fmt

examples/fib_prim.py

@@ -1,7 +1,6 @@
-from lyrt import native, from_prim
+from lyrt import from_prim, native
 from lyrt.prim import Int
 
-
 p1 = Int[32](1)
 p2 = Int[32](2)
 p35 = Int[32](35)

examples/tensor_prim.py

@@ -1,11 +1,10 @@
 from lyrt import from_prim
-from lyrt.prim import Float, Vector, Matrix, Tensor
+from lyrt.prim import Float, Matrix, Tensor, Vector
 
 # ゼロ初期化
 
 v = Vector[Float[32], 4].zeros()
 m = Matrix[Float[32], 3, 4].zeros()
-
 t = Tensor[Float[32], 2, 3, 4].zeros()
 
 print(from_prim(v))
@@ -16,24 +15,28 @@
 
 v2 = Vector[Float[32], 4]([0.3, -1.2, 4.5, 2.1])
 
-m2 = Matrix[Float[32], 3, 4]([
-    [1.1, -0.7, 3.3, 0.0],
-    [2.4, 5.6, -2.2, 1.9],
-    [0.5, 4.8, -1.1, 3.7],
-])
-
-t2 = Tensor[Float[32], 2, 3, 4]([
+m2 = Matrix[Float[32], 3, 4](
     [
-        [0.9, -1.3, 2.2, 4.4],
-        [3.1, 0.8, -0.5, 6.6],
-        [7.7, -2.4, 1.2, 0.3],
-    ],
+        [1.1, -0.7, 3.3, 0.0],
+        [2.4, 5.6, -2.2, 1.9],
+        [0.5, 4.8, -1.1, 3.7],
+    ]
+)
+
+t2 = Tensor[Float[32], 2, 3, 4](
     [
-        [5.5, 2.6, -3.3, 1.4],
-        [8.8, -0.9, 4.0, 2.2],
-        [6.1, 3.3, -1.7, 9.9],
-    ],
-])
+        [
+            [0.9, -1.3, 2.2, 4.4],
+            [3.1, 0.8, -0.5, 6.6],
+            [7.7, -2.4, 1.2, 0.3],
+        ],
+        [
+            [5.5, 2.6, -3.3, 1.4],
+            [8.8, -0.9, 4.0, 2.2],
+            [6.1, 3.3, -1.7, 9.9],
+        ],
+    ]
+)
 
 print(from_prim(v2))
 print(from_prim(m2))
@@ -43,17 +46,21 @@
 
 v3 = Vector[Float[32], ...]([1.0, 2.0, 3.0, 4.0, 5.0])
 
-m3 = Matrix[Float[32], ...]([
-    [1.0, 2.0, 3.0],
-    [4.0, 5.0, 6.0],
-])
+m3 = Matrix[Float[32], ...](
+    [
+        [1.0, 2.0, 3.0],
+        [4.0, 5.0, 6.0],
+    ]
+)
 
-t3 = Tensor[Float[32], ...]([
+t3 = Tensor[Float[32], ...](
     [
-        [1.0, 2.0],
-        [3.0, 4.0],
-    ],
-])
+        [
+            [1.0, 2.0],
+            [3.0, 4.0],
+        ],
+    ]
+)
 
 print(from_prim(v3))
 print(from_prim(m3))
@@ -62,4 +69,5 @@
 # 0階テンソルはスカラーと同値
 t4 = Tensor[Float[32], ...](3.14)
 assert from_prim(t4) == Float[32](3.14)
+
 print(from_prim(t4))

src/lyrt/__init__.pyi

@@ -6,11 +6,22 @@ __all__ = ["prim", "native", "to_prim", "from_prim"]
 
 T = TypeVar("T")
 PrimT = TypeVar("PrimT", bound=prim.Prim[prim.Int] | prim.Prim[prim.Float])
-PrimFunc = TypeVar("PrimFunc", bound=Callable[..., prim.Prim[prim.Int]] | Callable[..., prim.Prim[prim.Float]])
+PrimFunc = TypeVar(
+    "PrimFunc",
+    bound=Callable[..., prim.Prim[prim.Int]] | Callable[..., prim.Prim[prim.Float]],
+)
 
 def native(
     *,
     gc: Literal["none", "shadow-stack", "rc"] = "none",
 ) -> Callable[[PrimFunc], PrimFunc]: ...
 def to_prim(value: object, prim_type: type[PrimT]) -> PrimT: ...
-def from_prim(prim_value: prim.Prim[prim.Int] | prim.Prim[prim.Float] | prim.Vector | prim.Matrix | prim.Tensor) -> object: ...
+def from_prim(
+    prim_value: (
+        prim.Prim[prim.Int]
+        | prim.Prim[prim.Float]
+        | prim.Vector
+        | prim.Matrix
+        | prim.Tensor
+    ),
+) -> object: ...

src/lyrt/prim/__init__.pyi

@@ -23,8 +23,8 @@ class Prim(Generic[PrimT]):
     def __neg__(self: PrimT) -> PrimT: ...
     def __lt__(self: PrimT, other: PrimT) -> Int[1]: ...
     def __le__(self: PrimT, other: PrimT) -> Int[1]: ...
-    def __eq__(self: PrimT, other: PrimT) -> Int[1]: ... # type: ignore
-    def __ne__(self: PrimT, other: PrimT) -> Int[1]: ... # type: ignore
+    def __eq__(self: PrimT, other: PrimT) -> Int[1]: ...  # type: ignore
+    def __ne__(self: PrimT, other: PrimT) -> Int[1]: ...  # type: ignore
     def __gt__(self: PrimT, other: PrimT) -> Int[1]: ...
     def __ge__(self: PrimT, other: PrimT) -> Int[1]: ...
 
@@ -46,14 +46,12 @@ class Int(PrimInt[Int]):
     """Integer primitive type"""
 
     def __class_getitem__(cls, key: int) -> Type[Int]: ...
-
     def __init__(self, value: int) -> None: ...
 
 class Float(PrimFloat[Float]):
     """Floating-point primitive type"""
 
     def __class_getitem__(cls, key: int) -> Type[Float]: ...
-
     def __init__(self, value: int | float) -> None: ...
 
 type NumberLike = Prim[Int | Float] | int | float
@@ -62,9 +60,7 @@ Shape = TypeVarTuple("Shape")
 
 class Vector:
     def __class_getitem__(cls, key: tuple[Prim[Int | Float], int]) -> Type[Vector]: ...
-
     def __init__(self, value: List[NumberLike]) -> None: ...
-
     @classmethod
     def zeros(cls) -> Vector: ...
     @classmethod
@@ -73,10 +69,10 @@ class Vector:
     def full(cls, value: NumberLike) -> Vector: ...
 
 class Matrix:
-    def __class_getitem__(cls, key: tuple[Prim[Int | Float], int, int]) -> Type[Matrix]: ...
-
+    def __class_getitem__(
+        cls, key: tuple[Prim[Int | Float], int, int]
+    ) -> Type[Matrix]: ...
     def __init__(self, value: List[List[NumberLike]]) -> None: ...
-
     @classmethod
     def zeros(cls) -> Matrix: ...
     @classmethod
@@ -85,11 +81,12 @@ class Matrix:
     def full(cls, value: NumberLike) -> Matrix: ...
 
 class Tensor:
-    def __class_getitem__(cls, key: tuple[Prim[Int | Float], Unpack[Shape]]) -> Type[Tensor]: ...
+    def __class_getitem__(
+        cls, key: tuple[Prim[Int | Float], Unpack[Shape]]
+    ) -> Type[Tensor]: ...
 
     type _NestedNumberLike = NumberLike | List[_NestedNumberLike]
     def __init__(self, value: _NestedNumberLike) -> None: ...
-
     @classmethod
     def zeros(cls) -> Tensor: ...
     @classmethod

src/lython/visitors/_base.py

@@ -359,7 +359,9 @@ def get_primitive_type_from_spec(self, base_type: str, bits: int) -> ir.Type:
             return ir.IntegerType.get_signless(bits, context=self.ctx)
         elif base_type == "Float":
             if bits not in FLOAT_VALID_BITS:
-                raise ValueError(f"Float bit width must be one of {sorted(FLOAT_VALID_BITS)}, got {bits}")
+                raise ValueError(
+                    f"Float bit width must be one of {sorted(FLOAT_VALID_BITS)}, got {bits}"
+                )
             if bits == 16:
                 return ir.F16Type.get(context=self.ctx)
             elif bits == 32:
@@ -400,11 +402,15 @@ def annotation_to_primitive_type(
 
                 if base_type in PRIMITIVE_BASE_TYPES:
                     # Get the bit width from the slice
-                    if isinstance(annotation.slice, ast.Constant) and isinstance(annotation.slice.value, int):
+                    if isinstance(annotation.slice, ast.Constant) and isinstance(
+                        annotation.slice.value, int
+                    ):
                         bits = annotation.slice.value
                         return self.get_primitive_type_from_spec(base_type, bits)
                     else:
-                        raise ValueError(f"Primitive type {base_type} requires an integer bit width")
+                        raise ValueError(
+                            f"Primitive type {base_type} requires an integer bit width"
+                        )
 
         return None
 

src/lython/visitors/expr.py

@@ -602,7 +602,9 @@ def _handle_prim_constructor(
 
         # Get the bit width from the subscript
         assert isinstance(node.func, ast.Subscript)
-        if not isinstance(node.func.slice, ast.Constant) or not isinstance(node.func.slice.value, int):
+        if not isinstance(node.func.slice, ast.Constant) or not isinstance(
+            node.func.slice.value, int
+        ):
             raise ValueError(f"{base_type} requires an integer bit width")
         bits = node.func.slice.value
 

src/lython/visitors/stmt.py

@@ -136,7 +136,9 @@ def hoge(n: int) -> int:
             ir.StringAttr.get(arg.arg, self.ctx) for arg in node.args.args
         ]
         arg_names_attr = (
-            ir.ArrayAttr.get(arg_name_attrs, context=self.ctx)  # pyright: ignore[reportUnknownMemberType]
+            ir.ArrayAttr.get(
+                arg_name_attrs, context=self.ctx
+            )  # pyright: ignore[reportUnknownMemberType]
             if arg_name_attrs
             else None
         )
@@ -403,7 +405,9 @@ def _visit_method_def(
             ir.StringAttr.get(arg.arg, self.ctx) for arg in node.args.args
         ]
         arg_names_attr = (
-            ir.ArrayAttr.get(arg_name_attrs, context=self.ctx)  # pyright: ignore[reportUnknownMemberType]
+            ir.ArrayAttr.get(
+                arg_name_attrs, context=self.ctx
+            )  # pyright: ignore[reportUnknownMemberType]
             if arg_name_attrs
             else None
         )
@@ -692,9 +696,15 @@ def visit_If(self, node: ast.If) -> None:
 
         assert self.current_block is not None
         parent_region = self.current_block.region
-        true_block = parent_region.blocks.append()  # pyright: ignore[reportUnknownMemberType]
-        false_block = parent_region.blocks.append()  # pyright: ignore[reportUnknownMemberType]
-        merge_block = parent_region.blocks.append()  # pyright: ignore[reportUnknownMemberType]
+        true_block = (
+            parent_region.blocks.append()
+        )  # pyright: ignore[reportUnknownMemberType]
+        false_block = (
+            parent_region.blocks.append()
+        )  # pyright: ignore[reportUnknownMemberType]
+        merge_block = (
+            parent_region.blocks.append()
+        )  # pyright: ignore[reportUnknownMemberType]
         with self._loc(node), self.insertion_point():
             cf_ops.CondBranchOp(cond, [], [], true_block, false_block)
 

tools/CLI.cpp

@@ -111,24 +111,82 @@ if __name__ == "__main__":
   }
 
   LogicalResult runPipeline(ModuleOp module, MLIRContext& context) {
-    PassManager pm(&context);
-    // Verify @native functions before any transformations
-    // This enforces the modal logic separation (Primitive World vs Object World)
-    pm.addPass(py::createNativeVerificationPass());
-    // Insert reference counting operations using Affine SSA (Linear Type) logic
-    pm.addPass(py::createRefCountInsertionPass());
-    // Early canonicalization and CSE for arith/func ops
-    pm.addPass(mlir::createCanonicalizerPass());
-    pm.addPass(mlir::createCSEPass());
-    pm.addPass(py::createRuntimeLoweringPass());
-    pm.addPass(mlir::createConvertSCFToCFPass());
-    pm.addPass(mlir::createArithToLLVMConversionPass());
-    pm.addPass(mlir::createConvertControlFlowToLLVMPass());
-    pm.addPass(mlir::createConvertToLLVMPass());
-    pm.addPass(mlir::createReconcileUnrealizedCastsPass());
-    pm.addNestedPass<mlir::func::FuncOp>(mlir::createReconcileUnrealizedCastsPass());
-    pm.addPass(mlir::createCanonicalizerPass());
-    return pm.run(module);
+    bool dumpIR = static_cast<bool>(
+        llvm::sys::Process::GetEnv("LYTHON_DUMP_LOWERING_IR"));
+
+    if (dumpIR) {
+      llvm::errs() << "=== [Frontend Output (before any passes)] ===\n";
+      module.dump();
+    }
+
+    // Phase 1: Native verification
+    {
+      PassManager pm(&context);
+      pm.addPass(py::createNativeVerificationPass());
+      if (failed(pm.run(module)))
+        return failure();
+    }
+
+    if (dumpIR) {
+      llvm::errs() << "\n=== [After NativeVerificationPass] ===\n";
+      module.dump();
+    }
+
+    // Phase 2: Reference counting insertion
+    {
+      PassManager pm(&context);
+      pm.addPass(py::createRefCountInsertionPass());
+      if (failed(pm.run(module)))
+        return failure();
+    }
+
+    if (dumpIR) {
+      llvm::errs() << "\n=== [After RefCountInsertionPass] ===\n";
+      module.dump();
+    }
+
+    // Phase 3: Early canonicalization and CSE
+    {
+      PassManager pm(&context);
+      pm.addPass(mlir::createCanonicalizerPass());
+      pm.addPass(mlir::createCSEPass());
+      if (failed(pm.run(module)))
+        return failure();
+    }
+
+    if (dumpIR) {
+      llvm::errs() << "\n=== [After Canonicalizer + CSE] ===\n";
+      module.dump();
+    }
+
+    // Phase 4: Runtime lowering (Py dialect -> func/LLVM)
+    {
+      PassManager pm(&context);
+      pm.addPass(py::createRuntimeLoweringPass());
+      if (failed(pm.run(module)))
+        return failure();
+    }
+
+    // Phase 5: Final lowering to LLVM
+    {
+      PassManager pm(&context);
+      pm.addPass(mlir::createConvertSCFToCFPass());
+      pm.addPass(mlir::createArithToLLVMConversionPass());
+      pm.addPass(mlir::createConvertControlFlowToLLVMPass());
+      pm.addPass(mlir::createConvertToLLVMPass());
+      pm.addPass(mlir::createReconcileUnrealizedCastsPass());
+      pm.addNestedPass<mlir::func::FuncOp>(mlir::createReconcileUnrealizedCastsPass());
+      pm.addPass(mlir::createCanonicalizerPass());
+      if (failed(pm.run(module)))
+        return failure();
+    }
+
+    if (dumpIR) {
+      llvm::errs() << "\n=== [Final LLVM IR] ===\n";
+      module.dump();
+    }
+
+    return success();
   }
 
   void registerRuntimeSymbols(ExecutionEngine& engine) {