angr 9.2.103__py3-none-macosx_11_0_arm64.whl

angr/engines/vex/claripy/irop.py

@@ -0,0 +1,1279 @@
+"""
+This module contains symbolic implementations of VEX operations.
+"""
+
+from functools import partial
+import collections
+import itertools
+import operator
+import math
+import re
+
+import logging
+
+l = logging.getLogger(name=__name__)
+
+import pyvex
+import claripy
+
+#
+# The more sane approach
+#
+
+OP_ATTRS_PATTERN = re.compile(
+    r"^Iop_"
+    r"(?P<generic_name>\D+?)??"
+    r"(?P<from_type>[IFDV])??"
+    r"(?P<from_signed>[US])??"
+    r"(?P<from_size>\d+)??"
+    r"(?P<from_signed_back>[US])??"
+    # this screws up CmpLE: r'(?P<e_flag>E)??'
+    r"("
+    r"(?P<from_side>HL|HI|L|LO|lo)??"
+    r"(?P<conversion>to|as)"
+    r"(?P<to_type>Int|I|F|D|V)??"
+    r"(?P<to_size>\d+)??"
+    r"(?P<to_signed>[US])??"
+    r")??"
+    # special logic for SetV128lo32/64
+    r"("
+    r"(?P<set_side>lo)"
+    r"(?P<set_size>\d+)"
+    r")??"
+    r"(?P<vector_info>\d+U?S?F?0?x\d+)??"
+    r"(?P<rounding_mode>_R([ZPNM]))?$"
+)
+
+
+def op_attrs(p):
+    m = OP_ATTRS_PATTERN.match(p)
+
+    if not m:
+        return None
+    else:
+        attrs = m.groupdict()
+
+        attrs["from_signed"] = attrs["from_signed_back"] if attrs["from_signed"] is None else attrs["from_signed"]
+        attrs.pop("from_signed_back", None)
+        if attrs["generic_name"] == "CmpOR":
+            assert attrs["from_type"] == "D"
+            attrs["generic_name"] = "CmpORD"
+            attrs["from_type"] = None
+
+        # fix up vector stuff
+        vector_info = attrs.pop("vector_info", None)
+        if vector_info:
+            vm = re.match(
+                r"^(?P<vector_size>\d+)?"
+                r"(?P<vector_signed>[US])?"
+                r"(?P<vector_type>[FD])?"
+                r"(?P<vector_zero>0)?"
+                r"x"
+                r"(?P<vector_count>\d+)?$",
+                vector_info,
+            )
+            attrs.update(vm.groupdict())
+
+        return attrs
+
+
+all_operations = list(pyvex.irop_enums_to_ints.keys())
+operations = {}
+classified = set()
+unclassified = set()
+unsupported = set()
+explicit_attrs = {
+    "Iop_64x4toV256": {
+        "generic_name": "64x4",
+        "to_size": 256,
+    },
+    "Iop_Yl2xF64": {
+        "generic_name": "Yl2x",
+        "to_size": 64,
+    },
+    "Iop_Yl2xp1F64": {
+        "generic_name": "Yl2xp1",
+        "to_size": 64,
+    },
+    "Iop_V256to64_0": {
+        "generic_name": "unpack",
+        "to_size": 64,
+    },
+    "Iop_V256to64_1": {
+        "generic_name": "unpack",
+        "to_size": 64,
+    },
+    "Iop_V256to64_2": {
+        "generic_name": "unpack",
+        "to_size": 64,
+    },
+    "Iop_V256to64_3": {
+        "generic_name": "unpack",
+        "to_size": 64,
+    },
+    "Iop_V256toV128_0": {
+        "generic_name": "unpack",
+        "to_size": 128,
+    },
+    "Iop_V256toV128_1": {
+        "generic_name": "unpack",
+        "to_size": 128,
+    },
+}
+
+for _vec_lanewidth in (8, 16, 32, 64):
+    for _vec_width in (64, 128):
+        _vec_count = _vec_width // _vec_lanewidth
+        if _vec_count <= 1:
+            continue
+
+        # the regex thinks the I is an integral descriptor
+        explicit_attrs["Iop_InterleaveHI%dx%d" % (_vec_lanewidth, _vec_count)] = {
+            "generic_name": "InterleaveHI",
+            "to_size": _vec_width,
+            "vector_size": _vec_lanewidth,
+            "vector_count": _vec_count,
+        }
+
+
+def make_operations():
+    for p in all_operations:
+        if p in ("Iop_INVALID", "Iop_LAST"):
+            continue
+
+        try:
+            attrs = explicit_attrs[p]
+        except KeyError:
+            attrs = op_attrs(p)
+
+        if attrs is None:
+            unclassified.add(p)
+        else:
+            classified.add(p)
+            try:
+                operations[p] = SimIROp(p, **attrs)
+            except SimOperationError:
+                unsupported.add(p)
+
+    l.debug(
+        "%d matched (%d supported) and %d unmatched operations", len(classified), len(operations), len(unclassified)
+    )
+
+
+arithmetic_operation_map = {
+    "Add": "__add__",
+    "Sub": "__sub__",
+    "Mul": "__mul__",
+    "Div": "__floordiv__",
+    "Neg": "Neg",
+    "Abs": "Abs",
+    "Mod": "__mod__",
+}
+shift_operation_map = {
+    "Shl": "__lshift__",
+    "Shr": "LShR",
+    "Sar": "__rshift__",
+}
+bitwise_operation_map = {
+    "Xor": "__xor__",
+    "Or": "__or__",
+    "And": "__and__",
+    "Not": "__invert__",
+}
+
+operation_map = {}
+operation_map.update(arithmetic_operation_map)
+operation_map.update(shift_operation_map)
+operation_map.update(bitwise_operation_map)
+
+rm_map = {
+    0: claripy.fp.RM.RM_NearestTiesEven,
+    1: claripy.fp.RM.RM_TowardsNegativeInf,
+    2: claripy.fp.RM.RM_TowardsPositiveInf,
+    3: claripy.fp.RM.RM_TowardsZero,
+}
+
+generic_names = set()
+conversions = collections.defaultdict(list)
+unsupported_conversions = []
+add_operations = []
+other_operations = []
+vector_operations = []
+fp_ops = set()
+common_unsupported_generics = collections.Counter()
+
+
+def supports_vector(f):
+    f.supports_vector = True
+    return f
+
+
+class SimIROp:
+    """
+    A symbolic version of a Vex IR operation.
+    """
+
+    __slots__ = (
+        "name",
+        "op_attrs",
+        "_generic_name",
+        "_from_size",
+        "_from_side",
+        "_from_type",
+        "_from_signed",
+        "_to_size",
+        "_to_type",
+        "_to_signed",
+        "_set_side",
+        "_set_size",
+        "_conversion",
+        "_vector_size",
+        "_vector_signed",
+        "_vector_type",
+        "_vector_zero",
+        "_vector_count",
+        "_rounding_mode",
+        "_output_type",
+        "_output_size_bits",
+        "_float",
+        "_calculate",
+    )
+
+    def __init__(self, name, **attrs):
+        self.name = name
+        self.op_attrs = attrs
+
+        self._generic_name: str = None
+        self._from_size = None
+        self._from_side = None
+        self._from_type = None
+        self._from_signed = None
+        self._to_size = None
+        self._to_type = None
+        self._to_signed = None
+        self._set_side = None
+        self._set_size = None
+        self._conversion = None
+        self._vector_size = None
+        self._vector_signed = None
+        self._vector_type = None
+        self._vector_zero = None
+        self._vector_count = None
+
+        self._rounding_mode = None
+
+        for k, v in self.op_attrs.items():
+            if v is not None and ("size" in k or "count" in k):
+                v = int(v)
+            setattr(self, "_%s" % k, v)
+
+        # determine the output size
+        # pylint:disable=no-member
+        self._output_type = pyvex.get_op_retty(name)
+        # pylint:enable=no-member
+        self._output_size_bits = pyvex.const.get_type_size(self._output_type)
+
+        size_check = (
+            self._to_size is None
+            or (self._to_size * 2 if self._generic_name == "DivMod" else self._to_size) == self._output_size_bits
+        )
+        if not size_check:
+            raise SimOperationError("VEX output size doesn't match detected output size")
+
+        #
+        # Some categorization
+        #
+
+        generic_names.add(self._generic_name)
+        if self._conversion is not None:
+            conversions[(self._from_type, self._from_signed, self._to_type, self._to_signed)].append(self)
+
+        if len({self._vector_type, self._from_type, self._to_type} & {"F", "D"}) != 0:
+            self._float = True
+
+            if len({self._vector_type, self._from_type, self._to_type} & {"D"}) != 0:
+                # fp_ops.add(self.name)
+                raise UnsupportedIROpError("BCD ops aren't supported")
+        else:
+            self._float = False
+
+        #
+        # Now determine the operation
+        #
+
+        self._calculate = None
+
+        # is it explicitly implemented?
+        if hasattr(self, "_op_" + name):
+            self._calculate = getattr(self, "_op_" + name)
+        # if the generic name is None and there's a conversion present, this is a standard
+        # widening or narrowing or sign-extension
+        elif self._generic_name is None and self._conversion:
+            # convert int to float
+            if self._float and self._from_type == "I":
+                self._calculate = self._op_int_to_fp
+
+            # convert float to differently-sized float
+            elif self._from_type == "F" and self._to_type == "F":
+                self._calculate = self._op_fp_to_fp
+
+            elif self._from_type == "F" and self._to_type == "I":
+                self._calculate = self._op_fp_to_int
+
+            # this concatenates the args into the high and low halves of the result
+            elif self._from_side == "HL":
+                self._calculate = self._op_concat
+
+            # this just returns the high half of the first arg
+            elif self._from_size > self._to_size and self._from_side == "HI":
+                self._calculate = self._op_hi_half
+
+            # this just returns the high half of the first arg
+            elif self._from_size > self._to_size and self._from_side in ("L", "LO"):
+                self._calculate = self._op_lo_half
+
+            elif self._from_size > self._to_size and self._from_side is None:
+                self._calculate = self._op_extract
+
+            elif self._from_size < self._to_size and self.is_signed:
+                self._calculate = self._op_sign_extend
+
+            elif self._from_size < self._to_size and not self.is_signed:
+                self._calculate = self._op_zero_extend
+
+            else:
+                l.error("%s is an unexpected conversion operation configuration", self)
+                assert False
+
+        elif self._float and self._vector_zero:
+            # /* --- lowest-lane-only scalar FP --- */
+            f = getattr(claripy, "fp" + self._generic_name, None)
+            if f is not None:
+                f = partial(f, claripy.fp.RM.default())  # always? really?
+
+            f = f if f is not None else getattr(self, "_op_fgeneric_" + self._generic_name, None)
+            if f is None:
+                raise SimOperationError(f"no fp implementation found for operation {self._generic_name}")
+
+            self._calculate = partial(self._auto_vectorize, f)
+
+        # other conversions
+        elif self._conversion and self._generic_name not in {"Round", "Reinterp", "QNarrowBin"}:
+            if self._generic_name == "DivMod":
+                self._calculate = self._op_divmod
+            else:
+                unsupported_conversions.append(self.name)
+                common_unsupported_generics[self._generic_name] += 1
+
+        # generic bitwise
+        elif self._generic_name in bitwise_operation_map:
+            assert self._from_side is None
+            self._calculate = self._op_mapped
+
+        # generic mapping operations
+        elif self._generic_name in arithmetic_operation_map or self._generic_name in shift_operation_map:
+            assert self._from_side is None
+
+            if self._float and self._vector_count is None:
+                self._calculate = self._op_float_mapped
+            elif not self._float and self._vector_count is not None:
+                self._calculate = self._op_vector_mapped
+            elif self._float and self._vector_count is not None:
+                self._calculate = self._op_vector_float_mapped
+            else:
+                self._calculate = self._op_mapped
+
+        # TODO: clean up this mess
+        # specifically-implemented generics
+        elif self._float and hasattr(self, "_op_fgeneric_%s" % self._generic_name):
+            calculate = getattr(self, "_op_fgeneric_%s" % self._generic_name)
+            if self._vector_size is not None and not hasattr(calculate, "supports_vector"):
+                # NOTE: originally this branch just marked the op as unsupported but I think we can do better
+                # "marking unsupported" seems to include adding the op to the vector_operations list? why
+                self._calculate = partial(self._auto_vectorize, calculate)
+            else:
+                self._calculate = calculate
+
+        elif not self._float and hasattr(self, "_op_generic_%s" % self._generic_name):
+            calculate = getattr(self, "_op_generic_%s" % self._generic_name)
+            if self._vector_size is not None and not hasattr(calculate, "supports_vector"):
+                # NOTE: same as above
+                self._calculate = partial(self._auto_vectorize, calculate)
+            else:
+                self._calculate = calculate
+
+        else:
+            common_unsupported_generics[self._generic_name] += 1
+            other_operations.append(name)
+
+        # if we're here and calculate is None, we don't support this
+        if self._calculate is None:
+            raise UnsupportedIROpError("no calculate function identified for %s" % self.name)
+
+    def __repr__(self):
+        return "<SimIROp %s>" % self.name
+
+    def _dbg_print_attrs(self):
+        print("Operation: %s" % self.name)
+        for k, v in self.op_attrs.items():
+            if v is not None and v != "":
+                print(f"... {k}: {v}")
+
+    def calculate(self, *args):
+        if not all(isinstance(a, claripy.ast.Base) for a in args):
+            raise SimOperationError("IROp needs all args as claripy expressions")
+
+        if not self._float:
+            args = tuple(arg.raw_to_bv() for arg in args)
+
+        try:
+            if self._vector_size is None:
+                return self.extend_size(self._calculate(args))
+            else:
+                return self._calculate(args)
+        except (ZeroDivisionError, claripy.ClaripyZeroDivisionError) as e:
+            raise SimZeroDivisionException("divide by zero!") from e
+        except (TypeError, ValueError, SimValueError, claripy.ClaripyError) as e:
+            raise SimOperationError("%s._calculate() raised exception" % self.name) from e
+
+    def extend_size(self, o):
+        cur_size = o.size()
+        target_size = self._output_size_bits
+        if self._vector_count is not None:
+            # phrased this awkward way to account for vectorized widening multiply
+            target_size //= self._vector_count
+        if cur_size == target_size:
+            return o
+        if cur_size < target_size:
+            ext_size = target_size - cur_size
+            if (
+                self._to_signed == "S"
+                or (self._to_signed is None and self._from_signed == "S")
+                or (self._to_signed is None and self._vector_signed == "S")
+            ):
+                return claripy.SignExt(ext_size, o)
+            else:
+                return claripy.ZeroExt(ext_size, o)
+
+        # if cur_size > target_size:
+        # it should never happen!
+        raise SimOperationError("output of %s is too big" % self.name)
+
+    @property
+    def is_signed(self):
+        return self._from_signed == "S" or self._vector_signed == "S"
+
+    #
+    # The actual operation handlers go here.
+    #
+
+    # pylint:disable=no-self-use,unused-argument
+    def _op_mapped(self, args):
+        if self._from_size is not None:
+            sized_args = []
+            for a in args:
+                s = a.size()
+                if s == self._from_size:
+                    sized_args.append(a)
+                elif s < self._from_size:
+                    if self.is_signed:
+                        sized_args.append(claripy.SignExt(self._from_size - s, a))
+                    else:
+                        sized_args.append(claripy.ZeroExt(self._from_size - s, a))
+                elif s > self._from_size:
+                    raise SimOperationError("operation %s received too large an argument" % self.name)
+        else:
+            sized_args = args
+
+        if self._generic_name in operation_map:  # bitwise/arithmetic/shift operations
+            o = operation_map[self._generic_name]
+        else:
+            raise SimOperationError("op_mapped called with invalid mapping, for %s" % self.name)
+
+        if o == "__floordiv__" and self.is_signed:
+            # yikes!!!!!!!
+            return claripy.SDiv(*sized_args)
+        if o == "Abs":
+            return claripy.If(sized_args[0].SLT(0), -sized_args[0], sized_args[0])
+
+        return getattr(claripy.ast.BV, o)(*sized_args)
+
+    def _translate_rm(self, rm_num):
+        if not rm_num.symbolic:
+            return rm_map[rm_num.concrete_value]
+        else:
+            l.warning("symbolic rounding mode found, using default")
+            return claripy.fp.RM.default()
+
+    NO_RM = {"Neg", "Abs"}
+
+    def _op_float_mapped(self, args):
+        op = getattr(claripy, "fp" + self._generic_name)
+
+        if self._generic_name in self.NO_RM:
+            return op(*args)
+
+        rm = self._translate_rm(args[0])
+        return op(rm, *args[1:])
+
+    def _op_vector_mapped(self, args):
+        chopped_args = (
+            [claripy.Extract((i + 1) * self._vector_size - 1, i * self._vector_size, a) for a in args]
+            for i in reversed(range(self._vector_count))
+        )
+        return claripy.Concat(*(self._op_mapped(ca) for ca in chopped_args))
+
+    def _op_vector_float_mapped(self, args):
+        no_rm_arg = self._generic_name in self.NO_RM
+        rm_part = [] if no_rm_arg else [args[0]]
+        # wtf is up with these guys
+        if not no_rm_arg and self.name in {"Iop_Add32Fx2", "Iop_Sub32Fx2", "Iop_Mul32Fx2", "Iop_PwAdd32Fx2"}:
+            no_rm_arg = True
+            rm_part = [claripy.BVV(0, 8)]
+
+        chopped_args = (
+            [
+                claripy.Extract((i + 1) * self._vector_size - 1, i * self._vector_size, a).raw_to_fp()
+                for a in (args if no_rm_arg else args[1:])
+            ]
+            for i in reversed(range(self._vector_count))
+        )
+        return claripy.Concat(*(self._op_float_mapped(rm_part + ca).raw_to_bv() for ca in chopped_args))
+
+    @supports_vector
+    def _op_generic_Dup(self, args):
+        """
+        Vector duplication
+
+        Iop_Dup8x8
+        Iop_Dup8x16
+        Iop_Dup16x4
+        Iop_Dup16x8
+        Iop_Dup32x2
+        Iop_Dup32x4
+        """
+        arg_num = len(args)
+        if arg_num != 1:
+            raise SimOperationError("expect exactly one vector to be duplicated, got %d" % arg_num)
+        # Duplicate the vector for this many times
+        vector_count = self._vector_count
+        # Keep a copy of the vector to be duplicated
+        elem = args[0]
+        # Do the duplication
+        expr = elem
+        for _ in range(1, vector_count):
+            expr = claripy.Concat(elem, expr)
+        return expr
+
+    def _op_concat(self, args):
+        return claripy.Concat(*args)
+
+    def _op_hi_half(self, args):
+        return claripy.Extract(args[0].size() - 1, args[0].size() // 2, args[0])
+
+    def _op_lo_half(self, args):
+        return claripy.Extract(args[0].size() // 2 - 1, 0, args[0])
+
+    def _op_extract(self, args):
+        return claripy.Extract(self._to_size - 1, 0, args[0])
+
+    def _op_sign_extend(self, args):
+        return claripy.SignExt(self._to_size - args[0].size(), args[0])
+
+    def _op_zero_extend(self, args):
+        return claripy.ZeroExt(self._to_size - args[0].size(), args[0])
+
+    def vector_args(self, args):
+        """
+        Yields each of the individual lane pairs from the arguments, in
+        order from most significan to least significant
+        """
+        for i in reversed(range(self._vector_count)):
+            pieces = []
+            for vec in args:
+                piece = vec[(i + 1) * self._vector_size - 1 : i * self._vector_size]
+                if self._float:
+                    piece = piece.raw_to_fp()
+                pieces.append(piece)
+            yield pieces
+
+    @supports_vector
+    def _op_generic_GetElem(self, args):
+        """
+        Transfers one byte/half-word/word of a vector to a general-purpose register.
+
+        NOTE: the index should starts from the least significant bits.
+        For example, index 0 for Iop_GetElem32x2 returns the low half of a vector
+
+        Iop_GetElem8x8
+        Iop_GetElem16x4
+        Iop_GetElem32x2
+        Iop_GetElem8x16
+        Iop_GetElem16x8
+        Iop_GetElem32x4
+        Iop_GetElem64x2
+        """
+        # Size of the element
+        vector_size = self._vector_size
+        # Vector count
+        vector_count = self._vector_count
+        # Extension register value, element index
+        dReg, index = args
+        # Chopped elements; there should be `vector_count` elements in total
+        elements = dReg.chop(vector_size)
+
+        # Handle the index as symbolic
+        expr = elements[vector_count - 1]
+        for i in range(vector_count - 2, -1, -1):
+            # Iterate through the element from the second from LSB to the first from the MSB
+            expr = claripy.If(index == vector_count - i - 1, elements[i], expr)
+            # Example output: <BV32 if index == 0x1 then d0[63:32] else d0[31:0]>
+        return expr
+
+    @supports_vector
+    def _op_generic_SetElem(self, args):
+        """
+        Transfers one byte/half-word/word to a vector from a general-purpose register.
+
+        NOTE: the index should starts from the least significant bits.
+        For example, index 0 for Iop_SetElem32x2 sets the low half of a vector
+
+        Iop_SetElem8x8
+        Iop_SetElem16x4
+        Iop_SetElem32x2
+        Iop_SetElem8x16
+        Iop_SetElem16x8
+        Iop_SetElem32x4
+        Iop_SetElem64x2
+        """
+        # Size of the element
+        vector_size = self._vector_size
+        # Element count
+        vector_count = self._vector_count
+        # Extension register value, element index, element to set
+        dReg, index, element = args
+        # Chopped elements; there should be `vector_count` elements in total
+        elements = dReg.chop(vector_size)
+
+        # Generate new elements
+        new_elements = map(
+            lambda i: claripy.If(vector_count - 1 - i == index, element, elements[i]), range(vector_count)
+        )
+        return claripy.Concat(*new_elements)
+
+    def _op_generic_Mull(self, args):
+        op1, op2 = args
+        op1 = self.extend_size(op1)
+        op2 = self.extend_size(op2)
+        return op1 * op2
+
+    def _op_generic_Clz(self, args):
+        """Count the leading zeroes"""
+        piece_size = len(args[0])
+        wtf_expr = claripy.BVV(piece_size, piece_size)
+        for a in range(piece_size):
+            bit = claripy.Extract(a, a, args[0])
+            wtf_expr = claripy.If(bit == 1, claripy.BVV(piece_size - a - 1, piece_size), wtf_expr)
+        return wtf_expr
+
+    def _op_generic_Ctz(self, args):
+        """Count the trailing zeroes"""
+        piece_size = len(args[0])
+        wtf_expr = claripy.BVV(piece_size, piece_size)
+        for a in reversed(range(piece_size)):
+            bit = claripy.Extract(a, a, args[0])
+            wtf_expr = claripy.If(bit == 1, claripy.BVV(a, piece_size), wtf_expr)
+        return wtf_expr
+
+    def generic_minmax(self, args, cmp_op):
+        res_comps = []
+        for i in reversed(range(self._vector_count)):
+            a_comp = claripy.Extract((i + 1) * self._vector_size - 1, i * self._vector_size, args[0])
+            b_comp = claripy.Extract((i + 1) * self._vector_size - 1, i * self._vector_size, args[1])
+            res_comps.append(claripy.If(cmp_op(a_comp, b_comp), a_comp, b_comp))
+        return claripy.Concat(*res_comps)
+
+    @supports_vector
+    def _op_generic_Min(self, args):
+        return self.generic_minmax(args, claripy.SLT if self.is_signed else claripy.ULT)
+
+    @supports_vector
+    def _op_generic_Max(self, args):
+        return self.generic_minmax(args, claripy.SGT if self.is_signed else claripy.UGT)
+
+    @supports_vector
+    def _op_generic_GetMSBs(self, args):
+        size = self._vector_count * self._vector_size
+        bits = [claripy.Extract(i, i, args[0]) for i in range(size - 1, 6, -8)]
+        return claripy.Concat(*bits)
+
+    @supports_vector
+    def _op_generic_InterleaveLO(self, args):
+        s = self._vector_size
+        c = self._vector_count
+        left_vector = [args[0][(i + 1) * s - 1 : i * s] for i in range(c // 2)]
+        right_vector = [args[1][(i + 1) * s - 1 : i * s] for i in range(c // 2)]
+        return claripy.Concat(*itertools.chain.from_iterable(zip(reversed(left_vector), reversed(right_vector))))
+
+    @supports_vector
+    def _op_generic_InterleaveHI(self, args):
+        s = self._vector_size
+        c = self._vector_count
+        left_vector = [args[0][(i + 1) * s - 1 : i * s] for i in range(c // 2, c)]
+        right_vector = [args[1][(i + 1) * s - 1 : i * s] for i in range(c // 2, c)]
+        return claripy.Concat(*itertools.chain.from_iterable(zip(reversed(left_vector), reversed(right_vector))))
+
+    def generic_compare(self, args, comparison):
+        if self._vector_size is not None:
+            res_comps = []
+            for i in reversed(range(self._vector_count)):
+                a_comp = claripy.Extract((i + 1) * self._vector_size - 1, i * self._vector_size, args[0])
+                b_comp = claripy.Extract((i + 1) * self._vector_size - 1, i * self._vector_size, args[1])
+                res_comps.append(
+                    claripy.If(
+                        comparison(a_comp, b_comp),
+                        claripy.BVV(-1, self._vector_size),
+                        claripy.BVV(0, self._vector_size),
+                    )
+                )
+            return claripy.Concat(*res_comps)
+        else:
+            return claripy.If(comparison(args[0], args[1]), claripy.BVV(1, 1), claripy.BVV(0, 1))
+
+    @supports_vector
+    def _op_generic_CmpEQ(self, args):
+        return self.generic_compare(args, operator.eq)
+
+    _op_generic_CasCmpEQ = _op_generic_CmpEQ
+
+    def _op_generic_CmpNE(self, args):
+        return self.generic_compare(args, operator.ne)
+
+    _op_generic_ExpCmpNE = _op_generic_CmpNE
+    _op_generic_CasCmpNE = _op_generic_CmpNE
+
+    @supports_vector
+    def _op_generic_CmpNEZ(self, args):
+        assert len(args) == 1
+        args = [args[0], claripy.BVV(0, args[0].size())]
+        return self.generic_compare(args, operator.ne)  # TODO: Is this the correct action for scalars?
+
+    @supports_vector
+    def _op_generic_CmpGT(self, args):
+        return self.generic_compare(args, claripy.SGT if self.is_signed else claripy.UGT)
+
+    _op_generic_CasCmpGT = _op_generic_CmpGT
+
+    @supports_vector
+    def _op_generic_CmpGE(self, args):
+        return self.generic_compare(args, claripy.SGE if self.is_signed else claripy.UGE)
+
+    _op_generic_CasCmpGE = _op_generic_CmpGE
+
+    @supports_vector
+    def _op_generic_CmpLT(self, args):
+        return self.generic_compare(args, claripy.SLT if self.is_signed else claripy.ULT)
+
+    _op_generic_CasCmpLT = _op_generic_CmpLT
+
+    @supports_vector
+    def _op_generic_CmpLE(self, args):
+        return self.generic_compare(args, claripy.SLE if self.is_signed else claripy.ULE)
+
+    _op_generic_CasCmpLE = _op_generic_CmpLE
+
+    def _op_generic_CmpORD(self, args):
+        x = args[0]
+        y = args[1]
+        s = self._from_size
+        cond = claripy.SLT(x, y) if self.is_signed else claripy.ULT(x, y)
+        return claripy.If(x == y, claripy.BVV(0x2, s), claripy.If(cond, claripy.BVV(0x8, s), claripy.BVV(0x4, s)))
+
+    def generic_shift_thing(self, args, op):
+        if self._vector_size is not None:
+            shifted = []
+            if args[1].length != self._vector_size:
+                shift_by = args[1].zero_extend(self._vector_size - args[1].length)
+            else:
+                shift_by = args[1]
+            for i in reversed(range(self._vector_count)):
+                left = claripy.Extract((i + 1) * self._vector_size - 1, i * self._vector_size, args[0])
+                shifted.append(op(left, shift_by))
+            return claripy.Concat(*shifted)
+        else:
+            raise SimOperationError("you done fucked")
+
+    @supports_vector
+    def _op_generic_ShlN(self, args):
+        return self.generic_shift_thing(args, operator.lshift)
+
+    @supports_vector
+    def _op_generic_ShrN(self, args):
+        return self.generic_shift_thing(args, claripy.LShR)
+
+    @supports_vector
+    def _op_generic_SarN(self, args):
+        return self.generic_shift_thing(args, operator.rshift)
+
+    @supports_vector
+    def _op_generic_HAdd(self, args):
+        """
+        Halving add, for some ARM NEON instructions.
+        """
+        components = []
+        for a, b in self.vector_args(args):
+            if self.is_signed:
+                a = a.sign_extend(self._vector_size)
+                b = b.sign_extend(self._vector_size)
+            else:
+                a = a.zero_extend(self._vector_size)
+                b = b.zero_extend(self._vector_size)
+            components.append((a + b)[self._vector_size : 1])
+        return claripy.Concat(*components)
+
+    @supports_vector
+    def _op_generic_HSub(self, args):
+        """
+        Halving subtract, for some ARM NEON instructions.
+        """
+        components = []
+        for a, b in self.vector_args(args):
+            if self.is_signed:
+                a = a.sign_extend(self._vector_size)
+                b = b.sign_extend(self._vector_size)
+            else:
+                a = a.zero_extend(self._vector_size)
+                b = b.zero_extend(self._vector_size)
+            components.append((a - b)[self._vector_size : 1])
+        return claripy.Concat(*components)
+
+    @supports_vector
+    def _op_generic_QAdd(self, args):
+        """
+        Saturating add.
+        """
+        components = []
+        for a, b in self.vector_args(args):
+            top_a = a[self._vector_size - 1]
+            top_b = b[self._vector_size - 1]
+            res = a + b
+            top_r = res[self._vector_size - 1]
+            if self.is_signed:
+                big_top_r = (~top_r).zero_extend(self._vector_size - 1)
+                cap = (claripy.BVV(-1, self._vector_size) // 2) + big_top_r
+                cap_cond = ((~(top_a ^ top_b)) & (top_a ^ top_r)) == 1
+            else:
+                cap = claripy.BVV(-1, self._vector_size)
+                cap_cond = claripy.ULT(res, a)
+            components.append(claripy.If(cap_cond, cap, res))
+        return claripy.Concat(*components)
+
+    @supports_vector
+    def _op_generic_QSub(self, args):
+        """
+        Saturating subtract.
+        """
+        components = []
+        for a, b in self.vector_args(args):
+            top_a = a[self._vector_size - 1]
+            top_b = b[self._vector_size - 1]
+            res = a - b
+            top_r = res[self._vector_size - 1]
+            if self.is_signed:
+                big_top_r = (~top_r).zero_extend(self._vector_size - 1)
+                cap = (claripy.BVV(-1, self._vector_size) // 2) + big_top_r
+                cap_cond = ((top_a ^ top_b) & (top_a ^ top_r)) == 1
+            else:
+                cap = claripy.BVV(0, self._vector_size)
+                cap_cond = claripy.UGT(res, a)
+            components.append(claripy.If(cap_cond, cap, res))
+        return claripy.Concat(*components)
+
+    def _op_divmod(self, args):
+        if self.is_signed:
+            quotient = args[0].SDiv(claripy.SignExt(self._from_size - self._to_size, args[1]))
+            remainder = args[0].SMod(claripy.SignExt(self._from_size - self._to_size, args[1]))
+            quotient_size = self._to_size
+            remainder_size = self._to_size
+            return claripy.Concat(
+                claripy.Extract(remainder_size - 1, 0, remainder), claripy.Extract(quotient_size - 1, 0, quotient)
+            )
+        else:
+            quotient = args[0] // claripy.ZeroExt(self._from_size - self._to_size, args[1])
+            remainder = args[0] % claripy.ZeroExt(self._from_size - self._to_size, args[1])
+            quotient_size = self._to_size
+            remainder_size = self._to_size
+            return claripy.Concat(
+                claripy.Extract(remainder_size - 1, 0, remainder), claripy.Extract(quotient_size - 1, 0, quotient)
+            )
+
+    # pylint:enable=no-self-use,unused-argument
+
+    # FP!
+    def _op_int_to_fp(self, args):
+        rm_exists = self._from_size != 32 or self._to_size != 64
+        rm = self._translate_rm(args[0] if rm_exists else claripy.BVV(0, 32))
+        arg = args[1 if rm_exists else 0]
+
+        return arg.val_to_fp(claripy.fp.FSort.from_size(self._output_size_bits), signed=self._from_signed != "U", rm=rm)
+
+    def _op_fp_to_fp(self, args):
+        rm_exists = self._from_size != 32 or self._to_size != 64
+        rm = self._translate_rm(args[0] if rm_exists else claripy.BVV(0, 32))
+        arg = args[1 if rm_exists else 0].raw_to_fp()
+
+        return arg.raw_to_fp().to_fp(claripy.fp.FSort.from_size(self._output_size_bits), rm=rm)
+
+    def _op_fp_to_int(self, args):
+        if self.name.endswith("_RZ"):
+            rm = claripy.fp.RM_TowardsZero
+            arg = args[0]
+        elif self.name.endswith("_RN"):
+            rm = claripy.fp.RM_NearestTiesEven
+            arg = args[0]
+        elif self.name.endswith("_RM"):
+            rm = claripy.fp.RM_TowardsNegativeInf
+            arg = args[0]
+        elif self.name.endswith("_RP"):
+            rm = claripy.fp.RM_TowardsPositiveInf
+            arg = args[0]
+        else:
+            rm = self._translate_rm(args[0])
+            arg = args[1]
+
+        if not self._vector_size:
+            return self._compute_fp_to_int(rm, arg.raw_to_fp(), self._to_size)
+        else:
+            vector_args = arg.chop(self._vector_size)
+            return claripy.Concat(
+                *[self._compute_fp_to_int(rm, varg.raw_to_fp(), self._vector_size) for varg in vector_args]
+            )
+
+    def _compute_fp_to_int(self, rm, arg, to_size):
+        if self._to_signed == "S":
+            return claripy.fpToSBV(rm, arg, to_size)
+        else:
+            return claripy.fpToUBV(rm, arg, to_size)
+
+    def _op_fgeneric_Cmp(self, args):  # pylint:disable=no-self-use
+        # see https://github.com/angr/vex/blob/master/pub/libvex_ir.h#L580
+        a, b = args[0].raw_to_fp(), args[1].raw_to_fp()
+        return claripy.ite_cases(
+            (
+                (claripy.fpLT(a, b), claripy.BVV(0x01, 32)),
+                (claripy.fpGT(a, b), claripy.BVV(0x00, 32)),
+                (claripy.fpEQ(a, b), claripy.BVV(0x40, 32)),
+            ),
+            claripy.BVV(0x45, 32),
+        )
+
+    @staticmethod
+    def _fp_vector_comparison(cmp, a0, a1):
+        # for cmpps_eq stuff, i.e. Iop_CmpEQ32Fx4
+        return claripy.If(cmp(a0, a1), claripy.BVV(-1, len(a0)), claripy.BVV(0, len(a0)))
+
+    def _op_fgeneric_CmpEQ(self, a0, a1):
+        return self._fp_vector_comparison(claripy.fpEQ, a0, a1)
+
+    def _op_fgeneric_CmpLE(self, a0, a1):
+        return self._fp_vector_comparison(claripy.fpLT, a0, a1)
+
+    def _op_fgeneric_CmpLT(self, a0, a1):
+        return self._fp_vector_comparison(claripy.fpLEQ, a0, a1)
+
+    def _auto_vectorize(self, f, args, rm=None, rm_passed=False):
+        if rm is not None:
+            rm = self._translate_rm(rm)
+            if rm_passed:
+                f = partial(f, rm)
+
+        if self._vector_size is None:
+            return f(args)
+
+        if self._vector_zero:
+            chopped = [arg[(self._vector_size - 1) : 0].raw_to_fp() for arg in args]
+            result = f(*chopped).raw_to_bv()
+            return claripy.Concat(args[0][(args[0].length - 1) : self._vector_size], result)
+        else:
+            # I'm changing this behavior because I think this branch was never used otherwise
+            # before it only chopped the first argument but I'm going to make it chop all of them
+            result = []
+            for lane_args in self.vector_args(args):
+                if self._float:
+                    # HACK HACK HACK
+                    # this is such a weird divergence. why do the fp generics take several args and the int generics
+                    # take a list?
+                    result.append(f(*lane_args).raw_to_bv())
+                else:
+                    result.append(f(lane_args))
+            return claripy.Concat(*result)
+
+    @staticmethod
+    def _fgeneric_minmax(cmp_op, a, b):
+        a, b = a.raw_to_fp(), b.raw_to_fp()
+        return claripy.If(cmp_op(a, b), a, b)
+
+    def _op_fgeneric_Min(self, a, b):
+        return self._fgeneric_minmax(claripy.fpLT, a, b)
+
+    def _op_fgeneric_Max(self, a, b):
+        return self._fgeneric_minmax(claripy.fpGT, a, b)
+
+    def _op_fgeneric_Reinterp(self, args):
+        if self._to_type == "I":
+            return args[0].raw_to_bv()
+        elif self._to_type == "F":
+            return args[0].raw_to_fp()
+        else:
+            raise SimOperationError("unsupport Reinterp _to_type")
+
+    @supports_vector
+    def _op_fgeneric_Round(self, args):
+        if self._vector_size is not None:
+            rm = {
+                "RM": claripy.fp.RM.RM_TowardsNegativeInf,
+                "RP": claripy.fp.RM.RM_TowardsPositiveInf,
+                "RN": claripy.fp.RM.RM_NearestTiesEven,
+                "RZ": claripy.fp.RM.RM_TowardsZero,
+            }[self._rounding_mode]
+
+            rounded = []
+            for i in reversed(range(self._vector_count)):
+                # pylint:disable=no-member
+                left = claripy.Extract((i + 1) * self._vector_size - 1, i * self._vector_size, args[0]).raw_to_fp()
+                rounded.append(claripy.fpToSBV(rm, left, self._vector_size))
+            return claripy.Concat(*rounded)
+        else:
+            rm = self._translate_rm(args[0])
+            rounded_bv = claripy.fpToSBV(rm, args[1].raw_to_fp(), args[1].length)
+
+            # if exponent is large enough, floating points are always integers.
+            fsort = claripy.fp.FSort.from_size(args[1].length)
+            mantissa_bits = fsort.mantissa - 1  # -1 since FSort has mantissa value 1 higher than the number of bits
+            exp_bits = fsort.exp
+            rounded_fp = claripy.fpToFP(claripy.fp.RM.RM_NearestTiesEven, rounded_bv, fsort)
+            exp_bv = args[1].raw_to_bv()[exp_bits + mantissa_bits - 1 : mantissa_bits]
+            exp_threshold = (2 ** (exp_bits - 1) - 1) + mantissa_bits
+            return claripy.If(exp_bv >= exp_threshold, args[1].raw_to_fp(), rounded_fp)
+
+    def _generic_pack_saturation(self, args, src_size, dst_size, src_signed, dst_signed):
+        """
+        Generic pack with saturation.
+        Split args in chunks of src_size and then pack them into saturated chunks of dst_size bits.
+        Then chunks are concatenated.
+
+        So far as I know plain saturating conversion should only ever have to be performed on vectors.
+        """
+        if src_size <= 0 or dst_size <= 0:
+            raise SimOperationError("Can't pack from or to zero or negative size: %s" % self.name)
+        if src_size < dst_size:
+            raise SimOperationError("Can't pack from small size into larger size: %s" % self.name)
+        result = None
+
+        max_value = 2**dst_size - 1
+        if dst_signed:
+            max_value >>= 1
+        if not dst_signed or not src_signed:
+            min_value = 0
+        else:
+            min_value = -(2 ** (dst_size - 1))
+
+        gt = claripy.SGT if src_signed else claripy.UGT
+        lt = claripy.SLT if src_signed else claripy.ULT
+
+        for v in args:
+            for src_value in v.chop(src_size):
+                dst_value = claripy.If(
+                    gt(src_value, max_value),
+                    max_value,
+                    claripy.If(lt(src_value, min_value), min_value, src_value[dst_size - 1 : 0]),
+                )
+                if result is None:
+                    result = dst_value
+                else:
+                    result = self._op_concat((result, dst_value))
+        return result
+
+    @supports_vector
+    def _op_generic_QNarrowBin(self, args):
+        return self._generic_pack_saturation(
+            args, self._from_size, self._vector_size, self._from_signed == "S", self._vector_signed == "S"
+        )
+
+    def _op_Iop_64x4toV256(self, args):
+        return self._op_concat(args)
+
+    @staticmethod
+    def _op_Iop_V256to64_0(args):
+        return args[0][63:0]
+
+    @staticmethod
+    def _op_Iop_V256to64_1(args):
+        return args[0][127:64]
+
+    @staticmethod
+    def _op_Iop_V256to64_2(args):
+        return args[0][191:128]
+
+    @staticmethod
+    def _op_Iop_V256to64_3(args):
+        return args[0][255:192]
+
+    @staticmethod
+    def _op_Iop_V256toV128_0(args):
+        return args[0][127:0]
+
+    @staticmethod
+    def _op_Iop_V256toV128_1(args):
+        return args[0][255:128]
+
+    @staticmethod
+    def _op_Iop_MAddF64(args):
+        """
+        Ternary operation.
+            arg0 == 0
+            return arg1 * arg2 + arg3
+
+        :param args:    Arguments to this operation.
+        :return:        The operation result.
+        """
+
+        return args[1] * args[2] + args[3]
+
+    @supports_vector
+    def _op_generic_MulHi(self, args):
+        """
+        Sign-extend double each lane, multiply each lane, and store only the high half of the result
+        """
+        if self._vector_signed == "S":
+            lanes_0 = [lane.sign_extend(self._vector_size) for lane in args[0].chop(self._vector_size)]
+            lanes_1 = [lane.sign_extend(self._vector_size) for lane in args[1].chop(self._vector_size)]
+        else:
+            lanes_0 = [lane.zero_extend(self._vector_size) for lane in args[0].chop(self._vector_size)]
+            lanes_1 = [lane.zero_extend(self._vector_size) for lane in args[1].chop(self._vector_size)]
+        mulres = [a * b for a, b in zip(lanes_0, lanes_1)]
+        highparts = [x.chop(self._vector_size)[0] for x in mulres]
+        return claripy.Concat(*highparts)
+
+    @supports_vector
+    def _op_generic_Perm(self, args):
+        ordered_0 = list(reversed(args[0].chop(self._vector_size)))
+        ordered_1 = list(reversed(args[1].chop(self._vector_size)))
+        res = []
+        nbits = int(math.log2(self._vector_count))
+        for pword in ordered_1:
+            switch = pword[nbits - 1 : 0]
+            kill = pword[self._vector_size - 1]
+            switched = claripy.ite_cases([(switch == i, v) for i, v in enumerate(ordered_0[:-1])], ordered_0[-1])
+            killed = claripy.If(kill == 1, 0, switched)
+            res.append(killed)
+
+        return claripy.Concat(*reversed(res))
+
+    @supports_vector
+    def _op_generic_CatEvenLanes(self, args):
+        vec_0 = args[0].chop(self._vector_size)
+        vec_1 = args[1].chop(self._vector_size)
+        return claripy.Concat(*(vec_0[1::2] + vec_1[1::2]))
+
+    @supports_vector
+    def _op_generic_CatOddLanes(self, args):
+        vec_0 = args[0].chop(self._vector_size)
+        vec_1 = args[1].chop(self._vector_size)
+        return claripy.Concat(*(vec_0[::2] + vec_1[::2]))
+
+    def _op_generic_Set(self, args):
+        if self._set_side != "lo":
+            raise NotImplementedError(f"Unsupported set_side {self._set_side}")
+        if self._set_size in {32, 64}:
+            if self._set_size != args[1].size():
+                raise SimOperationError(f"Unexpected args[1] size {args[1].size()}")
+            v = claripy.Concat(args[0][args[0].size() - 1 : self._set_size], args[1])
+            return v
+        raise NotImplementedError(f"Unsupported set_size {self._set_size}")
+
+    # def _op_Iop_Yl2xF64(self, args):
+    #   rm = self._translate_rm(args[0])
+    #   arg2_bv = args[2].raw_to_bv()
+    #   # IEEE754 double looks like this:
+    #   # SEEEEEEEEEEEFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF
+    #   # thus, we extract the exponent bits, re-bias them, then
+    #   # (signed) convert them back into an FP value for the integer
+    #   # part of the log. then we make the approximation that log2(x)
+    #   # = x - 1 for 1.0 <= x < 2.0 to account for the mantissa.
+
+    #   # the bias for doubles is 1023
+    #   arg2_exp = (arg2_bv[62:52] - 1023).val_to_fp(claripy.fp.FSORT_DOUBLE, signed=True, rm=rm)
+    #   arg2_mantissa = claripy.Concat(claripy.BVV(int('001111111111', 2), 12), arg2_bv[51:0]).raw_to_fp()
+    #   # this is the hacky approximation:
+    #   log2_arg2_mantissa = claripy.fpSub(rm, arg2_mantissa, claripy.FPV(1.0, claripy.fp.FSORT_DOUBLE))
+    #   return claripy.fpMul(rm, args[1].raw_to_fp(), claripy.fpAdd(rm, arg2_exp, log2_arg2_mantissa))
+
+    # def _op_Iop_Yl2xp1F64(self, args):
+    #   rm_raw, arg1, arg2 = args
+    #   rm = self._translate_rm(rm_raw)
+    #   arg2_p1 = claripy.fpAdd(rm, arg2.raw_to_fp(), claripy.FPV(1.0, claripy.fp.FSORT_DOUBLE))
+    #   return self._op_Iop_Yl2xF64((rm_raw, arg1, arg2_p1))
+
+    @staticmethod
+    def pow(rm, arg, n):
+        out = claripy.FPV(1.0, arg.sort)
+        for _ in range(n):
+            out = claripy.fpMul(rm, arg, out)
+        return out
+
+    # def _op_Iop_SinF64(self, args):
+    #   rm, arg = args
+    #   rm = self._translate_rm(rm)
+    #   rounds = 15
+    #   accumulator = claripy.FPV(0.0, arg.sort)
+    #   factorialpart = 1.0
+    #   for i in range(1, rounds + 1):
+    #       term = claripy.fpDiv(rm, self.pow(rm, arg, 2*i - 1), claripy.FPV(float(factorialpart), arg.sort))
+    #       factorialpart *= ((i*2) + 1) * (i*2)
+    #       if i % 2 == 1:
+    #           accumulator = claripy.fpAdd(rm, accumulator, term)
+    #       else:
+    #           accumulator = claripy.fpSub(rm, accumulator, term)
+
+    #   return accumulator
+
+    # def _op_Iop_CosF64(self, args):
+    #   rm, arg = args
+    #   rm = self._translate_rm(rm)
+    #   rounds = 20
+    #   accumulator = claripy.FPV(1.0, arg.sort)
+    #   factorialpart = 2.0
+    #   for i in range(1, rounds + 1):
+    #       term = claripy.fpDiv(rm, self.pow(rm, arg, 2*i), claripy.FPV(float(factorialpart), arg.sort))
+    #       factorialpart *= (i*2 + 1) * (i*2 + 2)
+    #       if i % 2 == 1:
+    #           accumulator = claripy.fpSub(rm, accumulator, term)
+    #       else:
+    #           accumulator = claripy.fpAdd(rm, accumulator, term)
+
+    #   return accumulator
+
+
+#
+# Op Handler
+#
+
+
+def vexop_to_simop(op, extended=True, fp=True):
+    res = operations.get(op)
+    if res is None and extended:
+        attrs = op_attrs(op)
+        if attrs is None:
+            raise UnsupportedIROpError("Operation not implemented")
+        res = SimIROp(op, **attrs)
+    if res is None:
+        raise UnsupportedIROpError("Operation not implemented")
+    if res._float and not fp:
+        raise UnsupportedIROpError("Floating point support disabled")
+    return res
+
+
+from angr.errors import UnsupportedIROpError, SimOperationError, SimValueError, SimZeroDivisionException
+
+make_operations()