shader: replace CPU selector chains with direct image-table lookup

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

@@ -849,7 +849,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			return;
 		}
 		const auto key = ctx.Def(handle->Arg(0));
-		if (state.flattened_srt_variable == 0 || image.indirect_search_iterations == 0u ||
+		if (state.flattened_srt_variable == 0 ||
 		    image.indirect_resources.size() < 2u) {
 			ctx.Fail(inst, "has no indirect image runtime mapping");
 			return;

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

@@ -44,6 +44,16 @@ uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t m
 	};
 	const auto mapping = ConstantU32(state, mapping_offset);
 	const auto count = LoadMapping(mapping);
+	if (search_iterations == 0u) {
+		const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), key,
+		                          ConstantU32(state, 2u));
+		const auto valid = Binary(state, spv::OpULessThan, TypeBool(state), index, count);
+		const auto safe = Select(state, TypeU32(state), valid, index, ConstantU32(state, 0u));
+		const auto entry = Binary(state, spv::OpIAdd, TypeU32(state), safe,
+		                          ConstantU32(state, mapping_offset + 1u));
+		return Select(state, TypeU32(state), valid, LoadMapping(entry),
+		              ConstantU32(state, default_resource));
+	}
 	const auto entry_at = [&](uint32_t index) {
 		return Binary(state, spv::OpIAdd, TypeU32(state), mapping,
 		              Binary(state, spv::OpIAdd, TypeU32(state),

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

@@ -489,17 +489,16 @@ struct DescriptorSource {
 			uint32_t source = UINT32_MAX;
 			uint32_t stride = 0;
 			uint32_t offset = 0;
-			uint32_t immediate = 0;
-			uint32_t shift = 0;
-			uint32_t bits = UINT32_MAX;
 
 			bool operator==(const SelectorRead& other) const = default;
 		};
-		std::vector<SelectorRead> selectors;
+		std::optional<SelectorRead> selector;
 		uint32_t table_source    = 0;
 		uint32_t table_offset    = 0;
 		uint32_t table_immediate = 0;
 		uint32_t table_stride    = 0;
+		uint32_t table_record_bytes = 0;
+		bool     table_scalar = false;
 		uint32_t workgroup_axis  = UINT32_MAX;
 		Value    key_count;
 		Value                 selector_first;

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

@@ -101,7 +101,7 @@ SharedMemoryResources CollectMemoryResources(const Program& program, std::vector
 
 bool UsesFlattenedSrt(const Program& program) {
 	const auto uses_mapping = [](const auto& resource) {
-		return resource.indirect_search_iterations != 0u;
+		return resource.indirect_root != UINT32_MAX;
 	};
 	return std::ranges::any_of(program.blocks, [](const Block* block) {
 		return std::ranges::any_of(*block, [](const Inst& inst) {

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

@@ -249,8 +249,24 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
                                    uint32_t maximum_resources, Normalize&& normalize) {
 	const auto  descriptor_bytes = dword_count * sizeof(uint32_t);
 	const auto& sources = indirect.sources;
-	const auto* selector = indirect.selectors.empty() ? nullptr : &indirect.selectors.front();
+	const auto* selector = indirect.selector ? &*indirect.selector : nullptr;
 	auto& keys = program.material_keys;
+	const auto children_begin = descriptors.size();
+	const auto root_resource = specializations[resource_index];
+	const auto intern_candidate = [&](const DescriptorValue& candidate, uint32_t& ordinal) {
+		ordinal = 0u;
+		if (candidate == descriptors[resource_index]) return true;
+		const auto found = std::find(descriptors.begin() + children_begin, descriptors.end(), candidate);
+		ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
+		if (found == descriptors.end()) {
+			if (descriptors.size() >= maximum_resources) return false;
+			descriptors.push_back(candidate);
+			auto child = root_resource;
+			child.indirect_root = resource_index;
+			specializations.push_back(child);
+		}
+		return true;
+	};
 	const auto read_keys = [&](const ShaderBufferResource& material, uint64_t first,
 	                           uint64_t step, uint64_t count) {
 		keys.resize(count);
@@ -284,6 +300,48 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		} else {
 			return false;
 		}
+		if (selector == nullptr && table_value.dword_count == 4u &&
+		    indirect.workgroup_axis == UINT32_MAX) {
+			if (!indirect.table_scalar && table.Format() == Prospero::BufferFormat::kInvalid) {
+				descriptors[resource_index] = {.dword_count = dword_count};
+				return true;
+			}
+			if (table_size > UINT32_MAX ||
+			    (indirect.table_record_bytes != 0u &&
+			     (table.Type() != 0u || table.SwizzleEnabled() || table.AddTid() ||
+			      table.OutOfBounds() != 0u || (table_base & 3u) != 0u ||
+			      (table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes)))
+				return false;
+			// Descriptor bytes are refreshed together. The GPU selects the DWORD address;
+			// no material selector or unused table field is interpreted on the CPU.
+			keys.resize(table_size / 4u);
+			if (!ReadScalarTable(table_base, table_size, 0u, runtime, keys)) return false;
+			const auto mapping_offset = snapshot.flattened_srt.size();
+			descriptors[resource_index] = {.dword_count = dword_count};
+			snapshot.flattened_srt.resize(mapping_offset + 2u, 0u);
+			for (uint32_t word = 0; word < keys.size(); ++word) {
+				DescriptorValue candidate {.dword_count = dword_count};
+				std::copy_n(keys.begin() + word, std::min<size_t>(dword_count, keys.size() - word),
+				            candidate.dwords.begin());
+				if (!normalize(candidate)) return false;
+				uint32_t ordinal;
+				if (!intern_candidate(candidate, ordinal)) return false;
+				if (ordinal == 0u) continue;
+				snapshot.flattened_srt.resize(mapping_offset + 2u + word, 0u);
+				snapshot.flattened_srt[mapping_offset + 1u + word] = ordinal;
+			}
+			if (descriptors.size() == children_begin) {
+				snapshot.flattened_srt.resize(mapping_offset);
+			} else {
+				snapshot.flattened_srt[mapping_offset] =
+				    static_cast<uint32_t>(snapshot.flattened_srt.size() - mapping_offset - 1u);
+				auto& root = specializations[resource_index];
+				root.indirect_root = resource_index;
+				root.indirect_mapping_offset = static_cast<uint32_t>(mapping_offset);
+				root.indirect_search_iterations = 0u;
+			}
+			return true;
+		}
 		if (selector == nullptr) {
 			uint32_t key_count = 0;
 			if (indirect.workgroup_axis != UINT32_MAX) {
@@ -307,8 +365,6 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			ShaderBufferResource material;
 			uint32_t             first = 0, count = 0;
 			if (!DecodeBufferDescriptor(material_value, material) || material.Type() != 0u ||
-			    ((selector->shift != 0u || selector->bits != UINT32_MAX) &&
-			     (material.Base48() & 3u) != 0u) ||
 			    material.SwizzleEnabled() ||
 			    material.AddTid() || material.OutOfBounds() != 0u ||
 			    uint64_t {selector->offset} + 4u > material.Stride() ||
@@ -345,53 +401,12 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 				keys.push_back(key);
 				mask &= mask - 1u;
 			}
-		} else {
-			ShaderBufferResource material;
-			if (!DecodeBufferDescriptor(material_value, material) || table_value.dword_count != 4u) {
-				return false;
-			}
-			// The first aligned offset includes the immediate added after shader U32 arithmetic.
-			const auto step = std::max<uint64_t>(4u,
-			    std::gcd<uint64_t>(selector->stride, uint64_t {1} << 32u));
-			const uint64_t first = (selector->offset % step & ~uint64_t {3}) +
-			                       (selector->immediate & ~3u);
-			const auto size = material.GetSize();
-			const auto limit = std::min(first + (uint64_t {1} << 32u) - step, size >= 4u ? size - 4u : 0u);
-			const auto probe_count = size >= 4u && first <= limit ? (limit - first) / step + 1u : 0u;
-			if (probe_count > MaxIndirectDescriptorProbes) {
-				return false;
-			}
-			if (!read_keys(material, first, step, probe_count)) return false;
-			if (selector->shift == 0u) keys.push_back(0u);
-		}
-		for (uint32_t step = 0; step < indirect.selectors.size(); ++step) {
-			const auto& read = indirect.selectors[step];
-			if (step != 0u) {
-				DescriptorValue value;
-				ShaderBufferResource material;
-				if (!clean.EvaluateDescriptor(read.source, value) || !DecodeBufferDescriptor(value, material))
-					return false;
-				// Distinct keys can alias after U32 arithmetic and DWORD alignment.
-				for (auto& key: keys) key = (key * read.stride + read.offset) & ~3u;
-				std::ranges::sort(keys);
-				keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
-				for (auto& key: keys) {
-					const uint64_t offset = uint64_t {key} + (read.immediate & ~3u);
-					if (!ReadScalarTable(material.Base48(), material.GetSize(), offset, runtime, {&key, 1}))
-						return false;
-				}
-			}
-			for (auto& key: keys) key = (key >> read.shift) & read.bits;
-			if (step == 0u && (read.shift != 0u || read.bits != UINT32_MAX))
-				keys.push_back(0u); // An out-of-range root material load returns zero.
-			std::ranges::sort(keys);
-			keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
-		}
+		} else return false;
+		std::ranges::sort(keys);
+		keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
 	}
 
-	const auto children_begin = descriptors.size();
 	const auto mapping_offset = snapshot.flattened_srt.size();
-	const auto root_resource  = specializations[resource_index];
 	const auto key_count = sources.empty() ? keys.size() : sources.size();
 	snapshot.flattened_srt.resize(mapping_offset + 1u + key_count * 2u);
 	snapshot.flattened_srt[mapping_offset] = static_cast<uint32_t>(key_count);
@@ -414,19 +429,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		uint32_t ordinal = 0;
 		if (entry == 0) {
 			descriptors[resource_index] = candidate;
-		} else if (descriptors[resource_index] != candidate) {
-			const auto found =
-			    std::find(descriptors.begin() + children_begin, descriptors.end(), candidate);
-			ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
-			if (found == descriptors.end()) {
-				if (descriptors.size() >= maximum_resources) {
-					return false;
-				}
-				descriptors.push_back(candidate);
-				auto child          = root_resource;
-				child.indirect_root = resource_index;
-				specializations.push_back(child);
-			}
+		} else if (!intern_candidate(candidate, ordinal)) {
+			return false;
 		}
 		snapshot.flattened_srt[mapping_offset + 1u + entry * 2u] = key;
 		snapshot.flattened_srt[mapping_offset + 2u + entry * 2u] = ordinal;
@@ -584,9 +588,9 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 		const auto key_count = root.indirect_mapping_offset < snapshot.flattened_srt.size()
 		                           ? snapshot.flattened_srt[root.indirect_mapping_offset]
 		                           : 0u;
-		if (root.indirect_search_iterations == 0u || key_count < 2u ||
+		if (key_count == 0u ||
 		    static_cast<size_t>(root.indirect_mapping_offset) + 1u +
-		            static_cast<size_t>(key_count) * 2u >
+		            static_cast<size_t>(key_count) * (root.indirect_search_iterations == 0u ? 1u : 2u) >
 		        snapshot.flattened_srt.size()) {
 			return SpecializationFail("indirect image specialization has an invalid key mapping");
 		}
@@ -1028,9 +1032,9 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		if (!source.indirect_descriptor.has_value()) continue;
 		const auto& indirect                = *source.indirect_descriptor;
 		plan.requires_specialization_memory = true;
-		capture_indirect_reads |= !indirect.selectors.empty() || !indirect.sources.empty();
-		for (const auto& read: indirect.selectors)
-			MarkCleanFlatSlots(plan, Source(plan, read.source), plan.clean_flat_slots);
+		capture_indirect_reads = true;
+		if (indirect.selector)
+			MarkCleanFlatSlots(plan, Source(plan, indirect.selector->source), plan.clean_flat_slots);
 		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.selector_mask);
 		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.selector_first);
 		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.key_count);

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

@@ -8,7 +8,6 @@
 #include <algorithm>
 #include <fmt/format.h>
 #include <map>
-#include <numeric>
 #include <optional>
 #include <span>
 #include <tuple>
@@ -374,11 +373,68 @@ public:
 			}
 		}
 		for (const auto& plan: m_indirect_descriptors) {
+			auto key = plan.key;
+			if (plan.table_read != nullptr) {
+				const auto* anchor = plan.table_indexed ? plan.handle : plan.table_read;
+				auto* block = anchor->Parent();
+				const auto where = std::ranges::find_if(block->Instructions(),
+				    [&](const Inst& inst) { return &inst == anchor; });
+				const auto emit = [&](ValueOpcode op, std::initializer_list<Value> args) {
+					return Value(&*block->PrependNewInst(where, op, args));
+				};
+				const auto& read = *plan.table_read;
+				const auto& memory = m_program.memory_info[read.Flags<MemoryFlags>().index];
+				Value offset;
+				Value valid(true);
+				if (plan.table_indexed) {
+					const auto* table = read.Arg(0).ResolveInstruction();
+					const auto stride = emit(ValueOpcode::BitFieldUExtract,
+					    {table->Arg(1), Value(16u), Value(14u)});
+					const auto format = emit(ValueOpcode::BitFieldUExtract,
+					    {table->Arg(3), Value(12u), Value(7u)});
+					offset = emit(ValueOpcode::IAdd32,
+					    {emit(ValueOpcode::IMul32, {read.Arg(1), stride}), Value(memory.offset)});
+					valid = emit(ValueOpcode::LogicalAnd,
+					    {emit(ValueOpcode::ULessThan32, {read.Arg(1), table->Arg(2)}),
+					     emit(ValueOpcode::INotEqual32, {format, Value(0u)})});
+				} else {
+					const auto dynamic = emit(ValueOpcode::BitwiseAnd32, {read.Arg(1), Value(~3u)});
+					offset = emit(ValueOpcode::IAdd32, {dynamic, Value(memory.offset & ~3u)});
+					valid = emit(ValueOpcode::ULessThanEqual32, {dynamic, offset});
+				}
+				key = emit(ValueOpcode::SelectU32, {valid, offset, Value(UINT32_MAX)});
+				if (!plan.table_indexed) {
+					// NativeDescriptorSource proved this read reaches the image. Preserve its
+					// SSA Phi edges while retaining descriptor DWORDs for other consumers.
+					std::map<const Inst*, Value> projected;
+					const auto project = [&](auto&& self, Value value) -> Value {
+						value = value.Resolve();
+						const auto* phi = value.TryInstruction();
+						if (phi == plan.table_read) return key;
+						if (phi == nullptr || phi->GetOpcode() != ValueOpcode::Phi) return Value(UINT32_MAX);
+						if (const auto found = projected.find(phi); found != projected.end()) return found->second;
+						auto* parent = phi->Parent();
+						const auto position = std::ranges::find_if(parent->Instructions(),
+						    [&](const Inst& inst) { return &inst == phi; });
+						auto& selected = *parent->PrependNewInst(position, ValueOpcode::Phi);
+						selected.SetFlags(Type::U32);
+						projected.emplace(phi, Value(&selected));
+						for (uint32_t arm = 0; arm < phi->NumArgs(); ++arm)
+							selected.AddPhiOperand(phi->PhiBlock(arm), self(self, phi->Arg(arm)));
+						return Value(&selected);
+					};
+					key = project(project, plan.handle->Arg(0));
+				}
+				plan.handle->SetArg(0, key);
+				for (uint32_t word = 1; word < plan.handle->NumArgs(); ++word)
+					plan.handle->SetArg(word, Value(0u));
+				continue;
+			}
 			if (plan.handle->GetOpcode() == ValueOpcode::GetBufferResource) {
 				plan.handle->SetArg(0, plan.key);
 				for (uint32_t word = 1; word < 4u; ++word) plan.handle->SetArg(word, Value(0u));
 			} else {
-				if (plan.reads[0] == nullptr) plan.handle->SetArg(0, plan.key);
+				if (plan.reads[0] == nullptr) plan.handle->SetArg(0, key);
 				for (uint32_t dword = 0; dword < 4u; dword++) {
 					plan.handle->SetArg(dword + 1u, plan.roots[dword + 4u]);
 				}
@@ -439,6 +495,8 @@ private:
 		std::array<const Inst*, 8> reads {};
 		bool split_offsets = false;
 		bool retain_reads = false;
+		bool table_indexed = false;
+		const Inst* table_read = nullptr;
 	};
 
 	[[noreturn]] void Fail(uint32_t pc, const std::string& reason) const {
@@ -836,7 +894,7 @@ private:
 
 	uint32_t InternSource(const DescriptorSource& descriptor) {
 		for (uint32_t candidate = 0; candidate < m_sources.size(); candidate++) {
-			const auto& current = m_sources[candidate];
+			auto& current = m_sources[candidate];
 			if (current.dword_count != descriptor.dword_count ||
 			    current.indirect_descriptor.has_value() != descriptor.indirect_descriptor.has_value()) {
 				continue;
@@ -844,7 +902,7 @@ private:
 			if (current.indirect_descriptor.has_value()) {
 				const auto& a = *current.indirect_descriptor;
 				const auto& b = *descriptor.indirect_descriptor;
-				if (a.selectors != b.selectors || a.table_source != b.table_source ||
+				if (a.selector != b.selector || a.table_source != b.table_source ||
 				    a.table_offset != b.table_offset || a.table_immediate != b.table_immediate ||
 				    a.table_stride != b.table_stride ||
 				    a.workgroup_axis != b.workgroup_axis || a.sources != b.sources ||
@@ -861,6 +919,12 @@ private:
 				same = same && EquivalentValue(m_program, current.dwords[i], descriptor.dwords[i]);
 			}
 			if (same) {
+				if (current.indirect_descriptor) {
+					current.indirect_descriptor->table_scalar |= descriptor.indirect_descriptor->table_scalar;
+					current.indirect_descriptor->table_record_bytes = std::max(
+					    current.indirect_descriptor->table_record_bytes,
+					    descriptor.indirect_descriptor->table_record_bytes);
+				}
 				return candidate;
 			}
 		}
@@ -950,54 +1014,6 @@ private:
 		return ValidateSource(descriptor, bad_dword);
 	}
 
-	Value MaterialOffset(Value value, uint32_t& stride, uint32_t& offset) const {
-		stride = 1u;
-		offset = 0u;
-		value = value.Resolve();
-		while (const auto* inst = value.TryInstruction()) {
-			if (inst->NumArgs() != 2u) break;
-			uint32_t constant = 0;
-			Value remaining;
-			if (inst->GetOpcode() == ValueOpcode::ShiftLeftLogical32 &&
-			    ImmediateU32(inst->Arg(1), constant) && constant < 32u) {
-				stride *= 1u << constant;
-				remaining = inst->Arg(0);
-			} else if (inst->GetOpcode() == ValueOpcode::IAdd32 ||
-			           inst->GetOpcode() == ValueOpcode::IMul32) {
-				if (ImmediateU32(inst->Arg(0), constant)) remaining = inst->Arg(1);
-				else if (ImmediateU32(inst->Arg(1), constant)) remaining = inst->Arg(0);
-				else break;
-				if (inst->GetOpcode() == ValueOpcode::IAdd32) offset += stride * constant;
-				else stride *= constant;
-			} else {
-				break;
-			}
-			value = remaining.Resolve();
-		}
-		// Treat the remaining GPU-derived scalar as any U32 value.
-		return stride != 0u && value.GetType() == Type::U32 && !value.IsImmediate() ? value : Value {};
-	}
-
-	Value MaterialKey(Value key, DescriptorSource::IndirectDescriptor::SelectorRead& selector) const {
-		selector.shift = 0u;
-		selector.bits = UINT32_MAX;
-		const auto* inst = key.Resolve().TryInstruction();
-		uint32_t immediate;
-		if (inst != nullptr && inst->GetOpcode() == ValueOpcode::BitwiseAnd32) {
-			if (ImmediateU32(inst->Arg(1), immediate)) key = inst->Arg(0);
-			else if (ImmediateU32(inst->Arg(0), immediate)) key = inst->Arg(1);
-			else return key;
-			selector.bits = immediate;
-			inst = key.Resolve().TryInstruction();
-		}
-		if (inst != nullptr && inst->GetOpcode() == ValueOpcode::ShiftRightLogical32 &&
-		    ImmediateU32(inst->Arg(1), immediate) && immediate != 0u && immediate < 32u) {
-			selector.shift = immediate;
-			key = inst->Arg(0);
-		}
-		return key.Resolve();
-	}
-
 	enum class LaneQuantifier { Any, All };
 	struct EdgePredicate {
 		Value condition;
@@ -1444,7 +1460,7 @@ private:
 		        static_cast<uint64_t>(UINT32_MAX) + 1u) return false;
 		const auto mask = BoundedSetBitMask(offset.index, active);
 		if (mask.IsEmpty()) return false;
-		indirect.selectors.push_back({.source = InternSource(material_source),
+		indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead{.source = InternSource(material_source),
 		                              .stride = static_cast<uint32_t>(offset.stride),
 		                              .offset = static_cast<uint32_t>(offset.offset + memory.offset)});
 		indirect.key_count = Value(32u);
@@ -1609,6 +1625,7 @@ private:
 	                          DescriptorSource& table_source, Value& key, uint32_t& table_offset,
 	                          uint32_t& table_stride, uint32_t& table_immediate) {
 		Inst* table_handle = nullptr;
+		Value table_guard;
 		for (uint32_t dword = 0; dword < handle.NumArgs(); ++dword) {
 			auto* read = descriptor.dwords[dword].Resolve().TryInstruction();
 			if (read == nullptr) {
@@ -1616,6 +1633,30 @@ private:
 			}
 			uint32_t memory_index = 0;
 			const auto* memory = ScalarReadMemory(*read, memory_index);
+			uint32_t component = 0;
+			const bool indexed = memory == nullptr;
+			if (indexed) {
+				if (handle.NumArgs() != 8u) return false;
+				const auto* selected = UniformizedMaterialValue(descriptor.dwords[dword], handle);
+				if (selected == nullptr) return false;
+				const auto* extract = selected->Arg(1).Resolve().TryInstruction();
+				if (extract == nullptr || extract->GetOpcode() != ValueOpcode::CompositeExtractU32x4 ||
+				    !ImmediateU32(extract->Arg(1), component) || component >= 4u) return false;
+				read = extract->Arg(0).Resolve().TryInstruction();
+				if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32x4) return false;
+				memory_index = read->Flags<MemoryFlags>().index;
+				if (memory_index >= m_program.memory_info.size()) return false;
+				memory = &m_program.memory_info[memory_index];
+				uint32_t offset = 1u, scalar = 1u;
+				if (memory->kind != ResourceKind::Buffer || memory->typed || memory->formatted ||
+				    !memory->idxen || memory->offen || memory->data_bits != 32u ||
+				    memory->data_dwords != 4u ||
+				    !ImmediateU32(read->Arg(2), offset) || offset != 0u ||
+				    !ImmediateU32(read->Arg(3), scalar) || scalar != 0u ||
+				    !EquivalentValue(m_program, read->Arg(4), selected->Arg(0)) ||
+				    (dword != 0u && !EquivalentValue(m_program, table_guard, read->Arg(4)))) return false;
+				table_guard = read->Arg(4);
+			}
 			if (memory == nullptr || memory->offset > INT32_MAX || (memory->offset & 3u) != 0u ||
 			    !MemoryIndexBelongsTo(memory_index, *read)) {
 				return false;
@@ -1631,23 +1672,31 @@ private:
 			    (memory->kind == ResourceKind::ScalarAddress && read->Parent() != handle.Parent()) ||
 			    (table_handle != nullptr &&
 			     !EquivalentValue(m_program, Value(table_handle), Value(current_handle))) ||
-			    !MatchTableOffset(read->Arg(1), current_key, offset, stride)) {
+			    (!indexed && !MatchTableOffset(read->Arg(1), current_key, offset, stride))) {
 				return false;
 			}
+			if (indexed) {
+				current_key = read->Arg(1);
+				offset = component * 4u;
+			}
 			if (dword == 0u) {
+				if (component != 0u) return false;
+				plan.table_indexed = indexed;
 				key = current_key;
 				table_offset = offset;
 				table_immediate = memory->offset;
 				table_stride = stride;
-			} else if (table_stride != stride || !EquivalentValue(m_program, key, current_key) ||
+			} else if (plan.table_indexed != indexed || table_stride != stride ||
+			           !EquivalentValue(m_program, key, current_key) ||
 			           uint64_t {table_offset} + table_immediate + dword * sizeof(uint32_t) !=
 			               uint64_t {offset} + memory->offset) {
 				return false;
 			}
-			plan.split_offsets |= offset != table_offset;
+			plan.split_offsets |= !indexed && offset != table_offset;
 			table_handle = current_handle;
 			if (handle.NumArgs() == 8u) {
-				plan.retain_reads |= !UsesOnlyImageDescriptors(*read);
+				if (memory->kind == ResourceKind::ScalarAddress)
+					plan.retain_reads |= !UsesOnlyImageDescriptors(*read);
 			} else if (std::ranges::any_of(read->Uses(), [](const Use& use) {
 				return use.user->GetOpcode() != ValueOpcode::GetBufferResource ||
 				       std::ranges::any_of(use.user->Uses(), [](const Use& consumer) {
@@ -1666,36 +1715,6 @@ private:
 		return true;
 	}
 
-	bool MatchUniformizedBufferKey(Value key, const Inst& image,
-	                               DescriptorSource::IndirectDescriptor& indirect,
-	                               DescriptorSource& material_source) {
-		const auto* selected = UniformizedMaterialValue(key, image);
-		if (selected == nullptr) return false;
-		DescriptorSource::IndirectDescriptor::SelectorRead selector;
-		const auto value = MaterialKey(selected->Arg(1), selector);
-		if (selector.shift == 0u && selector.bits == UINT32_MAX) return false;
-		const auto* loaded = value.TryInstruction();
-		if (loaded == nullptr || loaded->GetOpcode() != ValueOpcode::SelectU32 ||
-		    !EquivalentValue(m_program, selected->Arg(0), loaded->Arg(0))) return false;
-		const auto* read = loaded->Arg(1).Resolve().TryInstruction();
-		if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32) return false;
-		const auto& memory = m_program.memory_info[read->Flags<MemoryFlags>().index];
-		uint32_t offset = 1u, scalar = 1u;
-		if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
-		    !memory.idxen || memory.offen ||
-		    !ImmediateU32(read->Arg(2), offset) || offset != 0u ||
-		    !ImmediateU32(read->Arg(3), scalar) || scalar != 0u ||
-		    !EquivalentValue(m_program, selected->Arg(0), read->Arg(4)) ||
-		    !MakeRuntimeTableSource(*read, material_source)) return false;
-		// GPU index arithmetic may wrap. Every record in the bounded V# is a candidate.
-		selector.source = InternSource(material_source);
-		selector.offset = memory.offset;
-		indirect.selectors.push_back(selector);
-		indirect.selector_first = Value(0u);
-		indirect.key_count = material_source.dwords[2];
-		return true;
-	}
-
 	bool TryMakeIndirectImage(Inst& handle, const DescriptorSource& descriptor,
 	                          IndirectDescriptorPlan& plan) {
 		if (handle.GetOpcode() != ValueOpcode::GetImageResource || handle.NumArgs() != 8u) return false;
@@ -1712,7 +1731,7 @@ private:
 		indirect.table_immediate = table_immediate;
 		indirect.table_stride = table_stride;
 		indirect.workgroup_axis = WorkgroupAxis(key);
-		if (indirect.workgroup_axis != UINT32_MAX) {
+		if (indirect.workgroup_axis != UINT32_MAX && table_source.dword_count == 2u) {
 			// The descriptor also supplies dimensions to shader arithmetic. Keep its reads;
 			// only the image handle is projected onto the bounded workgroup key.
 			plan.retain_reads = true;
@@ -1734,39 +1753,20 @@ private:
 			}
 			if ((table_offset & 3u) != 0u ||
 			    (bitscan && table_offset > UINT32_MAX - (32u * 32u - 1u))) return false;
-		} else if (!MatchUniformizedBufferKey(key, handle, indirect, material_source)) {
-			std::vector<const Inst*> visited;
-			auto selected = key;
-			for (;;) {
-				DescriptorSource::IndirectDescriptor::SelectorRead selector;
-				const auto* read = MaterialKey(selected, selector).TryInstruction();
-				uint32_t memory_index = 0;
-				const auto* memory = read != nullptr ? ScalarReadMemory(*read, memory_index) : nullptr;
-				DescriptorSource source;
-				if (memory == nullptr || memory->kind != ResourceKind::ScalarBuffer ||
-				    memory->offset > INT32_MAX || !MemoryIndexBelongsTo(memory_index, *read) ||
-				    !MakeRuntimeTableSource(*read, source)) break;
-				if (std::ranges::find(visited, read) != visited.end()) return false;
-				visited.push_back(read);
-				selected = MaterialOffset(read->Arg(1), selector.stride, selector.offset);
-				if (selected.IsEmpty()) return false;
-				selector.source = InternSource(source);
-				selector.immediate = memory->offset;
-				indirect.selectors.push_back(selector);
-				material_source = source;
-			}
-			if (indirect.selectors.empty()) return false;
-			std::ranges::reverse(indirect.selectors);
-		}
-		// Separate shader additions may only be merged into a contiguous descriptor
-		// when the selector cannot wrap between those additions.
-		if (plan.split_offsets && indirect.workgroup_axis == UINT32_MAX &&
-		    table_source.dword_count == 4u) {
-			const auto& selector = indirect.selectors.back();
-			const auto maximum_key = (UINT32_MAX >> selector.shift) & selector.bits;
-			if (uint64_t {maximum_key} * table_stride + table_offset + table_immediate + 28u >
-			    UINT32_MAX) return false;
+		} 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.
+			if (plan.split_offsets) return false;
+			indirect.table_stride = 0u;
+			indirect.table_offset = 0u;
+			indirect.table_immediate = 0u;
+			indirect.workgroup_axis = UINT32_MAX;
+			plan.table_read = plan.reads[0];
+			if (plan.table_indexed) indirect.table_record_bytes = table_immediate + 32u;
+			else indirect.table_scalar = true;
+			plan.retain_reads = true;
 		}
+
 		if (plan.retain_reads) plan.reads.fill(nullptr);
 		indirect.table_source = InternSource(table_source);
 		DescriptorSource image_source;
@@ -1836,7 +1836,7 @@ private:
 		indirect.table_stride = table_stride;
 		if (!MakeRuntimeTableSource(*read, material_source) ||
 		    !BoundedSelectedIndex(read->Arg(1), read->Arg(4), indirect)) return false;
-		indirect.selectors.push_back({.source = InternSource(material_source),
+		indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead{.source = InternSource(material_source),
 		                              .offset = memory.offset + lane * 4u});
 		indirect.table_source = InternSource(table_source);
 		DescriptorSource source;