shader: remove masked-value pass and normalize ancillary fields

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -651,9 +651,6 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	IR::ResolveControlFlowIdentities(ir);
 	IR::RemoveIdentities(ir.blocks);
 	IR::EliminateDeadCode(ir.blocks);
-	IR::EliminateMaskedValues(ir);
-	IR::RemoveIdentities(ir.blocks);
-	IR::EliminateDeadCode(ir.blocks);
 	const auto read_lane_stats = IR::EliminateReadLane(ir, ir.wave_size);
 	if (read_lane_stats.rewritten_reads != 0) {
 		LOGF("%s read-lane elimination: reads=%" PRIu32 "\n", GetDumpLabel(options),

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

@@ -268,20 +268,13 @@ void Translator::S_MOV_B64(const Decoder::Instruction& inst) {
 }
 
 void Translator::S_WQM_B64(const Decoder::Instruction& inst) {
-	const auto source_mask = ReadMask(inst.src0);
 	const auto mask_valid  = ReadMaskValid(inst.src0);
 	const auto result =
 	    IR::U64(ir.Emit(IR::ValueOpcode::WqmU64, {ReadOperand(inst.src0, IR::Type::U64)}));
-	// WQM activates whole quads and cannot clear an already active lane.
-	const auto widened_mask = ir.LogicalOr(source_mask, ThreadBit(ExtractU64(result)));
 	WriteOperand(DestinationOperand(inst), result);
-	if (inst.dst.kind == Decoder::OperandKind::ExecLo) {
-		ir.SetExec(widened_mask);
-	} else if (inst.dst.kind == Decoder::OperandKind::VccLo) {
-		ir.SetVcc(widened_mask);
-	} else if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
+	if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
 		const auto dst = static_cast<IR::ScalarReg>(inst.dst.reg);
-		ir.SetThreadBitScalarReg(dst, widened_mask);
+		ir.SetThreadBitScalarReg(dst, ThreadBit(ExtractU64(result)));
 		ir.SetScalarMaskTag(dst, mask_valid);
 	}
 	ir.SetScc(IR::U1(ir.Emit(IR::ValueOpcode::INotEqual64, {result, IR::Value(uint64_t {0})})));

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

@@ -4,7 +4,7 @@
 #include <algorithm>
 #include <bit>
 #include <cstdint>
-#include <unordered_map>
+#include <unordered_set>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 namespace {
@@ -118,95 +118,6 @@ bool FoldSelect(Inst& inst) {
 	return false;
 }
 
-bool IsSelect(ValueOpcode opcode) {
-	return opcode == ValueOpcode::SelectU1 || opcode == ValueOpcode::SelectU32 ||
-	       opcode == ValueOpcode::SelectF32;
-}
-
-bool Implies(Value condition, Value required, uint32_t depth = 0) {
-	condition = condition.Resolve();
-	required = required.Resolve();
-	if (condition == required) return true;
-	if (depth >= 64u) return false;
-	const auto* target = required.TryInstruction();
-	if (target != nullptr && target->GetOpcode() == ValueOpcode::LogicalOr &&
-	    (Implies(condition, target->Arg(0), depth + 1u) ||
-	     Implies(condition, target->Arg(1), depth + 1u))) {
-		return true;
-	}
-	const auto* inst = condition.TryInstruction();
-	return inst != nullptr && inst->GetOpcode() == ValueOpcode::LogicalAnd &&
-	       (Implies(inst->Arg(0), required, depth + 1u) ||
-	        Implies(inst->Arg(1), required, depth + 1u));
-}
-
-bool IsAcyclic(const BlockList& blocks) {
-	std::unordered_map<const Block*, bool> completed;
-	const auto visit = [&](auto&& self, const Block* block) -> bool {
-		const auto [entry, added] = completed.emplace(block, false);
-		if (!added) return entry->second;
-		for (const auto* successor: block->ImmSuccessors()) {
-			if (!self(self, successor)) return false;
-		}
-		completed[block] = true;
-		return true;
-	};
-	for (const auto* block: blocks) {
-		if (!visit(visit, block)) return false;
-	}
-	return true;
-}
-
-class MaskedUses {
-public:
-	MaskedUses(const Program& program, Value condition)
-	    : program(program), condition(condition.Resolve()) {}
-
-	bool OnlyWhenActive(const Inst& inst) {
-		const auto [entry, added] = completed.emplace(&inst, false);
-		if (!added) return entry->second;
-		for (const auto& use: inst.Uses()) {
-			const auto& user = *use.user;
-			const auto  op   = user.GetOpcode();
-			if (IsSelect(op)) {
-				if (use.operand == 0) return false;
-				if (use.operand == 1 && Implies(user.Arg(0), condition)) continue;
-			} else if (op == ValueOpcode::SetAttribute) {
-				if (use.operand == 0 && Implies(user.Arg(1), condition)) continue;
-				return false;
-			} else if (!IsLaneLocal(user, use.operand)) {
-				return false;
-			}
-			if (!OnlyWhenActive(user)) return false;
-		}
-		completed[&inst] = true;
-		return true;
-	}
-
-private:
-	bool IsLaneLocal(const Inst& inst, size_t operand) const {
-		const auto op = inst.GetOpcode();
-		// These opcode-table groups contain only construction, conversion and arithmetic.
-		if ((op >= ValueOpcode::BitCastU16F16 && op <= ValueOpcode::BitFieldSExtract) ||
-		    (op >= ValueOpcode::WqmU64 && op <= ValueOpcode::FPFract32) ||
-		    op == ValueOpcode::Identity || op == ValueOpcode::Phi ||
-		    op == ValueOpcode::MakeImageAddress) {
-			return true;
-		}
-		if (op != ValueOpcode::ImageSampleRaw || operand != 2) return false;
-		// Explicit LOD or gradients keep coordinates local; implicit derivatives read neighbors.
-		const auto index = inst.Flags<MemoryFlags>().index;
-		return index < program.memory_info.size() &&
-		       (program.memory_info[index].image_sample_flags &
-		        (Decoder::ImageSampleFlagLod | Decoder::ImageSampleFlagLevelZero |
-		         Decoder::ImageSampleFlagDerivative)) != 0;
-	}
-
-	const Program&                       program;
-	Value                                condition;
-	std::unordered_map<const Inst*, bool> completed;
-};
-
 bool FoldPhi(Inst& inst) {
 	Value same;
 	for (size_t index = 0; index < inst.NumArgs(); index++) {
@@ -283,7 +194,8 @@ bool FoldCompositeExtract(Inst& inst, ValueOpcode construct, size_t components)
 	return false;
 }
 
-void FoldInstruction(Block& block, Block::iterator instruction) {
+void FoldInstruction(Block& block, Block::iterator instruction,
+                      std::unordered_set<Inst*>& lowered_ancillary) {
 	auto& inst = *instruction;
 	switch (inst.GetOpcode()) {
 		case ValueOpcode::Phi: FoldPhi(inst); return;
@@ -315,7 +227,7 @@ void FoldInstruction(Block& block, Block::iterator instruction) {
 			const auto value  = Arg(inst, 0);
 			const auto offset = Arg(inst, 1);
 			const auto count  = Arg(inst, 2);
-			const auto* source = value.TryInstruction();
+			auto* source = value.TryInstruction();
 			if (source != nullptr && source->GetOpcode() == ValueOpcode::GetBuiltin &&
 			    source->Arg(0) == Value(static_cast<uint32_t>(StageInputKind::PackedAncillary)) &&
 			    IsImmediate(offset, Type::U32) && IsImmediate(count, Type::U32) && count.U32() != 0u) {
@@ -333,6 +245,7 @@ void FoldInstruction(Block& block, Block::iterator instruction) {
 						    {Value(static_cast<uint32_t>(field.kind)), Value(0u)});
 						inst.SetArg(0, Value(&*input));
 						inst.SetArg(1, Value(offset.U32() - field.start));
+						lowered_ancillary.insert(source);
 						return;
 					}
 				}
@@ -697,23 +610,25 @@ void FoldInstruction(Block& block, Block::iterator instruction) {
 } // namespace
 
 void ConstantPropagationPass(const BlockList& blocks) {
+	std::unordered_set<Inst*> lowered_ancillary;
 	for (auto* block: blocks) {
 		for (auto inst = block->begin(); inst != block->end(); ++inst) {
-			FoldInstruction(*block, inst);
+			FoldInstruction(*block, inst, lowered_ancillary);
 		}
 	}
-}
-
-void EliminateMaskedValues(Program& program) {
-	// A loop-carried value may have been produced under a previous iteration's predicate.
-	if (!IsAcyclic(program.blocks)) return;
-	for (auto block = program.blocks.rbegin(); block != program.blocks.rend(); ++block) {
-		auto& instructions = (*block)->Instructions();
-		for (auto inst = instructions.rbegin(); inst != instructions.rend(); ++inst) {
-			if (IsSelect(inst->GetOpcode()) &&
-			    MaskedUses(program, inst->Arg(0)).OnlyWhenActive(*inst)) {
-				Replace(*inst, inst->Arg(1));
+	// Normalize retained PHI/select values only after every supported field read has
+	// been lowered; direct raw consumers remain unsupported.
+	for (auto* source: lowered_ancillary) {
+		const bool retained_only = std::ranges::all_of(source->Uses(), [](const Use& use) {
+			const auto& user = *use.user;
+			if (!user.HasUses() && !user.MayHaveSideEffects()) {
+				return true;
 			}
+			return user.GetOpcode() == ValueOpcode::Phi ||
+			       (user.GetOpcode() == ValueOpcode::SelectU32 && use.operand == 2u);
+		});
+		if (retained_only) {
+			Replace(*source, Value(0u));
 		}
 	}
 }

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

@@ -4,9 +4,6 @@
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-struct Program;
-
 void ConstantPropagationPass(const BlockList& blocks);
-void EliminateMaskedValues(Program& program);
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

tests/shaderCfgTests.cpp

@@ -6007,6 +6007,30 @@ void TestPixelAncillaryLayerInput() {
     CheckSpirvBinaryValidates(wqm_result.spirv);
   }
 
+  const uint32_t loop[] = {
+      EncodeSop1(0x04, 8, 126),  // Save the original EXEC mask.
+      EncodeSop1(0x0a, 126, 8),   // Widen for helper lanes.
+      EncodeSop1(0x24, 12, 0),   // Narrow EXEC while saving the widened mask.
+      EncodeSop1(0x04, 10, 126), // Save the final export mask.
+      EncodeSopp(0x08, 6),       // Skip the body when EXEC is zero.
+      EncodeVop3Word0(0x148, 5), EncodeVop3Word1(5 + 256, 128 + 16, 128 + 11),
+      EncodeSMovB32(2, 128),
+      EncodeSop2(0x00, 2, 2, 129),
+      EncodeSopc(0x0a, 2, 131),
+      EncodeSopp(0x05, 0xfffdu), // Repeat the scalar loop three times.
+      EncodeSop1(0x04, 126, 12), // Pack under the widened mask.
+      EncodeVop2(0x2f, 6, 5 + 256, 5),
+      EncodeSop1(0x04, 126, 10), // Export only lanes that entered the body.
+      EncodeExp0(0x00, 0x1, true, false, true), EncodeExp1(6, 0, 0, 0), EncodeSopp(0x01),
+  };
+  auto loop_result = RecompileForTest(loop, options);
+  Check(ProgramHasInput(loop_result.program, StageInputKind::Layer) &&
+            !ProgramHasInput(loop_result.program, StageInputKind::PackedAncillary),
+        "loop and skipped body retained the symbolic ancillary input after field lowering");
+  Check(SpirvContainsOpcode(loop_result.spirv, 246u),
+        "ancillary regression did not retain its scalar loop");
+  CheckSpirvBinaryValidates(loop_result.spirv);
+
   const uint32_t unused[] = {EncodeSopp(0x01)};
   auto unused_result = RecompileForTest(unused, options);
   Check(!ProgramHasInput(unused_result.program, StageInputKind::Layer) &&
@@ -6031,122 +6055,14 @@ void TestPixelAncillaryLayerInput() {
   shader[1] = EncodeVop3Word1(5 + 256, 128 + 12, 128 + 4);
   ExpectFatal([&] { (void)RecompileForTest(shader, options); },
               "unsupported live ancillary field was silently replaced");
+  shader[1] = EncodeVop3Word1(5 + 256, 128 + 16, 128 + 11);
+  shader[4] = EncodeExp0(0x00, 0x7);
+  shader[5] = EncodeExp1(6, 7, 5, 0);
+  ExpectFatal([&] { (void)RecompileForTest(shader, options); },
+              "direct raw ancillary export was silently replaced after field lowering");
 #endif
 }
 
-void TestMaskedValueDemand() {
-  using namespace ShaderRecompiler::IR;
-  enum class Use { Guarded, Conjunction, Wqm, WqmConjunction, WqmOtherGuard,
-                   TrueArm, Unguarded, OtherGuard, Predicate,
-                   ReadLane, ExplicitLod, ExplicitGradient, LevelZero, ImplicitLod,
-                   Mixed, Phi, Loop };
-  for (const auto use : {Use::Guarded, Use::Conjunction, Use::Wqm,
-                         Use::WqmConjunction, Use::WqmOtherGuard,
-                         Use::TrueArm, Use::Unguarded,
-                         Use::OtherGuard, Use::Predicate, Use::ReadLane,
-                         Use::ExplicitLod, Use::ExplicitGradient, Use::LevelZero,
-                         Use::ImplicitLod, Use::Mixed, Use::Phi, Use::Loop}) {
-    Program program;
-    for (uint32_t i = 0; i < 4; ++i) {
-      program.block_storage.push_back(std::make_unique<Block>());
-      program.blocks.push_back(program.block_storage.back().get());
-    }
-    auto &entry = *program.blocks[0];
-    auto &join = *program.blocks[3];
-    entry.AddBranch(program.blocks[1]);
-    entry.AddBranch(program.blocks[2]);
-    program.blocks[1]->AddBranch(&join);
-    program.blocks[2]->AddBranch(&join);
-    if (use == Use::Loop) join.AddBranch(&entry);
-    const auto input = Value(&entry.AppendNewInst(ValueOpcode::GetUserData,
-                                                 {Value(static_cast<ScalarReg>(0))}));
-    const auto predicate = Value(&entry.AppendNewInst(ValueOpcode::INotEqual32,
-                                                      {input, Value(0u)}));
-    const auto other = Value(&entry.AppendNewInst(ValueOpcode::IEqual32,
-                                                  {input, Value(2u)}));
-    auto mask = predicate;
-    if (use == Use::Wqm || use == Use::WqmConjunction || use == Use::WqmOtherGuard) {
-      // WQM includes the original active lanes and additional helper lanes.
-      const auto helpers = Value(&entry.AppendNewInst(ValueOpcode::GetUserData,
-                                                       {Value(static_cast<ScalarReg>(1))}));
-      const auto helper = Value(&entry.AppendNewInst(ValueOpcode::INotEqual32,
-                                                      {helpers, Value(0u)}));
-      mask = Value(&entry.AppendNewInst(ValueOpcode::LogicalOr, {predicate, helper}));
-    }
-    auto &masked = entry.AppendNewInst(ValueOpcode::SelectU32,
-                                        {mask, Value(10u), Value(20u)});
-    Value value(&masked);
-    if (use == Use::Phi || use == Use::Loop) {
-      auto &phi = join.AppendNewInst(ValueOpcode::Phi, {}, static_cast<uint64_t>(Type::U32));
-      phi.AddPhiOperand(program.blocks[1], value);
-      phi.AddPhiOperand(program.blocks[2], Value(0u));
-      value = Value(&phi);
-    }
-    value = Value(&join.AppendNewInst(ValueOpcode::BitCastF32U32, {value}));
-    value = Value(&join.AppendNewInst(ValueOpcode::FPMul32, {value, Value::F32(2.f)}));
-    value = Value(&join.AppendNewInst(ValueOpcode::BitCastU32F32, {value}));
-    if (use == Use::ReadLane) {
-      value = Value(&join.AppendNewInst(ValueOpcode::ReadLane, {value, Value(0u)}));
-    }
-    if (use == Use::ExplicitLod || use == Use::ExplicitGradient ||
-        use == Use::LevelZero || use == Use::ImplicitLod) {
-      const auto zero = Value(0u);
-      auto &address = join.AppendNewInst(ValueOpcode::MakeImageAddress,
-          {value, zero, zero, zero, zero, zero, zero, zero, zero, zero, zero, zero, zero});
-      auto &image = entry.AppendNewInst(ValueOpcode::GetImageResource,
-          {zero, zero, zero, zero, zero, zero, zero, zero});
-      auto &sampler = entry.AppendNewInst(ValueOpcode::GetSamplerResource,
-          {zero, zero, zero, zero});
-      auto &memory = program.memory_info.emplace_back();
-      switch (use) {
-      case Use::ExplicitLod:
-        memory.image_sample_flags =
-            ShaderRecompiler::Decoder::ImageSampleFlagLod;
-        break;
-      case Use::ExplicitGradient:
-        memory.image_sample_flags =
-            ShaderRecompiler::Decoder::ImageSampleFlagDerivative;
-        break;
-      case Use::LevelZero:
-        memory.image_sample_flags =
-            ShaderRecompiler::Decoder::ImageSampleFlagLevelZero;
-        break;
-      default:
-        break;
-      }
-      value = Value(&join.AppendNewInst(ValueOpcode::ImageSampleRaw,
-                                         {Value(&image), Value(&sampler), Value(&address)}));
-      value = Value(&join.AppendNewInst(ValueOpcode::CompositeExtractU32x4, {value, Value(0u)}));
-    }
-    auto guard = predicate;
-    if (use == Use::Conjunction || use == Use::WqmConjunction) {
-      guard = Value(&join.AppendNewInst(ValueOpcode::LogicalAnd, {predicate, other}));
-    } else if (use == Use::WqmOtherGuard) {
-      guard = Value(&join.AppendNewInst(ValueOpcode::IEqual32, {input, Value(0u)}));
-    } else if (use == Use::OtherGuard) {
-      guard = other;
-    } else if (use == Use::Unguarded) {
-      guard = Value(true);
-    } else if (use == Use::Predicate) {
-      guard = Value(&join.AppendNewInst(ValueOpcode::INotEqual32, {value, Value(0u)}));
-    } else if (use == Use::TrueArm) {
-      value = Value(&join.AppendNewInst(ValueOpcode::SelectU32, {predicate, value, Value(0u)}));
-      guard = Value(true);
-    }
-    value = Value(&join.AppendNewInst(ValueOpcode::CompositeConstructU32x4,
-                                       {value, Value(0u), Value(0u), Value(0u)}));
-    join.AppendNewInst(ValueOpcode::SetAttribute, {value, guard});
-    if (use == Use::Mixed) join.AppendNewInst(ValueOpcode::SetAttribute, {value, Value(true)});
-    EliminateMaskedValues(program);
-    const bool removed = Value(&masked).Resolve() == Value(10u);
-    const bool safe = use == Use::Guarded || use == Use::Conjunction ||
-                      use == Use::Wqm || use == Use::WqmConjunction || use == Use::TrueArm ||
-                      use == Use::ExplicitLod || use == Use::ExplicitGradient ||
-                      use == Use::LevelZero || use == Use::Phi;
-    Check(removed == safe, "masked value demand crossed an unguarded or nonlocal use");
-  }
-}
-
 void TestGraphicsCreateInterpolantMapping() {
   ShaderRegister regs[32] = {};
 
@@ -12787,7 +12703,6 @@ int main() {
   TestNewShaderRecompilerRejectsDppOn64BitCompares();
   TestPsInputCountRegisterDecode();
   TestPixelAncillaryLayerInput();
-  TestMaskedValueDemand();
   TestNewShaderRecompilerUnbasedFlatUsesBda();
   TestNewShaderRecompilerFlatUserPointerUsesDma();
   TestNewShaderRecompilerFlatAddressDomainsUseDma();