numba-cuda 0.22.0__cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl

numba_cuda/numba/cuda/debuginfo.py

@@ -0,0 +1,997 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: BSD-2-Clause
+
+import abc
+import os
+from contextlib import contextmanager
+from enum import IntEnum
+
+import llvmlite
+from llvmlite import ir
+from numba.cuda import types
+from numba.cuda.core import config
+from numba.cuda import cgutils
+from numba.cuda.datamodel.models import ComplexModel, UnionModel, UniTupleModel
+from numba.cuda.types.ext_types import GridGroup
+from cuda.bindings import runtime
+
+
+# Check if CUDA Toolkit and llvmlite support polymorphic debug info
+def _get_llvmlite_version():
+    """Get llvmlite version as tuple (major, minor)."""
+    try:
+        version_str = llvmlite.__version__
+        # Parse version string like "0.46.0" or "0.46.0dev"
+        parts = version_str.split(".")
+        if len(parts) < 2:
+            return (0, 0)
+        major = int(parts[0])
+        minor = int(parts[1])
+        return (major, minor)
+    except (IndexError, AttributeError, ValueError):
+        return (0, 0)
+
+
+def _check_polymorphic_debug_info_support():
+    """Check if CTK and llvmlite support polymorphic debug info.
+
+    Returns:
+        tuple: (supported: bool, use_typed_const: bool)
+        - supported: Whether feature is supported at all
+        - use_typed_const: True for typed constant,
+                           False for node reference
+    """
+    # runtime.getLocalRuntimeVersion() returns (cudaError_t, version_int)
+    # Example: 13010 = CTK 13.1, 13020 = CTK 13.2
+    _, ctk_version_number = runtime.getLocalRuntimeVersion()
+    ctk_major = ctk_version_number // 1000
+    ctk_minor = (ctk_version_number % 1000) // 10
+    ctk_version = (ctk_major, ctk_minor)
+
+    llvmlite_version = _get_llvmlite_version()
+
+    # Support not available with CTK 13.1 or older
+    if ctk_version <= (13, 1):
+        return (False, False)
+
+    # llvmlite > 0.45: use typed constant
+    # llvmlite <= 0.45: use node reference
+    use_typed_const = llvmlite_version > (0, 45)
+    return (True, use_typed_const)
+
+
+# Check support and determine mode
+(DEBUG_POLY_SUPPORTED, DEBUG_POLY_USE_TYPED_CONST) = (
+    _check_polymorphic_debug_info_support()
+)
+
+# Set config based on polymorphic debug info support
+if not hasattr(config, "CUDA_DEBUG_POLY"):
+    config.CUDA_DEBUG_POLY = DEBUG_POLY_SUPPORTED
+if not hasattr(config, "CUDA_DEBUG_POLY_USE_TYPED_CONST"):
+    config.CUDA_DEBUG_POLY_USE_TYPED_CONST = DEBUG_POLY_USE_TYPED_CONST
+
+
+class DwarfAddressClass(IntEnum):
+    GENERIC = 0x00
+    GLOBAL = 0x01
+    REGISTER = 0x02
+    CONSTANT = 0x05
+    LOCAL = 0x06
+    PARAMETER = 0x07
+    SHARED = 0x08
+
+
+@contextmanager
+def suspend_emission(builder):
+    """Suspends the emission of debug_metadata for the duration of the context
+    managed block."""
+    ref = builder.debug_metadata
+    builder.debug_metadata = None
+    try:
+        yield
+    finally:
+        builder.debug_metadata = ref
+
+
+class AbstractDIBuilder(metaclass=abc.ABCMeta):
+    @abc.abstractmethod
+    def mark_variable(
+        self,
+        builder,
+        allocavalue,
+        name,
+        lltype,
+        size,
+        line,
+        datamodel=None,
+        argidx=None,
+    ):
+        """Emit debug info for the variable."""
+        pass
+
+    @abc.abstractmethod
+    def mark_location(self, builder, line):
+        """Emit source location information to the given IRBuilder."""
+        pass
+
+    @abc.abstractmethod
+    def mark_subprogram(self, function, qualname, argnames, argtypes, line):
+        """Emit source location information for the given function."""
+        pass
+
+    @abc.abstractmethod
+    def initialize(self):
+        """Initialize the debug info. An opportunity for the debuginfo to
+        prepare any necessary data structures.
+        """
+
+    @abc.abstractmethod
+    def finalize(self):
+        """Finalize the debuginfo by emitting all necessary metadata."""
+        pass
+
+
+class DummyDIBuilder(AbstractDIBuilder):
+    def __init__(self, module, filepath, cgctx, directives_only):
+        pass
+
+    def mark_variable(
+        self,
+        builder,
+        allocavalue,
+        name,
+        lltype,
+        size,
+        line,
+        datamodel=None,
+        argidx=None,
+    ):
+        pass
+
+    def mark_location(self, builder, line):
+        pass
+
+    def mark_subprogram(self, function, qualname, argnames, argtypes, line):
+        pass
+
+    def initialize(self):
+        pass
+
+    def finalize(self):
+        pass
+
+
+_BYTE_SIZE = 8
+
+
+class DIBuilder(AbstractDIBuilder):
+    DWARF_VERSION = 4
+    DEBUG_INFO_VERSION = 3
+    DBG_CU_NAME = "llvm.dbg.cu"
+    _DEBUG = False
+
+    def __init__(self, module, filepath, cgctx, directives_only):
+        self.module = module
+        self.filepath = os.path.abspath(filepath)
+        self.difile = self._di_file()
+        self.subprograms = []
+        self.cgctx = cgctx
+
+        if directives_only:
+            self.emission_kind = "DebugDirectivesOnly"
+        else:
+            self.emission_kind = "FullDebug"
+
+        self.initialize()
+
+    def initialize(self):
+        # Create the compile unit now because it is referenced when
+        # constructing subprograms
+        self.dicompileunit = self._di_compile_unit()
+
+    def get_dwarf_address_class(self, addrspace):
+        # Map NVVM address space to DWARF address class.
+        from numba.cuda.cudadrv import nvvm
+
+        addrspace_to_addrclass_dict = {
+            nvvm.ADDRSPACE_GENERIC: None,
+            nvvm.ADDRSPACE_GLOBAL: DwarfAddressClass.GLOBAL,
+            nvvm.ADDRSPACE_SHARED: DwarfAddressClass.SHARED,
+            nvvm.ADDRSPACE_CONSTANT: DwarfAddressClass.CONSTANT,
+            nvvm.ADDRSPACE_LOCAL: DwarfAddressClass.LOCAL,
+        }
+        return addrspace_to_addrclass_dict.get(addrspace)
+
+    def _var_type(self, lltype, size, datamodel=None):
+        if self._DEBUG:
+            print(
+                "-->",
+                lltype,
+                size,
+                datamodel,
+                getattr(datamodel, "fe_type", "NO FE TYPE"),
+            )
+        m = self.module
+        bitsize = _BYTE_SIZE * size
+
+        int_type = (ir.IntType,)
+        real_type = ir.FloatType, ir.DoubleType
+        # For simple numeric types, choose the closest encoding.
+        # We treat all integers as unsigned when there's no known datamodel.
+        if isinstance(lltype, int_type + real_type):
+            if datamodel is None:
+                # This is probably something like an `i8*` member of a struct
+                name = str(lltype)
+                if isinstance(lltype, int_type):
+                    ditok = "DW_ATE_unsigned"
+                else:
+                    ditok = "DW_ATE_float"
+            else:
+                # This is probably a known int/float scalar type
+                name = str(datamodel.fe_type)
+                if isinstance(datamodel.fe_type, types.Integer):
+                    if datamodel.fe_type.signed:
+                        ditok = "DW_ATE_signed"
+                    else:
+                        ditok = "DW_ATE_unsigned"
+                else:
+                    ditok = "DW_ATE_float"
+            mdtype = m.add_debug_info(
+                "DIBasicType",
+                {
+                    "name": name,
+                    "size": bitsize,
+                    "encoding": ir.DIToken(ditok),
+                },
+            )
+        elif isinstance(datamodel, ComplexModel):
+            # TODO: Is there a better way of determining "this is a complex
+            # number"?
+            #
+            # NOTE: Commented below is the way to generate the metadata for a
+            # C99 complex type that's directly supported by DWARF. Numba however
+            # generates a struct with real/imag cf. CPython to give a more
+            # pythonic feel to inspection.
+            #
+            # mdtype = m.add_debug_info('DIBasicType', {
+            #  'name': f"{datamodel.fe_type} ({str(lltype)})",
+            #  'size': bitsize,
+            # 'encoding': ir.DIToken('DW_ATE_complex_float'),
+            # })
+            meta = []
+            offset = 0
+            for ix, name in enumerate(("real", "imag")):
+                component = lltype.elements[ix]
+                component_size = self.cgctx.get_abi_sizeof(component)
+                component_basetype = m.add_debug_info(
+                    "DIBasicType",
+                    {
+                        "name": str(component),
+                        "size": _BYTE_SIZE * component_size,  # bits
+                        "encoding": ir.DIToken("DW_ATE_float"),
+                    },
+                )
+                derived_type = m.add_debug_info(
+                    "DIDerivedType",
+                    {
+                        "tag": ir.DIToken("DW_TAG_member"),
+                        "name": name,
+                        "baseType": component_basetype,
+                        "size": _BYTE_SIZE
+                        * component_size,  # DW_TAG_member size is in bits
+                        "offset": offset,
+                    },
+                )
+                meta.append(derived_type)
+                offset += _BYTE_SIZE * component_size  # offset is in bits
+            mdtype = m.add_debug_info(
+                "DICompositeType",
+                {
+                    "tag": ir.DIToken("DW_TAG_structure_type"),
+                    "name": f"{datamodel.fe_type} ({str(lltype)})",
+                    "identifier": str(lltype),
+                    "elements": m.add_metadata(meta),
+                    "size": offset,
+                },
+                is_distinct=True,
+            )
+        elif isinstance(datamodel, UniTupleModel):
+            element = lltype.element
+            el_size = self.cgctx.get_abi_sizeof(element)
+            basetype = self._var_type(element, el_size)
+            name = f"{datamodel.fe_type} ({str(lltype)})"
+            count = size // el_size
+            mdrange = m.add_debug_info(
+                "DISubrange",
+                {
+                    "count": count,
+                },
+            )
+            mdtype = m.add_debug_info(
+                "DICompositeType",
+                {
+                    "tag": ir.DIToken("DW_TAG_array_type"),
+                    "baseType": basetype,
+                    "name": name,
+                    "size": bitsize,
+                    "identifier": str(lltype),
+                    "elements": m.add_metadata([mdrange]),
+                },
+            )
+        elif isinstance(lltype, ir.PointerType):
+            model = getattr(datamodel, "_pointee_model", None)
+            basetype = self._var_type(
+                lltype.pointee, self.cgctx.get_abi_sizeof(lltype.pointee), model
+            )
+            mdtype = m.add_debug_info(
+                "DIDerivedType",
+                {
+                    "tag": ir.DIToken("DW_TAG_pointer_type"),
+                    "baseType": basetype,
+                    "size": _BYTE_SIZE * self.cgctx.get_abi_sizeof(lltype),
+                },
+            )
+        elif isinstance(lltype, ir.LiteralStructType):
+            # Struct type
+            meta = []
+            offset = 0
+            if datamodel is None or not datamodel.inner_models():
+                name = f"Anonymous struct ({str(lltype)})"
+                for field_id, element in enumerate(lltype.elements):
+                    size = self.cgctx.get_abi_sizeof(element)
+                    basetype = self._var_type(element, size)
+                    derived_type = m.add_debug_info(
+                        "DIDerivedType",
+                        {
+                            "tag": ir.DIToken("DW_TAG_member"),
+                            "name": f"<field {field_id}>",
+                            "baseType": basetype,
+                            "size": _BYTE_SIZE
+                            * size,  # DW_TAG_member size is in bits
+                            "offset": offset,
+                        },
+                    )
+                    meta.append(derived_type)
+                    offset += _BYTE_SIZE * size  # offset is in bits
+            else:
+                name = f"{datamodel.fe_type} ({str(lltype)})"
+                for element, field, model in zip(
+                    lltype.elements, datamodel._fields, datamodel.inner_models()
+                ):
+                    size = self.cgctx.get_abi_sizeof(element)
+                    basetype = self._var_type(element, size, datamodel=model)
+                    derived_type = m.add_debug_info(
+                        "DIDerivedType",
+                        {
+                            "tag": ir.DIToken("DW_TAG_member"),
+                            "name": field,
+                            "baseType": basetype,
+                            "size": _BYTE_SIZE
+                            * size,  # DW_TAG_member size is in bits
+                            "offset": offset,
+                        },
+                    )
+                    meta.append(derived_type)
+                    offset += _BYTE_SIZE * size  # offset is in bits
+
+            mdtype = m.add_debug_info(
+                "DICompositeType",
+                {
+                    "tag": ir.DIToken("DW_TAG_structure_type"),
+                    "name": name,
+                    "identifier": str(lltype),
+                    "elements": m.add_metadata(meta),
+                    "size": offset,
+                },
+                is_distinct=True,
+            )
+        elif isinstance(lltype, ir.ArrayType):
+            element = lltype.element
+            el_size = self.cgctx.get_abi_sizeof(element)
+            basetype = self._var_type(element, el_size)
+            count = size // el_size
+            mdrange = m.add_debug_info(
+                "DISubrange",
+                {
+                    "count": count,
+                },
+            )
+            mdtype = m.add_debug_info(
+                "DICompositeType",
+                {
+                    "tag": ir.DIToken("DW_TAG_array_type"),
+                    "baseType": basetype,
+                    "name": str(lltype),
+                    "size": bitsize,
+                    "identifier": str(lltype),
+                    "elements": m.add_metadata([mdrange]),
+                },
+            )
+        else:
+            # For all other types, describe it as sequence of bytes
+            count = size
+            mdrange = m.add_debug_info(
+                "DISubrange",
+                {
+                    "count": count,
+                },
+            )
+            mdbase = m.add_debug_info(
+                "DIBasicType",
+                {
+                    "name": "byte",
+                    "size": _BYTE_SIZE,
+                    "encoding": ir.DIToken("DW_ATE_unsigned_char"),
+                },
+            )
+            mdtype = m.add_debug_info(
+                "DICompositeType",
+                {
+                    "tag": ir.DIToken("DW_TAG_array_type"),
+                    "baseType": mdbase,
+                    "name": str(lltype),
+                    "size": bitsize,
+                    "identifier": str(lltype),
+                    "elements": m.add_metadata([mdrange]),
+                },
+            )
+
+        return mdtype
+
+    def mark_variable(
+        self,
+        builder,
+        allocavalue,
+        name,
+        lltype,
+        size,
+        line,
+        datamodel=None,
+        argidx=None,
+    ):
+        arg_index = 0 if argidx is None else argidx
+        m = self.module
+        fnty = ir.FunctionType(ir.VoidType(), [ir.MetaDataType()] * 3)
+        decl = cgutils.get_or_insert_function(m, fnty, "llvm.dbg.declare")
+
+        mdtype = self._var_type(lltype, size, datamodel=datamodel)
+        name = name.replace(".", "$")  # for gdb to work correctly
+        mdlocalvar = m.add_debug_info(
+            "DILocalVariable",
+            {
+                "name": name,
+                "arg": arg_index,
+                "scope": self.subprograms[-1],
+                "file": self.difile,
+                "line": line,
+                "type": mdtype,
+            },
+        )
+        mdexpr = m.add_debug_info("DIExpression", {})
+
+        return builder.call(decl, [allocavalue, mdlocalvar, mdexpr])
+
+    def mark_location(self, builder, line):
+        builder.debug_metadata = self._add_location(line)
+
+    def mark_subprogram(self, function, qualname, argnames, argtypes, line):
+        name = qualname
+        argmap = dict(zip(argnames, argtypes))
+        di_subp = self._add_subprogram(
+            name=name,
+            linkagename=function.name,
+            line=line,
+            function=function,
+            argmap=argmap,
+        )
+        function.set_metadata("dbg", di_subp)
+
+    def finalize(self):
+        dbgcu = cgutils.get_or_insert_named_metadata(
+            self.module, self.DBG_CU_NAME
+        )
+        dbgcu.add(self.dicompileunit)
+        self._set_module_flags()
+
+    #
+    # Internal APIs
+    #
+
+    def _set_module_flags(self):
+        """Set the module flags metadata"""
+        module = self.module
+        mflags = cgutils.get_or_insert_named_metadata(
+            module, "llvm.module.flags"
+        )
+        # Set *require* behavior to warning
+        # See http://llvm.org/docs/LangRef.html#module-flags-metadata
+        require_warning_behavior = self._const_int(2)
+        if self.DWARF_VERSION is not None:
+            dwarf_version = module.add_metadata(
+                [
+                    require_warning_behavior,
+                    "Dwarf Version",
+                    self._const_int(self.DWARF_VERSION),
+                ]
+            )
+            if dwarf_version not in mflags.operands:
+                mflags.add(dwarf_version)
+        debuginfo_version = module.add_metadata(
+            [
+                require_warning_behavior,
+                "Debug Info Version",
+                self._const_int(self.DEBUG_INFO_VERSION),
+            ]
+        )
+        if debuginfo_version not in mflags.operands:
+            mflags.add(debuginfo_version)
+
+    def _add_subprogram(self, name, linkagename, line, function, argmap):
+        """Emit subprogram metadata"""
+        subp = self._di_subprogram(name, linkagename, line, function, argmap)
+        self.subprograms.append(subp)
+        return subp
+
+    def _add_location(self, line):
+        """Emit location metatdaa"""
+        loc = self._di_location(line)
+        return loc
+
+    @classmethod
+    def _const_int(cls, num, bits=32):
+        """Util to create constant int in metadata"""
+        return ir.IntType(bits)(num)
+
+    @classmethod
+    def _const_bool(cls, boolean):
+        """Util to create constant boolean in metadata"""
+        return ir.IntType(1)(boolean)
+
+    #
+    # Helpers to emit the metadata nodes
+    #
+
+    def _di_file(self):
+        return self.module.add_debug_info(
+            "DIFile",
+            {
+                "directory": os.path.dirname(self.filepath),
+                "filename": os.path.basename(self.filepath),
+            },
+        )
+
+    def _di_compile_unit(self):
+        return self.module.add_debug_info(
+            "DICompileUnit",
+            {
+                "language": ir.DIToken("DW_LANG_C_plus_plus"),
+                "file": self.difile,
+                # Numba has to pretend to be clang to ensure the prologue is skipped
+                # correctly in gdb. See:
+                # https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=gdb/amd64-tdep.c;h=e563d369d8cb3eb3c2f732c2fa850ec70ba8d63b;hb=a4b0231e179607e47b1cdf1fe15c5dc25e482fad#l2521
+                # Note the "producer_is_llvm" call to specialise the prologue
+                # handling, this is defined here:
+                # https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=gdb/producer.c;h=cdfd80d904c09394febd18749bb90359b2d128cc;hb=a4b0231e179607e47b1cdf1fe15c5dc25e482fad#l124
+                # and to get a match for this condition the 'producer' must start
+                # with "clang ", hence the following...
+                "producer": "clang (Numba)",
+                "runtimeVersion": 0,
+                "isOptimized": config.OPT != 0,
+                "emissionKind": ir.DIToken(self.emission_kind),
+            },
+            is_distinct=True,
+        )
+
+    def _di_subroutine_type(self, line, function, argmap):
+        # The function call conv needs encoding.
+        llfunc = function
+        md = []
+
+        for idx, llarg in enumerate(llfunc.args):
+            if not llarg.name.startswith("arg."):
+                name = llarg.name.replace(".", "$")  # for gdb to work correctly
+                lltype = llarg.type
+                size = self.cgctx.get_abi_sizeof(lltype)
+                mdtype = self._var_type(lltype, size, datamodel=None)
+                md.append(mdtype)
+
+        for idx, (name, nbtype) in enumerate(argmap.items()):
+            name = name.replace(".", "$")  # for gdb to work correctly
+            datamodel = self.cgctx.data_model_manager[nbtype]
+            lltype = self.cgctx.get_value_type(nbtype)
+            size = self.cgctx.get_abi_sizeof(lltype)
+            mdtype = self._var_type(lltype, size, datamodel=datamodel)
+            md.append(mdtype)
+
+        return self.module.add_debug_info(
+            "DISubroutineType",
+            {
+                "types": self.module.add_metadata(md),
+            },
+        )
+
+    def _di_subprogram(self, name, linkagename, line, function, argmap):
+        return self.module.add_debug_info(
+            "DISubprogram",
+            {
+                "name": name,
+                "linkageName": linkagename,
+                "scope": self.difile,
+                "file": self.difile,
+                "line": line,
+                "type": self._di_subroutine_type(line, function, argmap),
+                "isLocal": False,
+                "isDefinition": True,
+                "scopeLine": line,
+                "isOptimized": config.OPT != 0,
+                "unit": self.dicompileunit,
+            },
+            is_distinct=True,
+        )
+
+    def _di_location(self, line):
+        return self.module.add_debug_info(
+            "DILocation",
+            {
+                "line": line,
+                "column": 1,
+                "scope": self.subprograms[-1],
+            },
+        )
+
+
+class CUDADIBuilder(DIBuilder):
+    def __init__(self, module, filepath, cgctx, directives_only):
+        super().__init__(module, filepath, cgctx, directives_only)
+        # Cache for local variable metadata type and line deduplication
+        self._vartypelinemap = {}
+        # Variable address space dictionary
+        self._var_addrspace_map = {}
+
+    def _set_addrspace_map(self, map):
+        self._var_addrspace_map = map
+
+    def _var_type(self, lltype, size, datamodel=None):
+        is_bool = False
+        is_int_literal = False
+        is_grid_group = False
+        m = self.module
+
+        if isinstance(lltype, ir.IntType):
+            if datamodel is None:
+                if size == 1:
+                    name = str(lltype)
+                    is_bool = True
+            else:
+                name = str(datamodel.fe_type)
+                if isinstance(datamodel.fe_type, types.Boolean):
+                    is_bool = True
+                    if isinstance(datamodel.fe_type, types.BooleanLiteral):
+                        name = "bool"
+                elif isinstance(datamodel.fe_type, types.Integer):
+                    if isinstance(datamodel.fe_type, types.IntegerLiteral):
+                        name = f"int{_BYTE_SIZE * size}"
+                        is_int_literal = True
+                elif isinstance(datamodel.fe_type, GridGroup):
+                    is_grid_group = True
+
+        if is_bool or is_int_literal or is_grid_group:
+            bitsize = _BYTE_SIZE * size
+            # Boolean type workaround until upstream Numba is fixed
+            if is_bool:
+                ditok = "DW_ATE_boolean"
+            elif is_int_literal:
+                ditok = "DW_ATE_signed"
+            # GridGroup type should use numba.cuda implementation
+            elif is_grid_group:
+                ditok = "DW_ATE_unsigned"
+
+            return m.add_debug_info(
+                "DIBasicType",
+                {
+                    "name": name,
+                    "size": bitsize,
+                    "encoding": ir.DIToken(ditok),
+                },
+            )
+
+        if isinstance(datamodel, UnionModel):
+            # UnionModel is handled here to represent polymorphic types
+            meta = []
+            maxwidth = 0
+            for field, model in zip(
+                datamodel._fields, datamodel.inner_models()
+            ):
+                # Ignore the "tag" field, focus on the "payload" field which
+                # contains the data types in memory
+                if field == "payload":
+                    # Store metadata dictionaries to create members later
+                    member_metadata_dicts = []
+
+                    for index, mod in enumerate(model.inner_models()):
+                        dtype = mod.get_value_type()
+                        membersize = self.cgctx.get_abi_sizeof(dtype)
+                        basetype = self._var_type(
+                            dtype, membersize, datamodel=mod
+                        )
+                        if isinstance(mod.fe_type, types.Literal):
+                            typename = str(mod.fe_type.literal_type)
+                        else:
+                            typename = str(mod.fe_type)
+                        # Use a prefix "_" on type names as field names
+                        membername = "_" + typename
+                        memberwidth = _BYTE_SIZE * membersize
+                        # Build the metadata dictionary
+                        metadata_dict = {
+                            "tag": ir.DIToken("DW_TAG_member"),
+                            "name": membername,
+                            "baseType": basetype,
+                            # DW_TAG_member size is in bits
+                            "size": memberwidth,
+                        }
+                        if config.CUDA_DEBUG_POLY:
+                            # Polymorphic debug info with DW_TAG_variant
+                            # extraData depends on llvmlite version
+                            if config.CUDA_DEBUG_POLY_USE_TYPED_CONST:
+                                metadata_dict["extraData"] = ir.IntType(8)(
+                                    index
+                                )
+                            else:
+                                # Use metadata node reference
+                                metadata_dict["extraData"] = m.add_metadata(
+                                    [ir.IntType(8)(index)]
+                                )
+                            # Add offset to each variant member
+                            # Offset equals the element's own width
+                            metadata_dict["offset"] = memberwidth
+                        member_metadata_dicts.append(metadata_dict)
+                        if memberwidth > maxwidth:
+                            maxwidth = memberwidth
+
+                    # Create the member DIDerivedTypes
+                    for metadata_dict in member_metadata_dicts:
+                        derived_type = m.add_debug_info(
+                            "DIDerivedType", metadata_dict
+                        )
+                        meta.append(derived_type)
+
+            if config.CUDA_DEBUG_POLY:
+                # Polymorphic variable debug info generation
+                wrapper_struct_size = 2 * maxwidth
+                # Generate unique discriminator name based on composite type
+                variant_elements_metadata = m.add_metadata(meta)
+                discriminator_unique_id = str(id(variant_elements_metadata))
+                discriminator_name = f"discriminator-{discriminator_unique_id}"
+                discriminator = m.add_debug_info(
+                    "DIDerivedType",
+                    {
+                        "tag": ir.DIToken("DW_TAG_member"),
+                        "name": discriminator_name,
+                        "baseType": m.add_debug_info(
+                            "DIBasicType",
+                            {
+                                "name": "int",
+                                "size": _BYTE_SIZE,
+                                "encoding": ir.DIToken("DW_ATE_unsigned"),
+                            },
+                        ),
+                        "size": _BYTE_SIZE,
+                        "flags": ir.DIToken("DIFlagArtificial"),
+                    },
+                )
+                # Create the final variant_part with actual members
+                variant_unique_identifier = discriminator_unique_id
+                variant_part_type = m.add_debug_info(
+                    "DICompositeType",
+                    {
+                        "file": self.difile,
+                        "tag": ir.DIToken("DW_TAG_variant_part"),
+                        "name": "variant_part",
+                        "identifier": variant_unique_identifier,
+                        "elements": variant_elements_metadata,
+                        "size": maxwidth,
+                        "discriminator": discriminator,
+                    },
+                )
+                # Create elements metadata for wrapper struct
+                elements_metadata = m.add_metadata(
+                    [discriminator, variant_part_type]
+                )
+                unique_identifier = str(id(elements_metadata))
+                wrapper_struct = m.add_debug_info(
+                    "DICompositeType",
+                    {
+                        "file": self.difile,
+                        "tag": ir.DIToken("DW_TAG_structure_type"),
+                        "name": "variant_wrapper_struct",
+                        "identifier": unique_identifier,
+                        "elements": elements_metadata,
+                        "size": wrapper_struct_size,
+                    },
+                )
+                return wrapper_struct
+            else:
+                fake_union_name = "dbg_poly_union"
+                return m.add_debug_info(
+                    "DICompositeType",
+                    {
+                        "file": self.difile,
+                        "tag": ir.DIToken("DW_TAG_union_type"),
+                        "name": fake_union_name,
+                        "identifier": str(lltype),
+                        "elements": m.add_metadata(meta),
+                        "size": maxwidth,
+                    },
+                    is_distinct=True,
+                )
+
+        # Check if there's actually address space info to handle
+        addrspace = getattr(self, "_addrspace", None)
+        if (
+            isinstance(lltype, ir.LiteralStructType)
+            and datamodel is not None
+            and datamodel.inner_models()
+            and addrspace not in (None, 0)
+        ):
+            # Process struct with datamodel that has address space info
+            meta = []
+            offset = 0
+            for element, field, model in zip(
+                lltype.elements, datamodel._fields, datamodel.inner_models()
+            ):
+                size_field = self.cgctx.get_abi_sizeof(element)
+                if isinstance(element, ir.PointerType) and field == "data":
+                    # Create pointer type with correct address space
+                    pointee_size = self.cgctx.get_abi_sizeof(element.pointee)
+                    pointee_model = getattr(model, "_pointee_model", None)
+                    pointee_type = self._var_type(
+                        element.pointee, pointee_size, datamodel=pointee_model
+                    )
+                    meta_ptr = {
+                        "tag": ir.DIToken("DW_TAG_pointer_type"),
+                        "baseType": pointee_type,
+                        "size": _BYTE_SIZE * size_field,
+                    }
+                    dwarf_addrclass = self.get_dwarf_address_class(addrspace)
+                    if dwarf_addrclass is not None:
+                        meta_ptr["dwarfAddressSpace"] = int(dwarf_addrclass)
+                    basetype = m.add_debug_info("DIDerivedType", meta_ptr)
+                else:
+                    basetype = self._var_type(
+                        element, size_field, datamodel=model
+                    )
+                derived_type = m.add_debug_info(
+                    "DIDerivedType",
+                    {
+                        "tag": ir.DIToken("DW_TAG_member"),
+                        "name": field,
+                        "baseType": basetype,
+                        "size": _BYTE_SIZE * size_field,
+                        "offset": offset,
+                    },
+                )
+                meta.append(derived_type)
+                offset += _BYTE_SIZE * size_field
+
+            return m.add_debug_info(
+                "DICompositeType",
+                {
+                    "tag": ir.DIToken("DW_TAG_structure_type"),
+                    "name": f"{datamodel.fe_type}",
+                    "elements": m.add_metadata(meta),
+                    "size": offset,
+                },
+                is_distinct=True,
+            )
+        # For other cases, use upstream Numba implementation
+        return super()._var_type(lltype, size, datamodel=datamodel)
+
+    def _di_subroutine_type(self, line, function, argmap):
+        # The function call conv needs encoding.
+        llfunc = function
+        md = []
+
+        # Create metadata type for return value
+        if len(llfunc.args) > 0:
+            lltype = llfunc.args[0].type
+            size = self.cgctx.get_abi_sizeof(lltype)
+            mdtype = self._var_type(lltype, size, datamodel=None)
+            md.append(mdtype)
+
+        # Create metadata type for arguments
+        for idx, (name, nbtype) in enumerate(argmap.items()):
+            datamodel = self.cgctx.data_model_manager[nbtype]
+            lltype = self.cgctx.get_value_type(nbtype)
+            size = self.cgctx.get_abi_sizeof(lltype)
+            mdtype = self._var_type(lltype, size, datamodel=datamodel)
+            md.append(mdtype)
+
+        return self.module.add_debug_info(
+            "DISubroutineType",
+            {
+                "types": self.module.add_metadata(md),
+            },
+        )
+
+    def mark_variable(
+        self,
+        builder,
+        allocavalue,
+        name,
+        lltype,
+        size,
+        line,
+        datamodel=None,
+        argidx=None,
+    ):
+        if name.startswith("$") or "." in name:
+            # Do not emit llvm.dbg.declare on user variable alias
+            return
+        else:
+            int_type = (ir.IntType,)
+            real_type = ir.FloatType, ir.DoubleType
+            if isinstance(lltype, int_type + real_type):
+                # Start with scalar variable, swtiching llvm.dbg.declare
+                # to llvm.dbg.value
+                return
+            else:
+                # Look up address space for this variable
+                self._addrspace = self._var_addrspace_map.get(name)
+                try:
+                    return super().mark_variable(
+                        builder,
+                        allocavalue,
+                        name,
+                        lltype,
+                        size,
+                        line,
+                        datamodel,
+                        argidx,
+                    )
+                finally:
+                    # Clean up address space info
+                    self._addrspace = None
+
+    def update_variable(
+        self,
+        builder,
+        value,
+        name,
+        lltype,
+        size,
+        line,
+        datamodel=None,
+        argidx=None,
+    ):
+        m = self.module
+        fnty = ir.FunctionType(ir.VoidType(), [ir.MetaDataType()] * 3)
+        decl = cgutils.get_or_insert_function(m, fnty, "llvm.dbg.value")
+
+        mdtype = self._var_type(lltype, size, datamodel)
+        index = name.find(".")
+        if index >= 0:
+            name = name[:index]
+        # Merge DILocalVariable nodes with same name and type but different
+        # lines. Use the cached [(name, type) -> line] info to deduplicate
+        # metadata. Use the lltype as part of key.
+        key = (name, lltype)
+        if key in self._vartypelinemap:
+            line = self._vartypelinemap[key]
+        else:
+            self._vartypelinemap[key] = line
+        arg_index = 0 if argidx is None else argidx
+        mdlocalvar = m.add_debug_info(
+            "DILocalVariable",
+            {
+                "name": name,
+                "arg": arg_index,
+                "scope": self.subprograms[-1],
+                "file": self.difile,
+                "line": line,
+                "type": mdtype,
+            },
+        )
+        mdexpr = m.add_debug_info("DIExpression", {})
+
+        return builder.call(decl, [value, mdlocalvar, mdexpr])