import builtins
import copy
import functools
import inspect
import math
import os
import warnings
import collections
from itertools import chain
from types import CodeType, FunctionType, ModuleType
from typing import (
    Any,
    Callable,
    Dict,
    List,
    NamedTuple,
    Optional,
    Set,
    Tuple,
    Type,
    Union,
)

import torch
import torch.utils._pytree as pytree
from torch._C import ScriptObject  # type: ignore[attr-defined]

from ._compatibility import compatibility
from .graph import _PyTreeCodeGen, _PyTreeInfo, Graph
from .graph_module import GraphModule
from .node import Argument, base_types, map_aggregate
from .proxy import ParameterProxy, Proxy, TracerBase, Scope, ScopeContextManager

HAS_VARSTUFF = inspect.CO_VARARGS | inspect.CO_VARKEYWORDS

# These need to run in global scope to handle nested calls correctly
_orig_module_call: Callable = torch.nn.Module.__call__
_orig_module_getattr: Callable = torch.nn.Module.__getattr__

_proxyable_classes: Dict[Type, None] = {}

_is_fx_tracing_flag = False


def is_fx_tracing():
    return _is_fx_tracing_flag

@compatibility(is_backward_compatible=True)
class ProxyableClassMeta(type):
    """
    ProxyableClassMeta allows you to make construction of a given Python class
    symbolically traceable. For example::

        import torch
        import torch.fx

        class TensorPair(metaclass=torch.fx.ProxyableClassMeta):
            def __init__(self, left, right):
                self.left, self.right = left, right

            def add(self, other):
                l = self.left + other.left
                r = self.right + other.right
                return TensorPair(l, r)

            def mul(self, other):
                l = self.left * other.left
                r = self.right * other.right
                return TensorPair(l, r)

        def use_tensor_pair_ctor(x : TensorPair, y : torch.Tensor):
            s = x.add(TensorPair(y, y))
            return s.mul(x)

        x = TensorPair(torch.randn(5, 3), torch.randn(5, 3))
        y = torch.randn(5, 3)
        ref_out = use_tensor_pair_ctor(x, y)

        traced = torch.fx.symbolic_trace(use_tensor_pair_ctor)
        print(traced.code)
        '''
        def forward(self, x : __main___TensorPair, y : torch.Tensor):
            tensor_pair = __main___TensorPair(y, y);  y = None
            add = x.add(tensor_pair);  tensor_pair = None
            mul = add.mul(x);  add = x = None
            return mul
        '''

    From this example, we can see that construction of a class (``TensorPair``)
    defined with ``ProxyableClassMeta`` as metaclass can be recorded in symbolic
    tracing.
    """

    def __init__(cls, name, bases, attrs):
        _proxyable_classes.setdefault(cls)
        super().__init__(name, bases, attrs)

    def __call__(cls, *args, **kwargs):
        instance = cls.__new__(cls)  # type: ignore[call-overload]

        found_proxies = []

        def check_proxy(a):
            if isinstance(a, Proxy):
                found_proxies.append(a)

        map_aggregate(args, check_proxy)
        map_aggregate(kwargs, check_proxy)

        if len(found_proxies) != 0:
            tracer = found_proxies[0].tracer
            return tracer.create_proxy("call_function", cls, args, kwargs)
        else:
            cls.__init__(instance, *args, **kwargs)  # type: ignore[misc]
            return instance


def _patch_function(fn: FunctionType, nargs: int) -> FunctionType:
    co = fn.__code__
    co_flags = co.co_flags & ~HAS_VARSTUFF
    co_args: tuple
    if hasattr(co, "co_qualname"):
        # Python-3.11+ code signature
        co_args = (
            nargs,
            0,
            0,
            co.co_nlocals,
            co.co_stacksize,
            co_flags,
            co.co_code,
            co.co_consts,
            co.co_names,
            co.co_varnames,
            co.co_filename,
            co.co_name,
            co.co_qualname,  # type: ignore[attr-defined]
            co.co_firstlineno,
            co.co_lnotab,
            co.co_exceptiontable,  # type: ignore[attr-defined]
            co.co_freevars,
            co.co_cellvars,
        )
    elif hasattr(co, "co_posonlyargcount"):
        co_args = (
            nargs,
            0,
            0,
            co.co_nlocals,
            co.co_stacksize,
            co_flags,
            co.co_code,
            co.co_consts,
            co.co_names,
            co.co_varnames,
            co.co_filename,
            co.co_name,
            co.co_firstlineno,
            co.co_lnotab,
            co.co_freevars,
            co.co_cellvars,
        )
    else:
        co_args = (
            nargs,
            0,
            co.co_nlocals,
            co.co_stacksize,
            co_flags,
            co.co_code,
            co.co_consts,
            co.co_names,
            co.co_varnames,
            co.co_filename,
            co.co_name,
            co.co_firstlineno,
            co.co_lnotab,
            co.co_freevars,
            co.co_cellvars,
        )
    new_code = CodeType(*co_args)  # type: ignore[arg-type]
    return FunctionType(
        new_code, fn.__globals__, fn.__name__, fn.__defaults__, fn.__closure__
    )

    # we need to insert placeholder nodes for *args and **kwargs
    # we can't call this function normally, otherwise it would try to unpack them
    # instead, let's make python think that args and kwargs are normal variables


@compatibility(is_backward_compatible=False)
class PHBase:
    """
    Object representing an input placeholder to `concrete_args`
    """

    def __repr__(self):
        return "PH"


PH = PHBase()


@compatibility(is_backward_compatible=True)
class Tracer(TracerBase):
    # Reference: https://github.com/pytorch/pytorch/issues/54354
    # The first line of this docstring overrides the one Sphinx generates for the
    # documentation. We need it so that Sphinx doesn't leak `math`s path from the
    # build environment (e.g. `<module 'math' from '/leaked/path').

    """Tracer(autowrap_modules=(math,), autowrap_functions=())

    ``Tracer`` is the class that implements the symbolic tracing functionality
    of ``torch.fx.symbolic_trace``. A call to ``symbolic_trace(m)`` is equivalent
    to ``Tracer().trace(m)``.

    Tracer can be subclassed to override various behaviors of the tracing
    process. The different behaviors that can be overridden are described
    in the docstrings of the methods on this class.
    """

    # Not checking BC on this API because the default value for `autowrap_modules`
    # includes the local filepath to the `math` module, which would jitter
    # across machines.
    @compatibility(is_backward_compatible=True)
    def __init__(
        self,
        autowrap_modules: Tuple[ModuleType] = (math,),
        autowrap_functions: Tuple[Callable, ...] = (),
        param_shapes_constant: bool = False,
    ) -> None:
        # This method's signature is overridden by the first line of this class'
        # docstring. If this method's signature is modified, the signature that
        # overrides it also should be modified accordingly.

        """
        Construct a Tracer object.

        Args:

            autowrap_modules (Tuple[ModuleType]): defaults to `(math, )`,
                Python modules whose functions should be wrapped automatically
                without needing to use fx.wrap(). Backward-compatibility for
                this parameter is guaranteed.

            autowrap_functions (Tuple[Callable, ...]): defaults to `()`,
                Python functions that should be wrapped automatically without
                needing to use fx.wrap(). Backward compatibility for this
                parameter is guaranteed.

            param_shapes_constant (bool): When this flag is set,  calls to shape,
                size and a few other shape like attributes of a module's parameter
                will be evaluated directly, rather than returning a new Proxy value
                for an attribute access. Backward compatibility for this parameter
                is guaranteed.
        """

        super().__init__()

        # Functions we will eagerly wrap when we see them while tracing
        # this captures both `math.sqrt()` and `from math import sqrt` automatically
        self._autowrap_function_ids: Set[int] = {
            id(value)
            for name, value in chain(*[m.__dict__.items() for m in autowrap_modules])
            if not name.startswith("_") and callable(value)
        }
        self._autowrap_function_ids.update({id(f) for f in autowrap_functions})

        # Python modules to apply autowrap to at the start, in addition to
        # modules we see while tracing
        self._autowrap_search: List[ModuleType] = list(autowrap_modules)
        self.param_shapes_constant = param_shapes_constant

        self.submodule_paths: Optional[Dict[torch.nn.Module, str]] = None
        self.root_module_name: str = ""
        # Maps the containing module's name to the operator name
        self.scope = Scope("", None)
        # Records the module call stack
        self.module_stack = collections.OrderedDict()
        # Mapping of node name to module scope
        self.node_name_to_scope: Dict[str, Tuple[str, type]] = {}

    @compatibility(is_backward_compatible=True)
    def create_arg(self, a: Any) -> "Argument":
        """
        A method to specify the behavior of tracing when preparing values to
        be used as arguments to nodes in the ``Graph``.

        By default, the behavior includes:

        #. Iterate through collection types (e.g. tuple, list, dict) and recursively
           call ``create_args`` on the elements.
        #. Given a Proxy object, return a reference to the underlying IR ``Node``
        #. Given a non-Proxy Tensor object, emit IR for various cases:

            * For a Parameter, emit a ``get_attr`` node referring to that Parameter
            * For a non-Parameter Tensor, store the Tensor away in a special
              attribute referring to that attribute.

        This method can be overridden to support more types.

        Args:

            a (Any): The value to be emitted as an ``Argument`` in the ``Graph``.


        Returns:

            The value ``a`` converted into the appropriate ``Argument``
        """
        # The base tracer is used to construct Graphs when there is no associated
        # module hierarchy, so it can never create parameter references.
        # The default tracer adds the ability to refer to parameters when
        # tracing modules.
        if isinstance(a, torch.nn.Parameter):
            for n, p in self.root.named_parameters():
                if a is p:
                    return self.create_node("get_attr", n, (), {})
            raise NameError("parameter is not a member of this module")
        elif isinstance(a, torch.Tensor):
            for n_, p_ in self.root.named_buffers():
                if a is p_:
                    return self.create_node("get_attr", n_, (), {})
        elif isinstance(a, torch.nn.Module):
            for n_, p_ in self.root.named_modules():
                if a is p_:
                    return self.create_node("get_attr", n_, (), {})
        # For NamedTuple instances that appear literally as args, we emit
        # a node to construct the NamedTuple and use that Node as the argument.
        if isinstance(a, tuple) and hasattr(a, "_fields"):
            args = tuple(self.create_arg(elem) for elem in a)
            return self.create_node("call_function", a.__class__, args, {})

        # Tensors do not have a reliable string repr() from which they can be
        # constructed (and we probably don't want to rely on that, either), so
        # for any constant Tensor values we encounter, first search for if they
        # are an attribute of some module in the module hierarchy. If so, emit
        # a get_attr to retrieve that tensor. Otherwise, we'll store away the
        # tensor value into a special attribute on the Module s.t. we can
        # retrieve it with a get_attr.
        if isinstance(a, (torch.Tensor, ScriptObject)):
            qualname: Optional[str] = self.tensor_attrs.get(a)

            # Tensor was not found in the Module hierarchy, stow it away in a
            # special attribute and set the qualname to refer to that