shader: batch dependent folds to avoid repeated SSA graph walks

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

@@ -51,11 +51,13 @@ public:
 	}
 
 	void QueueUsers(const std::vector<Use>& users) {
-		++generation;
-		affected.clear();
 		for (const auto& use: users) affected.push_back(use.user);
+	}
+
+	void FlushUsers() {
+		++generation;
 		// Pattern folds inspect nested operands even when intermediate operations
-		// stay dynamic. Visit consumers once per change, including cyclic Phi edges.
+		// stay dynamic. Visit their union once per drain, including cyclic Phi edges.
 		for (size_t i = 0; i < affected.size(); ++i) {
 			auto* inst = affected[i];
 			const auto found = locations.find(inst);
@@ -64,6 +66,7 @@ public:
 			Queue(inst);
 			for (const auto& use: inst->Uses()) affected.push_back(use.user);
 		}
+		affected.clear();
 	}
 
 	Block::iterator Prepend(Block& block, Block::iterator before, ValueOpcode opcode,
@@ -977,27 +980,30 @@ void ConstantPropagationPass(const BlockList& blocks, uint32_t wave_size) {
 	std::vector<Use> users;
 	std::vector<Value> operands;
 	const auto drain = [&] {
-		for (size_t i = 0; i < worklist.pending.size(); ++i) {
+		do {
-			auto* inst = worklist.pending[i];
+			for (size_t i = 0; i < worklist.pending.size(); ++i) {
-			auto& location = worklist.locations.at(inst);
+				auto* inst = worklist.pending[i];
-			location.queued = false;
+				auto& location = worklist.locations.at(inst);
-			const auto opcode = inst->GetOpcode();
+				location.queued = false;
-			if (opcode == ValueOpcode::Identity || opcode == ValueOpcode::Void) continue;
+				const auto opcode = inst->GetOpcode();
-			const auto where = location.instruction;
+				if (opcode == ValueOpcode::Identity || opcode == ValueOpcode::Void) continue;
-			users.assign(inst->Uses().begin(), inst->Uses().end());
+				const auto where = location.instruction;
-			operands.clear();
+				users.assign(inst->Uses().begin(), inst->Uses().end());
-			for (size_t arg = 0; arg < inst->NumArgs(); ++arg) operands.push_back(inst->Arg(arg));
+				operands.clear();
-			mask_projection.Fold(*inst, wave_size);
+				for (size_t arg = 0; arg < inst->NumArgs(); ++arg) operands.push_back(inst->Arg(arg));
-			FoldInstruction(*inst->Parent(), where, worklist);
+				mask_projection.Fold(*inst, wave_size);
-			bool changed = opcode != inst->GetOpcode() || operands.size() != inst->NumArgs();
+				FoldInstruction(*inst->Parent(), where, worklist);
-			for (size_t arg = 0; !changed && arg < operands.size(); ++arg)
+				bool changed = opcode != inst->GetOpcode() || operands.size() != inst->NumArgs();
-				changed = operands[arg] != inst->Arg(arg);
+				for (size_t arg = 0; !changed && arg < operands.size(); ++arg)
-			if (changed) {
+					changed = operands[arg] != inst->Arg(arg);
-				worklist.QueueUsers(users);
+				if (changed) {
-				worklist.Queue(inst);
+					worklist.QueueUsers(users);
+					worklist.Queue(inst);
+				}
 			}
-		}
+			worklist.pending.clear();
-		worklist.pending.clear();
+			worklist.FlushUsers();
+		} while (!worklist.pending.empty());
 	};
 	drain();
 	// Normalize retained PHI/select values only after every supported field read has