Add bool for writing to destination reg to RVFI output #61
Conversation
| @@ -783,7 +783,7 @@ function clause execute (CClear(q, m)) = { | |||
| if q_u == 0 & i == 0 then | |||
| DDC = null_cap | |||
| else | |||
| C(8 * q_u + i) = null_cap; | |||
| wC (false, 8 * q_u + i, null_cap); | |||
There was a problem hiding this comment.
This is the bit that will show in the ISA doc, and the no-context false may be confusing. I wonder if there's a clearer way, for example wrapping wc(false, ....) in a function. Not sure what to call that function though, since RVFI isn't really visible from the instruction-level description.
|
Why not have it auto detect when it’s been called twice? |
I thought about that, but what if you call it with an immediate where only one bit is set? |
|
Then it reports a single register write, which is accurate? |
|
Hmm, I guess that's one way of looking at it. It might be a little tricky to have Toooba detect that case, although if we're not hoping to synthesise with RVFI there's probably an easy enough hacky way to do it. |
In sail, when an instruction writes to a register, the register number will also be written to the
rvfi_rd_addrfield (if it has been built for RVFI output). When an instruction has no destination register, thervfi_rd_addrremains zero. The CClear instruction in the CHERI is an exception to the regular RISC-V instruction format. It can write to up to eight registers. However, CClear has no destination register, but rather a two-bit immediate determining which quarter of the register file this instruction will write to.Currently, the sail implementation for CClear writes the highest register number to
rvfi_rd_addr. This PR changes this behaviour and never writes a destination register to the field, which results in the field containing zeros. I think this is the correct behaviour for the model as thervfi_rd_addrfield cannot correctly represent the destination register behaviour of CClear.