The main Problem I see with this is that you still have to aware of the possibility of an arithmetic overflow. If we throw in that case, this could stall a smart contract.

The only solution I see is to use our formal verification system that automatically verifies whether overflow protection checks are in place (or not needed).

Could the same argument be made for the exception that is thrown for division by 0 and a negative array index? Those could have the same effect, but I can't see arguing that coming up with some default but incorrect behavior would be better (e.g. We could just wrap array indices around).

My concern is that the failures that can come from overflow go undetected.

I am definitely not suggesting that formal verification should not be done, there is no substitute. But I think allowing overflow offers no real advantage, unless you're hoping to hack a contract.

You have a valid point. Still, I think these are not exactly comparable:
Invalid array access and division by zero are more predictable and less "continuous" than arithmetic overflow.

But yes, it is probably better to stall a contract in case of overflow than to provoke undesired effects on state.

What are cases where overflow is actually desired? I think for most of them, we can still use mulmod and addmod.

Note that this is quite expensive to do, especially for types shorter than 256 bits.

Would you consider this for 0.4.0, or is it a bit involved for that?

Maybe this should be implemented in the EVM?

I think arithmetic overflow detection belongs to the EVM, but that would be a rather big change to it.

For the EVM I think it would need extra opcodes as overflowing math is still useful. In the meantime we can have it in Solidity and get the safety benefits at the expense of a bit of gas (completely worth it I think because you end up checking it in safe code anyway).

@rainbeam This is too involved for 0.4.0. Arithmetics is almost for free when compared to storage writes or even a transaction, so I think it is fine to do it. I do not really agree that this belongs to the EVM. Something like an overflow flag might be added to the EVM, but not a "force-stop" on overflow.

I think arithmetic overflow detection belongs to the EVM

By detection I meant the overflow flag. So perhaps that would then needs its on opcode for getting the overflow flag onto the stack and/or its own jump (JUMPOF).

@pirapira Can you please list all these other issues as tasks for this issue and close the others, or do you think that each issues is a substantial amount of work?

The main reason I opened separate issues is that each requires discussion if we want it or not, but if we are set to go over everything, here is a list.

• exception on overflow in unsigned->signed conversion
• exception on overflow in signed->unsigned conversion
• exception on overflow in size-decreasing implicit conversion
• exception on overflow in addition of two signed numbers
• exception on overflow in addition of two unsigned numbers
• exception on underflow in subtraction of two signed numbers
• exception on underflow in subtraction of two unsigned numbers
• exception on overflow in multiplication of two signed numbers
• exception on overflow in multiplication of two unsigned numbers
• exception on overflow in shifts
• exception on overflow in ++ on a signed number
• exception on overflow in ++ on an unsigned number
• exception on underflow in -- on a signed number
• exception on underflow in -- on an unsigned number
• exception on overflow in +=
• exception on overflow in -=
• exception on overflow in *=
• exception on overflow in /=
• make sure no optimizations are relying on (a + b - b == a); lest they remove overflow exceptions
• compiler error on an out-of-range constant expression when the constant expression is interpreted into a type (uint x = -1)
• exception on overflow in INT_MIN/-1 Breaking: overflow detection for SDIV - minInt / (-1) #1091
• exponentiation

This is labelled "soon" and how soon will this be? :) Does the long checklist need to be implemented at once, or can it be done iteratively?
Implementing some of them can already provide early safety/ROI imo, like:

• exception on overflow in addition of two unsigned numbers
• exception on underflow in subtraction of two unsigned numbers

@pirapira In your checklist, exception on overflow in /= is duplicated. I think for one of them you mean exception on underflow in -=.

Also, how about overflow of INT_MIN/-1 ?

@ethers Thank you. I fixed the checklist above.

These overflow checkings are breaking changes, and for some reasons, Solidity release 0.5.0 will be the next one that can contain breaking changes. I would support removing this rule (until we get to 1.0.0).

If Solidity doesn't already do this, a line item in the TODO list should be added for truncation during assignment.

uint16 a = 2^15;
uint8 b = a;

@axic yes, an unchecked region where also the SMT checker is disabled is certainly a good idea. Whether we can enable runtime checks by default in the non-unchecked areas should also depend on an assessment about how well the optimizer can handle such cases, and perhaps an improvement in that area.

The recently discovered overflow on batchTransfer in some ERC20 contracts is a good example of the type of issue that could be avoided if this is implemented.

https://medium.com/@peckshield/alert-new-batchoverflow-bug-in-multiple-erc20-smart-contracts-cve-2018-10299-511067db6536

For the syntax (mentioned in #796 (comment)) I propose:

unchecked "overflow" {
  a = b + c;
}

It follows the same syntax as not-so-well-known syntax of the assembly keyword: assembly "evmasm" {}

@axic yep, exactly my thought, that allows us to introduce other checks in the same way later.

that allows us to introduce other checks in the same way later.

That was my motivation :)

@axic what about making the current math operators safe and adding new operators for non-safe math

c = a + b // safe
c = a %+ b // non-safe

pick whatever symbol works, i chose %+ to imply modular arithmetic

Edit:
I guess this would not handle the conversion cases (e.g. signed->unsigned)
And the generic check-modifying block opens up interesting possibilities

In #796 (comment) I think there is one more case not mentioned, exponentiation.

The main reason I opened separate issues is that each requires discussion if we want it or not, but if we are set to go over everything, here is a list.

• exception on overflow in unsigned->signed conversion
• exception on overflow in signed->unsigned conversion
• exception on overflow in size-decreasing implicit conversion
• exception on overflow in addition of two signed numbers
• exception on overflow in addition of two unsigned numbers
• exception on underflow in subtraction of two signed numbers
• exception on underflow in subtraction of two unsigned numbers
• exception on overflow in multiplication of two signed numbers
• exception on overflow in multiplication of two unsigned numbers
• exception on overflow in shifts
• exception on overflow in ++ on a signed number
• exception on overflow in ++ on an unsigned number
• exception on underflow in -- on a signed number
• exception on underflow in -- on an unsigned number
• exception on overflow in +=
• exception on overflow in -=
• exception on overflow in *=
• exception on overflow in /=
• make sure no optimizations are relying on (a + b - b == a); lest they remove overflow exceptions
• compiler error on an out-of-range constant expression when the constant expression is interpreted into a type (uint x = -1)
• exception on overflow in INT_MIN/-1 Breaking: overflow detection for SDIV - minInt / (-1) #1091
• exponentiation

How many of these are completed till now?

@johnsoncarl we are implementing this along with the transition to yul as an intermediate language. The reason is that we hope that the yul optimizer can better cope with all the additional checks that are introduced and complicate the control-flow. It is still not 100% clear if we really want that on explicit type conversions. It is implemented (or at least PRs in progress) for increment, decrement, unsigned multiplication and addition.

We are thinking of using the existing yul implementation and adding it to 0.7.0.

Implemented in #9465