Support dumping memory

[Timmmm]

Once #1489 is completed we should augment it to allow dumping memory at the end of a test or at the end of an RVVI-DII run. Ideally we will dump to ELF format and set the entry point to the first instruction that was executed.

This is intended to allow generating ELFs from TestRig runs that can be run in lockstep without using DII (i.e. so the fetch unit is exercised).

[jrtc27]

Instruction streams from RVFI-DII can end up using different instructions for the same PC over time; DII just feeds in a sequence of instructions without regard for what the PC is. You can't represent that in a memory image.

[Timmmm]

Yeah we're aware. @nwf 's idea was that we just ignore that and hope that the ELF you get from dumping memory does some interesting things anyway, even if it doesn't exactly match what TestRig actually executed the first time.

If it doesn't work we have some other ideas, but it's very easy to try it so it seemed worth a shot!

[nwf]

To be clear, this is a conjecture: for a DII stream dii, which TestRIG would feed into, roughly, sail(dii) == dut(dii), it might still be useful to evaluate, instead, let mem = f(dii) in sail(mem) == dut(mem) for some processing function f. It happens that "the last value observed for each address in the sail trace of the execution of dii" is one such function.