Incorrect success value returned in vyper

Background

During the audit of Lido's Gate Seals code statemind team identified a weird behavior of the code that uses raw_call: https://github.com/lidofinance/gate-seals/blob/051593e74df01a4131c485b4fda52e691cd4b7d8/contracts/GateSeal.vy#L164 .

Construction like this:

success = raw_call(
    sealable,
    _abi_encode(SEAL_DURATION_SECONDS, method_id=method_id("pauseFor(uint256)")),
    revert_on_failure=False
)

was not fully documented: https://docs.vyperlang.org/en/v0.3.7/built-in-functions.html#raw_call .

The documentation says that: if max_outsize=0 it should return nothing and then it says that if revert_on_failure=False it should return a success flag in the tuple of response, but what if max_outsize=0 and revert_on_failure=False.

<img width="715" alt="image" src="https://user-images.githubusercontent.com/22330612/232125364-d2b3bbac-0b4f-40cb-80ff-f55d8eafef44.png">

So the team started researching what exactly happened in that case, after some research we found that the Vyper compiler generates the wrong bytecode in that case, it generates the sequence:

CALL // call
MLOAD // MLOAD is wrong since the CALL result is already stored in the stack

Impact

Example of buggy code:

@external
def returnSome(calling: address, a: uint256) -> bool:
    success: bool = false
    success = raw_call(
        calling,
        _abi_encode(a, method_id=method_id("a(uint256)")),
        revert_on_failure=False
        )

any contract that uses the raw_call with revert_on_failure=False and max_outsize=0 receives the wrong response from raw_call. Depending on the memory garbage, the result can be either True or False.

Patches

Fix by @charles-cooper https://github.com/vyperlang/vyper/commit/851f7a1b3aa2a36fd041e3d0ed38f9355a58c8ae

Workarounds

The simple workaround is always to put max_outsize>0. Workaround example https://github.com/lidofinance/gate-seals/pull/5/files

References

Lido's fix: https://github.com/lidofinance/gate-seals/pull/5/files