Simplify Zicsr AST

[nwf]

Split out from and sitting atop #617, the last commit here is the salient one. This transforms a bool * regidx under one ast data constructor into two ast data constructors, allowing us to stop punning between regidx and bits(5). This will be significant in light of #617 for supporting the E extension, where regidx will be (isomorphic to) bits(4); see #646.

[github-actions]

Test Results

396 tests  ±0   396 ✅ ±0   0s ⏱️ ±0s
  4 suites ±0     0 💤 ±0 
  1 files   ±0     0 ❌ ±0 

Results for commit 35eb511. ± Comparison against base commit 59c852c.

♻️ This comment has been updated with latest results.

[Timmmm]

Ah I understand the motivation now. In order to avoid losing the is_imm information could we use a union instead?

union RegOrImm = {
  Reg : regidx,
  Imm : bits(5),
}

[nwf]

I am... reasonably certain we haven't lost any information and that I haven't changed semantics. Rather than pass in a union and conditionally call X(), I just moved the discrimination into ast itself and had the appropriate arm unconditionally call X(). Can you say why you think a union is beneficial?

[arichardson]

I like the split into two separate AST clauses here but I think this changes the behaviour to no longer be correct.

[Timmmm]

Can you say why you think a union is beneficial?

Hey, sorry I forgot to reply to this. There are two cases I know of where knowing is_imm is needed:

1. CHERI CLEN CSRs can only have their metadata written CSRRW:

function handle_csr_op(csr, rs1, rd, is_imm, op, isWrite, isRead) = {
  if is_clen_csr(csr) then {
    // The only way to write metadata is with `csrrw`.
    // For extended CSRs, it also has to be in CapPtrMode mode.
    // Note `op == CSRRW` means `csrrw` or` csrrwi` which is why we also check `not(is_imm)`.
    let clen_write = (effective_cheri_mode() == CapPtrMode | csr_type(csr) == ClenNew) & op == CSRRW & not(is_imm);
    let write_val = if is_imm then {null_cap with address = zero_extend(rs1)}
      else if clen_write then C(rs1)
      else {null_cap with address = X(rs1)};
...

2. Some CLIC CSRs have weird behaviour that depends on is_imm. At least xnxti:

function access_xnxti(
  csr : csreg,
  rs1_val : xlenbits,
  op : csrop,
  is_imm : bool,
  isWrite : bool,
  priv : Privilege) -> xlenbits = {
...
    let min_pil = if is_imm then xcause[Xpil] else rs1_val[23..16];

(I haven't upstreamed that code yet; CLIC is a mess and I'm not sure how much we should want it in this repo... but it would be nice if we can still easily add it on.)

Neither of those bits of code are actually in this repo yet though, so maybe it is unfair to worry about them. I guess we could always add is_imm back as a parameter to doCSR when it is needed?