[Mlir-commits] [mlir] [MLIR][Python] Add a DSL for defining IRDL dialects in Python bindings (PR #169045)
Rolf Morel
llvmlistbot at llvm.org
Fri Dec 12 08:46:05 PST 2025
================
@@ -0,0 +1,384 @@
+# Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+from typing import Dict, List, Union, Callable, Tuple
+from dataclasses import dataclass
+from inspect import Parameter, Signature
+from types import SimpleNamespace
+from abc import ABC, abstractmethod
+from contextlib import nullcontext
+from ...dialects import irdl
+from .._ods_common import _cext, segmented_accessor
+from . import Variadicity
+
+ir = _cext.ir
+
+__all__ = [
+ "Variadicity",
+ "Is",
+ "AnyOf",
+ "AllOf",
+ "Any",
+ "BaseName",
+ "BaseRef",
+ "Operand",
+ "Result",
+ "Attribute",
+ "Dialect",
+]
+
+
+class ConstraintExpr(ABC):
+ @abstractmethod
+ def _lower(self, ctx: "ConstraintLoweringContext") -> ir.Value:
+ pass
+
+ def __or__(self, other: "ConstraintExpr") -> "ConstraintExpr":
+ return AnyOf(self, other)
+
+ def __and__(self, other: "ConstraintExpr") -> "ConstraintExpr":
+ return AllOf(self, other)
+
+
+class ConstraintLoweringContext:
+ def __init__(self):
+ # Cache so that the same ConstraintExpr instance reuses its SSA value.
+ self._cache: Dict[int, ir.Value] = {}
+
+ def lower(self, expr: ConstraintExpr) -> ir.Value:
+ key = id(expr)
+ if key in self._cache:
+ return self._cache[key]
+ v = expr._lower(self)
+ self._cache[key] = v
+ return v
+
+
+class Is(ConstraintExpr):
+ def __init__(self, val: Union[ir.Attribute, ir.Type]):
+ self.val = val
+
+ def _lower(self, ctx: ConstraintLoweringContext) -> ir.Value:
+ return irdl.is_(
+ ir.TypeAttr.get(self.val) if isinstance(self.val, ir.Type) else self.val
+ )
+
+
+class AnyOf(ConstraintExpr):
+ def __init__(self, *exprs: ConstraintExpr):
+ self.exprs = exprs
+
+ def _lower(self, ctx: ConstraintLoweringContext) -> ir.Value:
+ return irdl.any_of(ctx.lower(expr) for expr in self.exprs)
+
+
+class AllOf(ConstraintExpr):
+ def __init__(self, *exprs: ConstraintExpr):
+ self.exprs = exprs
+
+ def _lower(self, ctx: ConstraintLoweringContext) -> ir.Value:
+ return irdl.all_of(ctx.lower(expr) for expr in self.exprs)
+
+
+class Any(ConstraintExpr):
+ def _lower(self, ctx: ConstraintLoweringContext) -> ir.Value:
+ return irdl.any()
+
+
+class BaseName(ConstraintExpr):
+ def __init__(self, name: str):
+ self.name = name
+
+ def _lower(self, ctx: ConstraintLoweringContext) -> ir.Value:
+ return irdl.base(base_name=self.name)
+
+
+class BaseRef(ConstraintExpr):
+ def __init__(self, ref):
+ self.ref = ref
+
+ def _lower(self, ctx: ConstraintLoweringContext) -> ir.Value:
+ return irdl.base(base_ref=self.ref)
+
+
+class FieldDef:
+ pass
+
+
+ at dataclass
+class Operand(FieldDef):
+ constraint: ConstraintExpr
+ variadicity: Variadicity = Variadicity.single
+
+
+ at dataclass
+class Result(FieldDef):
+ constraint: ConstraintExpr
+ variadicity: Variadicity = Variadicity.single
+
+
+ at dataclass
+class Attribute(FieldDef):
+ constraint: ConstraintExpr
+
+ def __post_init__(self):
+ # just for unified processing,
+ # currently optional attribute is not supported by IRDL
+ self.variadicity = Variadicity.single
+
+
+def partition_fields(
+ fields: List[FieldDef],
+) -> Tuple[List[Operand], List[Attribute], List[Result]]:
+ operands = [i for i in fields if isinstance(i, Operand)]
+ attrs = [i for i in fields if isinstance(i, Attribute)]
+ results = [i for i in fields if isinstance(i, Result)]
+ return operands, attrs, results
+
+
+def normalize_value_range(
+ value_range: Union[ir.OpOperandList, ir.OpResultList],
+ variadicity: Variadicity,
+):
+ if variadicity == Variadicity.single:
+ return value_range[0]
+ if variadicity == Variadicity.optional:
+ return value_range[0] if len(value_range) > 0 else None
+ return value_range
+
+
+class Operation(ir.OpView):
+ @classmethod
+ def __init_subclass__(cls, *, name: str = None, **kwargs):
+ super().__init_subclass__(**kwargs)
+
+ # for subclasses without "name" parameter,
+ # just treat them as normal classes
+ if not name:
+ return
+
+ op_name = name
+ cls._op_name = op_name
+ dialect_name = cls._dialect_name
+ dialect_obj = cls._dialect_obj
+
+ fields = []
+ cls._fields = fields
+
+ for key, value in cls.__dict__.items():
+ if isinstance(value, FieldDef):
+ setattr(value, "name", key)
+ fields.append(value)
+
+ cls._generate_class_attributes(dialect_name, op_name, fields)
+ cls._generate_init_method(fields)
+ operands, attrs, results = partition_fields(fields)
+ cls._generate_attr_properties(attrs)
+ cls._generate_operand_properties(operands)
+ cls._generate_result_properties(results)
+
+ dialect_obj.operations.append(cls)
+
+ @staticmethod
+ def _variadicity_to_segment(variadicity: Variadicity) -> int:
+ if variadicity == Variadicity.variadic:
+ return -1
+ if variadicity == Variadicity.optional:
+ return 0
+ return 1
+
+ @staticmethod
+ def _generate_segments(
+ operands_or_results: List[Union[Operand, Result]],
+ ) -> List[int]:
+ if any(i.variadicity != Variadicity.single for i in operands_or_results):
+ return [
+ Operation._variadicity_to_segment(i.variadicity)
+ for i in operands_or_results
+ ]
+ return None
+
+ @staticmethod
+ def _generate_init_signature(fields: List[FieldDef]) -> Signature:
+ # results are placed at the beginning of the parameter list,
+ # but operands and attributes can appear in any relative order.
+ args = [i for i in fields if isinstance(i, Result)] + [
+ i for i in fields if not isinstance(i, Result)
+ ]
+ positional_args = [
+ i.name for i in args if i.variadicity != Variadicity.optional
+ ]
+ optional_args = [i.name for i in args if i.variadicity == Variadicity.optional]
+
+ params = [Parameter("self", Parameter.POSITIONAL_ONLY)]
+ for i in positional_args:
+ params.append(Parameter(i, Parameter.POSITIONAL_OR_KEYWORD))
+ for i in optional_args:
+ params.append(Parameter(i, Parameter.KEYWORD_ONLY, default=None))
+ params.append(Parameter("loc", Parameter.KEYWORD_ONLY, default=None))
+ params.append(Parameter("ip", Parameter.KEYWORD_ONLY, default=None))
+
+ return Signature(params)
+
+ @classmethod
+ def _generate_init_method(cls, fields: List[FieldDef]) -> None:
+ init_sig = cls._generate_init_signature(fields)
+ operands, attrs, results = partition_fields(fields)
+
+ def __init__(*args, **kwargs):
+ bound = init_sig.bind(*args, **kwargs)
+ bound.apply_defaults()
+ args = bound.arguments
+
+ _operands = [args[operand.name] for operand in operands]
+ _results = [args[result.name] for result in results]
+ _attributes = dict(
+ (attr.name, args[attr.name])
+ for attr in attrs
+ if args[attr.name] is not None
+ )
+ _regions = None
+ _ods_successors = None
+ self = args["self"]
+ super(Operation, self).__init__(
+ self.OPERATION_NAME,
+ self._ODS_REGIONS,
+ self._ODS_OPERAND_SEGMENTS,
+ self._ODS_RESULT_SEGMENTS,
+ attributes=_attributes,
+ results=_results,
+ operands=_operands,
+ successors=_ods_successors,
+ regions=_regions,
+ loc=args["loc"],
+ ip=args["ip"],
+ )
+
+ __init__.__signature__ = init_sig
+ cls.__init__ = __init__
+
+ @classmethod
+ def _generate_class_attributes(
+ cls, dialect_name: str, op_name: str, fields: List[FieldDef]
+ ) -> None:
+ operands, attrs, results = partition_fields(fields)
+
+ operand_segments = cls._generate_segments(operands)
+ result_segments = cls._generate_segments(results)
+
+ cls.OPERATION_NAME = f"{dialect_name}.{op_name}"
+ cls._ODS_REGIONS = (0, True)
+ cls._ODS_OPERAND_SEGMENTS = operand_segments
+ cls._ODS_RESULT_SEGMENTS = result_segments
+
+ @classmethod
+ def _generate_attr_properties(cls, attrs: List[Attribute]) -> None:
+ for attr in attrs:
+ setattr(
+ cls,
+ attr.name,
+ property(lambda self, name=attr.name: self.attributes[name]),
+ )
+
+ @classmethod
+ def _generate_operand_properties(cls, operands: List[Operand]) -> None:
+ for i, operand in enumerate(operands):
+ if cls._ODS_OPERAND_SEGMENTS:
+
+ def getter(self, i=i, operand=operand):
+ operand_range = segmented_accessor(
+ self.operation.operands,
+ self.operation.attributes["operandSegmentSizes"],
+ i,
+ )
+ return normalize_value_range(operand_range, operand.variadicity)
+
+ setattr(cls, operand.name, property(getter))
+ else:
+ setattr(cls, operand.name, property(lambda self, i=i: self.operands[i]))
+
+ @classmethod
+ def _generate_result_properties(cls, results: List[Result]) -> None:
+ for i, result in enumerate(results):
+ if cls._ODS_RESULT_SEGMENTS:
+
+ def getter(self, i=i, result=result):
+ result_range = segmented_accessor(
+ self.operation.results,
+ self.operation.attributes["resultSegmentSizes"],
+ i,
+ )
+ return normalize_value_range(result_range, result.variadicity)
+
+ setattr(cls, result.name, property(getter))
+ else:
+ setattr(cls, result.name, property(lambda self, i=i: self.results[i]))
+
+ @classmethod
+ def _emit_operation(cls) -> None:
+ ctx = ConstraintLoweringContext()
+ operands, attrs, results = partition_fields(cls._fields)
+
+ op = irdl.operation_(cls._op_name)
+ with ir.InsertionPoint(op.body):
+ if operands:
+ irdl.operands_(
+ [ctx.lower(i.constraint) for i in operands],
+ [i.name for i in operands],
+ [i.variadicity for i in operands],
+ )
+ if attrs:
+ irdl.attributes_(
+ [ctx.lower(i.constraint) for i in attrs],
+ [i.name for i in attrs],
+ )
+ if results:
+ irdl.results_(
+ [ctx.lower(i.constraint) for i in results],
+ [i.name for i in results],
+ [i.variadicity for i in results],
+ )
+
+
+class Dialect:
+ def __init__(self, name: str):
+ self.name = name
+ self.operations = []
+ self.Operation = type(
+ "Operation",
+ (Operation,),
+ {"_dialect_obj": self, "_dialect_name": name},
+ )
+
+ def _emit_dialect(self) -> None:
+ d = irdl.dialect(self.name)
+ with ir.InsertionPoint(d.body):
+ for op in self.operations:
+ op._emit_operation()
+
+ def _emit_module(self) -> ir.Module:
+ m = ir.Module.create()
+ with ir.InsertionPoint(m.body):
+ self._emit_dialect()
+
+ return m
+
+ def _make_dialect_class(self) -> type:
+ return type("Dialect", (ir.Dialect,), {"DIALECT_NAMESPACE": self.name})
----------------
rolfmorel wrote:
Could this class, i.e. the one `_make_dialect_class` belongs to, itself extend `ir.Dialect`?
One advantage is that this gives a path to users implementing methods on dialects. E.g. like in C++/tablegen, we could support
```python
class MyDialect(irdsl.Dialect):
def verifyOperationAttribute(self, op, attr):
...
my_dialect = MyDialect()
class MyOp(my_dialect.Operation):
...
```
Maybe we could adopt the Op definition syntax here as well:
```python
class RevolutionaryDialect(irdsl.Dialect, name="revolutionary"):
def verifyOperationAttribute(self, op, attr):
...
```
https://github.com/llvm/llvm-project/pull/169045
More information about the Mlir-commits
mailing list