make_setter is incorrect for complex types when the RHS references the LHS

[charles-cooper]

Version Information

• vyper Version (output of vyper --version): 0.2.15

What's your issue about?

This causes the memory location of x[0] to be overwritten before it gets copied to x[1].

@public
def bug(x_unsorted: uint256[2]) -> uint256:
    # Initial value
    x: uint256[2] = x_unsorted
    x = [x[1], x[0]] # bug!
    # x is now equal to [ x_unsorted[1], x_unsorted[1] ]
    return x[0]
# But when it's x = [x_unsorted[1], x_unsorted[0]] - all good

On 0.2.15 this produces code like

[seq,
  [mstore, 416, [mload, 448 <32+x>]],
  [mstore, 448, [mload, 416 <0+x>]]

This has been there for awhile, even in 0.2.0 we get

[with, _L, 416 <x>,
  [seq,
    [mstore, _L, [mload, 448 <32+x>]], <- still bug
    [mstore, [add, 32, _L], [mload, 416 <0+x>]]]],

I'm guessing a similar issue exists for contract variables.

How can it be fixed?

Well, fix make_setter, but it's unclear what the best approach is. The first obvious fix is just to process all the loads before all the stores, but this breaks for byte arrays (because those don't use the stack at all, they use memcpy macros).

[charles-cooper]

probably the most sane thing to do here is, first detect if the RHS depends on the LHS. if there is a dependency, then make_setter to a temporary buffer and then run make_setter on that to the LHS.