feat: Implement AffineToNeura pass with loop nest analysis and valid signal optimization

include/Conversion/AffineToNeura/LoopNestAnalysis.h

@@ -0,0 +1,80 @@
+//===- LoopNestAnalysis.h - Analyze affine loop nests ----------*- C++ -*-===//
+//
+// Loop nest analysis for affine loops.
+// 
+// Features:
+// 1. Build loop hierarchy tree (parent-child relationships, nesting depth)
+// 2. Identify perfect vs imperfect nesting
+// 3. Support valid signal reuse optimization for nested loops
+//
+//===----------------------------------------------------------------------===//
+#ifndef CONVERSION_AFFINE_TO_NEURA_LOOP_NEST_ANALYSIS_H
+#define CONVERSION_AFFINE_TO_NEURA_LOOP_NEST_ANALYSIS_H
+
+#include "mlir/Dialect/Affine/IR/AffineOps.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/IR/Operation.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallVector.h"
+#include <memory>
+
+namespace mlir {
+namespace neura {
+
+/// Loop information structure - Stores all analysis information for a single loop.
+struct LoopInfo {
+  affine::AffineForOp loop;              // The loop operation itself.
+  LoopInfo *parent = nullptr;            // Parent loop (nullptr if top-level).
+  llvm::SmallVector<LoopInfo *, 4> children;  // Child loops list.
+  unsigned depth = 0;                    // Nesting depth (0=top-level).
+  bool isPerfectNest = true;             // Whether it is a perfect nest.
+
+  // Operations list for imperfect nesting.
+  llvm::SmallVector<Operation *, 4> operationsBeforeChild;  // Operations before child loops.
+  llvm::SmallVector<Operation *, 4> operationsAfterChild;   // Operations after child loops.
+
+  LoopInfo(affine::AffineForOp loop) : loop(loop) {}
+};
+
+/// Loop nest analysis class.
+/// 
+/// Purpose: Provides loop hierarchy information for AffineToNeura pass to support optimization decisions.
+/// 
+/// Usage example:
+///   LoopNestAnalysis analysis(func_op);
+///   analysis.dump();  // Prints analysis results.
+///   LoopInfo *info = analysis.getLoopInfo(loop);
+///   if (info && info->parent) {
+///     // This is a nested loop, can reuse parent's valid signal.
+///   }
+class LoopNestAnalysis {
+public:
+  /// Constructor - Performs loop nest analysis on the given function.
+  explicit LoopNestAnalysis(func::FuncOp func);
+
+  /// Query interfaces.
+  LoopInfo *getLoopInfo(affine::AffineForOp loop) const;  // Gets loop information.
+  llvm::ArrayRef<LoopInfo *> getTopLevelLoops() const { return topLevelLoops; }  // Gets top-level loops.
+  llvm::ArrayRef<std::unique_ptr<LoopInfo>> getAllLoops() const { return allLoops; }  // Gets all loops.
+  bool isPerfectNest(affine::AffineForOp loop) const;  // Checks if perfect nest.
+  LoopInfo *getParentLoop(affine::AffineForOp loop) const;  // Gets parent loop.
+  llvm::ArrayRef<LoopInfo *> getChildLoops(affine::AffineForOp loop) const;  // Gets child loops.
+
+  /// Debug interface - Prints analysis results.
+  void dump() const;
+
+private:
+  /// Internal analysis methods.
+  void buildLoopNestTree(func::FuncOp func);  // Builds loop hierarchy tree.
+  void analyzePerfectNests();  // Analyzes perfect nest characteristics.
+
+  /// Data members.
+  llvm::DenseMap<Operation *, LoopInfo *> loopMap;  // Loop fast lookup table.
+  llvm::SmallVector<std::unique_ptr<LoopInfo>, 8> allLoops;  // All loops (owns ownership).
+  llvm::SmallVector<LoopInfo *, 4> topLevelLoops;  // Top-level loop pointers list.
+};
+
+} // namespace neura
+} // namespace mlir
+
+#endif

include/Conversion/ConversionPasses.h

@@ -18,6 +18,7 @@ std::unique_ptr<mlir::Pass> createLowerArithToNeuraPass();
 std::unique_ptr<mlir::Pass> createLowerLlvmToNeuraPass();
 std::unique_ptr<mlir::Pass> createLowerMemRefToNeuraPass();
 std::unique_ptr<mlir::Pass> createLowerBuiltinToNeuraPass();
+std::unique_ptr<mlir::Pass> createLowerAffineToNeuraPass();
 
 #define GEN_PASS_REGISTRATION
 #include "Conversion/ConversionPasses.h.inc"

include/Conversion/ConversionPasses.td

@@ -32,4 +32,16 @@ def LowerBuiltinToNeura : Pass<"lower-builtin-to-neura", "ModuleOp">{
   let constructor = "mlir::createLowerBuiltinToNeuraPass()";
 }
 
+def LowerAffineToNeura : Pass<"lower-affine-to-neura", "func::FuncOp">{
+  let summary = "Lower Affine perfect nested loops to Neura loop_control operations";
+  let description = [{
+    Converts perfectly nested affine.for loops directly to Neura dialect using 
+    loop_control operations, avoiding the need to flatten to LLVM IR first.
+    This preserves loop structure information for better optimization on 
+    dataflow architectures.
+  }];
+  let constructor = "mlir::createLowerAffineToNeuraPass()";
+  let dependentDialects = ["mlir::neura::NeuraDialect", "mlir::affine::AffineDialect"];
+}
+
 #endif // CONVERSION_PASSES_TD

include/NeuraDialect/Architecture/Architecture.h

@@ -57,7 +57,9 @@ enum OperationKind {
   // Loop control operations.
   ILoopControl = 34,
   // Constant operations.
-  IConstant = 35
+  IConstant = 35,
+  // Steering control fused operations.
+  ICarryInvariant = 36, IConditionalSelect = 37, IInvariantGroup = 38
 };
 
 //===----------------------------------------------------------------------===//

include/NeuraDialect/Mapping/mapping_util.h

@@ -12,6 +12,10 @@ OperationKind getOperationKindFromMlirOp(Operation *op);
 // Returns true if the operation does not need CGRA tile placement.
 bool is_non_materialized(Operation *op);
 
+// Returns true if the operation is a steering-mode operation that doesn't
+// require DataMovOp wrapping (e.g., constants, carry, invariant, etc.).
+bool is_steering_unwrapped_op(Operation *op);
+
 // Returns true if the operation is a materialized reserve user, i.e.,
 // phi, invariant, carry.
 bool isMaterializedReserveUser(Operation *op);

include/NeuraDialect/NeuraOps.td

@@ -657,4 +657,133 @@ def Neura_InvariantOp : Op<NeuraDialect, "invariant">{
   let arguments = (ins AnyType:$initial, AnyType:$condition);
   let results = (outs AnyType:$result);
   let assemblyFormat = "$initial `,` $condition attr-dict `:` type($initial) `,` type($condition) `->` type($result)";
+}
+
+// ============================================================================
+// FUSED OPERATIONS FOR RECMII OPTIMIZATION
+// ============================================================================
+
+// Defines the carry_invariant fused operation.
+def Neura_CarryInvariantOp : Op<NeuraDialect, "carry_invariant">{
+  let summary = "Fused carry and invariant operation for nested loops.";
+  let description = [{
+    Combines carry and invariant operations into a single operation to reduce RecMII.
+    This is optimized for nested loop patterns where an inner loop's carry result
+    is used as an invariant in the outer loop.
+
+    Semantics:
+    - If inner_condition is false (first inner iteration): return initial value
+    - Else if outer_condition is false (outer loop active, inner loop invariant): 
+        return initial value from inner carry
+    - Else: return carried value
+
+    Replaces the pattern:
+      %carry_result = neura.carry %init, %inner_cond, %carried
+      %inv_result = neura.invariant %carry_result, %outer_cond
+
+    With:
+      %result = neura.carry_invariant %init, %inner_cond, %outer_cond, %carried
+
+    RecMII Impact: Reduces 2 operations to 1 operation (-50% on critical path)
+
+    Example:
+      %out = neura.carry_invariant %init, %inner_cond, %outer_cond, %carried 
+             : i64, i1, i1, i64 -> i64
+  }];
+
+  let arguments = (ins 
+    AnyType:$initial,
+    AnyType:$inner_condition,
+    AnyType:$outer_condition,
+    AnyType:$carried
+  );
+  let results = (outs AnyType:$result);
+
+  let assemblyFormat = [{
+    $initial `,` $inner_condition `,` $outer_condition `,` $carried attr-dict 
+    `:` type($initial) `,` type($inner_condition) `,` type($outer_condition) `,` 
+    type($carried) `->` type($result)
+  }];
+}
+
+// Defines the conditional_select fused operation.
+def Neura_ConditionalSelectOp : Op<NeuraDialect, "cond_select">{
+  let summary = "Fused comparison and conditional selection operation.";
+  let description = [{
+    Combines comparison (icmp) and conditional selection (false_steer) into a 
+    single atomic operation to reduce RecMII.
+
+    Semantics:
+    - Performs comparison: result = (lhs <predicate> rhs)
+    - If result is false: return value
+    - If result is true: return default value (typically from hardware)
+
+    Replaces the pattern:
+      %cond = neura.icmp %lhs, %rhs <{cmpType = "slt"}>
+      %result = neura.false_steer %value, %cond
+
+    With:
+      %result = neura.cond_select %lhs, %rhs, %value <{predicate = "slt"}>
+
+    RecMII Impact: Reduces 2 operations to 1 operation (-50% on critical path)
+
+    Supported predicates: "eq", "ne", "slt", "sle", "sgt", "sge", "ult", "ule", "ugt", "uge"
+
+    Example:
+      %out = neura.cond_select %a, %b, %val <{predicate = "slt"}> 
+             : i64, i64, i64 -> i64
+  }];
+
+  let arguments = (ins 
+    AnyType:$lhs,
+    AnyType:$rhs,
+    AnyType:$value,
+    StrAttr:$predicate
+  );
+  let results = (outs AnyType:$result);
+
+  let assemblyFormat = [{
+    $lhs `,` $rhs `,` $value attr-dict `:` type($lhs) `,` type($rhs) `,` 
+    type($value) `->` type($result)
+  }];
+}
+
+// Defines the invariant_group batch operation.
+def Neura_InvariantGroupOp : Op<NeuraDialect, "invariant_group">{
+  let summary = "Batch invariant extraction for multiple values.";
+  let description = [{
+    Extracts multiple invariants with the same condition in a single operation.
+    This is optimized for nested loops where many values need to be marked as
+    invariant with respect to the outer loop.
+
+    Hardware can optimize this by:
+    - Sharing condition checking logic
+    - Parallel invariant extraction
+    - Reduced control overhead
+
+    Replaces multiple individual invariant operations:
+      %inv1 = neura.invariant %val1, %cond
+      %inv2 = neura.invariant %val2, %cond
+      %inv3 = neura.invariant %val3, %cond
+
+    With a single batch operation:
+      %inv1, %inv2, %inv3 = neura.invariant_group %val1, %val2, %val3, %cond
+
+    ResMII Impact: Reduces N operations to 1 operation (improves resource utilization)
+
+    Example:
+      %out1, %out2, %out3 = neura.invariant_group %in1, %in2, %in3, %cond
+             : i64, i64, i64, i1 -> i64, i64, i64
+  }];
+
+  let arguments = (ins 
+    Variadic<AnyType>:$inputs,
+    AnyType:$condition
+  );
+  let results = (outs Variadic<AnyType>:$outputs);
+
+  let assemblyFormat = [{
+    $inputs `,` $condition attr-dict `:` type($inputs) `,` type($condition) 
+    `->` type($outputs)
+  }];
 }

include/NeuraDialect/NeuraPasses.td

@@ -134,4 +134,5 @@ def RemovePredicatedType : Pass<"remove-predicated-type", "ModuleOp"> {
   }];
   let constructor = "neura::createRemovePredicatedTypePass()";
 }
+
 #endif // NEURA_PASSES_TD