[mlir] Expose Value hierarchy to Python bindings

[ftynse]

Summary

Values are ubiquitous in the IR, in particular block argument and operation
results are Values. Define Python classes for BlockArgument, OpResult and their
common ancestor Value. Define pseudo-container classes for lists of block
arguments and operation results, and use these containers to access the
corresponding values in blocks and operations.

[stellaraccident]

Nice!

mlir/lib/Bindings/Python/IRModules.cpp

781

I'm not sure this keep_alive is buying you anything, given the ownership model in use.

mlir/lib/Bindings/Python/IRModules.h

387

Op result values (plural)

mlir/include/mlir-c/IR.h

 
 /** Returns the position of the value in the list of results of the operation
  * that produced it. */
 intptr_t mlirOpResultGetResultNumber(MlirValue value);
 
 /** Returns the type of the value. */
 MlirType mlirValueGetType(MlirValue value);
 
+/** Prints the value to the standard error stream. */
+void mlirValueDump(MlirValue value);
+
 /** Prints a value by sending chunks of the string representation and
  * forwarding `userData to `callback`. Note that the callback may be called
  * several times with consecutive chunks of the string. */
 void mlirValuePrint(MlirValue value, MlirStringCallback callback,
                     void *userData);
 
 /*============================================================================*/
 /* Type API.                                                                  */

mlir/lib/Bindings/Python/IRModules.h

 //===- IRModules.h - IR Submodules of pybind module -----------------------===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef MLIR_BINDINGS_PYTHON_IRMODULES_H
 #define MLIR_BINDINGS_PYTHON_IRMODULES_H
 
 #include <pybind11/pybind11.h>
 
 #include "mlir-c/IR.h"
 #include "llvm/ADT/DenseMap.h"
 
 namespace mlir {
 namespace python {
 
 class PyBlock;
 class PyLocation;
 class PyMlirContext;
 class PyModule;
 class PyOperation;
 class PyType;
+class PyValue;
 
 /// Template for a reference to a concrete type which captures a python
 /// reference to its underlying python object.
 template <typename T>
 class PyObjectRef {
 public:
   PyObjectRef(T *referrent, pybind11::object object)
       : referrent(referrent), object(std::move(object)) {
public:
   PyType(PyMlirContextRef contextRef, MlirType type)
       : BaseContextObject(std::move(contextRef)), type(type) {}
   bool operator==(const PyType &other);
   operator MlirType() const { return type; }
 
   MlirType type;
 };
 
+/// Wrapper around the generic MlirValue.
+/// Values are managed completely by the operation that resulted in their
+/// definition. For op result value, this is the operation that defines the
+/// value. For block argument values, this is the operation that contains the
+/// block to which the value is an argument (blocks cannot be detached in Python
+/// bindings so such operation always exists).
+class PyValue {
+public:
+  PyValue(PyOperationRef parentOperation, MlirValue value)
+      : parentOperation(parentOperation), value(value) {}
+
+  MlirValue get() { return value; }
+  PyOperationRef &getParentOperation() { return parentOperation; }
+
+  void checkValid() { return parentOperation->checkValid(); }
+
+private:
+  PyOperationRef parentOperation;
+  MlirValue value;
+};
+
 void populateIRSubmodule(pybind11::module &m);
 
 } // namespace python
 } // namespace mlir
 
 #endif // MLIR_BINDINGS_PYTHON_IRMODULES_H

mlir/lib/Bindings/Python/IRModules.cpp

 //===- IRModules.cpp - IR Submodules of pybind module ---------------------===//
 //
 // 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
 //
 //===----------------------------------------------------------------------===//
 
 #include "IRModules.h"

 
 static const char kAppendBlockDocstring[] =
     R"(Appends a new block, with argument types as positional args.
 
 Returns:
   The created block.
 )";
 
+static const char kValueDunderStrDocstring[] =
+    R"(Returns the string form of the value.
+
+If the value is a block argument, this is the assembly form of its type and the
+position in the argument list. If the value is an operation result, this is
+equivalent to printing the operation that produced it.
+)";
+
 //------------------------------------------------------------------------------
 // Conversion utilities.
 //------------------------------------------------------------------------------
 
 namespace {
 
 /// Accumulates into a python string from a method that accepts an
 /// MlirStringCallback.

 // PyType.
 //------------------------------------------------------------------------------
 
 bool PyType::operator==(const PyType &other) {
   return mlirTypeEqual(type, other.type);
 }
 
 //------------------------------------------------------------------------------
+// PyValue and subclases.
+//------------------------------------------------------------------------------
+
+namespace {
+/// CRTP base class for Python MLIR values that subclass Value and should be
+/// castable from it. The value hierarchy is one level deep and is not supposed
+/// to accommodate other levels unless core MLIR changes.
+template <typename DerivedTy> class PyConcreteValue : public PyValue {
+public:
+  // Derived classes must define statics for:
+  //   IsAFunctionTy isaFunction
+  //   const char *pyClassName
+  // and redefine bindDerived.
+  using ClassTy = py::class_<DerivedTy, PyValue>;
+  using IsAFunctionTy = int (*)(MlirValue);
+
+  PyConcreteValue() = default;
+  PyConcreteValue(PyOperationRef operationRef, MlirValue value)
+      : PyValue(operationRef, value) {}
+  PyConcreteValue(PyValue &orig)
+      : PyConcreteValue(orig.getParentOperation(), castFrom(orig)) {}
+
+  /// Attempts to cast the original value to the derived type and throws on
+  /// type mismatches.
+  static MlirValue castFrom(PyValue &orig) {
+    if (!DerivedTy::isaFunction(orig.get())) {
+      auto origRepr = py::repr(py::cast(orig)).cast<std::string>();
+      throw SetPyError(PyExc_ValueError, llvm::Twine("Cannot cast value to ") +
+                                             DerivedTy::pyClassName +
+                                             " (from " + origRepr + ")");
+    }
+    return orig.get();
+  }
+
+  /// Binds the Python module objects to functions of this class.
+  static void bind(py::module &m) {
+    auto cls = ClassTy(m, DerivedTy::pyClassName);
+    cls.def(py::init<PyValue &>(), py::keep_alive<0, 1>());
+    DerivedTy::bindDerived(cls);
+  }
+
+  /// Implemented by derived classes to add methods to the Python subclass.
+  static void bindDerived(ClassTy &m) {}
+};
+
+/// Python wrapper for MlirBlockArgument.
+class PyBlockArgument : public PyConcreteValue<PyBlockArgument> {
+public:
+  static constexpr IsAFunctionTy isaFunction = mlirValueIsABlockArgument;
+  static constexpr const char *pyClassName = "BlockArgument";
+  using PyConcreteValue::PyConcreteValue;
+
+  static void bindDerived(ClassTy &c) {
+    c.def_property_readonly("owner", [](PyBlockArgument &self) {
+      return PyBlock(self.getParentOperation(),
+                     mlirBlockArgumentGetOwner(self.get()));
+    });
+    c.def_property_readonly("arg_number", [](PyBlockArgument &self) {
+      return mlirBlockArgumentGetArgNumber(self.get());
+    });
+    c.def("set_type", [](PyBlockArgument &self, PyType type) {
+      return mlirBlockArgumentSetType(self.get(), type);
+    });
+  }
+};
+
+/// Python wrapper for MlirOpResult.
+class PyOpResult : public PyConcreteValue<PyOpResult> {
+public:
+  static constexpr IsAFunctionTy isaFunction = mlirValueIsAOpResult;
+  static constexpr const char *pyClassName = "OpResult";
+  using PyConcreteValue::PyConcreteValue;
+
+  static void bindDerived(ClassTy &c) {
+    c.def_property_readonly("owner", [](PyOpResult &self) {
+      assert(
+          mlirOperationEqual(self.getParentOperation()->get(),
+                             mlirOpResultGetOwner(self.get())) &&
+          "expected the owner of the value in Python to match that in the IR");
+      return self.getParentOperation();
+    });
+    c.def_property_readonly("result_number", [](PyOpResult &self) {
+      return mlirOpResultGetResultNumber(self.get());
+    });
+  }
+};
+
+/// A list of block arguments. Internally, these are stored as consecutive
+/// elements, random access is cheap. The argument list is associated with the
+/// operation that contains the block (detached blocks are not allowed in
+/// Python bindings) and extends its lifetime.
+class PyBlockArgumentList {
+public:
+  PyBlockArgumentList(PyOperationRef operation, MlirBlock block)
+      : operation(std::move(operation)), block(block) {}
+
+  /// Returns the length of the block argument list.
+  intptr_t dunderLen() {
+    operation->checkValid();
+    return mlirBlockGetNumArguments(block);
+  }
+
+  /// Returns `index`-th element of the block argument list.
+  PyBlockArgument dunderGetItem(intptr_t index) {
+    if (index < 0 || index >= dunderLen()) {
+      throw SetPyError(PyExc_IndexError,
+                       "attempt to access out of bounds region");
+    }
+    PyValue value(operation, mlirBlockGetArgument(block, index));
+    return PyBlockArgument(value);
+  }
+
+  /// Defines a Python class in the bindings.
+  static void bind(py::module &m) {
+    py::class_<PyBlockArgumentList>(m, "BlockArgumentList")
+        .def("__len__", &PyBlockArgumentList::dunderLen)
+        .def("__getitem__", &PyBlockArgumentList::dunderGetItem);
+  }
+
+private:
+  PyOperationRef operation;
+  MlirBlock block;
+};
+
+/// A list of operation results. Internally, these are stored as consecutive
+/// elements, random access is cheap. The result list is associated with the
+/// operation whose results these are, and extends the lifetime of this
+/// operation.
+class PyOpResultList {
+public:
+  PyOpResultList(PyOperationRef operation) : operation(operation) {}
+
+  /// Returns the length of the result list.
+  intptr_t dunderLen() {
+    operation->checkValid();
+    return mlirOperationGetNumResults(operation->get());
+  }
+
+  /// Returns `index`-th element in the result list.
+  PyOpResult dunderGetItem(intptr_t index) {
+    if (index < 0 || index >= dunderLen()) {
+      throw SetPyError(PyExc_IndexError,
+                       "attempt to access out of bounds region");
+    }
+    PyValue value(operation, mlirOperationGetResult(operation->get(), index));
+    return PyOpResult(value);
+  }
+
+  /// Defines a Python class in the bindings.
+  static void bind(py::module &m) {
+    py::class_<PyOpResultList>(m, "OpResultList")
+        .def("__len__", &PyOpResultList::dunderLen)
+        .def("__getitem__", &PyOpResultList::dunderGetItem);
+  }
+
+private:
+  PyOperationRef operation;
+};
+
+} // end namespace
+
+//------------------------------------------------------------------------------
 // Standard attribute subclasses.
 //------------------------------------------------------------------------------
 
 namespace {
 
 /// CRTP base classes for Python attributes that subclass Attribute and should
 /// be castable from it (i.e. via something like StringAttr(attr)).
 /// By default, attribute class hierarchies are one level deep (i.e. a
void mlir::python::populateIRSubmodule(py::module &m) {
   py::class_<PyOperation>(m, "Operation")
       .def_property_readonly(
           "context",
           [](PyOperation &self) { return self.getContext().getObject(); },
           "Context that owns the Operation")
       .def_property_readonly(
           "regions",
           [](PyOperation &self) { return PyRegionList(self.getRef()); })
+      .def_property_readonly(
+          "results",
+          [](PyOperation &self) { return PyOpResultList(self.getRef()); },
+          "Returns the list of Operation results.")
       .def("__iter__",
            [](PyOperation &self) { return PyRegionIterator(self.getRef()); })
       .def(
           "__str__",
           [](PyOperation &self) {
             self.checkValid();
             PyPrintAccumulator printAccum;
             mlirOperationPrint(self.get(), printAccum.getCallback(),
py::class_<PyRegion>(m, "Region")
         } catch (std::exception &e) {
           return false;
         }
       });
 
   // Mapping of PyBlock.
   py::class_<PyBlock>(m, "Block")
       .def_property_readonly(
+          "arguments",
+          [](PyBlock &self) {
+            return PyBlockArgumentList(self.getParentOperation(), self.get());
+          },
+          "Returns a list of block arguments.")
+      .def_property_readonly(
           "operations",
           [](PyBlock &self) {
             return PyOperationList(self.getParentOperation(), self.get());
           },
           "Returns a forward-optimized sequence of operations.")
       .def(
           "__iter__",
           [](PyBlock &self) {
void mlir::python::populateIRSubmodule(py::module &m) {
   PyVectorType::bind(m);
   PyRankedTensorType::bind(m);
   PyUnrankedTensorType::bind(m);
   PyMemRefType::bind(m);
   PyUnrankedMemRefType::bind(m);
   PyTupleType::bind(m);
   PyFunctionType::bind(m);
 
+  // Mapping of Value.
+  py::class_<PyValue>(m, "Value")
+      .def_property_readonly(
+          "context",
+          [](PyValue &self) { return self.getParentOperation()->getContext(); },
+          "Context in which the value lives.")
+      .def(
+          "dump", [](PyValue &self) { mlirValueDump(self.get()); },
+          kDumpDocstring)
+      .def(
+          "__str__",
+          [](PyValue &self) {
+            PyPrintAccumulator printAccum;
+            printAccum.parts.append("Value(");
+            mlirValuePrint(self.get(), printAccum.getCallback(),
+                           printAccum.getUserData());
+            printAccum.parts.append(")");
+            return printAccum.join();
+          },
+          kValueDunderStrDocstring)
+      .def_property_readonly("type", [](PyValue &self) {
+        return PyType(self.getParentOperation()->getContext(),
+                      mlirValueGetType(self.get()));
+      });
+  PyBlockArgument::bind(m);
+  PyOpResult::bind(m);
+
   // Container bindings.
+  PyBlockArgumentList::bind(m);
   PyBlockIterator::bind(m);
   PyBlockList::bind(m);
   PyOperationIterator::bind(m);
   PyOperationList::bind(m);
+  PyOpResultList::bind(m);
   PyRegionIterator::bind(m);
   PyRegionList::bind(m);
 }

mlir/lib/CAPI/IR/IR.cpp

intptr_t mlirOpResultGetResultNumber(MlirValue value) {
   return static_cast<intptr_t>(
       unwrap(value).cast<OpResult>().getResultNumber());
 }
 
 MlirType mlirValueGetType(MlirValue value) {
   return wrap(unwrap(value).getType());
 }
 
+void mlirValueDump(MlirValue value) { unwrap(value).dump(); }
+
 void mlirValuePrint(MlirValue value, MlirStringCallback callback,
                     void *userData) {
   detail::CallbackOstream stream(callback, userData);
   unwrap(value).print(stream);
   stream.flush();
 }
 
 /* ========================================================================== */

mlir/test/Bindings/Python/ir_operation.py

def testTraverseOpRegionBlockIndices():
   # CHECK:           OP 0: %0 = "custom.addi"
   # CHECK:           OP 1: return
   # CHECK:    OP 1: "module_terminator"
   walk_operations("", module.operation)
 
 run(testTraverseOpRegionBlockIndices)
 
 
+# CHECK-LABEL: TEST: testBlockArgumentList
+def testBlockArgumentList():
+  ctx = mlir.ir.Context()
+  module = ctx.parse_module(r"""
+    func @f1(%arg0: i32, %arg1: f64, %arg2: index) {
+      return
+    }
+  """)
+  func = module.operation.regions[0].blocks[0].operations[0]
+  entry_block = func.regions[0].blocks[0]
+  assert len(entry_block.arguments) == 3
+  # CHECK: Argument 0, type i32
+  # CHECK: Argument 1, type f64
+  # CHECK: Argument 2, type index
+  for arg in entry_block.arguments:
+    print(f"Argument {arg.arg_number}, type {arg.type}")
+    new_type = mlir.ir.IntegerType.get_signless(ctx, 8 * (arg.arg_number + 1))
+    arg.set_type(new_type)
+
+  # CHECK: Argument 0, type i8
+  # CHECK: Argument 1, type i16
+  # CHECK: Argument 2, type i24
+  for arg in entry_block.arguments:
+    print(f"Argument {arg.arg_number}, type {arg.type}")
+
+
+run(testBlockArgumentList)
+
+
 # CHECK-LABEL: TEST: testDetachedOperation
 def testDetachedOperation():
   ctx = mlir.ir.Context()
   ctx.allow_unregistered_dialects = True
   loc = ctx.get_unknown_location()
   i32 = mlir.ir.IntegerType.get_signed(ctx, 32)
   op1 = ctx.create_operation(
       "custom.op1", loc, results=[i32, i32], regions=1, attributes={
def testOperationWithRegion():
   entry_block.operations.insert(0, op1)
   # CHECK: func @f1
   # CHECK: "custom.op1"()
   # CHECK:   "custom.terminator"
   # CHECK: %0 = "custom.addi"
   print(module)
 
 run(testOperationWithRegion)
+
+
+# CHECK-LABEL: TEST: testOperationResultList
+def testOperationResultList():
+  ctx = mlir.ir.Context()
+  module = ctx.parse_module(r"""
+    func @f1() {
+      %0:3 = call @f2() : () -> (i32, f64, index)
+      return
+    }
+    func @f2() -> (i32, f64, index)
+  """)
+  caller = module.operation.regions[0].blocks[0].operations[0]
+  call = caller.regions[0].blocks[0].operations[0]
+  assert len(call.results) == 3
+  # CHECK: Result 0, type i32
+  # CHECK: Result 1, type f64
+  # CHECK: Result 2, type index
+  for res in call.results:
+    print(f"Result {res.result_number}, type {res.type}")
+
+
+run(testOperationResultList)