(Compiler) Incorrect bytecode generation leading to incorrect behavior
(Compiler) Unexpected behavior
Description
Brief/Intro
The slice precompile skips length != 0 check when the first argument is an is_adhoc_slice. However, if length == 0, the make_setter function will discard all operations except mstore dst_size_field 0 regardless of any potential side effect, leading to incorrect compilation.
Vulnerability Details
The slice buildin skips a lot of checks when the src to be slices is an adhoc_slice (msg.data or (address).code). Notably, the length_literal < 1 check is skipped.
class Slice(BuiltinFunctionT):
...
def fetch_call_return(self, node):
...
arg = node.args[0]
...
length_expr = node.args[2].reduced()
# CMC 2022-03-22 NOTE slight code duplication with semantics/analysis/local
is_adhoc_slice = arg.get("attr") == "code" or (
arg.get("value.id") == "msg" and arg.get("attr") == "data"
)
start_literal = start_expr.value if isinstance(start_expr, vy_ast.Int) else None
length_literal = length_expr.value if isinstance(length_expr, vy_ast.Int) else None
if not is_adhoc_slice:
if length_literal is not None:
if length_literal < 1:
raise ArgumentException("Length cannot be less than 1", length_expr)
if length_literal > arg_type.length:
raise ArgumentException(f"slice out of bounds for {arg_type}", length_expr)
if start_literal is not None:
if start_literal > arg_type.length:
raise ArgumentException(f"slice out of bounds for {arg_type}", start_expr)
if length_literal is not None and start_literal + length_literal > arg_type.length:
raise ArgumentException(f"slice out of bounds for {arg_type}", node)
...
During code generation, the IR for the slice builtin is created by the following snippet
def _build_adhoc_slice_node(sub: IRnode, start: IRnode, length: IRnode, context: Context) -> IRnode:
assert length.is_literal, "typechecker failed"
assert isinstance(length.value, int) # mypy hint
dst_typ = BytesT(length.value)
# allocate a buffer for the return value
buf = context.new_internal_variable(dst_typ)
with scope_multi((start, length), ("start", "length")) as (b1, (start, length)):
# `msg.data` by `calldatacopy`
if sub.value == "~calldata":
node = [
"seq",
check_buffer_overflow_ir(start, length, "calldatasize"),
["mstore", buf, length],
["calldatacopy", add_ofst(buf, 32), start, length],
buf,
]
elif sub.value == "~selfcode":
...
else:
...
ret = IRnode.from_list(node, typ=BytesT(length.value), location=MEMORY)
return b1.resolve(ret)
And then returned to parse_AnnAssign as rhs
class Stmt:
...
def parse_AnnAssign(self):
ltyp = self.stmt.target._metadata["type"]
varname = self.stmt.target.id
lhs = self.context.new_variable(varname, ltyp)
assert self.stmt.value is not None
rhs = Expr(self.stmt.value, self.context).ir_node
return make_setter(lhs, rhs)
...
Then make_setter is called to generate the final ir for this statement.
def make_setter(left, right, hi=None):
check_assign(left, right)
if potential_overlap(left, right):
raise CompilerPanic("overlap between src and dst!")
assert (hi is not None) == _dirty_read_risk(right)
if left.typ._is_prim_word:
...
elif isinstance(left.typ, _BytestringT):
# TODO rethink/streamline the clamp_basetype logic
if needs_clamp(right.typ, right.encoding):
...
else:
ret = make_byte_array_copier(left, right)
return IRnode.from_list(ret)
elif isinstance(left.typ, DArrayT):
...
...
def make_byte_array_copier(dst, src):
assert isinstance(src.typ, _BytestringT)
assert isinstance(dst.typ, _BytestringT)
_check_assign_bytes(dst, src)
if src.value == "~empty" or src.typ.maxlen == 0:
# set length word to 0.
return STORE(dst, 0) #drops all rhs (src) ir if src.typ.maxlen == 0
...
Unfortunately, rhs (slice result) ir is completely dropped when its src.typ.maxlen == 0, along with any side effects that it may have. This in turn leads to incorrect compilation.
While we use slice as an entry point for this issue, we consider the actual root cause to be make_setter and make_byte_array_copier which eliminates ir without checking whether it contains side effect. There are potentially other entry points that we haven't explored yet, but can lead to the same issue.
Additionally, the is_adhoc_slice checker is also not precise, since it doesn't check whether the code attribute comes from an address, so if user defined a struct that contains a field calls code, then it'll also be treated as an is_adhoc_slice in Slice.fetch_call_return.
is_adhoc_slice = arg.get("attr") == "code" or (
arg.get("value.id") == "msg" and arg.get("attr") == "data"
)
Impact Details
Incorrectly optimizing out side effects can result in all kinds of unexpected behaviors in a contract, such as skipping critical checks performed in the incorrectly eliminated code, skipping important business logic, and more. This exposes contract developers to substantial risks.
