From @holiman:

It's here: ethereum/tests#249 (comment) . But I'll add my tldr; (which also became quite long):

Input

The evm should take the following inputs:

• --code <code> - code to be executed.
• --codeFile <file> - file containing code to be executed. Sometimes really large chunks of input cannot be passed through bash.
• --gas <int>
• --price <int>
• --sender <address> - address of ORIGIN
• --receiver <address - address of ADDRESS
• --input <code> : CALLDATA
• --value <int>
• --json - boolean flag, output json output for each opcode or not (it's useful to disable json when benchmarking)
• --nomemory - disable showing the full memory output for each op
• --create - if specified, it's executed as initcode
• --prestate - a chain specification, the same one that the client normally would use.

Basically, the evm should be able to run things very simply, like so:

#evm --code 6040 --json run
{"pc":0,"op":96,"gas":"0x2540be400","gasCost":"0x3","memory":"0x","memSize":0,"stack":[],"depth":1,"error":null,"opName":"PUSH1"}
{"pc":2,"op":0,"gas":"0x2540be3fd","gasCost":"0x0","memory":"0x","memSize":0,"stack":["0x40"],"depth":1,"error":null,"opName":"STOP"}
{"output":"","gasUsed":"0x3","time":141485}

But it should also be able to reconstruct an actual on-chain transaction, with complex options including prestate, where no code is passed, since it's already been showed into the prestate:

./evm --prestate /home/martin/workspace/evmlab/output//0xd6d519-genesis-geth_wq38zsy5.json --gas 150000 --sender 0x69ea6b31ef305d6b99bb2d4c9d99456fa108b02a --receiver 0xb97048628db6b661d4c2aa833e95dbe1a905b280 --input a9059cbb0000000000000000000000008eef795fd9150f118bddeca556a5a2a2438ab865000000000000000000000000000000000000000000000081ebd8ffd6b2a58000 --json run

Output

The evm should output a json object for each operation. Example:

{"pc":0,"op":96,"gas":"0x2540be400","gasCost":"0x3","memory":"0x","memSize":0,"stack":[],"depth":1,"error":null,"opName":"PUSH1"}

Required: pc, op, gas, stack, depth
Optional: opName, gasCost, error

The stack, memory and memSize are the values before execution of the op.

At the end of execution, some summarical info is good, e.g.

{"output":"","gasUsed":"0x3","time":141485}

When errors occur, geth and parity handles them differently.

Minor changes to how things work is ok, we can handle discrepancies in format and minor quirks.