget final: public QWidget {
+	Q_OBJECT
+
+public:
+	explicit DualSenseWidget(QWidget* parent = nullptr);
+	[[nodiscard]] QSize sizeHint() const override;
+	void                SetPressed(const QSet<QString>& pressed);
+	void                SetSelected(const QString& selected);
+	void                SetSticks(const QPointF& left, const QPointF& right);
+
+signals:
+	void ControlClicked(const QString& id);
+
+protected:
+	void paintEvent(QPaintEvent* event) override;
+	void mouseMoveEvent(QMouseEvent* event) override;
+	void mousePressEvent(QMouseEvent* event) override;
+	void leaveEvent(QEvent* event) override;
+
+private:
+	[[nodiscard]] QTransform ViewTransform() const;
+	[[nodiscard]] QString    HitTest(const QPointF& position) const;
+
+	QSet<QString> m_pressed;
+	QString       m_selected;
+	QString       m_hovered;
+	QPointF       m_left_stick;
+	QPointF       m_right_stick;
+};
+
+#endif /* LAUNCHER_INCLUDE_DUALSENSE_WIDGET_H_ */

src/launcher/include/inputMappingDialog.h

@@ -2,32 +2,38 @@
 #define LAUNCHER_INCLUDE_INPUT_MAPPING_DIALOG_H_
 
 #include <QDialog>
+#include <QHash>
 #include <QStringList>
 
+class ControllerPreview;
+class DualSenseWidget;
+class QLabel;
 class QPushButton;
 class QDoubleSpinBox;
-class QTreeWidget;
-class QTreeWidgetItem;
-class QWidget;
 
 class InputMappingDialog final: public QDialog {
 public:
-	explicit InputMappingDialog(const QStringList& mapping, QWidget* parent = nullptr);
+	explicit InputMappingDialog(const QStringList& mapping, ControllerPreview* preview,
+	                            QWidget* parent = nullptr);
+	~InputMappingDialog() override;
 
 	[[nodiscard]] QStringList Mapping() const;
 
 private:
-	void ChangeBinding();
-	void ClearBinding();
+	void ChangeBinding(const QString& id);
 	void RestoreDefaults();
-	void SetBinding(QTreeWidgetItem* item, const QString& binding);
 	void UpdateButtons();
+	void UpdatePreview();
+	void UpdateHighlights();
 
-	QTreeWidget*    m_bindings        = nullptr;
-	QPushButton*    m_change_button   = nullptr;
-	QPushButton*    m_clear_button    = nullptr;
-	QDoubleSpinBox* m_sensitivity     = nullptr;
-	bool            m_custom_bindings = false;
+	ControllerPreview*           m_preview;
+	DualSenseWidget*             m_controller = nullptr;
+	QHash<QString, QString>      m_bindings;
+	QHash<QString, QPushButton*> m_buttons;
+	QString                      m_selected;
+	QLabel*                      m_status          = nullptr;
+	QDoubleSpinBox*              m_sensitivity     = nullptr;
+	bool                         m_custom_bindings = false;
 };
 
 #endif /* LAUNCHER_INCLUDE_INPUT_MAPPING_DIALOG_H_ */

src/launcher/resources.qrc

@@ -4,6 +4,7 @@
   <file alias="global-settings.svg">icons/global-settings.svg</file>
   <file alias="input-mapping.svg">icons/input-mapping.svg</file>
   <file alias="refresh.svg">icons/refresh.svg</file>
+  <file alias="grid-view.svg">icons/grid-view.svg</file>
   <file alias="remove-configuration.svg">icons/remove-configuration.svg</file>
   <file alias="trophy.svg">icons/trophy.svg</file>
   <file alias="hidden-trophy.png">../../assets/images/hidden-trophy.png</file>

src/launcher/src/configurationListWidget.cpp

@@ -34,6 +34,7 @@
 #include <QJsonParseError>
 #include <QKeySequence>
 #include <QLineEdit>
+#include <QListView>
 #include <QMenu>
 #include <QMessageBox>
 #include <QPainter>
@@ -46,7 +47,9 @@
 #include <QSet>
 #include <QSettings>
 #include <QSignalBlocker>
+#include <QSlider>
 #include <QStyle>
+#include <QStyledItemDelegate>
 #include <QToolButton>
 #include <QTreeWidget>
 #include <QUrl>
@@ -141,6 +144,26 @@ static QStringList SettingsStringList(const QVariant& value) {
 	return text.isEmpty() ? QStringList() : QStringList({text});
 }
 
+class GameGridDelegate: public QStyledItemDelegate {
+public:
+	using QStyledItemDelegate::QStyledItemDelegate;
+
+	QSize sizeHint(const QStyleOptionViewItem& option, const QModelIndex&) const override {
+		return option.decorationSize + QSize(16, 24 + option.fontMetrics.lineSpacing() * 2);
+	}
+
+protected:
+	void initStyleOption(QStyleOptionViewItem* option, const QModelIndex& index) const override {
+		const auto image_size = option->decorationSize;
+		QStyledItemDelegate::initStyleOption(option, index);
+		option->decorationSize      = image_size;
+		option->decorationPosition  = QStyleOptionViewItem::Top;
+		option->decorationAlignment = Qt::AlignCenter;
+		option->displayAlignment    = Qt::AlignHCenter | Qt::AlignTop;
+		option->features |= QStyleOptionViewItem::WrapText;
+	}
+};
+
 static void ConfigureGameList(Ui::ConfigurationListWidget* ui) {
 	ui->cfgs_list->setAlternatingRowColors(false);
 	ui->cfgs_list->setAllColumnsShowFocus(true);
@@ -152,6 +175,21 @@ static void ConfigureGameList(Ui::ConfigurationListWidget* ui) {
 	ui->cfgs_list->header()->setDefaultAlignment(Qt::AlignLeft | Qt::AlignVCenter);
 	ui->cfgs_list->header()->setHighlightSections(false);
 	ui->cfgs_list->header()->setStretchLastSection(true);
+
+	// Both presentations use the same items, icons and current game.
+	ui->cfgs_grid->setModel(ui->cfgs_list->model());
+	ui->cfgs_grid->setSelectionModel(ui->cfgs_list->selectionModel());
+	ui->cfgs_grid->setViewMode(QListView::IconMode);
+	ui->cfgs_grid->setMovement(QListView::Static);
+	ui->cfgs_grid->setResizeMode(QListView::Adjust);
+	ui->cfgs_grid->setSelectionMode(QAbstractItemView::SingleSelection);
+	ui->cfgs_grid->setSelectionBehavior(QAbstractItemView::SelectRows);
+	ui->cfgs_grid->setEditTriggers(QAbstractItemView::NoEditTriggers);
+	ui->cfgs_grid->setSpacing(8);
+	ui->cfgs_grid->setUniformItemSizes(true);
+	ui->cfgs_grid->setWordWrap(true);
+	ui->cfgs_grid->setItemDelegate(new GameGridDelegate(ui->cfgs_grid));
+	ui->cfgs_grid->setContextMenuPolicy(Qt::CustomContextMenu);
 }
 
 static void AddSaveDataDir(QStringList* dirs, QSet<QString>* seen, const QString& root,
@@ -252,7 +290,27 @@ ConfigurationListWidget::ConfigurationListWidget(QWidget* parent)
 	connect(m_ui->cfgs_list, &QTreeWidget::itemDoubleClicked, this,
 	        &ConfigurationListWidget::list_itemDoubleClicked);
 	connect(m_ui->cfgs_list, &QTreeWidget::customContextMenuRequested, this,
-	        &ConfigurationListWidget::show_context_menu);
+	        [this](const QPoint& pos) { ShowContextMenu(m_ui->cfgs_list->itemAt(pos)); });
+	connect(m_ui->cfgs_grid, &QListView::doubleClicked, this, [this](const QModelIndex& index) {
+		list_itemDoubleClicked(m_ui->cfgs_list->itemFromIndex(index), GAME_NAME_COLUMN);
+	});
+	connect(m_ui->cfgs_grid, &QListView::customContextMenuRequested, this,
+	        [this](const QPoint& pos) {
+		        ShowContextMenu(m_ui->cfgs_list->itemFromIndex(m_ui->cfgs_grid->indexAt(pos)));
+	        });
+	connect(m_ui->grid_view_button, &QToolButton::toggled, this, [this](bool grid) {
+		m_ui->grid_size_slider->setVisible(grid);
+		auto* view = grid ? static_cast<QAbstractItemView*>(m_ui->cfgs_grid) : m_ui->cfgs_list;
+		m_ui->game_views->setCurrentWidget(view);
+		view->setCurrentIndex(view->currentIndex().siblingAtColumn(GAME_NAME_COLUMN));
+		view->scrollTo(view->currentIndex());
+		view->setFocus();
+	});
+	connect(m_ui->grid_view_button, &QToolButton::clicked, this,
+	        &ConfigurationListWidget::WriteSettings);
+	connect(m_ui->cfgs_list->model(), &QAbstractItemModel::layoutChanged, this, [this]() {
+		filter_configurations(m_ui->search_line_edit->text());
+	});
 	connect(m_ui->search_line_edit, &QLineEdit::textChanged, this,
 	        &ConfigurationListWidget::filter_configurations);
 	connect(m_compatibility, &CompatibilityDatabase::Updated, this,
@@ -261,6 +319,19 @@ ConfigurationListWidget::ConfigurationListWidget(QWidget* parent)
 	m_ui->cfgs_list->setDragDropMode(QAbstractItemView::NoDragDrop);
 
 	ReadSettings();
+	m_ui->grid_size_slider->setVisible(m_ui->grid_view_button->isChecked());
+	const auto resize_grid = [this](int width) {
+		m_ui->cfgs_grid->setIconSize(QSize(width, width * 9 / 16));
+	};
+	resize_grid(m_ui->grid_size_slider->value());
+	connect(m_ui->grid_size_slider, &QSlider::valueChanged, this, [this, resize_grid](int width) {
+		resize_grid(width);
+		if (!m_ui->grid_size_slider->isSliderDown()) {
+			WriteSettings();
+		}
+	});
+	connect(m_ui->grid_size_slider, &QSlider::sliderReleased, this,
+	        &ConfigurationListWidget::WriteSettings);
 	if (m_compatibility->IsLocal()) {
 		m_compatibility->Load();
 	}
@@ -300,6 +371,7 @@ void ConfigurationListWidget::UpdateToolbarIcons() {
 	set_icon(m_ui->edit_button, QStringLiteral(":/icons/edit-configuration.svg"));
 	set_icon(m_ui->delete_button, QStringLiteral(":/icons/remove-configuration.svg"));
 	set_icon(m_ui->trophy_overview_button, QStringLiteral(":/icons/trophy.svg"));
+	set_icon(m_ui->grid_view_button, QStringLiteral(":/icons/grid-view.svg"));
 }
 
 void ConfigurationListWidget::WriteSettings() {
@@ -323,6 +395,8 @@ void ConfigurationListWidget::WriteSettings() {
 	s->beginGroup(CONF_LAUNCHER);
 	m_game_dirs = NormalizeGameDirectories(m_game_dirs);
 	s->setValue(CONF_GAME_DIRS, m_game_dirs);
+	s->setValue("grid_view", m_ui->grid_view_button->isChecked());
+	s->setValue("grid_icon_width", m_ui->grid_size_slider->value());
 	s->remove(CONF_GAME_DIR);
 	s->endGroup();
 
@@ -363,6 +437,9 @@ void ConfigurationListWidget::ReadSettings() {
 	ConfigurationEditDialog::ReadSettings(*s);
 
 	s->beginGroup(CONF_LAUNCHER);
+	m_ui->grid_view_button->setChecked(s->value("grid_view", false).toBool());
+	m_ui->grid_size_slider->setValue(
+	    s->value("grid_icon_width", m_ui->grid_size_slider->value()).toInt());
 	m_game_dirs = NormalizeGameDirectories(SettingsStringList(s->value(CONF_GAME_DIRS)));
 	if (m_game_dirs.isEmpty()) {
 		m_game_dirs = NormalizeGameDirectories(SettingsStringList(s->value(CONF_GAME_DIR)));
@@ -959,7 +1036,7 @@ void ConfigurationListWidget::ExportGameSettings(QWidget* parent) const {
 }
 
 void ConfigurationListWidget::edit_input_mapping() {
-	InputMappingDialog dialog(m_global_info.host_input_mapping, this);
+	InputMappingDialog dialog(m_global_info.host_input_mapping, m_controller_preview, this);
 	if (dialog.exec() == QDialog::Accepted) {
 		m_global_info.host_input_mapping = dialog.Mapping();
 		WriteSettings();
@@ -1097,6 +1174,7 @@ void ConfigurationListWidget::filter_configurations(const QString& text) {
 		                   item->text(GAME_NAME_COLUMN).contains(query, Qt::CaseInsensitive) ||
 		                   item->text(GAME_SERIAL_COLUMN).contains(query, Qt::CaseInsensitive);
 		item->setHidden(!match);
+		m_ui->cfgs_grid->setRowHidden(item_index, !match);
 
 		if (match && item == m_selected_item) {
 			selection_is_shown = true;
@@ -1137,8 +1215,8 @@ void ConfigurationListWidget::list_itemDoubleClicked(QTreeWidgetItem* witem, int
 	}
 }
 
-void ConfigurationListWidget::show_context_menu(const QPoint& pos) {
-	auto* item = static_cast<ConfigurationItem*>(m_ui->cfgs_list->itemAt(pos));
+void ConfigurationListWidget::ShowContextMenu(QTreeWidgetItem* witem) {
+	auto* item = static_cast<ConfigurationItem*>(witem);
 
 	if (item != nullptr) {
 		m_ui->cfgs_list->setCurrentItem(item);

src/launcher/src/controllerLightbar.cpp

@@ -1,84 +0,0 @@
-#include "controllerLightbar.h"
-
-#include <QDebug>
-
-#include <SDL3/SDL.h>
-
-ControllerLightbar::ControllerLightbar(QObject* parent): QObject(parent) {
-	m_timer.setInterval(100);
-	connect(&m_timer, &QTimer::timeout, this, &ControllerLightbar::Refresh);
-}
-
-ControllerLightbar::~ControllerLightbar() {
-	Stop();
-}
-
-void ControllerLightbar::SetColor(const QString& color) {
-	const QColor parsed(color);
-	if (!parsed.isValid()) {
-		Stop();
-		return;
-	}
-	if (m_color == parsed) {
-		return;
-	}
-	if (!m_timer.isActive()) {
-		if (!SDL_InitSubSystem(SDL_INIT_GAMEPAD)) {
-			qWarning() << "Cannot initialize controller color preview:" << SDL_GetError();
-			return;
-		}
-		m_timer.start();
-	}
-	m_color = parsed;
-	for (auto* gamepad: m_gamepads) {
-		ApplyColor(gamepad);
-	}
-	Refresh();
-}
-
-void ControllerLightbar::Stop() {
-	if (!m_timer.isActive()) {
-		return;
-	}
-	m_timer.stop();
-	for (auto* gamepad: m_gamepads) {
-		SDL_CloseGamepad(gamepad);
-	}
-	m_gamepads.clear();
-	SDL_QuitSubSystem(SDL_INIT_GAMEPAD);
-	m_color = {};
-}
-
-void ControllerLightbar::Refresh() {
-	// Bluetooth DualSense may defer an accepted LED command until its connection
-	// animation ends. Polling events also delivers those deferred commands.
-	SDL_Event event;
-	while (SDL_PollEvent(&event)) {
-		if (event.type == SDL_EVENT_GAMEPAD_ADDED && !m_gamepads.contains(event.gdevice.which)) {
-			auto* gamepad = SDL_OpenGamepad(event.gdevice.which);
-			if (gamepad == nullptr) {
-				qWarning() << "Cannot open controller color preview:" << SDL_GetError();
-				continue;
-			}
-			if (SDL_GetBooleanProperty(SDL_GetGamepadProperties(gamepad),
-			                           SDL_PROP_GAMEPAD_CAP_RGB_LED_BOOLEAN, false)) {
-				m_gamepads.insert(event.gdevice.which, gamepad);
-				ApplyColor(gamepad);
-			} else {
-				SDL_CloseGamepad(gamepad);
-			}
-		} else if (event.type == SDL_EVENT_GAMEPAD_REMOVED) {
-			if (auto* gamepad = m_gamepads.take(event.gdevice.which); gamepad != nullptr) {
-				SDL_CloseGamepad(gamepad);
-			}
-		}
-	}
-}
-
-void ControllerLightbar::ApplyColor(SDL_Gamepad* gamepad) const {
-	if (!SDL_SetGamepadLED(gamepad, static_cast<Uint8>(m_color.red()),
-	                       static_cast<Uint8>(m_color.green()),
-	                       static_cast<Uint8>(m_color.blue()))) {
-		qWarning() << "Cannot set controller LED:" << SDL_GetError();
-	}
-}

src/launcher/src/controllerPreview.cpp

@@ -0,0 +1,202 @@
+#include "controllerPreview.h"
+
+#include <QDebug>
+
+#include <SDL3/SDL.h>
+#include <cmath>
+
+namespace {
+
+constexpr struct {
+	SDL_GamepadButton button;
+	const char*       control;
+} BUTTONS[] = {
+    {SDL_GAMEPAD_BUTTON_SOUTH, "Cross"},      {SDL_GAMEPAD_BUTTON_EAST, "Circle"},
+    {SDL_GAMEPAD_BUTTON_WEST, "Square"},      {SDL_GAMEPAD_BUTTON_NORTH, "Triangle"},
+    {SDL_GAMEPAD_BUTTON_DPAD_UP, "Up"},       {SDL_GAMEPAD_BUTTON_DPAD_DOWN, "Down"},
+    {SDL_GAMEPAD_BUTTON_DPAD_LEFT, "Left"},   {SDL_GAMEPAD_BUTTON_DPAD_RIGHT, "Right"},
+    {SDL_GAMEPAD_BUTTON_LEFT_SHOULDER, "L1"}, {SDL_GAMEPAD_BUTTON_RIGHT_SHOULDER, "R1"},
+    {SDL_GAMEPAD_BUTTON_LEFT_STICK, "L3"},    {SDL_GAMEPAD_BUTTON_RIGHT_STICK, "R3"},
+    {SDL_GAMEPAD_BUTTON_START, "Options"},
+};
+
+double Axis(SDL_Gamepad* gamepad, SDL_GamepadAxis axis) {
+	const int value = SDL_GetGamepadAxis(gamepad, axis);
+	return value / (value < 0 ? 32768.0 : 32767.0);
+}
+
+} // namespace
+
+ControllerPreview::ControllerPreview(QObject* parent): QObject(parent) {
+	m_timer.setInterval(100);
+	connect(&m_timer, &QTimer::timeout, this, &ControllerPreview::Refresh);
+}
+
+ControllerPreview::~ControllerPreview() {
+	Stop();
+}
+
+void ControllerPreview::SetColor(const QString& color) {
+	const QColor parsed(color);
+	if (m_color == parsed && m_timer.isActive()) {
+		return;
+	}
+	m_color = parsed;
+	UpdateRunning();
+	for (auto* gamepad: m_gamepads) {
+		ApplyColor(gamepad);
+	}
+}
+
+void ControllerPreview::SetPreviewEnabled(bool enabled) {
+	m_preview_enabled = enabled;
+	m_timer.setInterval(enabled ? 16 : 100);
+	UpdateRunning();
+}
+
+void ControllerPreview::Stop() {
+	m_preview_enabled = false;
+	m_color           = {};
+	UpdateRunning();
+}
+
+void ControllerPreview::SetSuspended(bool suspended) {
+	if (m_suspended != suspended) {
+		m_suspended = suspended;
+		UpdateRunning();
+	}
+}
+
+void ControllerPreview::UpdateRunning() {
+	if (!m_suspended && (m_preview_enabled || m_color.isValid())) {
+		if (m_timer.isActive()) {
+			return;
+		}
+		SDL_SetHint(SDL_HINT_JOYSTICK_ALLOW_BACKGROUND_EVENTS, "1");
+		if (!SDL_InitSubSystem(SDL_INIT_GAMEPAD)) {
+			m_error = QString::fromUtf8(SDL_GetError());
+			qWarning() << "Cannot initialize controller preview:" << m_error;
+			emit StateChanged();
+			return;
+		}
+		m_error.clear();
+		int   count = 0;
+		auto* ids   = SDL_GetGamepads(&count);
+		for (int index = 0; index < count; ++index) {
+			OpenGamepad(ids[index]);
+		}
+		SDL_free(ids);
+		m_timer.start();
+		Refresh();
+		emit StateChanged();
+		return;
+	}
+	if (m_timer.isActive()) {
+		m_timer.stop();
+		for (auto* gamepad: m_gamepads) {
+			SDL_CloseGamepad(gamepad);
+		}
+		m_gamepads.clear();
+		SDL_QuitSubSystem(SDL_INIT_GAMEPAD);
+	}
+	m_pressed.clear();
+	m_left_stick  = {};
+	m_right_stick = {};
+	m_device_name.clear();
+	m_error = m_suspended ? tr("Preview is paused while a game is running.") : QString();
+	emit StateChanged();
+}
+
+void ControllerPreview::OpenGamepad(SDL_JoystickID id) {
+	if (!m_gamepads.contains(id)) {
+		if (auto* gamepad = SDL_OpenGamepad(id); gamepad != nullptr) {
+			m_gamepads.insert(id, gamepad);
+			ApplyColor(gamepad);
+		}
+	}
+}
+
+void ControllerPreview::Refresh() {
+	// One SDL event pump serves both the live diagram and lightbar preview.
+	// It also delivers deferred Bluetooth DualSense LED commands.
+	SDL_Event event;
+	while (SDL_PollEvent(&event)) {
+		if (event.type == SDL_EVENT_GAMEPAD_ADDED) {
+			OpenGamepad(event.gdevice.which);
+		} else if (event.type == SDL_EVENT_GAMEPAD_REMOVED) {
+			if (auto* gamepad = m_gamepads.take(event.gdevice.which); gamepad != nullptr) {
+				SDL_CloseGamepad(gamepad);
+			}
+		}
+	}
+
+	QSet<QString> pressed;
+	QPointF       left;
+	QPointF       right;
+	QString       name;
+	if (!m_gamepads.isEmpty()) {
+		auto* gamepad = m_gamepads.first();
+		name          = QString::fromUtf8(SDL_GetGamepadName(gamepad));
+		for (const auto& binding: BUTTONS) {
+			if (SDL_GetGamepadButton(gamepad, binding.button)) {
+				pressed.insert(QString::fromLatin1(binding.control));
+			}
+		}
+		const auto stick = [&](SDL_GamepadAxis x, SDL_GamepadAxis y, const char* prefix) {
+			QPointF position(Axis(gamepad, x), Axis(gamepad, y));
+			if (std::hypot(position.x(), position.y()) < 0.15) {
+				position = {};
+			}
+			const auto id = QString::fromLatin1(prefix);
+			if (position.x() < -0.25) pressed.insert(id + QStringLiteral("Left"));
+			if (position.x() > 0.25) pressed.insert(id + QStringLiteral("Right"));
+			if (position.y() < -0.25) pressed.insert(id + QStringLiteral("Up"));
+			if (position.y() > 0.25) pressed.insert(id + QStringLiteral("Down"));
+			return position;
+		};
+		left  = stick(SDL_GAMEPAD_AXIS_LEFTX, SDL_GAMEPAD_AXIS_LEFTY, "LeftStick");
+		right = stick(SDL_GAMEPAD_AXIS_RIGHTX, SDL_GAMEPAD_AXIS_RIGHTY, "RightStick");
+		if (Axis(gamepad, SDL_GAMEPAD_AXIS_LEFT_TRIGGER) > 0.05) {
+			pressed.insert(QStringLiteral("L2"));
+		}
+		if (Axis(gamepad, SDL_GAMEPAD_AXIS_RIGHT_TRIGGER) > 0.05) {
+			pressed.insert(QStringLiteral("R2"));
+		}
+		bool touching = false;
+		if (SDL_GetNumGamepadTouchpads(gamepad) > 0) {
+			for (int finger = 0; finger < SDL_GetNumGamepadTouchpadFingers(gamepad, 0); ++finger) {
+				bool  down = false;
+				float x    = 0;
+				if (SDL_GetGamepadTouchpadFinger(gamepad, 0, finger, &down, &x, nullptr, nullptr) &&
+				    down) {
+					pressed.insert(x < 0.5f ? QStringLiteral("TouchPad")
+					                        : QStringLiteral("TouchPadRight"));
+					touching = true;
+				}
+			}
+		}
+		if (!touching && SDL_GetGamepadButton(gamepad, SDL_GAMEPAD_BUTTON_TOUCHPAD)) {
+			pressed.insert(QStringLiteral("TouchPad"));
+			pressed.insert(QStringLiteral("TouchPadRight"));
+		}
+	}
+	if (pressed != m_pressed || left != m_left_stick || right != m_right_stick ||
+	    name != m_device_name) {
+		m_pressed     = pressed;
+		m_left_stick  = left;
+		m_right_stick = right;
+		m_device_name = name;
+		emit StateChanged();
+	}
+}
+
+void ControllerPreview::ApplyColor(SDL_Gamepad* gamepad) const {
+	if (m_color.isValid() &&
+	    SDL_GetBooleanProperty(SDL_GetGamepadProperties(gamepad),
+	                           SDL_PROP_GAMEPAD_CAP_RGB_LED_BOOLEAN, false) &&
+	    !SDL_SetGamepadLED(gamepad, static_cast<Uint8>(m_color.red()),
+	                       static_cast<Uint8>(m_color.green()),
+	                       static_cast<Uint8>(m_color.blue()))) {
+		qWarning() << "Cannot set controller LED:" << SDL_GetError();
+	}
+}

src/launcher/src/dualsenseWidget.cpp

@@ -0,0 +1,300 @@
+#include "dualsenseWidget.h"
+
+#include <QMouseEvent>
+#include <QPainter>
+#include <QPainterPath>
+
+#include <algorithm>
+#include <cmath>
+
+namespace {
+
+enum class Shape { Button, Shoulder, Touchpad, Face, Dpad, Stick, Direction };
+
+struct Control {
+	const char* id;
+	QRectF      bounds;
+	Shape       shape = Shape::Button;
+	int         angle = 0;
+};
+
+// Painting and hit testing share a canvas with room above the shell for the shoulders.
+const QRectF CANVAS(0, -20, 480, 330);
+
+const Control CONTROLS[] = {
+    {"L2", {90, -12, 58, 18}, Shape::Shoulder},
+    {"R2", {332, -12, 58, 18}, Shape::Shoulder},
+    {"L1", {84, 12, 64, 12}, Shape::Shoulder},
+    {"R1", {332, 12, 64, 12}, Shape::Shoulder},
+    {"TouchPad", {150, 25, 90, 100}, Shape::Touchpad},
+    {"TouchPadRight", {240, 25, 90, 100}, Shape::Touchpad},
+    {"Options", {330, 60, 9, 20}},
+    {"Up", {105, 72, 22, 25}, Shape::Dpad, 0},
+    {"Right", {124, 94, 25, 22}, Shape::Dpad, 90},
+    {"Down", {105, 113, 22, 25}, Shape::Dpad, 180},
+    {"Left", {83, 94, 25, 22}, Shape::Dpad, 270},
+    {"Triangle", {353, 67, 26, 26}, Shape::Face},
+    {"Circle", {382, 96, 26, 26}, Shape::Face},
+    {"Cross", {353, 125, 26, 26}, Shape::Face},
+    {"Square", {324, 96, 26, 26}, Shape::Face},
+    {"L3", {143, 131, 66, 66}, Shape::Stick},
+    {"R3", {271, 131, 66, 66}, Shape::Stick},
+    {"LeftStickUp", {143, 131, 66, 66}, Shape::Direction, 0},
+    {"LeftStickRight", {143, 131, 66, 66}, Shape::Direction, 90},
+    {"LeftStickDown", {143, 131, 66, 66}, Shape::Direction, 180},
+    {"LeftStickLeft", {143, 131, 66, 66}, Shape::Direction, 270},
+    {"RightStickUp", {271, 131, 66, 66}, Shape::Direction, 0},
+    {"RightStickRight", {271, 131, 66, 66}, Shape::Direction, 90},
+    {"RightStickDown", {271, 131, 66, 66}, Shape::Direction, 180},
+    {"RightStickLeft", {271, 131, 66, 66}, Shape::Direction, 270},
+};
+
+QPainterPath TouchpadPath() {
+	QPainterPath path;
+	path.moveTo(158, 42);
+	path.cubicTo(156, 33, 168, 32, 178, 31);
+	path.cubicTo(219, 28, 261, 28, 302, 31);
+	path.cubicTo(312, 32, 324, 33, 322, 42);
+	path.cubicTo(318, 64, 315, 87, 309, 105);
+	path.cubicTo(306, 117, 299, 122, 286, 122);
+	path.lineTo(194, 122);
+	path.cubicTo(181, 122, 174, 117, 171, 105);
+	path.cubicTo(165, 87, 162, 64, 158, 42);
+	path.closeSubpath();
+	return path;
+}
+
+QPainterPath ControlPath(const Control& control) {
+	QPainterPath path;
+	if (control.shape == Shape::Direction) {
+		const auto inner = control.bounds.adjusted(7, 7, -7, -7);
+		const int  start = 49 - control.angle;
+		path.arcMoveTo(control.bounds, start);
+		path.arcTo(control.bounds, start, 82);
+		path.arcTo(inner, start + 82, -82);
+		path.closeSubpath();
+	} else if (control.shape == Shape::Touchpad) {
+		path.addRect(control.bounds);
+		return TouchpadPath().intersected(path);
+	} else if (control.shape == Shape::Face || control.shape == Shape::Stick) {
+		path.addEllipse(control.shape == Shape::Stick ? control.bounds.adjusted(7, 7, -7, -7)
+		                                             : control.bounds);
+	} else if (control.shape == Shape::Dpad) {
+		path.moveTo(-8, -12);
+		path.quadTo(-11, -12, -11, -8);
+		path.lineTo(-10, 3);
+		path.quadTo(-10, 6, -7, 8);
+		path.lineTo(0, 12);
+		path.lineTo(7, 8);
+		path.quadTo(10, 6, 10, 3);
+		path.lineTo(11, -8);
+		path.quadTo(11, -12, 8, -12);
+		path.closeSubpath();
+		QTransform transform;
+		transform.translate(control.bounds.center().x(), control.bounds.center().y());
+		transform.rotate(control.angle);
+		return transform.map(path);
+	} else if (control.shape == Shape::Shoulder) {
+		const auto& bounds = control.bounds;
+		path.moveTo(bounds.bottomLeft());
+		path.lineTo(bounds.left() + 2, bounds.top() + bounds.height() / 2);
+		path.cubicTo(bounds.left() + 10, bounds.top(), bounds.right() - 10, bounds.top(),
+		             bounds.right() - 2, bounds.top() + bounds.height() / 2);
+		path.lineTo(bounds.bottomRight());
+		path.closeSubpath();
+	} else {
+		path.addRoundedRect(control.bounds, 4, 4);
+	}
+	return path;
+}
+
+void DrawSymbol(QPainter& painter, const Control& control) {
+	const auto id = QLatin1String(control.id);
+	painter.save();
+	painter.translate(control.bounds.center());
+	painter.setBrush(Qt::NoBrush);
+	if (control.shape == Shape::Dpad) {
+		painter.rotate(control.angle);
+		painter.drawPolyline(QPolygonF({{-4, 2}, {0, -2}, {4, 2}}));
+	} else if (id == QLatin1String("Triangle")) {
+		painter.drawPolygon(QPolygonF({{0, -7}, {7, 5}, {-7, 5}}));
+	} else if (id == QLatin1String("Circle")) {
+		painter.drawEllipse(QPointF(), 7, 7);
+	} else if (id == QLatin1String("Square")) {
+		painter.drawRect(QRectF(-6, -6, 12, 12));
+	} else if (id == QLatin1String("Cross")) {
+		painter.drawLine(QPointF(-5, -5), QPointF(5, 5));
+		painter.drawLine(QPointF(-5, 5), QPointF(5, -5));
+	}
+	painter.restore();
+}
+
+} // namespace
+
+DualSenseWidget::DualSenseWidget(QWidget* parent): QWidget(parent) {
+	setMinimumSize(330, 230);
+	setMouseTracking(true);
+	setSizePolicy(QSizePolicy::Expanding, QSizePolicy::Expanding);
+	setAccessibleName(tr("DualSense controller"));
+	setAccessibleDescription(tr("Click a control to select its binding. Connected controller input "
+	                            "lights up the controls."));
+}
+
+QSize DualSenseWidget::sizeHint() const {
+	return CANVAS.size().toSize();
+}
+
+void DualSenseWidget::SetPressed(const QSet<QString>& pressed) {
+	if (m_pressed != pressed) {
+		m_pressed = pressed;
+		update();
+	}
+}
+
+void DualSenseWidget::SetSelected(const QString& selected) {
+	if (m_selected != selected) {
+		m_selected = selected;
+		update();
+	}
+}
+
+void DualSenseWidget::SetSticks(const QPointF& left, const QPointF& right) {
+	if (m_left_stick != left || m_right_stick != right) {
+		m_left_stick  = left;
+		m_right_stick = right;
+		update();
+	}
+}
+
+QTransform DualSenseWidget::ViewTransform() const {
+	const qreal scale = std::min(width() / CANVAS.width(), height() / CANVAS.height());
+	QTransform  transform;
+	transform.translate(width() / 2.0, height() / 2.0);
+	transform.scale(scale, scale);
+	transform.translate(-CANVAS.center().x(), -CANVAS.center().y());
+	return transform;
+}
+
+QString DualSenseWidget::HitTest(const QPointF& position) const {
+	const auto point = ViewTransform().inverted().map(position);
+	for (const auto& control: CONTROLS) {
+		if (ControlPath(control).contains(point)) {
+			return QString::fromLatin1(control.id);
+		}
+	}
+	return {};
+}
+
+void DualSenseWidget::paintEvent(QPaintEvent*) {
+	const QColor black("#111111");
+	const QColor white("#eeeeee");
+	QPainter     painter(this);
+	painter.setRenderHint(QPainter::Antialiasing);
+	painter.setTransform(ViewTransform());
+
+	QPainterPath shell;
+	shell.moveTo(81, 48);
+	shell.cubicTo(151, 21, 329, 21, 399, 48);
+	shell.cubicTo(420, 94, 445, 166, 445, 215);
+	shell.cubicTo(447, 253, 437, 291, 419, 296);
+	shell.cubicTo(403, 302, 395, 297, 389, 283);
+	shell.cubicTo(381, 268, 373, 236, 354, 213);
+	shell.cubicTo(347, 205, 341, 205, 325, 205);
+	shell.lineTo(155, 205);
+	shell.cubicTo(139, 205, 133, 205, 126, 213);
+	shell.cubicTo(107, 236, 99, 268, 91, 283);
+	shell.cubicTo(85, 297, 77, 302, 61, 296);
+	shell.cubicTo(43, 291, 33, 253, 35, 215);
+	shell.cubicTo(35, 166, 60, 94, 81, 48);
+	shell.closeSubpath();
+	painter.setBrush(black);
+	painter.setPen(QPen(white, 1.5));
+	painter.drawPath(shell);
+
+	// Panel seams flow into the inner grips without a second outline across the belly.
+	QPainterPath seams;
+	seams.moveTo(153, 35);
+	seams.cubicTo(153, 69, 176, 105, 158, 127);
+	seams.cubicTo(123, 166, 108, 226, 76, 297);
+	seams.moveTo(327, 35);
+	seams.cubicTo(327, 69, 304, 105, 322, 127);
+	seams.cubicTo(357, 166, 372, 226, 404, 297);
+	painter.setBrush(Qt::NoBrush);
+	painter.drawPath(seams);
+	painter.setBrush(black);
+	painter.drawPath(TouchpadPath());
+	painter.drawRoundedRect(QRectF(141, 60, 9, 20), 4, 4); // Create button, currently unmapped.
+	painter.drawEllipse(QPointF(240, 166), 8, 8);
+	QFont font = painter.font();
+	font.setPixelSize(9);
+	font.setBold(true);
+	painter.setFont(font);
+
+	for (const auto& control: CONTROLS) {
+		const auto id       = QString::fromLatin1(control.id);
+		const bool pressed  = m_pressed.contains(id);
+		const bool selected = m_selected == id;
+		const bool hovered  = m_hovered == id;
+		const bool active   = pressed || selected || hovered;
+		if ((control.shape == Shape::Direction || control.shape == Shape::Touchpad) && !active) {
+			continue;
+		}
+		const QColor fill = pressed ? QColor("#2eb6df") : active ? QColor("#3a6179") : black;
+		painter.setBrush(fill);
+		painter.setPen(QPen(active ? QColor("#79e3ff") : white, selected ? 2 : 1.3));
+		if (control.shape == Shape::Touchpad) {
+			// Clip one shared surface so independent bindings never add a permanent center seam.
+			painter.save();
+			painter.setClipRect(control.bounds);
+			painter.drawPath(TouchpadPath());
+			painter.restore();
+		} else if (control.shape == Shape::Stick) {
+			painter.setBrush(black);
+			painter.drawEllipse(control.bounds);
+			const auto    position = id == QLatin1String("L3") ? m_left_stick : m_right_stick;
+			QPointF       offset(std::clamp(position.x(), -1.0, 1.0),
+			                     std::clamp(position.y(), -1.0, 1.0));
+			offset *= 5 / std::max(1.0, std::hypot(offset.x(), offset.y()));
+			painter.setBrush(fill);
+			painter.drawEllipse(control.bounds.adjusted(12, 12, -12, -12).translated(offset));
+		} else {
+			painter.drawPath(ControlPath(control));
+			painter.setPen(QPen(white, 1.3));
+			if (control.shape == Shape::Face || control.shape == Shape::Dpad) {
+				DrawSymbol(painter, control);
+			} else if (control.shape == Shape::Shoulder) {
+				painter.drawText(control.bounds.adjusted(0, 2, 0, 0), Qt::AlignCenter, id);
+			} else if (id == QLatin1String("Options")) {
+				painter.drawLine(control.bounds.center() - QPointF(0, 4),
+				                 control.bounds.center() + QPointF(0, 4));
+			}
+		}
+	}
+}
+
+void DualSenseWidget::mouseMoveEvent(QMouseEvent* event) {
+	const auto hovered = HitTest(event->position());
+	if (hovered != m_hovered) {
+		m_hovered = hovered;
+		setCursor(hovered.isEmpty() ? Qt::ArrowCursor : Qt::PointingHandCursor);
+		setToolTip(hovered);
+		update();
+	}
+}
+
+void DualSenseWidget::mousePressEvent(QMouseEvent* event) {
+	if (event->button() == Qt::LeftButton) {
+		const auto id = HitTest(event->position());
+		if (!id.isEmpty()) {
+			emit ControlClicked(id);
+			return;
+		}
+	}
+	QWidget::mousePressEvent(event);
+}
+
+void DualSenseWidget::leaveEvent(QEvent*) {
+	m_hovered.clear();
+	unsetCursor();
+	update();
+}

src/launcher/src/inputMappingDialog.cpp

@@ -1,61 +1,69 @@
 #include "inputMappingDialog.h"
 
-#include <QAbstractItemView>
+#include "controllerPreview.h"
+#include "dualsenseWidget.h"
+
 #include <QDialogButtonBox>
 #include <QDoubleSpinBox>
+#include <QGridLayout>
+#include <QGroupBox>
 #include <QHBoxLayout>
 #include <QHash>
-#include <QHeaderView>
 #include <QKeyEvent>
 #include <QKeySequence>
 #include <QLabel>
 #include <QMouseEvent>
 #include <QPushButton>
+#include <QScreen>
+#include <QScrollArea>
 #include <QStringList>
-#include <QTreeWidget>
 #include <QVBoxLayout>
 
 namespace {
 
-constexpr int  BINDING_COLUMN            = 1;
 constexpr auto DEFAULT_MOUSE_SENSITIVITY = 1.0;
 constexpr char MOUSE_SENSITIVITY[]       = "MouseSensitivity=";
 
+enum Group { Dpad, LeftStick, RightStick, Face, Shoulders, Center, Shortcuts, GroupCount };
+
 struct PadControl {
 	const char* id;
 	const char* label;
 	const char* default_binding;
+	Group       group;
+	int         row;
+	int         column;
 };
 
 constexpr PadControl PAD_CONTROLS[] = {
-    {"Up", "D-pad Up", "Up"},
-    {"Down", "D-pad Down", "Down"},
-    {"Left", "D-pad Left", "Left"},
-    {"Right", "D-pad Right", "Right"},
-    {"LeftStickUp", "Left stick Up", "W"},
-    {"LeftStickDown", "Left stick Down", "S"},
-    {"LeftStickLeft", "Left stick Left", "A"},
-    {"LeftStickRight", "Left stick Right", "D"},
-    {"RightStickUp", "Right stick Up", "T"},
-    {"RightStickDown", "Right stick Down", "G"},
-    {"RightStickLeft", "Right stick Left", "F"},
-    {"RightStickRight", "Right stick Right", "H"},
-    {"Triangle", "Triangle", "I"},
-    {"Circle", "Circle", "L"},
-    {"Cross", "Cross", "J"},
-    {"Square", "Square", "K"},
-    {"L1", "L1", "Q"},
-    {"R1", "R1", "E"},
-    {"L2", "L2", "Z"},
-    {"R2", "R2", "C"},
-    {"L3", "L3", "Left Shift"},
-    {"R3", "R3", "Left Ctrl"},
-    {"Options", "Options", "Return"},
-    {"TouchPad", "Touch pad left (SELECT)", "Backspace"},
-    {"TouchPadRight", "Touch pad right (START)", "Tab"},
-    {"SpeakerVolume", "Speaker volume (cycle)", "1"},
-    {"VibrationIntensity", "Vibration intensity (cycle)", "2"},
-    {"TriggerEffectIntensity", "Trigger effect intensity (cycle)", "3"},
+    {"Up", "Up", "Up", Dpad, 0, 1},
+    {"Down", "Down", "Down", Dpad, 2, 1},
+    {"Left", "Left", "Left", Dpad, 1, 0},
+    {"Right", "Right", "Right", Dpad, 1, 2},
+    {"LeftStickUp", "Up", "W", LeftStick, 0, 1},
+    {"LeftStickDown", "Down", "S", LeftStick, 2, 1},
+    {"LeftStickLeft", "Left", "A", LeftStick, 1, 0},
+    {"LeftStickRight", "Right", "D", LeftStick, 1, 2},
+    {"RightStickUp", "Up", "T", RightStick, 0, 1},
+    {"RightStickDown", "Down", "G", RightStick, 2, 1},
+    {"RightStickLeft", "Left", "F", RightStick, 1, 0},
+    {"RightStickRight", "Right", "H", RightStick, 1, 2},
+    {"Triangle", "Triangle", "I", Face, 0, 1},
+    {"Circle", "Circle", "L", Face, 1, 2},
+    {"Cross", "Cross", "J", Face, 2, 1},
+    {"Square", "Square", "K", Face, 1, 0},
+    {"L1", "L1", "Q", Shoulders, 0, 1},
+    {"R1", "R1", "E", Shoulders, 0, 2},
+    {"L2", "L2", "Z", Shoulders, 0, 0},
+    {"R2", "R2", "C", Shoulders, 0, 3},
+    {"L3", "Press / L3", "Left Shift", LeftStick, 1, 1},
+    {"R3", "Press / R3", "Left Ctrl", RightStick, 1, 1},
+    {"Options", "Options", "Return", Center, 0, 1},
+    {"TouchPad", "Touch pad left", "Backspace", Center, 0, 0},
+    {"TouchPadRight", "Touch pad right", "Tab", Center, 0, 2},
+    {"SpeakerVolume", "Speaker volume", "1", Shortcuts, 0, 0},
+    {"VibrationIntensity", "Vibration intensity", "2", Shortcuts, 0, 1},
+    {"TriggerEffectIntensity", "Trigger intensity", "3", Shortcuts, 0, 2},
 };
 
 QString KeypadName(int key) {
@@ -183,6 +191,17 @@ private:
 	QString m_binding;
 };
 
+void AssignBinding(QHash<QString, QString>& bindings, const QString& id, const QString& binding) {
+	for (auto item = bindings.begin(); item != bindings.end();) {
+		if (item.value().compare(binding, Qt::CaseInsensitive) == 0) {
+			item = bindings.erase(item);
+		} else {
+			++item;
+		}
+	}
+	bindings.insert(id, binding);
+}
+
 QHash<QString, QString> ParseMapping(const QStringList& mapping) {
 	QHash<QString, QString> result;
 	for (const auto& entry: mapping) {
@@ -191,15 +210,7 @@ QHash<QString, QString> ParseMapping(const QStringList& mapping) {
 		}
 		const auto separator = entry.indexOf(QLatin1Char('='));
 		if (separator > 0 && separator + 1 < entry.size()) {
-			const auto binding = entry.mid(separator + 1);
-			for (auto item = result.begin(); item != result.end();) {
-				if (item.value().compare(binding, Qt::CaseInsensitive) == 0) {
-					item = result.erase(item);
-				} else {
-					++item;
-				}
-			}
-			result.insert(entry.left(separator), binding);
+			AssignBinding(result, entry.left(separator), entry.mid(separator + 1));
 		}
 	}
 	return result;
@@ -216,82 +227,146 @@ double ParseMouseSensitivity(const QStringList& mapping) {
 
 } // namespace
 
-InputMappingDialog::InputMappingDialog(const QStringList& mapping, QWidget* parent)
-    : QDialog(parent) {
-	setWindowTitle(tr("Input Mapping"));
-	resize(460, 650);
+InputMappingDialog::InputMappingDialog(const QStringList& mapping, ControllerPreview* preview,
+                                       QWidget* parent)
+    : QDialog(parent), m_preview(preview) {
+	setWindowTitle(tr("Controls"));
+	resize(QSize(1100, 820).boundedTo(screen()->availableGeometry().size()));
 
 	auto* layout = new QVBoxLayout(this);
-	layout->addWidget(
-	    new QLabel(tr("Map keyboard or mouse buttons to DualSense controls.\n"
-	                  "Press F7 in-game to toggle mouse movement on the right stick."),
-	               this));
+	layout->setSpacing(10);
+	auto* title = new QLabel(tr("Controller setup"), this);
+	auto  font  = title->font();
+	font.setPointSize(font.pointSize() + 7);
+	font.setBold(true);
+	title->setFont(font);
+	layout->addWidget(title);
+	auto* hint = new QLabel(
+	    tr("Select a button to remap. Press controller buttons to test."), this);
+	hint->setWordWrap(true);
+	layout->addWidget(hint);
+	m_status = new QLabel(this);
+	m_status->setObjectName(QStringLiteral("controllerStatus"));
+	m_status->setWordWrap(true);
+	layout->addWidget(m_status);
 
-	const auto parsed = ParseMapping(mapping);
-	m_custom_bindings = !parsed.isEmpty();
+	auto* scroll = new QScrollArea(this);
+	scroll->setWidgetResizable(true);
+	scroll->setFrameShape(QFrame::NoFrame);
+	auto* content = new QWidget(scroll);
+	auto* grid    = new QGridLayout(content);
+	grid->setContentsMargins(0, 0, 0, 0);
+	grid->setSpacing(12);
+	QGroupBox*   groups[GroupCount];
+	QGridLayout* cells[GroupCount];
+	const char*  titles[] = {"D-Pad",
+	                         "Left stick",
+	                         "Right stick",
+	                         "Face buttons",
+	                         "Shoulders and triggers",
+	                         "Touch pad and options",
+	                         "In-game shortcuts (cycle)"};
+	for (int index = 0; index < GroupCount; ++index) {
+		groups[index] = new QGroupBox(tr(titles[index]), content);
+		cells[index]  = new QGridLayout(groups[index]);
+		cells[index]->setSpacing(5);
+	}
+
+	m_bindings        = ParseMapping(mapping);
+	m_custom_bindings = !m_bindings.isEmpty();
+	for (const auto& control: PAD_CONTROLS) {
+		const auto id          = QString::fromLatin1(control.id);
+		auto*      cell        = new QGroupBox(tr(control.label), groups[control.group]);
+		auto*      cell_layout = new QVBoxLayout(cell);
+		cell_layout->setContentsMargins(3, 5, 3, 3);
+		auto* button = new QPushButton(cell);
+		button->setObjectName(QStringLiteral("binding_") + id);
+		button->setAutoDefault(false);
+		button->setMinimumWidth(64);
+		button->setSizePolicy(QSizePolicy::Expanding, QSizePolicy::Fixed);
+		button->setAccessibleName(tr(titles[control.group]) + QStringLiteral(" / ") +
+		                          tr(control.label));
+		cell_layout->addWidget(button);
+		cells[control.group]->addWidget(cell, control.row, control.column);
+		m_buttons.insert(id, button);
+		if (!m_custom_bindings) {
+			m_bindings.insert(id, QString::fromLatin1(control.default_binding));
+		}
+		connect(button, &QPushButton::clicked, this, [this, id] { ChangeBinding(id); });
+	}
+
+	auto* left = new QVBoxLayout;
+	left->addWidget(groups[Dpad]);
+	left->addWidget(groups[LeftStick]);
+	left->addStretch();
+	grid->addLayout(left, 0, 0);
+	auto* center = new QVBoxLayout;
+	center->addWidget(groups[Shoulders]);
+	m_controller = new DualSenseWidget(content);
+	m_controller->setObjectName(QStringLiteral("controllerDiagram"));
+	center->addWidget(m_controller, 1);
+	center->addWidget(groups[Center]);
+	grid->addLayout(center, 0, 1);
+	auto* right = new QVBoxLayout;
+	right->addWidget(groups[Face]);
+	right->addWidget(groups[RightStick]);
+	right->addStretch();
+	grid->addLayout(right, 0, 2);
+	grid->setColumnStretch(1, 1);
+	grid->addWidget(groups[Shortcuts], 1, 0, 1, 3);
+	scroll->setWidget(content);
+	layout->addWidget(scroll, 1);
 
 	auto* sensitivity_layout = new QHBoxLayout;
 	sensitivity_layout->addWidget(new QLabel(tr("Mouse sensitivity"), this));
 	m_sensitivity = new QDoubleSpinBox(this);
+	m_sensitivity->setObjectName(QStringLiteral("mouseSensitivity"));
 	m_sensitivity->setRange(0.1, 5.0);
 	m_sensitivity->setSingleStep(0.1);
 	m_sensitivity->setDecimals(1);
-	m_sensitivity->setSuffix(QStringLiteral("x"));
+	m_sensitivity->setSuffix(QStringLiteral("×"));
 	m_sensitivity->setValue(ParseMouseSensitivity(mapping));
 	sensitivity_layout->addWidget(m_sensitivity);
+	sensitivity_layout->addWidget(
+	    new QLabel(tr("F7 in-game toggles mouse movement on the right stick."), this));
 	sensitivity_layout->addStretch();
 	layout->addLayout(sensitivity_layout);
 
-	m_bindings = new QTreeWidget(this);
-	m_bindings->setColumnCount(2);
-	m_bindings->setHeaderLabels({tr("DualSense control"), tr("Host input")});
-	m_bindings->setRootIsDecorated(false);
-	m_bindings->setSelectionMode(QAbstractItemView::SingleSelection);
-	m_bindings->header()->setSectionResizeMode(0, QHeaderView::Stretch);
-	m_bindings->header()->setSectionResizeMode(1, QHeaderView::Stretch);
-	layout->addWidget(m_bindings);
-
-	for (const auto& control: PAD_CONTROLS) {
-		auto* item = new QTreeWidgetItem(m_bindings);
-		item->setText(0, tr(control.label));
-		item->setData(0, Qt::UserRole, QString::fromLatin1(control.id));
-		SetBinding(item, m_custom_bindings ? parsed.value(QString::fromLatin1(control.id))
-		                                   : QString::fromLatin1(control.default_binding));
-	}
-	m_bindings->setCurrentItem(m_bindings->topLevelItem(0));
-
-	auto* controls  = new QHBoxLayout;
-	m_change_button = new QPushButton(tr("Change..."), this);
-	m_clear_button  = new QPushButton(tr("Clear"), this);
-	auto* defaults  = new QPushButton(tr("Defaults"), this);
-	controls->addWidget(m_change_button);
-	controls->addWidget(m_clear_button);
-	controls->addWidget(defaults);
-	controls->addStretch();
-	layout->addLayout(controls);
-
-	auto* buttons = new QDialogButtonBox(QDialogButtonBox::Ok | QDialogButtonBox::Cancel, this);
+	auto* buttons  = new QDialogButtonBox(QDialogButtonBox::RestoreDefaults |
+	                                          QDialogButtonBox::Save | QDialogButtonBox::Cancel,
+	                                      this);
+	auto* defaults = buttons->button(QDialogButtonBox::RestoreDefaults);
+	defaults->setObjectName(QStringLiteral("defaults"));
+	defaults->setAutoDefault(false);
 	layout->addWidget(buttons);
-
-	connect(m_bindings, &QTreeWidget::itemDoubleClicked, this,
-	        [this](QTreeWidgetItem*, int) { ChangeBinding(); });
-	connect(m_bindings, &QTreeWidget::itemSelectionChanged, this, [this]() { UpdateButtons(); });
-	connect(m_change_button, &QPushButton::clicked, this, [this]() { ChangeBinding(); });
-	connect(m_clear_button, &QPushButton::clicked, this, [this]() { ClearBinding(); });
-	connect(defaults, &QPushButton::clicked, this, [this]() { RestoreDefaults(); });
+	connect(m_controller, &DualSenseWidget::ControlClicked, this,
+	        &InputMappingDialog::ChangeBinding);
+	connect(defaults, &QPushButton::clicked, this, &InputMappingDialog::RestoreDefaults);
 	connect(buttons, &QDialogButtonBox::accepted, this, &QDialog::accept);
 	connect(buttons, &QDialogButtonBox::rejected, this, &QDialog::reject);
+	if (m_preview != nullptr) {
+		connect(m_preview, &ControllerPreview::StateChanged, this,
+		        &InputMappingDialog::UpdatePreview);
+		m_preview->SetPreviewEnabled(true);
+	}
 	UpdateButtons();
+	UpdatePreview();
+}
+
+InputMappingDialog::~InputMappingDialog() {
+	if (m_preview != nullptr) {
+		m_preview->SetPreviewEnabled(false);
+	}
 }
 
 QStringList InputMappingDialog::Mapping() const {
 	QStringList result;
 	if (m_custom_bindings) {
-		for (int index = 0; index < m_bindings->topLevelItemCount(); index++) {
-			const auto* item    = m_bindings->topLevelItem(index);
-			const auto  binding = item->data(BINDING_COLUMN, Qt::UserRole).toString();
+		for (const auto& control: PAD_CONTROLS) {
+			const auto id      = QString::fromLatin1(control.id);
+			const auto binding = m_bindings.value(id);
 			if (!binding.isEmpty()) {
-				result.append(item->data(0, Qt::UserRole).toString() + QLatin1Char('=') + binding);
+				result.append(id + QLatin1Char('=') + binding);
 			}
 		}
 	}
@@ -302,58 +377,73 @@ QStringList InputMappingDialog::Mapping() const {
 	return result;
 }
 
-void InputMappingDialog::ChangeBinding() {
-	auto* item = m_bindings->currentItem();
-	if (item == nullptr) {
+void InputMappingDialog::ChangeBinding(const QString& id) {
+	if (!m_buttons.contains(id)) {
 		return;
 	}
-
+	m_selected = id;
+	UpdateHighlights();
 	InputCaptureDialog dialog(this);
-	if (dialog.exec() != QDialog::Accepted) {
-		return;
+	dialog.setWindowTitle(tr("Set binding — %1").arg(m_buttons.value(id)->accessibleName()));
+	if (dialog.exec() == QDialog::Accepted) {
+		AssignBinding(m_bindings, id, dialog.Binding());
+		m_custom_bindings = true;
 	}
-
-	for (int index = 0; index < m_bindings->topLevelItemCount(); index++) {
-		auto* other = m_bindings->topLevelItem(index);
-		if (other != item && other->data(BINDING_COLUMN, Qt::UserRole)
-		                             .toString()
-		                             .compare(dialog.Binding(), Qt::CaseInsensitive) == 0) {
-			SetBinding(other, {});
-		}
-	}
-	SetBinding(item, dialog.Binding());
-	m_custom_bindings = true;
-}
-
-void InputMappingDialog::ClearBinding() {
-	SetBinding(m_bindings->currentItem(), {});
-	m_custom_bindings = true;
+	m_selected.clear();
+	UpdateButtons();
 }
 
 void InputMappingDialog::RestoreDefaults() {
-	for (int index = 0; index < m_bindings->topLevelItemCount(); index++) {
-		SetBinding(m_bindings->topLevelItem(index),
-		           QString::fromLatin1(PAD_CONTROLS[index].default_binding));
+	m_bindings.clear();
+	for (const auto& control: PAD_CONTROLS) {
+		m_bindings.insert(QString::fromLatin1(control.id),
+		                  QString::fromLatin1(control.default_binding));
 	}
 	m_sensitivity->setValue(DEFAULT_MOUSE_SENSITIVITY);
 	m_custom_bindings = false;
-}
-
-void InputMappingDialog::SetBinding(QTreeWidgetItem* item, const QString& binding) {
-	if (item == nullptr) {
-		return;
-	}
-	item->setData(BINDING_COLUMN, Qt::UserRole, binding);
-	item->setText(BINDING_COLUMN, binding.isEmpty() ? tr("None") : binding);
 	UpdateButtons();
 }
 
 void InputMappingDialog::UpdateButtons() {
-	if (m_change_button == nullptr) {
-		return;
+	for (auto item = m_buttons.begin(); item != m_buttons.end(); ++item) {
+		auto*      button  = item.value();
+		const auto binding = m_bindings.value(item.key());
+		const auto label   = binding.isEmpty() ? tr("None") : binding;
+		button->setText(QString(label).replace(QLatin1Char('&'), QLatin1String("&&")));
+		button->setToolTip(tr("%1: %2. Click to change.").arg(button->accessibleName(), label));
+	}
+	UpdateHighlights();
+}
+
+void InputMappingDialog::UpdatePreview() {
+	const auto pressed = m_preview != nullptr ? m_preview->Pressed() : QSet<QString>();
+	m_controller->SetPressed(pressed);
+	m_controller->SetSticks(m_preview != nullptr ? m_preview->LeftStick() : QPointF(),
+	                        m_preview != nullptr ? m_preview->RightStick() : QPointF());
+	UpdateHighlights();
+	if (m_preview != nullptr && !m_preview->Error().isEmpty()) {
+		m_status->setText(tr("Controller preview unavailable: %1").arg(m_preview->Error()));
+	} else if (m_preview != nullptr && !m_preview->DeviceName().isEmpty()) {
+		m_status->setText(tr("Connected: %1").arg(m_preview->DeviceName()));
+	} else {
+		m_status->setText(tr("No controller connected"));
+	}
+}
+
+void InputMappingDialog::UpdateHighlights() {
+	m_controller->SetSelected(m_selected);
+	for (auto item = m_buttons.begin(); item != m_buttons.end(); ++item) {
+		const bool down   = m_preview != nullptr && m_preview->Pressed().contains(item.key());
+		auto       colors = palette();
+		if (down) {
+			colors.setColor(QPalette::Button, colors.color(QPalette::Highlight));
+			colors.setColor(QPalette::ButtonText, colors.color(QPalette::HighlightedText));
+		} else if (item.key() == m_selected) {
+			colors.setColor(QPalette::ButtonText, colors.color(QPalette::Highlight));
+		}
+		auto label_font = font();
+		label_font.setBold(down);
+		item.value()->setPalette(colors);
+		item.value()->setFont(label_font);
 	}
-	const auto* item = m_bindings->currentItem();
-	m_change_button->setEnabled(item != nullptr);
-	m_clear_button->setEnabled(item != nullptr &&
-	                           !item->data(BINDING_COLUMN, Qt::UserRole).toString().isEmpty());
 }

src/launcher/src/mainDialog.cpp

@@ -4,7 +4,7 @@
 #include "configuration.h"
 #include "configurationItem.h"
 #include "configurationListWidget.h"
-#include "controllerLightbar.h"
+#include "controllerPreview.h"
 #include "gameContent.h"
 #include "updateChecker.h"
 
@@ -89,7 +89,7 @@ private:
 	Ui::MainDialog* m_ui             = {nullptr};
 	MainDialog*     m_main_dialog    = nullptr;
 	UpdateChecker*  m_update_checker = nullptr;
-	ControllerLightbar m_lightbar;
+	ControllerPreview m_controller_preview;
 	QString         m_interpreter;
 
 	QProcess m_process;
@@ -112,6 +112,7 @@ MainDialogPrivate::~MainDialogPrivate() {
 void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 	m_ui = new Ui::MainDialog;
 	m_ui->setupUi(main_dialog);
+	m_ui->widget->SetControllerPreview(&m_controller_preview);
 
 	m_main_dialog = main_dialog;
 	m_update_checker = new UpdateChecker(main_dialog);
@@ -127,7 +128,7 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 	connect(m_ui->widget, &ConfigurationListWidget::PreviewControllerColor, this,
 	        [this](const QString& color) {
 		        if (m_process.state() == QProcess::NotRunning) {
-			        m_lightbar.SetColor(color);
+			        m_controller_preview.SetColor(color);
 		        }
 	        });
 	connect(m_ui->widget, &ConfigurationListWidget::Run, this, &MainDialogPrivate::Run);
@@ -514,7 +515,7 @@ void MainDialogPrivate::Run() {
 	}
 
 	m_running_item->SetRunning(true);
-	m_lightbar.Stop();
+	m_controller_preview.SetSuspended(true);
 
 	auto info = m_ui->widget->CreateConfiguration(*m_running_item);
 	m_main_dialog->RunInterpreter(&m_process, *info);
@@ -535,10 +536,11 @@ void MainDialogPrivate::Update() {
 
 	m_ui->widget->SetRunEnabled(run_enabled);
 	if (m_process.state() != QProcess::NotRunning) {
-		m_lightbar.Stop();
+		m_controller_preview.SetSuspended(true);
 		return;
 	}
-	m_lightbar.SetColor(m_ui->widget->GetGlobalControllerColor());
+	m_controller_preview.SetSuspended(false);
+	m_controller_preview.SetColor(m_ui->widget->GetGlobalControllerColor());
 }
 
 #include "mainDialog.moc"

src/launcher/styles/launcher.qss

@@ -1,3 +1,13 @@
+QListView#cfgs_grid {
+    background: palette(base);
+    border: 1px solid palette(mid);
+    color: palette(text);
+}
+
+QListView#cfgs_grid::item {
+    padding: 8px;
+}
+
 GameListTreeWidget#cfgs_list {
     background: transparent;
     border: 1px solid palette(mid);
@@ -39,6 +49,7 @@ QLineEdit#search_line_edit:focus {
     border-color: palette(highlight);
 }
 
+QToolButton#grid_view_button,
 QToolButton#refresh_button,
 QToolButton#global_settings_button,
 QToolButton#input_mapping_button,
@@ -51,6 +62,7 @@ QToolButton#delete_button {
     padding: 3px;
 }
 
+QToolButton#grid_view_button:hover,
 QToolButton#refresh_button:hover,
 QToolButton#global_settings_button:hover,
 QToolButton#input_mapping_button:hover,
@@ -61,6 +73,8 @@ QToolButton#delete_button:hover {
     border-color: palette(mid);
 }
 
+QToolButton#grid_view_button:checked,
+QToolButton#grid_view_button:pressed,
 QToolButton#refresh_button:pressed,
 QToolButton#global_settings_button:pressed,
 QToolButton#input_mapping_button:pressed,

src/libs/agc.cpp

@@ -20,6 +20,7 @@
 #include "graphics/presentation/videoOut.h"
 #include "graphics/presentation/window.h"
 #include "graphics/shader/shader.h"
+#include "kernel/memory.h"
 #include "kernel/pthread.h"
 #include "libs/errno.h"
 #include "libs/libs.h"
@@ -44,6 +45,23 @@ namespace Libs::Graphics {
 
 static RenderContext* g_renderer = nullptr;
 
+namespace Gen5Driver {
+
+static constexpr uint64_t DriverDmemBase = 0x0fe0000000ull;
+static constexpr uint32_t DriverDmemSize = 0x200000;
+static constexpr uint32_t AgcDmemOffset  = 0x40000;
+
+struct DriverDmem {
+	std::mutex mutex;
+	int64_t    physical_offset = -1;
+};
+
+static DriverDmem g_driver_dmem;
+static int InitializeDmem();
+static void ReleaseDmem();
+
+} // namespace Gen5Driver
+
 template <typename... Args>
 static void AgcTrace(const char* format, const Args&... args) {
 	if (Config::GraphicsDebugDumpEnabled()) {
@@ -70,6 +88,7 @@ void Initialize() {
 void Shutdown() {
 	EXIT_IF(g_renderer == nullptr);
 	g_renderer->ShutdownGpu();
+	Gen5Driver::ReleaseDmem();
 	VideoOut::VideoOutShutdown();
 	WindowShutdown();
 	g_renderer = nullptr;
@@ -337,6 +356,38 @@ int KYTY_SYSV_ABI AgcInit(uint32_t* state, uint32_t ver) {
 	}
 
 	printf("version = %u\n", ver);
+	std::scoped_lock lock {Gen5Driver::g_driver_dmem.mutex};
+	if (Gen5Driver::g_driver_dmem.physical_offset >= 0) {
+		return OK;
+	}
+	if (const auto result = Gen5Driver::InitializeDmem(); result != OK) {
+		return result;
+	}
+	struct FunctionShaderBindings {
+		uint64_t code;
+		uint32_t user_data[2];
+	};
+	static_assert(sizeof(FunctionShaderBindings) == 16);
+	// Ordered-count queue validator, including its return to the caller.
+	static constexpr std::array<uint32_t, 16> validator {
+	    0xbeeb03ff, 0x00000021, 0xb96a1818, 0xbf06106a,
+	    0xbf800000, 0x8590807e, 0xb96a0a18, 0xbf066a80,
+	    0xbf800000, 0x85ea807e, 0x88ea106a, 0xbf870003,
+	    0xbefc03ff, 0x8a6ca000, 0xbf920009, 0xbefd210e,
+	};
+	const auto base = Gen5Driver::DriverDmemBase + Gen5Driver::AgcDmemOffset;
+	const auto function_table = base + 0x60;
+	const auto code_address = base + 0x100; // Shader code requires 256-byte alignment.
+	std::array<FunctionShaderBindings, 8> functions {};
+	functions[2].code = code_address;
+	const std::array<uint32_t, 4> descriptor {
+	    static_cast<uint32_t>(function_table),
+	    static_cast<uint32_t>(function_table >> 32u) | 0x00100000u,
+	    static_cast<uint32_t>(functions.size()), 0x5204u,
+	};
+	std::memcpy(reinterpret_cast<void*>(function_table), functions.data(), sizeof(functions));
+	std::memcpy(reinterpret_cast<void*>(code_address), validator.data(), sizeof(validator));
+	std::memcpy(reinterpret_cast<void*>(base), descriptor.data(), sizeof(descriptor));
 
 	return OK;
 }
@@ -4224,6 +4275,36 @@ namespace Gen5Driver {
 
 LIB_NAME("Graphics5Driver", "Graphics5Driver");
 
+static int InitializeDmem() {
+	int64_t physical_offset = -1;
+	auto result = LibKernel::Memory::KernelAllocateDirectMemory(
+	    0, LibKernel::Memory::KernelGetDirectMemorySize(), DriverDmemSize, DriverDmemSize,
+	    12, &physical_offset);
+	if (result != OK) {
+		return result;
+	}
+	void* address = reinterpret_cast<void*>(DriverDmemBase);
+	result = LibKernel::Memory::KernelMapNamedDirectMemory(
+	    &address, DriverDmemSize, 0x33, 0, physical_offset, DriverDmemSize, "SceAgcDriver");
+	if (result != OK) {
+		EXIT_IF(LibKernel::Memory::KernelCheckedReleaseDirectMemory(physical_offset, DriverDmemSize) != OK);
+		return result;
+	}
+	EXIT_IF(reinterpret_cast<uint64_t>(address) != DriverDmemBase);
+	std::memset(address, 0, DriverDmemSize);
+	g_driver_dmem.physical_offset = physical_offset;
+	return OK;
+}
+
+static void ReleaseDmem() {
+	std::scoped_lock lock {g_driver_dmem.mutex};
+	if (g_driver_dmem.physical_offset >= 0) {
+		EXIT_IF(LibKernel::Memory::KernelCheckedReleaseDirectMemory(
+		            g_driver_dmem.physical_offset, DriverDmemSize) != OK);
+		g_driver_dmem.physical_offset = -1;
+	}
+}
+
 struct TessellationDriverState {
 	uint64_t tf_ring_base      = 0;
 	uint32_t tf_ring_size      = 0;

tests/ResourceMaterializationTests.cpp

@@ -435,7 +435,7 @@ void TestUnbasedFlatCacheHitMaterializes() {
         "unbased FLAT plan produced unexpected descriptors");
 }
 
-void TestWrittenDescriptorUsesStrictReaderOnce() {
+void TestWrittenDescriptorPredicateReads(bool memory_condition, bool loop) {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   Program program;
   program.stage = Libs::Graphics::ShaderType::Compute;
@@ -456,10 +456,10 @@ void TestWrittenDescriptorUsesStrictReaderOnce() {
   source.dword_count = 4;
   program.descriptor_sources.push_back(source);
   program.info.buffers.push_back({.source = 0, .written = true});
-  // A host-evaluable branch captures resource reads and needs the writable
-  // descriptor's clean provenance for the renderer's disjointness proof.
+  // Only memory-derived host decisions need clean, disjoint descriptor reads.
   auto &condition = block.AppendNewInst(ValueOpcode::IEqual32,
-                                        {Value(&offset), Value(4u)});
+      {memory_condition ? Value(&read) : Value(&offset),
+       Value(memory_condition ? 0x8000u : 4u)});
   auto &store_block = AddValueBlock(program);
   AddValueBlock(program);
   program.blocks[0]->condition = Value(&condition);
@@ -467,12 +467,26 @@ void TestWrittenDescriptorUsesStrictReaderOnce() {
       Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind::ConditionalBranch;
   program.blocks[0]->terminator.true_block = program.blocks[1];
   program.blocks[0]->terminator.false_block = program.blocks[2];
+
   program.blocks[1]->id = 1;
   program.blocks[1]->terminator.kind =
       Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind::Return;
   program.blocks[2]->id = 2;
   program.blocks[2]->terminator.kind =
       Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind::Return;
+  if (loop) {
+    // The predicate reaches the store through a continue block, then loops back.
+    auto &continued = AddValueBlock(program);
+    program.blocks[0]->terminator.true_block = &continued;
+    continued.terminator.kind = Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind::Branch;
+    continued.terminator.true_block = &store_block;
+    continued.AddBranch(&store_block);
+    store_block.terminator.kind = Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind::Branch;
+    store_block.terminator.true_block = █
+    store_block.AddBranch(&block);
+  }
+  block.AddBranch(program.blocks[0]->terminator.true_block);
+  block.AddBranch(program.blocks[2]);
   program.memory_info.push_back({.kind = ResourceKind::Buffer, .resource = 0});
   auto &output = store_block.AppendNewInst(ValueOpcode::GetBufferResource,
       {source.dwords[0], source.dwords[1], source.dwords[2], source.dwords[3]});
@@ -480,8 +494,8 @@ void TestWrittenDescriptorUsesStrictReaderOnce() {
       {Value(&output), Value(0u), Value(0u), Value(0u), Value(1u), Value(true)})
       .SetFlags(MemoryFlags{.index = 1});
   auto plan = ExtractResourcePlan(program);
-  Check(plan.capture_specialization_reads,
-        "conditional writable descriptor lost its alias proof");
+  Check(plan.capture_specialization_reads == memory_condition && !plan.control_flow.empty(),
+        "conditional writable descriptor used the wrong memory dependency policy");
   struct Reads { uint32_t ordinary = 0; uint32_t strict = 0; bool clean = false; } reads;
   const std::array<uint32_t, 1> user_data{4u};
   const SrtRuntime runtime{
@@ -501,16 +515,21 @@ void TestWrittenDescriptorUsesStrictReaderOnce() {
       }};
   ResourceSnapshot snapshot;
   ResourceSpecialization specialization;
-  Check(!MaterializeResources(plan, runtime, snapshot, specialization) &&
-            reads.ordinary == 0 && reads.strict == 1,
-        "GPU-dirty dynamic writable descriptor bypassed strict provenance");
-  reads.clean = true;
-  reads.strict = 0;
-  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
-            reads.ordinary == 0 && reads.strict == 1 && snapshot.buffers[0].dwords[0] == 0x8000u &&
-            snapshot.specialization_reads ==
-                std::vector<std::pair<uint64_t, uint64_t>>{{0x1004u, 4u}},
-        "writable descriptor was evaluated twice or scalar EXEC suppressed its read");
+  for (const bool clean : {false, true}) {
+    reads = {.clean = clean};
+    // A failed predicate explores both edges and the writable descriptor retries its clean read.
+    const bool expected = !memory_condition || clean;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) == expected &&
+              reads.ordinary == (memory_condition ? 0u : 1u) &&
+              reads.strict == (memory_condition ? (clean ? 1u : 2u) : 0u),
+          "writable descriptor changed reader policy or bypassed a dirty memory predicate");
+    if (!expected) continue;
+    Check(snapshot.buffers[0].dwords[0] == 0x8000u &&
+              snapshot.specialization_reads == (memory_condition
+                  ? std::vector<std::pair<uint64_t, uint64_t>>{{0x1004u, 4u}}
+                  : std::vector<std::pair<uint64_t, uint64_t>>{}),
+          "writable descriptor lost its address or captured an immutable predicate");
+  }
 }
 
 void TestFailedMaterializationRejectsStage() {
@@ -671,7 +690,9 @@ int main() {
   TestUniformVectorDescriptorRead();
   TestExactReciprocalDescriptorArithmetic();
   TestUnbasedFlatCacheHitMaterializes();
-  TestWrittenDescriptorUsesStrictReaderOnce();
+  for (const bool memory_condition : {false, true})
+    for (const bool loop : {false, true})
+      TestWrittenDescriptorPredicateReads(memory_condition, loop);
   TestFailedMaterializationRejectsStage();
   TestFiniteImageRefreshReusesScalarReads();
   TestMixedSamplerVariantsShareRuntimeDescriptor();

tests/ResourceTrackingTests.cpp

@@ -742,7 +742,9 @@ void TestBoundedComputeImageLoop() {
   enum class Variant {
     Bounded, Plain, Nonzero, TrueEdge, WrongGuard, EntryBypass, ExitBypass, GuardBlock,
     WrongStep, DivergentBound, DivergentKey, Disjunction, WrongPolarity,
-    IncrementBypass, PreviousBound, Masked, MaskedWrongGuard, MaskResurrection, StatusOverwrite,
+    IncrementBypass, PreviousBound, Forwarded, ForwardedChain, ForwardedJoin,
+    ForwardedBranch, ForwardedIncrementBypass,
+    Masked, MaskedWrongGuard, MaskResurrection, StatusOverwrite,
     GuardedDiamond, GuardedDiamondBypass
   };
   const auto make_plan = [](Variant variant) {
@@ -759,16 +761,21 @@ void TestBoundedComputeImageLoop() {
     auto *guard = masked ? fixture.AddBlock() : header;
     auto *increment = masked ? fixture.AddBlock() : latch;
     auto *final_exit = variant == Variant::PreviousBound ? fixture.AddBlock() : exit;
+    const bool forwarded = variant >= Variant::Forwarded && variant <= Variant::ForwardedIncrementBypass;
+    auto *forward = forwarded ? fixture.AddBlock() : increment;
+    auto *incoming = variant == Variant::ForwardedChain ? fixture.AddBlock() : forward;
+    auto *detour = variant == Variant::ForwardedJoin ? fixture.AddBlock() : nullptr;
     entry->AddBranch(header);
     if (!masked) {
       header->AddBranch(exit);
       header->AddBranch(body);
       body->AddBranch(latch);
-      latch->AddBranch(header);
+      latch->AddBranch(forwarded ? forward : header);
       if (variant == Variant::PreviousBound) latch->AddBranch(final_exit);
       if (variant == Variant::EntryBypass) entry->AddBranch(body);
       if (variant == Variant::ExitBypass) exit->AddBranch(body);
-      if (variant == Variant::IncrementBypass || variant == Variant::PreviousBound)
+      if (variant == Variant::IncrementBypass || variant == Variant::PreviousBound ||
+          variant == Variant::ForwardedIncrementBypass)
         exit->AddBranch(latch);
       fixture.program.blocks[0]->terminator = {
           .kind = variant == Variant::EntryBypass
@@ -789,16 +796,16 @@ void TestBoundedComputeImageLoop() {
           .kind = variant == Variant::PreviousBound
                       ? CFG::TerminatorKind::ConditionalBranch
                       : CFG::TerminatorKind::Branch,
-          .true_block = fixture.program.blocks[1u],
+          .true_block = forwarded ? forward : fixture.program.blocks[1u],
           .false_block = final_exit};
       fixture.program.blocks[4]->terminator = {
           .kind = final_exit != exit || variant == Variant::ExitBypass ||
-                          variant == Variant::IncrementBypass
+                          variant == Variant::IncrementBypass || variant == Variant::ForwardedIncrementBypass
                       ? CFG::TerminatorKind::Branch
                       : CFG::TerminatorKind::Return,
           .true_block =
               fixture.program.blocks[final_exit != exit ||
-                                             variant == Variant::IncrementBypass
+                                             variant == Variant::IncrementBypass || variant == Variant::ForwardedIncrementBypass
                                          ? 3u
                                          : 2u]};
       if (final_exit != exit)
@@ -822,6 +829,29 @@ void TestBoundedComputeImageLoop() {
                                         count}, 0, compare);
     const auto initial_active = fixture.Emit(ValueOpcode::INotEqual32,
                                              {local, Value(0u)}, 0, entry);
+    if (forwarded) {
+      forward->AddBranch(incoming == forward ? header : incoming);
+      forward->terminator = {.kind = CFG::TerminatorKind::Branch,
+                            .true_block = incoming == forward ? header : incoming};
+      if (incoming != forward) {
+        incoming->AddBranch(header);
+        incoming->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = header};
+      }
+      if (variant == Variant::ForwardedBranch) {
+        forward->AddBranch(exit);
+        forward->condition = initial_active;
+        forward->terminator = {.kind = CFG::TerminatorKind::ConditionalBranch,
+                              .true_block = header, .false_block = exit};
+      }
+      if (detour != nullptr) {
+        latch->AddBranch(detour);
+        latch->condition = initial_active;
+        latch->terminator = {.kind = CFG::TerminatorKind::ConditionalBranch,
+                            .true_block = forward, .false_block = detour};
+        detour->AddBranch(forward);
+        detour->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = forward};
+      }
+    }
     if (masked) {
       auto *active_phi = diamond ? nullptr : &header->AppendNewInst(
           ValueOpcode::Phi, {}, static_cast<uint64_t>(Type::U1));
@@ -934,7 +964,9 @@ void TestBoundedComputeImageLoop() {
                                    {key, Value(variant == Variant::WrongStep ? 2u : 1u)},
                                    0, increment);
     phi.AddPhiOperand(entry, variant == Variant::DivergentKey ? local : Value(0u));
-    phi.AddPhiOperand(increment, step);
+    phi.AddPhiOperand(forwarded ? incoming : increment, step);
+    Check(!forwarded || step.ResolveInstruction()->Parent() != phi.PhiBlock(1),
+          "forwarded loop fixture did not separate the increment from its Phi incoming edge");
 
     fixture.block = variant == Variant::GuardBlock ? header : body;
     const auto table = fixture.Address(fixture.UserData(0), fixture.UserData(1));
@@ -975,7 +1007,11 @@ void TestBoundedComputeImageLoop() {
   make_plan(Variant::TrueEdge);
   make_plan(Variant::Masked);
   make_plan(Variant::GuardedDiamond);
+  make_plan(Variant::Forwarded);
+  make_plan(Variant::ForwardedChain);
   for (const auto variant : {Variant::IncrementBypass, Variant::PreviousBound,
+                             Variant::ForwardedJoin, Variant::ForwardedBranch,
+                             Variant::ForwardedIncrementBypass,
                              Variant::MaskedWrongGuard,
                              Variant::MaskResurrection, Variant::StatusOverwrite,
                              Variant::GuardedDiamondBypass}) {
@@ -2612,6 +2648,129 @@ void TestFiniteImagePhiCycle() {
   }
 }
 
+void TestFiniteInlineSampler() {
+  namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
+  Fixture fixture;
+  auto* entry = fixture.block;
+  auto* defined = fixture.AddBlock();
+  auto* merge = fixture.AddBlock();
+  entry->AddBranch(defined);
+  entry->AddBranch(merge);
+  defined->AddBranch(merge);
+  const auto lane = fixture.Emit(ValueOpcode::GetBuiltin,
+      {Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(0u)});
+  entry->condition = fixture.Emit(ValueOpcode::IEqual32, {lane, Value(0u)});
+  entry->terminator = {.kind = CFG::TerminatorKind::ConditionalBranch,
+                      .true_block = defined, .false_block = merge};
+  defined->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = merge};
+  merge->terminator.kind = CFG::TerminatorKind::Return;
+  fixture.block = merge;
+  auto& lod = merge->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U32));
+  auto& filter = merge->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U32));
+  lod.AddPhiOperand(entry, Value(0u));
+  lod.AddPhiOperand(defined, Value(0x00fff000u));
+  filter.AddPhiOperand(defined, Value(0x0a500000u));
+  filter.AddPhiOperand(entry, Value(0u));
+  const auto sampler = fixture.Sampler({Value(0u), Value(&lod), Value(&filter), Value(0u)});
+  const auto direct = fixture.Sampler({Value(0u), Value(0u), Value(0u), Value(0u)});
+  const auto previous = fixture.Sampler(
+      {Value(0u), Value(0x00fff000u), Value(0x0a500000u), Value(0u)});
+  using F = Libs::Graphics::Prospero::BufferFormat;
+  for (const auto format : {F::k32Float, F::k32SInt}) {
+    const auto image = fixture.Image({Value(0x1000u), Value(uint32_t(format) << 20u),
+        Value(0u), Value(Libs::Graphics::DstSel(4, 5, 6, 7) |
+            (uint32_t(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u)),
+        Value(0u), Value(0u), Value(0u), Value(0u)});
+    MemoryInfo memory;
+    memory.kind = ResourceKind::Image;
+    memory.image_dimension = Decoder::ImageDimension::Dim2D;
+    if (format == F::k32Float)
+      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, previous, fixture.ImageAddress()},
+                   fixture.AddMemory(memory, 0x440u));
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+                 fixture.AddMemory(memory, 0x1dd0u));
+    if (format == F::k32Float)
+      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, direct, fixture.ImageAddress()},
+                   fixture.AddMemory(memory, 0x1dd8u));
+  }
+  fixture.PlanAndTrack();
+  const auto root = fixture.program.memory_info[1].sampler;
+  const auto source = fixture.program.info.samplers[root].source;
+  const auto& finite = fixture.program.descriptor_sources[source].indirect_descriptor;
+  Check(finite && finite->sources.size() == 2 && fixture.program.info.samplers.size() == 3,
+        "finite inline sampler lost either state or aliased the direct sampler");
+  const auto* key = sampler.Instruction()->Arg(0).Instruction();
+  Check(key->GetOpcode() == ValueOpcode::Phi && key->Parent() == merge &&
+            key->PhiBlock(0) == entry && key->PhiBlock(1) == defined &&
+            key->Arg(0).U32() == 0 && key->Arg(1).U32() == 1,
+        "inline sampler DWORD correlation did not preserve its GPU selector");
+  const auto plan = ExtractResourcePlan(fixture.program);
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  const SrtRuntime runtime{.read_specialization_memory = ReadTestMemory};
+  Check(MaterializeResources(plan, runtime, snapshot, specialization),
+        "finite inline sampler could not materialize literal choices");
+  const auto candidates = fixture.program.info.samplers[root].indirect_resources;
+  Check(candidates.size() == 2 && snapshot.samplers[candidates[0]].dwords[1] == 0 &&
+            snapshot.samplers[candidates[0]].dwords[2] == 0 &&
+            snapshot.samplers[candidates[1]].dwords[1] == 0x00fff000u &&
+            snapshot.samplers[candidates[1]].dwords[2] == 0x0a500000u,
+        "finite sampler materialization mixed the two correlated DWORD states");
+  Check(candidates[1] == fixture.program.memory_info[0].sampler && candidates[1] < root,
+        "finite sampler did not reuse the earlier constant candidate");
+  ApplyResourceSpecialization(fixture.program, specialization);
+  for (const uint32_t operation : {1u, 3u}) {
+    const auto selected = fixture.program.memory_info[operation].sampler;
+    const auto& choices = fixture.program.info.samplers[selected].indirect_resources;
+    Check(choices.size() == 2, "sampler class specialization lost a finite candidate");
+    for (uint32_t ordinal = 0; ordinal < choices.size(); ++ordinal) {
+      const auto& choice = fixture.program.info.samplers[choices[ordinal]];
+      Check(choice.snapshot_index == candidates[ordinal] &&
+                choice.force_point_filtering == (operation == 3u) &&
+                choice.integer_border == (operation == 3u),
+            "finite sampler class remapping changed candidate identity or filtering");
+    }
+  }
+  Check(fixture.program.info.samplers[fixture.program.memory_info[2].sampler]
+            .indirect_resources.empty(),
+        "direct sampler was changed into a dynamic selection");
+  auto invalid = ExtractResourcePlan(fixture.program);
+  invalid.info.samplers[0].source = static_cast<uint32_t>(invalid.descriptor_sources.size());
+  Check(!MaterializeResources(invalid, runtime, snapshot, specialization),
+        "invalid sampler source bypassed materialization validation");
+
+  Fixture adjusted;
+  auto* initial = adjusted.block;
+  auto* alternate = adjusted.AddBlock();
+  auto* join = adjusted.AddBlock();
+  initial->AddBranch(alternate);
+  initial->AddBranch(join);
+  alternate->AddBranch(join);
+  initial->condition = adjusted.Emit(ValueOpcode::IEqual32,
+      {adjusted.Emit(ValueOpcode::LaneId), Value(0u)});
+  initial->terminator = {.kind = CFG::TerminatorKind::ConditionalBranch,
+                        .true_block = alternate, .false_block = join};
+  alternate->terminator = {.kind = CFG::TerminatorKind::Branch, .true_block = join};
+  join->terminator.kind = CFG::TerminatorKind::Return;
+  auto& selected = join->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U32));
+  selected.AddPhiOperand(initial, Value(0u));
+  selected.AddPhiOperand(alternate, Value(0x00fff000u));
+  adjusted.block = join;
+  const auto shared = adjusted.Sampler({Value(0u), Value(&selected), Value(0u), Value(0u)});
+  const auto image = adjusted.Image({Value(0u), Value(0u), Value(0u), Value(0u),
+                                    Value(0u), Value(0u), Value(0u), Value(0u)});
+  for (const uint32_t flags : {0u, uint32_t(Decoder::ImageSampleFlagAdjust)}) {
+    MemoryInfo memory;
+    memory.kind = ResourceKind::Image;
+    memory.image_dimension = Decoder::ImageDimension::Dim2D;
+    memory.image_sample_flags = flags;
+    adjusted.Emit(ValueOpcode::ImageSampleRaw, {image, shared, adjusted.ImageAddress()},
+                  adjusted.AddMemory(memory, 0x1dd0u));
+  }
+  CheckFatal([&] { adjusted.PlanAndTrack(); }, "not a valid runtime value",
+             "finite sampler planning accepted a shared SampleAdjust handle");
+}
+
 void TestFiniteImageBitScanSentinel() {
   namespace CFG = Libs::Graphics::ShaderRecompiler::CFG;
   for (const bool nonzero : {false, true}) {
@@ -3572,6 +3731,7 @@ int main() {
     Run("writable descriptor phi", TestWritableDescriptorPhi);
     Run("conditional sampler phi", TestConditionalSamplerPhi);
     Run("finite image phi cycle", TestFiniteImagePhiCycle);
+    Run("finite inline sampler", TestFiniteInlineSampler);
     Run("finite image bit scan sentinel", TestFiniteImageBitScanSentinel);
     Run("runtime-rooted loop", TestLoopCycleEnteredThroughRuntimeValue);
     Run("invariant loop phi", TestInvariantLoopPhi);

tests/ShaderRecompilerComputeTests.cpp

@@ -5055,7 +5055,7 @@ public:
     for (const auto mode : {0x41u, 0x61u}) {
       for (const auto &groups : empty_dispatches) {
         context.GetRenderExecutor().DispatchDirect(
-            0, scheduler.Current(), groups[0], groups[1], groups[2], mode);
+            0, scheduler.Current(), groups[0], groups[1], groups[2], mode, false);
       }
     }
     Require(name, "HTile fixture",
@@ -9950,6 +9950,113 @@ public:
     return result;
   }
 
+  void CheckScalarLeafCall() {
+    constexpr const char* name = "ScalarLeafCall";
+    constexpr uintptr_t base = 0x0000000204800000ull;
+    constexpr uint64_t size = 0x10000;
+    constexpr auto table = base + 0x100;
+    constexpr auto first = base + 0x1000;
+    constexpr auto second = base + 0x2000;
+    constexpr auto output = base + 0x4000;
+    std::vector<u32> code{
+        EncodeSMovB32(20, InlineU32(0)),
+        0xbeea03ffu, static_cast<u32>(base), 0xbeeb03ffu, static_cast<u32>(base >> 32u),
+        0xf4080135u, 0xfa000000u, 0xbf8cc07fu, 0xf4240382u, 0xfa000020u,
+        0xbf8cc07fu, 0xbf130e80u, 0xbf840004u, 0xf4240402u, 0xfa000028u,
+        0xbf8cc07fu, 0xbe8e210eu, EncodeVop1(0x01, 1, 20),
+        EncodeMubuf0(0x1c, 0, false), EncodeMubuf1(1, 0, 0)};
+    AppendEnd(&code);
+    ShaderMapUserData(reinterpret_cast<uint64_t>(code.data()),
+        {.type = Prospero::ShaderBinaryType::kCs,
+         .code_size_bytes = static_cast<u32>(code.size() * sizeof(u32))});
+    EnsureRuntimeContext();
+    int64_t physical = -1;
+    Require(name, "allocate", LibKernel::Memory::KernelAllocateDirectMemory(
+        0, LibKernel::Memory::KernelGetDirectMemorySize(), size, size, 0, &physical) == 0,
+        "call fixture allocation failed");
+    void* mapped = reinterpret_cast<void*>(base);
+    Require(name, "map", LibKernel::Memory::KernelMapDirectMemory(
+        &mapped, size, 0x3, 0x10, physical, size) == 0 && mapped == reinterpret_cast<void*>(base),
+        "call fixture mapping failed");
+    std::memset(mapped, 0, size);
+    const ShaderBufferResource table_desc{{static_cast<u32>(table),
+        static_cast<u32>(table >> 32u) | (16u << 16u), 8, 0x5204u}};
+    LibKernel::Memory::WriteBacking(base, table_desc.fields, sizeof(table_desc.fields));
+    const std::array leaf{EncodeSMovB32(20, 255), 41u, EncodeSop1(0x21, 125, 14)};
+    for (const auto address : {first, second})
+      LibKernel::Memory::WriteBacking(address, leaf.data(), sizeof(leaf));
+
+    RenderContext context(m_runtime_context);
+    context.InitializeGpu(nullptr);
+    LibKernel::Memory::InstallGpuResources(&context);
+    context.GetGpu().SendCommandSync([&] {
+      CommandProcessor processor(context, 0);
+      processor.BufferInit();
+      context.MapMemory(base, size);
+      auto& cache = context.GetBufferCache();
+      auto& shaders = processor.GetShCtx();
+      shaders.SetCsShader({.data_addr = reinterpret_cast<uint64_t>(code.data()),
+                           .num_thread_x = 1, .num_thread_y = 1, .num_thread_z = 1,
+                           .wave_size = 64, .user_sgpr = 4});
+      ShaderBufferResource out{};
+      out.UpdateAddress48(output);
+      out.fields[2] = 4;
+      out.fields[3] = DstSel(4, 5, 6, 7) |
+          (static_cast<u32>(Prospero::BufferFormat::k32UInt) << 12u);
+      for (u32 i = 0; i < 4; ++i)
+        shaders.SetCsUserSgpr(i, out.fields[i], HW::UserSgprType::Unknown);
+      uint64_t original_id = 0;
+      struct Case { uint64_t address; u32 value; bool update; u32 field = 0; };
+      for (const auto test : {Case{first, 41, false}, Case{first, 41, false},
+                              Case{second, 41, false}, Case{second, 73, true},
+                              Case{first, 41, false}, Case{first, 6, false, 0x1818},
+                              Case{first, 0, false, 0x0a18}}) {
+        Require(name, "refresh function pointer", context.InvalidateMemory(table + 0x20, 8),
+                "function table escaped mapped memory");
+        const uint64_t target = test.address | 3u;
+        LibKernel::Memory::WriteBacking(table + 0x20, &target, sizeof(target));
+        if (test.field != 0) {
+          Require(name, "refresh hardware-field leaf", context.InvalidateMemory(first, sizeof(leaf)),
+                  "hardware-field leaf escaped mapped memory");
+          const std::array field_leaf{0xb96a0000u | test.field,
+              EncodeSMovB32(20, 106), EncodeSop1(0x21, 125, 14)};
+          LibKernel::Memory::WriteBacking(first, field_leaf.data(), sizeof(field_leaf));
+        }
+        if (test.update) {
+          auto [buffer, offset] = cache.ObtainBuffer(test.address + 4, 4, true);
+          buffer->Fill(offset, 4, test.value);
+          Require(name, "GPU-written function code", cache.HasGpuDirtyBytes(test.address + 4, 4),
+                  "code mutation did not remain GPU-owned before refresh");
+        }
+        ShaderComputeInputInfo input{};
+        const auto program = context.GetPipelineCache().GetComputeProgram(
+            shaders.GetCs(), processor.GetCtx().GetShaderRegisters(), input);
+        if (original_id == 0) original_id = program.id;
+        Require(name, "exact function identity",
+                (program.id == original_id) == (test.value == 41),
+                "helper bytes were stale or identical relocated/restored code recompiled");
+        Require(name, "current code dependency",
+                std::ranges::find(input.stage.resources->specialization_reads,
+                    std::pair<uint64_t, uint64_t>{test.address, sizeof(leaf)}) !=
+                    input.stage.resources->specialization_reads.end(),
+                "function relocation retained an old code dependency");
+        processor.DispatchDirect(1, 1, 1, 0x41u);
+        cache.ReadMemory(output, sizeof(u32));
+        u32 actual = 0;
+        Require(name, "scalar leaf execution and return",
+                LibKernel::Memory::TryReadBacking(output, &actual, sizeof(actual)) && actual == test.value,
+                "caller did not observe the leaf's scalar result after return");
+      }
+      context.UnmapMemory(base, size);
+      context.GetCommandScheduler().Finish();
+    });
+    LibKernel::Memory::InstallGpuResources(nullptr);
+    context.ShutdownGpu();
+    Require(name, "release", LibKernel::Memory::KernelCheckedReleaseDirectMemory(physical, size) == 0,
+            "call fixture release failed");
+    std::printf("[gpu]     %-32s ok\n", name);
+  }
+
   void CheckNativeIndirectDispatch() {
     constexpr const char *name = "NativeIndirectDispatch";
     constexpr uintptr_t base = 0x0000000204600000ull;
@@ -10604,7 +10711,7 @@ public:
           for (uint32_t i = 0; i < user_data.size(); i++) {
             shaders.SetCsUserSgpr(i, user_data[i], HW::UserSgprType::Unknown);
           }
-          executor.DispatchDirect(0, scheduler.Current(), (count + 63) / 64, 1, 1, 0x41u);
+          executor.DispatchDirect(0, scheduler.Current(), (count + 63) / 64, 1, 1, 0x41u, false);
         };
         const auto metadata_words = static_cast<uint32_t>(metadata_size / 4);
         fill_metadata(metadata_words);
@@ -15249,7 +15356,8 @@ public:
                 const Image *storage_image = nullptr,
                 const Image *storage_image_uint = nullptr,
                 vk::Sampler sampler = nullptr,
-                std::span<const Image> sampled_resources = {}) {
+                std::span<const Image> sampled_resources = {},
+                std::span<const vk::Sampler> sampler_resources = {}) {
     using Kind = ShaderRecompiler::IR::DescriptorBindingKind;
     const auto &layout = compiled.program.bindings;
     auto shader_data = compiled.packed_user_data;
@@ -15564,11 +15672,12 @@ public:
       writes.push_back(write);
     }
     if (Count(Kind::Samplers) != 0) {
-      Require(test.name, "dispatch", sampler != nullptr,
-              "sampler descriptor requested but no sampler was provided");
+      Require(test.name, "dispatch", sampler_resources.empty()
+                  ? sampler != nullptr : sampler_resources.size() == Count(Kind::Samplers),
+              "sampler resources must match the native sampler array");
       sampler_infos.resize(Count(Kind::Samplers));
-      for (auto &info : sampler_infos) {
-        info.sampler = sampler;
+      for (u32 index = 0; index < sampler_infos.size(); ++index) {
+        sampler_infos[index].sampler = sampler_resources.empty() ? sampler : sampler_resources[index];
       }
       vk::WriteDescriptorSet write{};
       write.dstSet = descriptor_set;
@@ -28387,7 +28496,7 @@ TestCase ScalarMemoryLoadVariants() {
   std::vector<u32> expected = initial;
   expected.insert(expected.end(), initial.begin(), initial.end());
 
-  return {"ScalarMemoryLoadVariants",
+  TestCase test{"ScalarMemoryLoadVariants",
           code,
           initial,
           expected,
@@ -28396,6 +28505,9 @@ TestCase ScalarMemoryLoadVariants() {
            O::S_BUFFER_LOAD_DWORDX2, O::S_BUFFER_LOAD_DWORDX4,
            O::S_BUFFER_LOAD_DWORDX8, O::S_BUFFER_LOAD_DWORDX16, O::V_MOV_B32,
            O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
+  test.bda_mappings = {{0, 0}};
+  test.required_spirv = {"get_bda_pointer"};
+  return test;
 }
 
 TestCase ScalarBufferOffsetAlignmentAndCarry() {
@@ -28439,11 +28551,43 @@ TestCase ScalarLoadSignedImmediateOffsetAddsSoffset() {
   AppendStoreSgpr(&code, 1, 0);
   AppendEnd(&code);
 
-  return {"ScalarLoadSignedImmediateOffsetAddsSoffset",
+  TestCase test{"ScalarLoadSignedImmediateOffsetAddsSoffset",
           code,
           {0x11111111u, 0x22222222u},
           {0x22222222u, 0x22222222u},
           {O::S_MOV_B32, O::S_LOAD_DWORD, O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
+  test.bda_mappings = {{0, 0}};
+  test.required_spirv = {"get_bda_pointer"};
+  return test;
+}
+
+TestCase ScalarPayloadBeforeAliasedAtomic() {
+  using O = ShaderOpcode;
+  constexpr u32 GuestBase = 0x10000u;
+  std::vector<u32> code;
+  AppendSMovLiteral(&code, 8, GuestBase);
+  AppendSMovLiteral(&code, 9, 0u);
+  code.insert(code.end(), {EncodeSmem0(0x00, 12, 4), EncodeSmem1(24, 125),
+                           EncodeSopp(0x0c, 0xc07f)});
+  AppendVMovU32(&code, 0, 1u << 20u);
+  AppendVMovU32(&code, 1, 0u);
+  // Same 64-bit OR and byte offset as SAROS b62b494cb2567011.
+  code.insert(code.end(), {0xe1680018u, 0x80000000u});
+  AppendStoreSgpr(&code, 12, 0);
+  AppendEnd(&code);
+  TestCase test{"ScalarPayloadBeforeAliasedAtomic", code,
+                {0u, 0u, 0u, 0u, 0u, 0u, 0x10u, 0x11223344u},
+                {0x10u, 0u, 0u, 0u, 0u, 0u, 0x100010u, 0x11223344u},
+                {O::S_MOV_B32, O::S_LOAD_DWORD, O::S_WAITCNT, O::V_MOV_B32,
+                 O::BUFFER_ATOMIC_OR_X2, O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
+  test.bda_mappings = {{GuestBase, 0}};
+  test.required_spirv = {"get_bda_pointer"};
+  test.forbidden_spirv = {"flattened_srt"};
+  test.compute_info.threads_num[0] = 1;
+  test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;
+  test.compute_info.workgroup_register = 4;
+  test.has_compute_info = true;
+  return test;
 }
 
 TestCase BufferLoadStore() {
@@ -31034,23 +31178,26 @@ TestCase BranchVccnzUsesCarryProducedWaveMask() {
   return test;
 }
 
-TestCase ScalarLoadAlignsComponentsAndMasksAddress() {
+TestCase ScalarLoadAlignsComponents() {
   using O = ShaderOpcode;
 
   std::vector<u32> code;
   AppendSMovLiteral(&code, 0, 1u);
   AppendSMovLiteral(&code, 2, 3u);
-  AppendSMovLiteral(&code, 3, 0xffff0000u);
+  AppendSMovLiteral(&code, 3, 0u); // S_LOAD uses the full base; only buffer descriptors mask it.
   code.push_back(EncodeSmem0(0x00, 1, 1));
   code.push_back(EncodeSmem1(3, 0));
   AppendStoreSgpr(&code, 1, 0);
   AppendEnd(&code);
 
-  return {"ScalarLoadAlignsComponentsAndMasksAddress",
+  TestCase test{"ScalarLoadAlignsComponents",
           code,
           {0x11111111u, 0x22222222u},
           {0x11111111u, 0x22222222u},
           {O::S_MOV_B32, O::S_LOAD_DWORD, O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
+  test.bda_mappings = {{0, 0}};
+  test.required_spirv = {"get_bda_pointer"};
+  return test;
 }
 
 TestCase ScalarLoadAlignsDynamicBase() {
@@ -34802,6 +34949,122 @@ void CheckIndirectBufferStore(VulkanHarness &vulkan) {
   std::printf("[compute] %-32s ok\n", test.name);
 }
 
+void CheckFiniteInlineSamplerPhi(VulkanHarness &vulkan) {
+  TestCase test;
+  test.name = "FiniteInlineSamplerPhi";
+  auto &code = test.code;
+  code.insert(code.end(), {EncodeSop1(0x04, 36, InlineU32(0)),
+      EncodeSop1(0x04, 38, InlineU32(0)), EncodeSop1(0x04, 24, 126),
+      EncodeVop1(0x01, 10, 16), EncodeVopc(0xd5, InlineU32(0), 10)});
+  const auto skip_sampler = code.size();
+  code.push_back(EncodeSopp(0x08));
+  // Actual correlated inline sampler words from 905b2e2c: zero or {0,fff000,a500000,0}.
+  code.insert(code.end(), {EncodeSop2(0x24, 37, InlineU32(12), InlineU32(12)),
+      EncodeSop1(0x04, 38, 255), 0x0a500000u});
+  code[skip_sampler] = EncodeSopp(0x08, code.size() - skip_sampler - 1u);
+  code.push_back(EncodeSop1(0x04, 126, 24));
+  AppendVMovLiteral(&code, 20, std::bit_cast<u32>(0.5f));
+  AppendVMovLiteral(&code, 21, std::bit_cast<u32>(0.5f));
+  AppendVMovLiteral(&code, 1, 0x12345678u);
+  AppendVMovLiteral(&code, 2, 0x87654321u);
+  code.insert(code.end(), {EncodeVopc(0xd4, InlineU32(2), 10),
+      EncodeMimg0(0x27, 1), EncodeMimg1(1, 20, 0, 9),
+      EncodeMimg0(0x27, 1), EncodeMimg1(2, 20, 2, 9),
+      EncodeSop1(0x04, 126, 24), EncodeVop1(0x01, 30, 16),
+      EncodeVop2(0x1a, 31, InlineU32(3), 30),
+      EncodeMubuf0(0x1c), EncodeMubuf1(1, 12, 31),
+      EncodeMubuf0(0x1c, 4), EncodeMubuf1(2, 12, 31)});
+  AppendEnd(&code);
+  test.initial.assign(7, 0xa5a5a5a5u);
+  test.expected = {std::bit_cast<u32>(4.0f), 0xfffffff9u,
+      std::bit_cast<u32>(2.0f), 0xfffffff9u, 0x12345678u, 0x87654321u, 0xa5a5a5a5u};
+  test.user_data = MakeSampledTextureData(Prospero::BufferFormat::k32_32_32_32Float);
+  test.user_data[3] |= static_cast<u32>(Prospero::ImageType::kColor2D) << 28u;
+  const auto signed_image = MakeSampledTextureData(Prospero::BufferFormat::k32SInt);
+  std::copy_n(signed_image.begin(), 8, test.user_data.begin() + 8);
+  test.user_data[11] |= static_cast<u32>(Prospero::ImageType::kColor2D) << 28u;
+  test.user_data[50] = test.initial.size() * sizeof(u32);
+  test.user_data[51] = 3u << 28u;
+  test.has_user_data = true;
+  test.compute_info.threads_num[0] = test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;
+  test.compute_info.wave_size = 32;
+  test.compute_info.group_id[0] = true;
+  test.compute_info.workgroup_register = 16;
+  test.has_compute_info = true;
+  test.dispatch_x = 3;
+  test.opcodes = {ShaderOpcode::S_MOV_B64, ShaderOpcode::S_BFM_B32,
+      ShaderOpcode::S_CBRANCH_EXECZ, ShaderOpcode::V_MOV_B32, ShaderOpcode::V_CMPX_NE_U32,
+      ShaderOpcode::V_CMPX_GT_U32, ShaderOpcode::V_LSHLREV_B32,
+      ShaderOpcode::IMAGE_SAMPLE, ShaderOpcode::BUFFER_STORE_DWORD, ShaderOpcode::S_ENDPGM};
+  test.required_spirv = {"OpPhi", "NonUniform", "OpImageSampleExplicitLod"};
+  auto compiled = CompileCase(test, vulkan.SubgroupSize());
+  std::set<u32> nonuniform_ids;
+  std::vector<std::pair<u32, u32>> sampled_ids;
+  for (size_t offset = 5; offset < compiled.spirv.size();) {
+    const auto words = std::span(compiled.spirv).subspan(offset, compiled.spirv[offset] >> 16u);
+    if (static_cast<spv::Op>(words[0] & 0xffffu) == spv::OpDecorate &&
+        words[2] == spv::DecorationNonUniform) nonuniform_ids.insert(words[1]);
+    if (static_cast<spv::Op>(words[0] & 0xffffu) == spv::OpSampledImage)
+      sampled_ids.emplace_back(words[2], words[4]);
+    offset += words.size();
+  }
+  Require(test.name, "dynamic sampler decorations", sampled_ids.size() >= 2 &&
+              std::ranges::all_of(sampled_ids, [&](const auto &ids) {
+                return nonuniform_ids.contains(ids.first) && nonuniform_ids.contains(ids.second);
+              }),
+          "fetched sampler or sampled image lost its NonUniform decoration");
+  const auto &samplers = compiled.program.info.samplers;
+  u32 roots = 0;
+  for (const auto &root : samplers) {
+    if (root.indirect_resources.empty()) continue;
+    ++roots;
+    Require(test.name, "correlated sampler candidates", root.indirect_resources.size() == 2,
+            "finite sampler was folded or lost a correlated branch state");
+    for (const auto child : root.indirect_resources)
+      Require(test.name, "sampler class remapping", child < samplers.size() &&
+                  samplers[child].force_point_filtering == root.force_point_filtering &&
+                  samplers[child].integer_border == root.integer_border,
+              "a finite candidate retained another image class's sampler");
+  }
+  Require(test.name, "finite sampler class variants", roots == 2 && samplers.size() == 4 &&
+              compiled.resources.samplers.size() == 2,
+          "the two-state sampler did not retain both float and signed image classes");
+  constexpr std::array<u32, 4> zero_sampler{};
+  constexpr std::array<u32, 4> linear_sampler{0, 0x00fff000u, 0x0a500000u, 0};
+  Require(test.name, "inline sampler states",
+          std::ranges::all_of(compiled.resources.samplers, [&](const auto &value) {
+            return std::equal(zero_sampler.begin(), zero_sampler.end(), value.dwords.begin()) ||
+                std::equal(linear_sampler.begin(), linear_sampler.end(), value.dwords.begin());
+          }) && compiled.resources.samplers[0] != compiled.resources.samplers[1],
+          "descriptor words were mixed between the two native sampler states");
+  auto pixels = MakeRgbaImage(4, 4);
+  for (u32 y = 0; y < 4; ++y)
+    for (u32 x = 2; x < 4; ++x) SetRgbaPixel(&pixels, 4, x, y, std::bit_cast<u32>(4.0f), 0, 0, 0);
+  std::array<VulkanHarness::Image, 2> images{
+      vulkan.CreateImageMips(test.name, 4, 4, vk::Format::eR32G32B32A32Sfloat,
+          vk::ImageUsageFlagBits::eSampled, {pixels}, 4, vk::ImageLayout::eShaderReadOnlyOptimal,
+          vk::ImageType::e2D, vk::ImageViewType::e2D, 1),
+      vulkan.CreateImageMips(test.name, 4, 4, vk::Format::eR32Sint,
+          vk::ImageUsageFlagBits::eSampled, {std::vector<u32>(16, 0xfffffff9u)}, 1,
+          vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageType::e2D, vk::ImageViewType::e2D, 1)};
+  std::vector<vk::Sampler> native_samplers;
+  for (const auto &resource : samplers) {
+    ShaderSamplerResource descriptor{};
+    std::copy_n(compiled.resources.samplers[resource.snapshot_index].dwords.begin(), 4, descriptor.fields);
+    const auto filter = !resource.force_point_filtering && descriptor.XyMagFilter() == 1
+        ? vk::Filter::eLinear : vk::Filter::eNearest;
+    native_samplers.push_back(vulkan.CreateSampler(test.name, 1, filter));
+  }
+  auto output = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
+  vulkan.Dispatch(test, compiled, output, nullptr, nullptr, nullptr, nullptr, nullptr, images, native_samplers);
+  CompareWords(test, "point, linear, remapped signed sampler and inactive output", test.expected,
+      vulkan.ReadBuffer(test.name, output, test.expected.size()));
+  vulkan.DestroyBuffer(&output);
+  for (const auto sampler : native_samplers) vulkan.Device().destroySampler(sampler);
+  for (auto &image : images) vulkan.DestroyImage(&image);
+  std::printf("[compute] %-32s ok\n", test.name);
+}
+
 void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
   constexpr const char *name = "IndirectImageKeySwitch";
   constexpr uint32_t mapping_capacity = 1793u;
@@ -35047,42 +35310,87 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
   std::vector<VulkanHarness::Image> textures;
   for (u32 resource = 0; resource < 7u; ++resource) {
     const auto layers = resource == 4u ? 6u : 1u;
-    std::vector<u32> pixels(layers * 4u, std::bit_cast<u32>(float(resource)));
+    std::vector<u32> pixels(layers * 16u, std::bit_cast<u32>(float(resource)));
+    std::vector<u32> mip(layers * 4u,
+                         std::bit_cast<u32>(resource == 0u ? 0.0f : float(resource + 10u)));
     textures.push_back(vulkan.CreateImageMips(
-        test.name, 1, 1, vk::Format::eR32G32B32A32Sfloat,
-        vk::ImageUsageFlagBits::eSampled, {pixels}, 4u,
+        test.name, 2, 2, vk::Format::eR32G32B32A32Sfloat,
+        vk::ImageUsageFlagBits::eSampled, {pixels, mip}, 4u,
         vk::ImageLayout::eShaderReadOnlyOptimal, vk::ImageType::e2D,
         resource == 4u ? vk::ImageViewType::e2DArray : vk::ImageViewType::e2D, layers));
   }
   const auto native_sampler = vulkan.CreateSampler(test.name);
-  for (const u32 wave_size : {32u, 64u}) {
-    compute.wave_size = wave_size;
-    compute.host_subgroup_size = vulkan.SubgroupSize();
-    compute.threads_num[0] = wave_size;
-    compute.threads_num[1] = compute.threads_num[2] = 1u;
-    compiled.program.wave_size = wave_size;
-    compiled.spirv = ShaderRecompiler::Spirv::EmitProgram(compiled.program, {.compute = &compute});
-    compiled.packed_user_data.resize(compiled.program.bindings.ShaderDataDwords());
-    ValidateSpirv(test.name, compiled.spirv);
-    const auto halves = wave_size > compute.host_subgroup_size ? 2u : 1u;
-    Require(test.name, "native array run samples", tools.Disassemble(compiled.spirv, &text) &&
-                CountText(text, "OpImageSampleExplicitLod") == 3u * halves,
-            "mixed runs expanded into per-image samples");
-    test.initial.assign(wave_size, 0xdeadbeefu);
-    test.expected.assign(wave_size, 0u);
-    for (u32 lane = 0; lane < ordinals.size(); ++lane) {
-      test.expected[lane] = std::bit_cast<u32>(float(root.indirect_resources[ordinals[lane]]));
+  for (const bool fetch : {false, true}) {
+    if (fetch) {
+      test.name = "IndirectImageFetchContiguousRuns";
+      const auto position = std::ranges::find_if(*block, [&](const Inst &inst) {
+        return &inst == &sample;
+      });
+      auto &read = *block->PrependNewInst(
+          position, ValueOpcode::ImageRead,
+          {Value(&image), Value(&address), Value(true)}, memory_flag_bits);
+      sample.ReplaceUsesWith(Value(&read), false);
+      sampler.Invalidate();
+      compiled.program.info.samplers.clear();
+      compiled.program.info.sampled_pairs.clear();
+      compiled.program.info.images[4].cube = false;
+      compiled.program.memory_info[0].image_sample_flags = 0;
+      compiled.program.memory_info[0].image_has_mip = true;
+      for (u32 component = 0; component < 4u; ++component)
+        address.SetArg(component, Value(0u));
+      // Native DIM=2DArray places LOD after the layer, even for a 2D candidate.
+      address.SetArg(3, Value(1u));
+      compiled.program.shader_info_complete = false;
+      CollectShaderInfo(compiled.program, {.compute = &compute});
     }
-    auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
-    vulkan.Dispatch(test, compiled, buffer, nullptr, nullptr, nullptr, nullptr,
-                    native_sampler, textures);
-    const auto actual = vulkan.ReadBuffer(test.name, buffer, test.expected.size());
-    vulkan.DestroyBuffer(&buffer);
-    CompareWords(test, "root, native slots, cube and mapping bounds", test.expected, actual);
+    for (const u32 wave_size : {32u, 64u}) {
+      compute.wave_size = wave_size;
+      compute.host_subgroup_size = vulkan.SubgroupSize();
+      compute.threads_num[0] = wave_size;
+      compute.threads_num[1] = compute.threads_num[2] = 1u;
+      compiled.program.wave_size = wave_size;
+      compiled.spirv = ShaderRecompiler::Spirv::EmitProgram(compiled.program, {.compute = &compute});
+      compiled.packed_user_data.resize(compiled.program.bindings.ShaderDataDwords());
+      ValidateSpirv(test.name, compiled.spirv);
+      const auto halves = wave_size > compute.host_subgroup_size ? 2u : 1u;
+      Require(test.name, "native array image runs", tools.Disassemble(compiled.spirv, &text) &&
+                  CountText(text, fetch ? "OpImageFetch" : "OpImageSampleExplicitLod") == 3u * halves,
+              "mixed runs expanded into per-image accesses");
+      if (fetch) {
+        std::vector<u32> decorated;
+        std::vector<u32> fetched_images;
+        for (size_t offset = 5; offset < compiled.spirv.size();) {
+          const auto words = std::span<const u32>(compiled.spirv).subspan(
+              offset, compiled.spirv[offset] >> 16u);
+          const auto opcode = static_cast<spv::Op>(words[0] & 0xffffu);
+          if (opcode == spv::OpDecorate && words[2] == spv::DecorationNonUniform)
+            decorated.push_back(words[1]);
+          if (opcode == spv::OpImageFetch) fetched_images.push_back(words[3]);
+          offset += words.size();
+        }
+        Require(test.name, "nonuniform fetch operands",
+                std::ranges::count_if(fetched_images, [&](u32 operand) {
+                  return std::ranges::find(decorated, operand) != decorated.end();
+                }) == 2u * halves,
+                "a dynamically selected fetch image lacks its nonuniform decoration");
+      }
+      test.initial.assign(wave_size, 0xdeadbeefu);
+      test.expected.assign(wave_size, 0u);
+      for (u32 lane = 0; lane < ordinals.size(); ++lane) {
+        const auto resource = root.indirect_resources[ordinals[lane]];
+        test.expected[lane] = std::bit_cast<u32>(float(resource + (fetch && resource != 0u ? 10u : 0u)));
+      }
+      auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
+      vulkan.Dispatch(test, compiled, buffer, nullptr, nullptr, nullptr, nullptr,
+                      fetch ? vk::Sampler{} : native_sampler, textures);
+      const auto actual = vulkan.ReadBuffer(test.name, buffer, test.expected.size());
+      vulkan.DestroyBuffer(&buffer);
+      CompareWords(test, "root, native slots, cube and mapping bounds", test.expected, actual);
+    }
+    std::printf("[compute] %-32s ok\n", test.name);
   }
   vulkan.Device().destroySampler(native_sampler);
   for (auto &texture : textures) vulkan.DestroyImage(&texture);
-  std::printf("[compute] %-32s ok\n", test.name);
 }
 
 TestCase ImageStoreMipSelectsPpsa01340Descriptor() {
@@ -36691,7 +36999,8 @@ std::vector<TestCase> MakeCases() {
   AddCase(ScalarMemoryLoadVariants);
   AddCase(ScalarBufferOffsetAlignmentAndCarry);
   AddCase(ScalarLoadSignedImmediateOffsetAddsSoffset);
-  AddCase(ScalarLoadAlignsComponentsAndMasksAddress);
+  AddCase(ScalarPayloadBeforeAliasedAtomic);
+  AddCase(ScalarLoadAlignsComponents);
   AddCase(ScalarLoadAlignsDynamicBase);
   cases.push_back(ScalarBufferFromLoopReadlane(32));
   cases.push_back(ScalarBufferFromLoopReadlane(64));
@@ -39615,17 +39924,19 @@ void CheckResourcePlanHandoff() {
     ValidateSpirv(name, compiled.spirv);
     const auto &bindings = compiled.program.bindings;
     Require(name, "GPU data requirements",
-            compiled.program.info.user_data_registers.empty() && bindings.descriptor_counts[0] == 1 &&
-                (bindings.descriptor_counts[static_cast<size_t>(DescriptorBindingKind::FlattenedSrt)] != 0) ==
+            compiled.program.info.user_data_registers ==
+                (numeric_read ? std::vector<u32>{0, 1} : std::vector<u32>{}) &&
+                bindings.descriptor_counts[0] == 1 &&
+                bindings.descriptor_counts[static_cast<size_t>(DescriptorBindingKind::FlattenedSrt)] == 0 &&
+                (bindings.descriptor_counts[static_cast<size_t>(DescriptorBindingKind::BdaPagetable)] != 0) ==
                     numeric_read,
-            "descriptor-only shader retained SRT uploads or live scalar data was removed");
+            "descriptor-only shader retained SRT uploads or native scalar payload lost its address");
 
     memory[0] = 0x2000u;
-    memory[4] = 13u;
     Require(name, "independent host plan",
             MaterializeResources(plan, runtime, snapshot, specialization) &&
                 snapshot.buffers[0].dwords[0] == 0x2000u &&
-                (!numeric_read || snapshot.flattened_srt.back() == 13u),
+                snapshot.flattened_srt.size() == 4u,
             "compiled shader cleanup invalidated or froze the host resource plan");
   }
   std::printf("[host]    %-32s ok\n", name);
@@ -41240,6 +41551,53 @@ void CheckShaderFusion() {
   std::printf("[host]    %-32s ok\n", "ShaderFusion");
 }
 
+void CheckAgcSystemTable(RenderContext &renderer) {
+  constexpr const char *name = "AgcSystemTable";
+  constexpr uint64_t driver_base = 0x0fe0000000ull;
+  constexpr uint64_t table_address = driver_base + 0x40000;
+  constexpr std::array<uint32_t, 16> native_validator{
+      0xbeeb03ff, 0x00000021, 0xb96a1818, 0xbf06106a,
+      0xbf800000, 0x8590807e, 0xb96a0a18, 0xbf066a80,
+      0xbf800000, 0x85ea807e, 0x88ea106a, 0xbf870003,
+      0xbefc03ff, 0x8a6ca000, 0xbf920009, 0xbefd210e};
+  LibKernel::Memory::InstallGpuResources(&renderer);
+  uint32_t state = 0xabcdef01u;
+  Require(name, "driver startup", Gen5::AgcInit(&state, 13) == 0 && state == 0xabcdef01u,
+          "Agc startup failed or changed its ignored public state pointer");
+  LibKernel::Memory::VirtualQueryInfo mapping{};
+  Require(name, "shared driver backing",
+          LibKernel::Memory::KernelVirtualQuery(reinterpret_cast<void *>(driver_base), 0,
+                                                 &mapping, sizeof(mapping)) == 0 &&
+              mapping.start == driver_base && mapping.end == driver_base + 0x200000 &&
+              mapping.memory_type == 12 && mapping.protection == 0x33 && mapping.is_direct &&
+              std::strcmp(mapping.name, "SceAgcDriver") == 0 &&
+              renderer.IsMapped(driver_base, 0x200000),
+          "driver startup bypassed shared DMEM metadata or GPU range registration");
+  ShaderBufferResource descriptor{};
+  std::array<uint64_t, 16> functions{};
+  std::array<uint32_t, 16> code{};
+  Require(name, "ordered-count function backing",
+          LibKernel::Memory::TryReadBacking(table_address, &descriptor, sizeof(descriptor)) &&
+              descriptor.Base48() == table_address + 0x60 && descriptor.Stride() == 16 &&
+              descriptor.NumRecords() == 8 && descriptor.fields[3] == 0x5204 &&
+              LibKernel::Memory::TryReadBacking(descriptor.Base48(), functions.data(),
+                                                 sizeof(functions)) &&
+              functions[4] != 0 && (functions[4] & 0xffu) == 0 && functions[5] == 0 &&
+              LibKernel::Memory::TryReadBacking(functions[4], code.data(), sizeof(code)) &&
+              code == native_validator,
+          "system descriptor did not reach the queue validator and its user data");
+  auto *instrumentation = reinterpret_cast<uint64_t *>(table_address + 0x18);
+  Require(name, "default instrumentation", *instrumentation == 0,
+          "unused vertex validation was enabled at startup");
+  *instrumentation = 0x12340000u;
+  Require(name, "repeat startup", Gen5::AgcInit(nullptr, 13) == 0 &&
+              *instrumentation == 0x12340000u,
+          "repeat Agc initialization reset existing system state");
+  *instrumentation = 0;
+  LibKernel::Memory::InstallGpuResources(nullptr);
+  std::printf("[host]    %-32s ok\n", name);
+}
+
 struct CommandBufferLayout {
   using Callback = KYTY_SYSV_ABI bool (*)(Gen5::CommandBuffer *, uint32_t,
                                           void *);
@@ -42591,6 +42949,10 @@ int main(int argc, char **argv) {
     VulkanHarness vulkan;
     RunCase(&vulkan, BufferOffsetsUsePackedWords(false));
     RunCase(&vulkan, BufferOffsetsUsePackedWords(true));
+    RunCase(&vulkan, ScalarMemoryLoadVariants());
+    RunCase(&vulkan, ScalarLoadSignedImmediateOffsetAddsSoffset());
+    RunCase(&vulkan, ScalarLoadAlignsComponents());
+    RunCase(&vulkan, ScalarPayloadBeforeAliasedAtomic());
     return 0;
   }
   if (argc == 2 && std::strcmp(argv[1], "--buffer-publication-only") == 0) {
@@ -42659,6 +43021,8 @@ int main(int argc, char **argv) {
   if (argc == 2 && std::strcmp(argv[1], "--gpu-command-lane-only") == 0) {
     VulkanHarness vulkan;
     vulkan.CheckGpuCommandLane();
+    vulkan.CheckScalarLeafCall();
+    CheckAgcSystemTable(vulkan.RuntimeRenderer());
     CheckPm4IndirectControlFlow(vulkan.RuntimeRenderer());
     CheckPm4WaitResume(vulkan.RuntimeRenderer());
     CheckPm4RewindResume(vulkan.RuntimeRenderer());
@@ -42945,6 +43309,7 @@ int main(int argc, char **argv) {
     CheckImageSamplerSpecialization();
     VulkanHarness vulkan;
     CheckIndirectImageKeySwitch(vulkan);
+    CheckFiniteInlineSamplerPhi(vulkan);
     RunCase(&vulkan, ImageCubeGradientsPreserveDerivatives());
     return 0;
   }
@@ -43091,6 +43456,7 @@ int main(int argc, char **argv) {
   CheckPm4DepthControlHighBits(vulkan.RuntimeRenderer());
   CheckPm4DepthRenderOverride(vulkan.RuntimeRenderer());
   CheckShaderFusion();
+  CheckAgcSystemTable(vulkan.RuntimeRenderer());
   CheckPm4WaitPackets(vulkan.RuntimeRenderer());
   CheckPm4DrawIndirectMultiPacket(vulkan.RuntimeRenderer());
   CheckPm4ContextStateOperations(vulkan.RuntimeRenderer());
@@ -43107,6 +43473,7 @@ int main(int argc, char **argv) {
   CheckRuntimeBufferRecords(vulkan);
   CheckComputeThreadDimensions(vulkan);
   CheckIndirectImageKeySwitch(vulkan);
+  CheckFiniteInlineSamplerPhi(vulkan);
   CheckWave64WholeWaveResults();
   CheckPs5GameExampleImageClearRuntimeShape();
   vulkan.CheckSchedulerTimeline();
@@ -43151,6 +43518,7 @@ int main(int argc, char **argv) {
   }
   vulkan.CheckGpuSuspendPoint();
   vulkan.CheckGpuCommandLane();
+  vulkan.CheckScalarLeafCall();
   if (skipped_device_checks) {
     std::printf(
         "ShaderRecompilerComputeTests: device rasterization checks skipped, this "

tests/shaderCfgTests.cpp

@@ -15,6 +15,7 @@
 #include "graphics/shader/recompiler/frontend/cfg/ShaderCFG.h"
 #include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 #include "graphics/shader/recompiler/frontend/translate/Translate.h"
+#include "graphics/shader/recompiler/frontend/translate/Translator.h"
 #include "graphics/shader/recompiler/ir/IREmitter.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/recompiler/ir/passes/ConstantPropagation.h"
@@ -10308,6 +10309,218 @@ void TestNewShaderRecompilerBranchConditionForms() {
   }
 }
 
+void TestGraphicsComputeGetreg() {
+  using namespace ShaderRecompiler;
+  for (const auto [word, expected] : {std::pair{0xb96a1818u, 6u}, std::pair{0xb96a0a18u, 0u}}) {
+    Decoder::Instruction decoded;
+    Decoder::DecodeInstruction(std::span(&word, 1), 0, decoded);
+    Check(decoded.opcode == Decoder::Opcode::S_GETREG_B32 &&
+              decoded.dst.kind == Decoder::OperandKind::VccLo &&
+              decoded.src0.value == (word & 0xffffu), "GETREG field or VCC destination decoded incorrectly");
+    for (const auto wave : {32u, 64u}) {
+      IR::Program program;
+      program.wave_size = wave;
+      IR::Block block;
+      Frontend::Translator translator(program, &block, 1, false, true);
+      translator.TranslateInstruction(decoded);
+      bool wrote_vcc = false;
+      for (const auto& inst : block) {
+        Check(inst.GetOpcode() != IR::ValueOpcode::SetScc, "GETREG changed SCC");
+        if (inst.GetOpcode() == IR::ValueOpcode::SetVccLo) {
+          Check(inst.Arg(0).IsImmediate() && inst.Arg(0).U32() == expected,
+                "graphics compute hardware field has the wrong value");
+          wrote_vcc = true;
+        }
+      }
+      Check(wrote_vcc, "GETREG did not write VCC_LO");
+    }
+  }
+  const uint32_t shader[] = {0xb96a1818u, EncodeSopp(0x01)};
+  HW::ComputeShaderInfo regs{};
+  regs.cs_regs.data_addr = reinterpret_cast<uint64_t>(shader);
+  ShaderMappedData mapped{};
+  mapped.code_size_bytes = sizeof(shader);
+  ShaderMapUserData(regs.cs_regs.data_addr, mapped);
+  HW::ShaderRegisters sh{};
+  ShaderComputeInputInfo graphics{}, async{};
+  async.async_compute = true;
+  (void)PrepareProgram(regs, sh, graphics);
+  (void)PrepareProgram(regs, sh, async);
+  Check(!graphics.async_compute && async.async_compute &&
+            MakeStageStaticKey(graphics) != MakeStageStaticKey(async) &&
+            MakeStageStaticKey(graphics).size() == MakeStageStaticKey(async).size(),
+        "compute input reset or cache identity lost queue provenance");
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
+  auto options = MakeCompileOptions(ShaderType::Compute);
+  options.input_info.compute = &async;
+  ExpectFatal([&] { (void)TranslateProgram(shader, options); }, "async GETREG used graphics constants");
+  auto vertex = MakeCompileOptions(ShaderType::Vertex);
+  ExpectFatal([&] { (void)TranslateProgram(shader, vertex); }, "vertex GETREG used compute constants");
+  options.input_info.compute = &graphics;
+  const uint32_t unsupported[] = {0xb96af818u, EncodeSopp(0x01)};
+  Decoder::Program decoded;
+  Decoder::DecodeProgram(unsupported, decoded);
+  Check(decoded.instructions[0].src0.value == 0xf818u, "GETREG selector was sign extended");
+  ExpectFatal([&] { (void)TranslateProgram(unsupported, options); }, "unsupported GETREG field was accepted");
+#endif
+}
+
+void TestMemoryFedScalarLeafCall() {
+  using namespace ShaderRecompiler;
+  constexpr uint64_t base = 0xfe0040000ull;
+  const uint32_t shader[] = {
+      0xbeea03ffu, 0xe0040000u, 0xbeeb03ffu, 0x0000000fu,
+      0xf4080135u, 0xfa000000u, 0xbf8cc07fu, 0xf4240382u, 0xfa000020u,
+      0xbf8cc07fu, 0xbf130e80u, 0xbf840004u, 0xf4240402u, 0xfa000028u,
+      0xbf8cc07fu, 0xbe8e210eu, 0xbf810000u};
+  Decoder::Program decoded;
+  Decoder::DecodeProgram(shader, decoded);
+  Check(decoded.has_swap_pc, "decoded scalar call was not recorded");
+  const uint32_t literal[] = {EncodeSMovB32(0, 255), 0xbe8e210eu, EncodeSopp(0x01)};
+  Decoder::DecodeProgram(literal, decoded);
+  Check(!decoded.has_swap_pc && decoded.instructions.size() == 2,
+        "reused decoder retained a call or treated a literal as SWAPPC");
+  const uint32_t null_call[] = {EncodeSop1(0x21, 125, 14), EncodeSopp(0x01)};
+  Decoder::DecodeProgram(null_call, decoded);
+  Check(!decoded.has_swap_pc && decoded.instructions.front().opcode == Decoder::Opcode::S_SETPC_B64,
+        "NULL-destination SWAPPC was recorded as a call");
+  struct Memory {
+    std::array<uint32_t, 128> words{};
+    uint64_t end = base + 0x10c;
+    bool clean = true;
+  } memory;
+  std::copy_n(std::array{0xe0040060u, 0x0010000fu, 8u, 0x5204u}.begin(),
+              4, memory.words.begin());
+  memory.words[32] = static_cast<uint32_t>(base + 0x103); // SWAPPC aligns the old pair.
+  memory.words[33] = static_cast<uint32_t>(base >> 32u);
+  memory.words[64] = EncodeSMovB32(3, 255);
+  memory.words[65] = 0x12345678;
+  memory.words[66] = EncodeSop1(0x21, 125, 14);
+  const auto read = [](void *data, uint64_t address, std::span<uint32_t> out) {
+    const auto& m = *static_cast<Memory*>(data);
+    if (!m.clean || address < base || address > m.end || out.size_bytes() > m.end - address)
+      return false;
+    std::memcpy(out.data(), reinterpret_cast<const uint8_t*>(m.words.data()) + address - base,
+                out.size_bytes());
+    return true;
+  };
+  IR::SrtRuntime runtime{.read_memory = read, .userdata = &memory, .read_specialization_memory = read};
+  const auto options = MakeCompileOptions(ShaderType::Compute);
+  auto source = PrepareShaderSource(shader, options);
+  Check(source.call && source.decoded.has_swap_pc &&
+            source.decoded.instructions[source.call->instruction].branch_target == UINT32_MAX,
+        "memory-fed SWAPPC did not retain an unresolved target");
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
+  ExpectFatal([&] { (void)TranslateProgram(source.decoded, options); },
+              "unresolved SWAPPC silently fell through");
+#endif
+  const auto& linked = RefreshShaderSource(source, runtime);
+  Check(linked.has_swap_pc && source.revision == 1 && linked.instructions.back().branch_target == 0x40 &&
+            linked.instructions[source.call->instruction].branch_target == 0x44 &&
+            source.reads.back() == std::pair<uint64_t, uint64_t>{base + 0x100, 12},
+        "aliased call target, exact leaf range, or continuation was lost");
+  const auto* storage = linked.instructions.data();
+  const auto* scratch = source.linked->observed_function.data();
+  (void)RefreshShaderSource(source, runtime);
+  Check(source.revision == 1 && source.linked->decoded.instructions.data() == storage &&
+            source.linked->observed_function.data() == scratch,
+        "unchanged leaf rebuilt decoded instructions or refresh storage");
+  auto translated = TranslateProgram(linked, options);
+  translated.program.source_reads = source.reads;
+  const auto plan = IR::ExtractResourcePlan(translated.program);
+  Check(plan.source_reads == source.reads && plan.capture_specialization_reads,
+        "linked code dependencies were omitted from resource alias validation");
+
+  std::vector<uint32_t> writer(std::begin(shader), std::end(shader) - 1);
+  writer.insert(writer.end(), {EncodeVop1(0x01, 1, 3), EncodeMubuf0(0x1c, 0, false),
+                              EncodeMubuf1(1, 6, 0), EncodeSopp(0x01)});
+  std::array<uint32_t, 64> user_data{};
+  user_data[25] = static_cast<uint32_t>(base >> 32u);
+  user_data[26] = 4;
+  user_data[27] = 3u << 28u;
+  auto write_options = options;
+  write_options.user_data = user_data;
+  auto write_runtime = runtime;
+  write_runtime.user_data = user_data;
+  auto write_source = PrepareShaderSource(writer, write_options);
+  auto write_program = TranslateProgram(RefreshShaderSource(write_source, write_runtime), write_options);
+  write_program.program.source_reads = write_source.reads;
+  auto write_plan = IR::ExtractResourcePlan(write_program.program);
+  IR::ResourceSnapshot snapshot;
+  IR::ResourceSpecialization specialization;
+  const auto written = std::ranges::find_if(write_plan.info.buffers,
+      [](const auto& buffer) { return buffer.written; });
+  Check(written != write_plan.info.buffers.end(), "caller store lost its writable binding");
+  const auto output_index = static_cast<size_t>(written - write_plan.info.buffers.begin());
+  for (const auto offset : {0x80u, 0x100u, 0x1f0u}) {
+    user_data[24] = static_cast<uint32_t>(base + offset);
+    Check(IR::MaterializeResources(write_plan, write_runtime, snapshot, specialization) &&
+              snapshot.buffers[output_index].dwords[0] == user_data[24] &&
+              snapshot.buffers[output_index].dwords[2] == 4 &&
+              std::ranges::find(snapshot.specialization_reads,
+                  std::pair<uint64_t, uint64_t>{base + 0x80, 4}) != snapshot.specialization_reads.end() &&
+              std::ranges::find(snapshot.specialization_reads,
+                  std::pair<uint64_t, uint64_t>{base + 0x100, 12}) != snapshot.specialization_reads.end(),
+          "call dependencies or writable ranges were lost before binding validation");
+  }
+
+  // Relocating identical code updates dependencies without rebuilding the source.
+  std::copy_n(memory.words.begin() + 64, 3, memory.words.begin() + 96);
+  memory.words[32] = static_cast<uint32_t>(base + 0x180);
+  memory.end = base + 0x18c;
+  (void)RefreshShaderSource(source, runtime);
+  Check(source.revision == 1 && source.reads.back().first == base + 0x180,
+        "identical relocated leaf recompiled or retained its old memory dependency");
+  memory.words[97] = 0x87654321;
+  (void)RefreshShaderSource(source, runtime);
+  Check(source.revision == 2 && source.linked->code.back() == EncodeSop1(0x21, 125, 14),
+        "changed leaf bytes retained stale source");
+  // A newly selected shorter function may end before the old bulk-read length.
+  memory.words[96] = EncodeSMovB32(3, 129);
+  memory.words[97] = EncodeSop1(0x21, 125, 14);
+  memory.end = base + 0x188;
+  (void)RefreshShaderSource(source, runtime);
+  Check(source.revision == 3 && source.reads.back().second == 8,
+        "shorter exact-boundary leaf was not decoded after a failed old-span read");
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
+  memory.clean = false;
+  ExpectFatal([&] { (void)RefreshShaderSource(source, runtime); },
+              "unreadable call dependencies reused stale linked code");
+  memory.clean = true;
+  auto ordinary = runtime;
+  ordinary.read_specialization_memory = nullptr;
+  ExpectFatal([&] { (void)RefreshShaderSource(source, ordinary); },
+              "call target query accepted an ordinary non-strict memory reader");
+  std::vector<uint32_t> escaping(std::begin(shader), std::end(shader) - 1);
+  escaping.insert(escaping.end(), {EncodeVop1(0x01, 1, 14),
+      EncodeMubuf0(0x1c, 0, false), EncodeMubuf1(1, 0, 0), EncodeSopp(0x01)});
+  ExpectFatal([&] {
+    auto escaping_source = PrepareShaderSource(escaping, options);
+    (void)TranslateProgram(RefreshShaderSource(escaping_source, runtime), options);
+  }, "saved native return PC escaped into GPU data");
+  memory.words[96] = EncodeSMovB32(14, 129);
+  ExpectFatal([&] { (void)RefreshShaderSource(source, runtime); },
+              "leaf clobbering its return pair was linked");
+  // AND_B64 writes the high half despite the decoder's generic scalar width.
+  const uint32_t high_writer[] = {
+      EncodeSMovB32(15, 129), EncodeSop2(0x0f, 14, 128, 128),
+      EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
+  auto pair_source = PrepareShaderSource(high_writer, options);
+  IR::SrtWalker pair_walker(pair_source.call_targets, runtime);
+  uint32_t high = 1;
+  Check(pair_walker.Evaluate(pair_source.call->target[1], high) && high == 0,
+        "scalar pair producer used a stale preceding high-half definition");
+  const uint32_t joined[] = {
+      EncodeSopp(0x04, 2), EncodeSMovB32(14, 129), EncodeSopp(0x02, 1),
+      EncodeSMovB32(14, 130), EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
+  ExpectFatal([&] { (void)PrepareShaderSource(joined, options); },
+              "ambiguous reaching definitions were accepted for a call target");
+  const uint32_t multiple[] = {EncodeSop1(0x21, 14, 14), EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
+  ExpectFatal([&] { (void)PrepareShaderSource(multiple, options); },
+              "call metadata bypassed the multiple-call rejection");
+#endif
+}
+
 void TestNewShaderRecompilerSetpcBranch() {
   const uint32_t shader[] = {
       EncodeSop1(0x1f, 4, 0),      // s_getpc_b64 s[4:5]
@@ -10350,7 +10563,10 @@ void TestFusedShaderHandoffPreservesRegisters() {
   options.back_code = back;
   for (const auto handoff: {EncodeSop1(0x20, 0, 6), 0xbefd2106u}) {
     front[2] = handoff; // SETPC or captured SWAPPC with NULL destination
-    auto translated = TranslateProgram(front, options);
+    auto source = PrepareShaderSource(front, options);
+    Check(!source.decoded.has_swap_pc && !source.call,
+          "fused SETPC handoff entered scalar call preparation");
+    auto translated = TranslateProgram(source.decoded, options);
     uint32_t allocations = 0;
     for (const auto* block: translated.program.blocks) {
       for (const auto& inst: *block) {
@@ -10365,6 +10581,17 @@ void TestFusedShaderHandoffPreservesRegisters() {
     }
     Check(allocations == 1u, "fused shader omitted the back shader allocation");
   }
+  const uint32_t call_front[] = {EncodeSop1(0x21, 14, 14), EncodeSop1(0x20, 0, 6)};
+  Check(Decoder::DecodeFrontProgram(call_front).has_swap_pc,
+        "front decoder omitted its scalar call");
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
+  ExpectFatal([&] { (void)PrepareShaderSource(call_front, options); },
+              "fused front call bypassed the non-compute rejection");
+  const uint32_t call_back[] = {EncodeSop1(0x21, 14, 14), EncodeSopp(0x01)};
+  options.back_code = call_back;
+  ExpectFatal([&] { (void)PrepareShaderSource(front, options); },
+              "fused back call metadata was not propagated");
+#endif
 }
 
 void TestMeshExportStorage() {
@@ -13682,6 +13909,331 @@ void TestGpuProducedWritableDescriptor() {
         "GPU-updated address or record count created a shader permutation");
 }
 
+void TestImmutableDescriptorPredicate() {
+  using namespace ShaderRecompiler::IR;
+  ResourcePlan plan;
+  const auto emit = [&](ValueOpcode op, std::initializer_list<Value> args) {
+    auto &inst = plan.value_storage.emplace_back(op);
+    uint32_t index = 0;
+    for (const auto arg : args) inst.SetArg(index++, arg);
+    return Value(&inst);
+  };
+  const auto handle = emit(ValueOpcode::GetBufferResource,
+                           {Value(0x1000u), Value(4u << 16u), Value(16u), Value(0u)});
+  for (const bool coherent : {false, true}) {
+    plan.memory_info.push_back({.kind = ResourceKind::Buffer,
+                                .idxen = !coherent, .coherent = coherent});
+    const auto loaded = emit(ValueOpcode::LoadBufferU32,
+                              {handle, Value(0u), Value(0u), Value(0u), Value(true)});
+    loaded.ResolveInstruction()->SetFlags(MemoryFlags{
+        .index = static_cast<uint32_t>(plan.memory_info.size() - 1u)});
+    const auto mask = emit(ValueOpcode::INotEqual32, {loaded, Value(0u)});
+    const auto selected = emit(ValueOpcode::SelectU32, {mask, Value(1u), Value(0u)});
+    const auto reduced = emit(ValueOpcode::ReadFirstLane, {selected, mask});
+    Check(ValidateRuntimeValue(plan, reduced, RuntimeValueType::Integer),
+          "integer EXEC reduction lost its active-lane proof");
+    Check(!ValidateRuntimeValue(plan, reduced, RuntimeValueType::ImmutableInteger),
+          "memory-dependent EXEC reduction was classified as immutable");
+  }
+}
+
+void TestUniformSelectedWritableDescriptor() {
+  using namespace ShaderRecompiler::IR;
+  // SAROS b9da5e64f4c5b10a builds s[8:11] from two scalar address expressions,
+  // then sets the stride in the high address word after the branch merge.
+  const uint32_t shader[] = {
+      0xbfa00002u, 0xbf070280u, 0xbf84000du, 0xf4001a82u,
+      0xfa00009cu, 0xbf8cc07fu, 0x8f6b856au, 0xf4001a82u,
+      0xfa000080u, 0xbf8cc07fu, 0x98eb6b6au, 0xb7eb0080u,
+      0x80086b03u, 0xbf800000u, 0x82098000u, 0xbf820006u,
+      0xf4041a82u, 0xfa000040u, 0xbf8cc07fu, 0x8008066au,
+      0xbf800000u, 0x8209016bu, 0xd7460001u, 0x04010c07u,
+      0x7e000280u, 0xf4001a82u, 0xfa000084u, 0xbf8cc07fu,
+      0x810a846au, 0xbe8b03ffu, 0x00016204u, 0x4a020284u,
+      0xbe891d92u, 0xe0702000u, 0x80020001u, 0xbf810000u,
+  };
+  std::array<uint32_t, 8> user_data{0x30u, 0x1cu, 1u, 0xfffffff0u,
+                                     0x1000u, 0u, 0xfffffffcu, 0u};
+  auto options = MakeCompileOptions(ShaderType::Compute);
+  options.user_data = user_data;
+  auto translated = ShaderRecompiler::TranslateProgram(shader, options);
+  auto plan = ExtractResourcePlan(translated.program);
+  Check(plan.info.buffers.size() == 1 && plan.info.buffers[0].written,
+        "uniform address selection lost its writable buffer");
+  const auto read = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+    const bool alternate = *static_cast<uint32_t *>(data) == 0;
+    if (words.size() != 1) return false;
+    switch (address) {
+    case 0x1040: words[0] = 8u; return alternate;
+    case 0x1044: words[0] = 1u; return alternate;
+    case 0x1080: words[0] = 3u; return !alternate;
+    case 0x109c: words[0] = 2u; return !alternate;
+    case 0x1084: words[0] = 10u; return true;
+    default: return false;
+    }
+  };
+  const SrtRuntime runtime{.user_data = user_data, .read_memory = read,
+                           .userdata = &user_data[2], .read_specialization_memory = read};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  for (const uint32_t flag : {1u, 0u, 1u}) {
+    user_data[2] = flag;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization),
+          "uniform selection evaluated an unavailable untaken scalar source");
+    const auto &descriptor = snapshot.buffers[0].dwords;
+    Check(descriptor[0] == (flag ? 0xe0u : 4u) &&
+              descriptor[1] == (flag ? 0x40031u : 0x4001eu) &&
+              descriptor[2] == 14u && descriptor[3] == 0x16204u,
+          "uniform writable descriptor lost its branch, carry, or post-merge stride");
+  }
+}
+
+void TestUniformSelectedDescriptorLoadAddress() {
+  using namespace ShaderRecompiler::IR;
+  const uint32_t shader[] = {
+      EncodeSopc(0x06, 0, 128),     // s_cmp_eq_u32 s0, 0
+      EncodeSopp(0x04, 2),         // s_cbranch_scc0 alternate
+      EncodeSop1(0x04, 6, 2),      // s_mov_b64 s[6:7], s[2:3]
+      EncodeSopp(0x02, 1),         // s_branch load
+      EncodeSop1(0x04, 6, 4),      // s_mov_b64 s[6:7], s[4:5]
+      EncodeSmem0(0x02, 8, 3), 125u << 25u, // s_load_dwordx4 s[8:11], s[6:7], 0
+      EncodeMubuf0(0x1c), EncodeMubuf1(0, 2, 0),
+      EncodeSopp(0x01),
+  };
+  std::array<uint32_t, 6> user_data{0u, 0u, 0x1000u, 0u, 0x2000u, 0u};
+  auto options = MakeCompileOptions(ShaderType::Compute);
+  options.user_data = user_data;
+  auto translated = ShaderRecompiler::TranslateProgram(shader, options);
+  auto plan = ExtractResourcePlan(translated.program);
+  Check(plan.srt_reads.size() == 4u && plan.info.buffers.size() == 1u &&
+            plan.info.buffers[0].written && !plan.capture_specialization_reads,
+        "conditional scalar address duplicated or detached descriptor reads");
+  const auto read = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+    const auto flag = *static_cast<uint32_t *>(data);
+    const uint64_t base = flag ? 0x2000u : 0x1000u;
+    const std::array<uint32_t, 4> descriptor{0x3000u + flag * 0x1000u,
+                                            4u << 16u, 16u, 0x16204u};
+    if (words.size() != 1u || address < base || address >= base + 16u ||
+        (address & 3u) != 0u) return false;
+    words[0] = descriptor[(address - base) / 4u];
+    return true;
+  };
+  const SrtRuntime runtime{.user_data = user_data, .read_memory = read,
+                           .userdata = &user_data[0]};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  for (const uint32_t flag : {0u, 1u, 0u}) {
+    user_data[0] = flag;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              snapshot.buffers[0].dwords[0] == 0x3000u + flag * 0x1000u &&
+              snapshot.specialization_reads.empty(),
+          "conditional descriptor load read the untaken pointer");
+  }
+}
+
+void TestNativeScalarAtomicPayloadStaysOnGpu() {
+  using namespace ShaderRecompiler::IR;
+  // Exact final block from SAROS b62b494cb2567011: read a mutable flag,
+  // atomically update that address, then test the loaded flag before S_TRAP.
+  const uint32_t shader[] = {
+      0x7e0002ffu, 0x00100000u, 0x7e020280u, 0xf4040100u,
+      0xfa000018u, 0xbf8cc07fu, 0x8801ff05u, 0x01000000u,
+      0xbe800304u, 0xbe820381u, 0xbe8303ffu, 0x00016204u,
+      0xf4001a82u, 0xfa000018u, 0xe1680018u, 0x80000000u,
+      0xbf8cc07fu, 0xbf0d846au, 0xbf840001u, 0xbf920001u,
+      0xbf810000u,
+  };
+  const std::array<uint32_t, 2> user_data{0x1000u, 0u};
+  auto options = MakeCompileOptions(ShaderType::Compute);
+  options.user_data = user_data;
+  auto translated = ShaderRecompiler::TranslateProgram(shader, options);
+  const auto &program = translated.program;
+  const Inst *payload = nullptr;
+  for (const auto *block : program.blocks) {
+    for (const auto &inst : *block) {
+      if (inst.GetOpcode() == ValueOpcode::LoadAddressU32 &&
+          program.memory_info[inst.Flags<MemoryFlags>().index].kind == ResourceKind::ScalarAddress)
+        payload = &inst;
+    }
+  }
+  Check(payload != nullptr && payload->Parent() != nullptr &&
+            !program.memory_info[payload->Flags<MemoryFlags>().index].planning_only &&
+            program.info.uses_dma && program.info.buffers.size() == 1 &&
+            program.info.buffers[0].atomic,
+        "native scalar payload was replaced by a host snapshot before its atomic write");
+  auto plan = ExtractResourcePlan(program);
+  Check(plan.srt_reads.size() == 2 && plan.control_flow.empty() &&
+            !plan.capture_specialization_reads,
+        "terminal native shader assertion entered resource planning");
+  struct Reads { uint32_t descriptors = 0; uint32_t payload = 0; } reads;
+  const auto read = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+    auto &reads = *static_cast<Reads *>(data);
+    if (address == 0x2018u) { ++reads.payload; return false; }
+    if (words.size() != 1 || (address != 0x1018u && address != 0x101cu)) return false;
+    ++reads.descriptors;
+    words[0] = address == 0x1018u ? 0x2000u : 0u;
+    return true;
+  };
+  const SrtRuntime runtime{.user_data = user_data, .read_memory = read,
+                           .userdata = &reads, .read_specialization_memory = read};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            reads.descriptors == 2 && reads.payload == 0 &&
+            snapshot.specialization_reads.empty() && snapshot.buffers[0].dwords[0] == 0x2000u,
+        "native resource materialization read the mutable flag or lost its atomic descriptor");
+}
+
+void TestBoundedScalarMaterialImageKeys() {
+  using namespace ShaderRecompiler::IR;
+  // SAROS 9fba2edffc549531: min(header count,64), scalar rows of 160 bytes,
+  // and separate float/uint image fields at 0x7c/0x94 in the same image heap.
+  std::vector<uint32_t> shader {
+      0xf4080106u, 0xfa0000a0u, 0xf4080806u, 0xfa0000b0u,
+      0xbf8cc07fu, 0xf4201550u, 0xfa000024u, 0xbf8cc07fu,
+      EncodeSop2(0x07, 85, 85, 0xc0), EncodeSMovB32(86, 128),
+      EncodeSMovB32(87, 126),
+  };
+  const auto loop = shader.size();
+  shader.insert(shader.end(), {0x7da2aaf9u, 0x86860056u, EncodeSopp(0x08),
+      0x931cff56u, 0x000000a0u, 0xf42006c2u, 0x38000094u,
+      0xf4200682u, 0x3800007cu, 0xbf8cc07fu, 0x8f58851bu, 0x8f3c851au,
+      0xf42c1304u, 0xb0000000u, 0xf42c0d04u, 0x78000000u, 0xbf8cc07fu,
+      0xf0000110u, 0x00130106u, 0xf0000110u, 0x000d0206u,
+      EncodeVop1(0x01, 3, 78), // Preserve a T# dimension word as native scalar payload.
+      EncodeMubuf0(0x1c, 0, false), EncodeMubuf1(1, 0, 0),
+      EncodeMubuf0(0x1c, 4, false), EncodeMubuf1(2, 0, 0),
+      EncodeMubuf0(0x1c, 8, false), EncodeMubuf1(3, 0, 0),
+      EncodeSop2(0x02, 86, 86, 129), EncodeSMovB32(126, 87),
+      EncodeSopp(0x02)});
+  shader.back() = EncodeSopp(0x02, static_cast<uint16_t>(loop - shader.size()));
+  shader[loop + 2u] = EncodeSopp(0x08, shader.size() - loop - 3u);
+  shader.push_back(EncodeSopp(0x01));
+  std::array<uint32_t, 14> user_data{};
+  user_data[8] = 0x100000u;
+  user_data[9] = 32u << 16u;
+  user_data[10] = 198656u;
+  user_data[11] = 0x5204u;
+  user_data[12] = 0x1000u;
+  auto options = MakeCompileOptions(ShaderType::Compute);
+  options.user_data = user_data;
+  auto translated = ShaderRecompiler::TranslateProgram(shader, options);
+  auto plan = ExtractResourcePlan(translated.program);
+  Check(plan.info.images.size() == 2 && plan.capture_specialization_reads,
+        "scalar material fields lost their distinct image roots or clean read capture");
+  for (const auto &image : plan.info.images) {
+    const auto &source = plan.descriptor_sources[image.source].indirect_descriptor;
+    Check(source && source->selector && source->selector->stride == 160u &&
+              (source->selector->offset == 0x7cu || source->selector->offset == 0x94u) &&
+              source->table_stride == 32u && !source->table_scalar &&
+              source->key_count.ResolveInstruction()->GetOpcode() == ValueOpcode::UMin32,
+          "scalar material image selection fell back to the whole heap");
+  }
+  Check(std::ranges::any_of(translated.program.blocks, [&](const auto *block) {
+    return std::ranges::any_of(*block, [&](const auto &inst) {
+      if (inst.GetOpcode() != ValueOpcode::ReadConstBuffer) return false;
+      const auto &memory = translated.program.memory_info[inst.template Flags<MemoryFlags>().index];
+      return memory.component_count == 8u && memory.component_index == 2u &&
+             !memory.planning_only && inst.HasUses();
+    });
+  }), "scalar material projection removed a live T# dimension payload");
+  struct Memory {
+    uint32_t count = 2;
+    uint32_t stride = 160;
+    uint32_t records = 409;
+    uint64_t unavailable = 0;
+    bool clean = true;
+    bool mixed = false;
+    bool same_key = false;
+    uint32_t keys = 0;
+    uint32_t images = 0;
+  } memory;
+  const auto read = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+    auto &m = *static_cast<Memory *>(data);
+    if (!m.clean || address == m.unavailable) return false;
+    if (words.size() == 1 && address >= 0x10a0u && address < 0x10c0u) {
+      const uint32_t descriptors[] = {0x2000u, m.stride << 16u, m.records, 0x5204u,
+                                     0x4000u, 16u << 16u, 8u, 0x4dfacu};
+      words[0] = descriptors[(address - 0x10a0u) / 4u];
+      return true;
+    }
+    if (address == 0x4024u && words.size() == 1) { words[0] = m.count; return true; }
+    for (const uint32_t row : {0u, 1u}) {
+      for (const uint32_t field : {0x7cu, 0x94u}) {
+        if (address != 0x2000u + row * 160u + field || words.size() != 1) continue;
+        ++m.keys;
+        words[0] = (row == 0 || m.same_key ? 5582u : 5567u) + (field == 0x94u ? 6u : 0u);
+        return true;
+      }
+    }
+    for (const uint32_t key : {5567u, 5582u, 5573u, 5588u}) {
+      if (address != 0x100000u + key * 32u || words.size() != 8) continue;
+      ++m.images;
+      std::ranges::fill(words, 0u);
+      words[0] = key;
+      const bool integer = key == 5573u || key == 5588u;
+      words[1] = static_cast<uint32_t>(integer && !(m.mixed && key == 5573u)
+          ? Prospero::BufferFormat::k32UInt : Prospero::BufferFormat::k32Float) << 20u;
+      words[3] = DstSel(4, 5, 6, 7) |
+          (static_cast<uint32_t>(Prospero::ImageType::kColor3D) << 28u);
+      return true;
+    }
+    return false;
+  };
+  const SrtRuntime runtime{.user_data = user_data, .userdata = &memory,
+                           .read_specialization_memory = read};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  for (const uint32_t stride : {160u, 16u}) {
+    memory = {.stride = stride, .records = stride == 160u ? 409u : 20u};
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              memory.keys == 4 && memory.images == 4 && snapshot.images.size() == 4,
+          "scalar material selection used V# stride as its instruction stride or scanned unused keys");
+    for (uint32_t root = 0; root < plan.info.images.size(); ++root) {
+      const auto &selector = *plan.descriptor_sources[plan.info.images[root].source].indirect_descriptor->selector;
+      const uint32_t first = selector.offset == 0x94u ? 5573u : 5567u;
+      const auto offset = specialization.images[root].indirect_mapping_offset;
+      Check(snapshot.flattened_srt[offset] == 2u &&
+                snapshot.flattened_srt[offset + 1u] == first &&
+                snapshot.flattened_srt[offset + 3u] == first + 15u,
+            "scalar material image mapping changed descriptor-index keys into byte offsets");
+    }
+  }
+  for (const bool same_key : {false, true}) {
+    memory = {.count = same_key ? 2u : 1u, .same_key = same_key};
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              memory.keys == (same_key ? 4u : 2u) && memory.images == 2 &&
+              snapshot.images.size() == 2 &&
+              std::ranges::all_of(specialization.images, [](const auto &image) {
+                return image.indirect_root == ImageResource::NoIndirectImage;
+              }), "single-candidate scalar material loop retained an indirect image mapping");
+  }
+  memory = {.count = 0};
+  Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+            memory.keys == 0 && memory.images == 0 && snapshot.images.size() == 2,
+        "empty scalar material loop retained image probes");
+  memory = {.clean = false};
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
+        "dirty scalar material proof was accepted");
+  memory = {.unavailable = 0x2094u};
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
+        "unavailable selected material key was accepted");
+  memory = {.stride = 1, .records = 311};
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization) && memory.keys == 0,
+        "scalar material byte range escaped V# bounds");
+  memory = {.mixed = true};
+  Check(!MaterializeResources(plan, runtime, snapshot, specialization),
+        "selected material fields bypassed image compatibility");
+  memory = {};
+  Check(MaterializeResources(plan, runtime, snapshot, specialization),
+        "native scalar material image selection did not recover after rejected candidates");
+  ApplyResourceSpecialization(translated.program, specialization);
+  Check(translated.program.info.images.size() == 4 &&
+            translated.program.info.images[0].indirect_resources.size() == 2 &&
+            translated.program.info.images[1].indirect_resources.size() == 2,
+        "native image reads lost their distinct material field dispatches");
+}
+
 void TestTypedDescriptorRealCarryAndScalarLoads() {
   const uint32_t carry_shader[] = {
       EncodeSop1(0x1f, 0, 0),      // s_getpc_b64 s[0:1]
@@ -15120,6 +15672,8 @@ int main() {
   TestNewShaderRecompilerPixelImageSampleLodSelection();
   TestNewShaderRecompilerBranchConditionForms();
   TestNewShaderRecompilerSetpcBranch();
+  TestGraphicsComputeGetreg();
+  TestMemoryFedScalarLeafCall();
   TestFusedShaderHandoffPreservesRegisters();
   TestMeshExportStorage();
   TestMergedShaderUserDataSnapshot();
@@ -15153,6 +15707,11 @@ int main() {
   TestTypedDescriptorRealWideMoveTranslation();
   TestComputeImageFill();
   TestGpuProducedWritableDescriptor();
+  TestUniformSelectedWritableDescriptor();
+  TestUniformSelectedDescriptorLoadAddress();
+  TestNativeScalarAtomicPayloadStaysOnGpu();
+  TestBoundedScalarMaterialImageKeys();
+  TestImmutableDescriptorPredicate();
   TestTypedDescriptorRealCarryAndScalarLoads();
   TestSrtWalkerRealSmemTranslation();
   TestSrtWalkerVccBaseTranslation();