REVERT instruction

[axic]

Replaces #140.

[axic]

I assume we did assign 140 for this, given issue 140 was accepted.

[Souptacular]

You can either assign the issue or PR number. In this case it would be better to assign the issue # since that is what it has been referred to most widely.

[axic]

I wonder if it would make sense adjusting CALL (and CALLCODE, DELEGATECALL) to indicate a REVERT.

Currently, CALL (& co) will return 1 on success and 0 in in case it failed. Perhaps code 2 could be returned in case of REVERT. Downside is that Solidity contracts would not be backwards compatible.

Likely it would make sense to take this into consideration when designing the successor of CALL and not touching the current one.

[chriseth]

For the sake of backwards compatibility, anything that reverts state changes should signal a "failure" to the caller. Since we cannot distinguish the two cases where the callee did provide a return value and where it didn't, I would propose to postpone that feature until we have a means to signal this e.g. in a new call opcode: If the caller used the CALL opcode and the callee provided failure data, the data is discarded.

[Arachnid]

Returning 0 from a CALL that ends in a REVERT seems reasonable, though it would be easier to check returned data if contracts had a way to detect the length of returned data; at present they could at least zero the return memory area before calling (or we could HF to specify CALL does that).

[axic]

I agree it is better to make this low impact by not touching CALL at the moment. External callers (i.e. users of the RPC) will still be able distinguish and retrieve and error message.

[Arachnid]

If we add it without any return support initially, then it'll be even more complex to add it later. Shouldn't we try and solve this once-and-for-all?

[axic]

I mean that CALL wouldn't differentiate and would ignore the return value. However, the return value would still be available through external calls, such as a transaction or anything invoked via an RPC.

[Arachnid]

Wouldn't it make more sense to have CALL copy the return value, though? Does Solidity currently make any assumptions about the return memory area in case of a failed call?

[chriseth]

@Arachnid if revert writes to the same memory area as return then the caller has to be able to distinguish the two cases. Since the ABI does not have an unused bit, I don't see a good way to put the distinction into the data. The only other option (apart from a new call opcode) I see is returning 2 in case of reversion, but that would signal a success condition to older code, which is probably not what you would want.

[Arachnid]

@Arachnid if revert writes to the same memory area as return then the caller has to be able to distinguish the two cases. Since the ABI does not have an unused bit, I don't see a good way to put the distinction into the data.

If CALL returns 0, the data area contains any return data from a REVERT. If it returns 1, the data area contains any return data from a normal return. Both have the issue that you can't directly determine the return data length, but that's orthogonal.

[chriseth]

That's not correct: If CALL returns 0, the data area either contains the data from the revert or whatever was there before (because it was an OOG).

[Arachnid]

Right; just like if call returns 1 - the data area either contains the return data, or whatever was there before (if the return length was 0).

[chriseth]

@Arachnid the difference is that currently, with contracts adhering to the ABI, you always know the type of the contents from the type of the contract.

[Arachnid]

@chriseth That could still be the case - just extend the ABI spec to specify an ABI definition for error returns, too. The default would be 'null'.

[VoR0220]

@chriseth there is a good way around this by just creating a new type in the ABI, that type being the "error" type. I fully agree with @Arachnid here. An error type has been sorely needed and missed here and has created much confusion for newcomers as to why the only error they ever get with their solidity code is "invalid jump dest.".

[chriseth]

@Arachnid @VoR0220 of course that can be done but we still need a flag to indicate whether the data is an error or regular return data.

[Arachnid]

@chriseth But we already have that flag - the return code of CALL. If it's zero, the return data - if any - follows the error ABI. If it's nonzero, the return data - if any - follows the return ABI.

[chriseth]

@Arachnid sure, but we have three situations instead of two: OOG, revert, success. There is currently no way to distinguish between OOG and revert, so new code that wants to read the revert error code has to clear the output area beforehand. In any case, it is really messy and complicated and thus I would propose to move the feature of actually reading the error data (not supplying it) to a new call opcode.

[Arachnid]

But what I'm saying is that the issue of "was any data returned" is orthogonal to returning the data; we already have this problem (and the need to zero out the data before calling) with regular returns, where the called contract could return 0 bytes of data and the caller wouldn't know.

It makes far more sense to me to just return error data using the existing mechanism, and then add a 'RETURNDATALEN' (or similar) opcode afterwards.

[chriseth]

Sure, but I would say that something like "returndatalen" should be a feature of the call opcode. I don't think it is a nice design if you can check the size of the returned data at any time after the call. This is another word of state the VM (and all related tools) have to keep.

[Arachnid]

I agree - a better CALL opcode would be an improvement. But I don't see why that means that CALL shouldn't return error data with the introduction of REVERT. We shouldn't make it impossible just because the current design is not ideal.

[vbuterin]

Looks good to me.

[Arachnid]

Update post all-core-devs meeting: The consensus was that REVERT should take the same arguments as RETURN, and CALL should return data provided to REVERT, with another EIP to be opened on improving handling of return data from CALL and other related opcodes.

[pirapira]

@frangio you are correct with how you interpret the section. I'll leave a comment there.

[pirapira]

This is not true if those contracts contain 0xfd as an instruction.

[pirapira]

I'll add unless they contain 0xfd as an instruction after a week of silence.

[frangio]

Okay, that is a pretty important difference! IIRC Solidity had been emitting 0xfd opcodes several months before Byzantium, so there were many contracts out there whose behavior changed after being created.

[maurelian]

1 week of silence. :)

[pirapira]

Added.

[pirapira]

I think this EIP should say

If block.number >= BYZANTIUM_FORK_BLKNUM

I wait for another week and apply the change.

[pirapira]

Looks good to me.

[cleanunicorn]

If REVERT is called with a message where is this returned?
Will this be returned in eth_getTransactionReceipt for that transaction?

[pirapira]

The message is returned to the caller if the caller is also a program in EVM.

Transaction receipts do not contain the REVERTed messages (that was once considered).