Solidity makes it possible to define variables that extend past the last (2**256-th) slot of storage, which results in wrap-around back to slot zero. Since EVM uses 256-bit integer arithmetic, most operations on such variables just work. The only situation which requires special attention is iteration against absolute slot addresses: the invariant that the last slot belonging to a variable has the highest address does not hold. When implemented incorrectly, a loop over an array will immediately terminate if the container spans the end of storage - due to the initial position already being greater than the end position. This affected storage array clearing loops generated by both evmasm and IR pipelines. Additionally, (only in the evmasm pipeline) copying operations whose source was an array straddling the end of storage were also affected. At the language level, the buggy code would be generated for array assignment, array initialization, delete operator, <array>.pop() and <array>.push(). Note that a clearing loop is inserted by the compiler not only for invocations of the delete operator, but also to zero storage when overwriting a longer array with a shorter one, popping an element or even pushing an empty element to a dynamic array. Since clearing is a separate loop, it is possible for the bug to only affect it and not the copy operation it follows (which is always the case in the IR pipeline). The bug is extremely unlikely to be triggered accidentally due to the probabilistic impossibility of a short dynamic array being allocated right at the storage boundary. On the other hand, scenarios in which a user may place a static array there intentionally do not seem realistic and are limited to unusual layouts, in which a contract does not place any storage variables at slot zero (otherwise they would overlap the array).

Function call arguments in Yul are evaluated right to left. This order matters when the argument expressions have side-effects, and changing it may change contract behavior. FullInliner is an optimizer step that can replace a function call with the body of that function. The transformation involves assigning argument expressions to temporary variables, which imposes an explicit evaluation order. FullInliner was written with the assumption that this order does not necessarily have to match usual argument evaluation order because the argument expressions have no side-effects. In most circumstances this assumption is true because the default optimization step sequence contains the ExpressionSplitter step. ExpressionSplitter ensures that the code is in *expression-split form*, which means that function calls cannot appear nested inside expressions, and all function call arguments have to be variables. The assumption is, however, not guaranteed to be true in general. Version 0.6.7 introduced a setting allowing users to specify an arbitrary optimization step sequence, making it possible for the FullInliner to actually encounter argument expressions with side-effects, which can result in behavior differences between optimized and unoptimized bytecode. Contracts compiled without optimization or with the default optimization sequence are not affected. To trigger the bug the user has to explicitly choose compiler settings that contain a sequence with FullInliner step not preceded by ExpressionSplitter.

When accessing the ``.selector`` member on an expression with side-effects, like an assignment, a function call or a conditional, the expression would not be evaluated in the legacy code generation. This would happen in expressions where the functions used in the expression were all known at compilation time, regardless of whether the whole expression could be evaluated at compilation time or not. Note that the code generated by the IR pipeline was unaffected and would behave as expected.

When ABI-encoding a statically-sized calldata array, the compiler always pads the data area to a multiple of 32-bytes and ensures that the padding bytes are zeroed. In some cases, this cleanup used to be performed by always writing exactly 32 bytes, regardless of how many needed to be zeroed. This was done with the assumption that the data that would eventually occupy the area past the end of the array had not yet been written, because the encoder processes tuple components in the order they were given. While this assumption is mostly true, there is an important corner case: dynamically encoded tuple components are stored separately from the statically-sized ones in an area called the *tail* of the encoding and the tail immediately follows the *head*, which is where the statically-sized components are placed. The aforementioned cleanup, if performed for the last component of the head would cross into the tail and overwrite up to 32 bytes of the first component stored there with zeros. The only array type for which the cleanup could actually result in an overwrite were arrays with ``uint256`` or ``bytes32`` as the base element type and in this case the size of the corrupted area was always exactly 32 bytes. The problem affected tuples at any nesting level. This included also structs, which are encoded as tuples in the ABI. Note also that lists of parameters and return values of functions, events and errors are encoded as tuples.

Copying ``bytes`` arrays from memory or calldata to storage is done in chunks of 32 bytes even if the length is not a multiple of 32. Thereby, extra bytes past the end of the array may be copied from calldata or memory to storage. These dirty bytes may then become observable after a ``.push()`` without arguments to the bytes array in storage, i.e. such a push will not result in a zero value at the end of the array as expected. This bug only affects the legacy code generation pipeline, the new code generation pipeline via IR is not affected.

When calling external functions, it is irrelevant if the data location of the parameters is ``calldata`` or ``memory``, the encoding of the data does not change. Because of that, changing the data location when overriding external functions is allowed. The compiler incorrectly also allowed a change in the data location for overriding public and internal functions. Since public functions can be called internally as well as externally, this causes invalid code to be generated when such an incorrectly overridden function is called internally through the base contract. The caller provides a memory pointer, but the called function interprets it as a calldata pointer or vice-versa.

Calldata validation for nested dynamic types is deferred until the first access to the nested values. Such an access may for example be a copy to memory or an index or member access to the outer type. While in most such accesses calldata validation correctly checks that the data area of the nested array is completely contained in the passed calldata (i.e. in the range [0, calldatasize()]), this check may not be performed, when ABI encoding such nested types again directly from calldata. For instance, this can happen, if a value in calldata with a nested dynamic array is passed to an external call, used in ``abi.encode`` or emitted as event. In such cases, if the data area of the nested array extends beyond ``calldatasize()``, ABI encoding it did not revert, but continued reading values from beyond ``calldatasize()`` (i.e. zero values).

When immutable variables of signed integer type shorter than 256 bits are read, their higher order bits were unconditionally set to zero. The correct operation would be to sign-extend the value, i.e. set the higher order bits to one if the sign bit is one. This sign-extension is performed by Solidity just prior to when it matters, i.e. when a value is stored in memory, when it is compared or when a division is performed. Because of that, to our knowledge, the only way to access the value in its unclean state is by reading it through inline assembly.

The ABI specification uses pointers to data areas for everything that is dynamically-sized. When decoding data from memory (instead of calldata), the ABI decoder did not properly validate some of these pointers. More specifically, it was possible to use large values for the pointers inside arrays such that computing the offset resulted in an undetected overflow. This could lead to these pointers targeting areas in memory outside of the actual area to be decoded. This way, it was possible for ``abi.decode`` to return different values for the same encoded byte array.

Solidity's bytecode optimizer has a step that can compute Keccak-256 hashes, if the contents of the memory are known during compilation time. This step also has a mechanism to determine that two Keccak-256 hashes are equal even if the values in memory are not known during compile time. This mechanism had a bug where Keccak-256 of the same memory content, but different sizes were considered equal. More specifically, ``keccak256(mpos1, length1)`` and ``keccak256(mpos2, length2)`` in some cases were considered equal if ``length1`` and ``length2``, when rounded up to nearest multiple of 32 were the same, and when the memory contents at ``mpos1`` and ``mpos2`` can be deduced to be equal. You maybe affected if you compute multiple Keccak-256 hashes of the same content, but with different lengths inside inline assembly. You are unaffected if your code uses ``keccak256`` with a length that is not a compile-time constant or if it is always a multiple of 32.