[mlir python] Port in-tree dialects to nanobind.

This is a companion to #118583, although it can be landed independently
because since #117922 dialects do not have to use the same Python
binding framework as the Python core code.

This PR ports all of the in-tree dialect and pass extensions to nanobind,
with the exception of those that remain for testing pybind11 support. It
would make sense to merge this PR after merging #118583, if we have
agreed that we are migrating the core to nanobind.

This PR also:
* removes CollectDiagnosticsToStringScope from NanobindAdaptors.h. This
  was overlooked in a previous PR and it is duplicated in Diagnostics.h.

Author: hawkinsp

mlir/cmake/modules/AddMLIRPython.cmake

@@ -142,6 +142,9 @@ function(declare_mlir_python_extension name)
     mlir_python_DEPENDS ""
     mlir_python_BINDINGS_LIBRARY "${ARG_PYTHON_BINDINGS_LIBRARY}"
   )
+  if (LLVM_COMPILER_IS_GCC_COMPATIBLE OR CLANG_CL)
+    set_target_properties(${name} PROPERTIES INTERFACE_COMPILE_OPTIONS "-Wno-cast-qual;-Wno-zero-length-array;-Wno-extra-semi;-Wno-nested-anon-types;-Wno-pedantic")
+  endif()
 
   # Set the interface source and link_libs properties of the target
   # These properties support generator expressions and are automatically exported

mlir/include/mlir/Bindings/Python/NanobindAdaptors.h

@@ -631,40 +631,6 @@ class mlir_value_subclass : public pure_subclass {
 
 } // namespace nanobind_adaptors
 
-/// RAII scope intercepting all diagnostics into a string. The message must be
-/// checked before this goes out of scope.
-class CollectDiagnosticsToStringScope {
-public:
-  explicit CollectDiagnosticsToStringScope(MlirContext ctx) : context(ctx) {
-    handlerID = mlirContextAttachDiagnosticHandler(ctx, &handler, &errorMessage,
-                                                   /*deleteUserData=*/nullptr);
-  }
-  ~CollectDiagnosticsToStringScope() {
-    assert(errorMessage.empty() && "unchecked error message");
-    mlirContextDetachDiagnosticHandler(context, handlerID);
-  }
-
-  [[nodiscard]] std::string takeMessage() { return std::move(errorMessage); }
-
-private:
-  static MlirLogicalResult handler(MlirDiagnostic diag, void *data) {
-    auto printer = +[](MlirStringRef message, void *data) {
-      *static_cast<std::string *>(data) +=
-          llvm::StringRef(message.data, message.length);
-    };
-    MlirLocation loc = mlirDiagnosticGetLocation(diag);
-    *static_cast<std::string *>(data) += "at ";
-    mlirLocationPrint(loc, printer, data);
-    *static_cast<std::string *>(data) += ": ";
-    mlirDiagnosticPrint(diag, printer, data);
-    return mlirLogicalResultSuccess();
-  }
-
-  MlirContext context;
-  MlirDiagnosticHandlerID handlerID;
-  std::string errorMessage = "";
-};
-
 } // namespace python
 } // namespace mlir
 

mlir/lib/Bindings/Python/AsyncPasses.cpp

@@ -8,14 +8,13 @@
 
 #include "mlir-c/Dialect/Async.h"
 
-#include <pybind11/detail/common.h>
-#include <pybind11/pybind11.h>
+#include <nanobind/nanobind.h>
 
 // -----------------------------------------------------------------------------
 // Module initialization.
 // -----------------------------------------------------------------------------
 
-PYBIND11_MODULE(_mlirAsyncPasses, m) {
+NB_MODULE(_mlirAsyncPasses, m) {
   m.doc() = "MLIR Async Dialect Passes";
 
   // Register all Async passes on load.

mlir/lib/Bindings/Python/DialectGPU.cpp

@@ -9,21 +9,23 @@
 #include "mlir-c/Dialect/GPU.h"
 #include "mlir-c/IR.h"
 #include "mlir-c/Support.h"
-#include "mlir/Bindings/Python/PybindAdaptors.h"
+#include "mlir/Bindings/Python/NanobindAdaptors.h"
 
-#include <pybind11/detail/common.h>
-#include <pybind11/pybind11.h>
+#include <nanobind/nanobind.h>
+#include <nanobind/stl/optional.h>
+
+namespace nb = nanobind;
+using namespace nanobind::literals;
 
-namespace py = pybind11;
 using namespace mlir;
 using namespace mlir::python;
-using namespace mlir::python::adaptors;
+using namespace mlir::python::nanobind_adaptors;
 
 // -----------------------------------------------------------------------------
 // Module initialization.
 // -----------------------------------------------------------------------------
 
-PYBIND11_MODULE(_mlirDialectsGPU, m) {
+NB_MODULE(_mlirDialectsGPU, m) {
   m.doc() = "MLIR GPU Dialect";
   //===-------------------------------------------------------------------===//
   // AsyncTokenType
@@ -34,11 +36,11 @@ PYBIND11_MODULE(_mlirDialectsGPU, m) {
 
   mlirGPUAsyncTokenType.def_classmethod(
       "get",
-      [](py::object cls, MlirContext ctx) {
+      [](nb::object cls, MlirContext ctx) {
         return cls(mlirGPUAsyncTokenTypeGet(ctx));
       },
-      "Gets an instance of AsyncTokenType in the same context", py::arg("cls"),
-      py::arg("ctx") = py::none());
+      "Gets an instance of AsyncTokenType in the same context", nb::arg("cls"),
+      nb::arg("ctx").none() = nb::none());
 
   //===-------------------------------------------------------------------===//
   // ObjectAttr
@@ -47,12 +49,12 @@ PYBIND11_MODULE(_mlirDialectsGPU, m) {
   mlir_attribute_subclass(m, "ObjectAttr", mlirAttributeIsAGPUObjectAttr)
       .def_classmethod(
           "get",
-          [](py::object cls, MlirAttribute target, uint32_t format,
-             py::bytes object, std::optional<MlirAttribute> mlirObjectProps,
+          [](nb::object cls, MlirAttribute target, uint32_t format,
+             nb::bytes object, std::optional<MlirAttribute> mlirObjectProps,
              std::optional<MlirAttribute> mlirKernelsAttr) {
-            py::buffer_info info(py::buffer(object).request());
-            MlirStringRef objectStrRef =
-                mlirStringRefCreate(static_cast<char *>(info.ptr), info.size);
+            MlirStringRef objectStrRef = mlirStringRefCreate(
+                static_cast<char *>(const_cast<void *>(object.data())),
+                object.size());
             return cls(mlirGPUObjectAttrGetWithKernels(
                 mlirAttributeGetContext(target), target, format, objectStrRef,
                 mlirObjectProps.has_value() ? *mlirObjectProps
@@ -61,7 +63,7 @@ PYBIND11_MODULE(_mlirDialectsGPU, m) {
                                             : MlirAttribute{nullptr}));
           },
           "cls"_a, "target"_a, "format"_a, "object"_a,
-          "properties"_a = py::none(), "kernels"_a = py::none(),
+          "properties"_a.none() = nb::none(), "kernels"_a.none() = nb::none(),
           "Gets a gpu.object from parameters.")
       .def_property_readonly(
           "target",
@@ -73,18 +75,18 @@ PYBIND11_MODULE(_mlirDialectsGPU, m) {
           "object",
           [](MlirAttribute self) {
             MlirStringRef stringRef = mlirGPUObjectAttrGetObject(self);
-            return py::bytes(stringRef.data, stringRef.length);
+            return nb::bytes(stringRef.data, stringRef.length);
           })
       .def_property_readonly("properties",
-                             [](MlirAttribute self) {
+                             [](MlirAttribute self) -> nb::object {
                                if (mlirGPUObjectAttrHasProperties(self))
-                                 return py::cast(
+                                 return nb::cast(
                                      mlirGPUObjectAttrGetProperties(self));
-                               return py::none().cast<py::object>();
+                               return nb::none();
                              })
-      .def_property_readonly("kernels", [](MlirAttribute self) {
+      .def_property_readonly("kernels", [](MlirAttribute self) -> nb::object {
         if (mlirGPUObjectAttrHasKernels(self))
-          return py::cast(mlirGPUObjectAttrGetKernels(self));
-        return py::none().cast<py::object>();
+          return nb::cast(mlirGPUObjectAttrGetKernels(self));
+        return nb::none();
       });
 }

mlir/lib/Bindings/Python/DialectLLVM.cpp

@@ -12,15 +12,23 @@
 #include "mlir-c/IR.h"
 #include "mlir-c/Support.h"
 #include "mlir/Bindings/Python/Diagnostics.h"
-#include "mlir/Bindings/Python/PybindAdaptors.h"
+#include "mlir/Bindings/Python/NanobindAdaptors.h"
+
+#include <nanobind/nanobind.h>
+#include <nanobind/stl/optional.h>
+#include <nanobind/stl/string.h>
+#include <nanobind/stl/vector.h>
+
+namespace nb = nanobind;
+
+using namespace nanobind::literals;
 
-namespace py = pybind11;
 using namespace llvm;
 using namespace mlir;
 using namespace mlir::python;
-using namespace mlir::python::adaptors;
+using namespace mlir::python::nanobind_adaptors;
 
-void populateDialectLLVMSubmodule(const pybind11::module &m) {
+void populateDialectLLVMSubmodule(const nanobind::module_ &m) {
 
   //===--------------------------------------------------------------------===//
   // StructType
@@ -31,58 +39,58 @@ void populateDialectLLVMSubmodule(const pybind11::module &m) {
 
   llvmStructType.def_classmethod(
       "get_literal",
-      [](py::object cls, const std::vector<MlirType> &elements, bool packed,
+      [](nb::object cls, const std::vector<MlirType> &elements, bool packed,
          MlirLocation loc) {
         CollectDiagnosticsToStringScope scope(mlirLocationGetContext(loc));
 
         MlirType type = mlirLLVMStructTypeLiteralGetChecked(
             loc, elements.size(), elements.data(), packed);
         if (mlirTypeIsNull(type)) {
-          throw py::value_error(scope.takeMessage());
+          throw nb::value_error(scope.takeMessage().c_str());
         }
         return cls(type);
       },
-      "cls"_a, "elements"_a, py::kw_only(), "packed"_a = false,
-      "loc"_a = py::none());
+      "cls"_a, "elements"_a, nb::kw_only(), "packed"_a = false,
+      "loc"_a.none() = nb::none());
 
   llvmStructType.def_classmethod(
       "get_identified",
-      [](py::object cls, const std::string &name, MlirContext context) {
+      [](nb::object cls, const std::string &name, MlirContext context) {
         return cls(mlirLLVMStructTypeIdentifiedGet(
             context, mlirStringRefCreate(name.data(), name.size())));
       },
-      "cls"_a, "name"_a, py::kw_only(), "context"_a = py::none());
+      "cls"_a, "name"_a, nb::kw_only(), "context"_a.none() = nb::none());
 
   llvmStructType.def_classmethod(
       "get_opaque",
-      [](py::object cls, const std::string &name, MlirContext context) {
+      [](nb::object cls, const std::string &name, MlirContext context) {
         return cls(mlirLLVMStructTypeOpaqueGet(
             context, mlirStringRefCreate(name.data(), name.size())));
       },
-      "cls"_a, "name"_a, "context"_a = py::none());
+      "cls"_a, "name"_a, "context"_a.none() = nb::none());
 
   llvmStructType.def(
       "set_body",
       [](MlirType self, const std::vector<MlirType> &elements, bool packed) {
         MlirLogicalResult result = mlirLLVMStructTypeSetBody(
             self, elements.size(), elements.data(), packed);
         if (!mlirLogicalResultIsSuccess(result)) {
-          throw py::value_error(
+          throw nb::value_error(
               "Struct body already set to different content.");
         }
       },
-      "elements"_a, py::kw_only(), "packed"_a = false);
+      "elements"_a, nb::kw_only(), "packed"_a = false);
 
   llvmStructType.def_classmethod(
       "new_identified",
-      [](py::object cls, const std::string &name,
+      [](nb::object cls, const std::string &name,
          const std::vector<MlirType> &elements, bool packed, MlirContext ctx) {
         return cls(mlirLLVMStructTypeIdentifiedNewGet(
             ctx, mlirStringRefCreate(name.data(), name.length()),
             elements.size(), elements.data(), packed));
       },
-      "cls"_a, "name"_a, "elements"_a, py::kw_only(), "packed"_a = false,
-      "context"_a = py::none());
+      "cls"_a, "name"_a, "elements"_a, nb::kw_only(), "packed"_a = false,
+      "context"_a.none() = nb::none());
 
   llvmStructType.def_property_readonly(
       "name", [](MlirType type) -> std::optional<std::string> {
@@ -93,12 +101,12 @@ void populateDialectLLVMSubmodule(const pybind11::module &m) {
         return StringRef(stringRef.data, stringRef.length).str();
       });
 
-  llvmStructType.def_property_readonly("body", [](MlirType type) -> py::object {
+  llvmStructType.def_property_readonly("body", [](MlirType type) -> nb::object {
     // Don't crash in absence of a body.
     if (mlirLLVMStructTypeIsOpaque(type))
-      return py::none();
+      return nb::none();
 
-    py::list body;
+    nb::list body;
     for (intptr_t i = 0, e = mlirLLVMStructTypeGetNumElementTypes(type); i < e;
          ++i) {
       body.append(mlirLLVMStructTypeGetElementType(type, i));
@@ -119,24 +127,24 @@ void populateDialectLLVMSubmodule(const pybind11::module &m) {
   mlir_type_subclass(m, "PointerType", mlirTypeIsALLVMPointerType)
       .def_classmethod(
           "get",
-          [](py::object cls, std::optional<unsigned> addressSpace,
+          [](nb::object cls, std::optional<unsigned> addressSpace,
              MlirContext context) {
             CollectDiagnosticsToStringScope scope(context);
             MlirType type = mlirLLVMPointerTypeGet(
                 context, addressSpace.has_value() ? *addressSpace : 0);
             if (mlirTypeIsNull(type)) {
-              throw py::value_error(scope.takeMessage());
+              throw nb::value_error(scope.takeMessage().c_str());
             }
             return cls(type);
           },
-          "cls"_a, "address_space"_a = py::none(), py::kw_only(),
-          "context"_a = py::none())
+          "cls"_a, "address_space"_a.none() = nb::none(), nb::kw_only(),
+          "context"_a.none() = nb::none())
       .def_property_readonly("address_space", [](MlirType type) {
         return mlirLLVMPointerTypeGetAddressSpace(type);
       });
 }
 
-PYBIND11_MODULE(_mlirDialectsLLVM, m) {
+NB_MODULE(_mlirDialectsLLVM, m) {
   m.doc() = "MLIR LLVM Dialect";
 
   populateDialectLLVMSubmodule(m);

mlir/lib/Bindings/Python/DialectLinalg.cpp

@@ -6,22 +6,24 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include <nanobind/nanobind.h>
+
 #include "mlir-c/Dialect/Linalg.h"
 #include "mlir-c/IR.h"
-#include "mlir/Bindings/Python/PybindAdaptors.h"
+#include "mlir/Bindings/Python/NanobindAdaptors.h"
 
-namespace py = pybind11;
+namespace nb = nanobind;
 
-static void populateDialectLinalgSubmodule(py::module m) {
+static void populateDialectLinalgSubmodule(nb::module_ m) {
   m.def(
       "fill_builtin_region",
       [](MlirOperation op) { mlirLinalgFillBuiltinNamedOpRegion(op); },
-      py::arg("op"),
+      nb::arg("op"),
       "Fill the region for `op`, which is assumed to be a builtin named Linalg "
       "op.");
 }
 
-PYBIND11_MODULE(_mlirDialectsLinalg, m) {
+NB_MODULE(_mlirDialectsLinalg, m) {
   m.doc() = "MLIR Linalg dialect.";
 
   populateDialectLinalgSubmodule(m);

mlir/lib/Bindings/Python/DialectNVGPU.cpp

@@ -6,35 +6,36 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include <nanobind/nanobind.h>
+
 #include "mlir-c/Dialect/NVGPU.h"
 #include "mlir-c/IR.h"
-#include "mlir/Bindings/Python/PybindAdaptors.h"
-#include <pybind11/pybind11.h>
+#include "mlir/Bindings/Python/NanobindAdaptors.h"
 
-namespace py = pybind11;
+namespace nb = nanobind;
 using namespace llvm;
 using namespace mlir;
 using namespace mlir::python;
-using namespace mlir::python::adaptors;
+using namespace mlir::python::nanobind_adaptors;
 
-static void populateDialectNVGPUSubmodule(const pybind11::module &m) {
+static void populateDialectNVGPUSubmodule(const nb::module_ &m) {
   auto nvgpuTensorMapDescriptorType = mlir_type_subclass(
       m, "TensorMapDescriptorType", mlirTypeIsANVGPUTensorMapDescriptorType);
 
   nvgpuTensorMapDescriptorType.def_classmethod(
       "get",
-      [](py::object cls, MlirType tensorMemrefType, int swizzle, int l2promo,
+      [](nb::object cls, MlirType tensorMemrefType, int swizzle, int l2promo,
          int oobFill, int interleave, MlirContext ctx) {
         return cls(mlirNVGPUTensorMapDescriptorTypeGet(
             ctx, tensorMemrefType, swizzle, l2promo, oobFill, interleave));
       },
       "Gets an instance of TensorMapDescriptorType in the same context",
-      py::arg("cls"), py::arg("tensor_type"), py::arg("swizzle"),
-      py::arg("l2promo"), py::arg("oob_fill"), py::arg("interleave"),
-      py::arg("ctx") = py::none());
+      nb::arg("cls"), nb::arg("tensor_type"), nb::arg("swizzle"),
+      nb::arg("l2promo"), nb::arg("oob_fill"), nb::arg("interleave"),
+      nb::arg("ctx").none() = nb::none());
 }
 
-PYBIND11_MODULE(_mlirDialectsNVGPU, m) {
+NB_MODULE(_mlirDialectsNVGPU, m) {
   m.doc() = "MLIR NVGPU dialect.";
 
   populateDialectNVGPUSubmodule(m);