Add B extension (Zba, Zbb, Zbs) #29
Conversation
- Update `insn_shimm` and `insn_alu` to support unsigned words (`uwmode`). This is used when the result is of width XLEN but the lower 32 bits from one of the inputs (rs1) is zero-extended. - Update `isa_propagate` to match multi-letter extensions e.g. `Zba`. - Add Zba instructions and the beginnings of the rest of the B extension.
- Add B-type format. - Implement (most) Zbb instructions. - Add `insnt_count` and `insn_ext` for bit count and extension instructions. - Be unconvinced that the count checks will work as intended. - Be confused about `orc_b` and `rev8` instructions and leave them commented out for now.
Everything but `clz` `ctz` and `cpop` should work.
CLZ, CTZ, CPOP, ORC.B, and REV8.
Having the csr_spec option enables CSR testing, including the misa CSR which uses the define `RISCV_FORMAL_CSR_MISA`. If the CSR_MISA flag is defined then instructions from extensions will check the misa CSR includes the bit flag for the instruction. We don't want that, since NERV misa is all zeros.
| isa_propagate_pair("rv32i"+ext, "rv32ib") | ||
| isa_propagate_pair("rv64i"+ext, "rv64ib") | ||
| isa_propagate_pair("rv32ib", "rv64ib") | ||
|
|
There was a problem hiding this comment.
It would be nice to generate all variants of extensions here, you can do something like:
isa_suffixes = [
("m", []),
("c", []),
("b", ["Zba", "Zbb", "Zbs"])
]
for suffix, sub_suffixes in isa_suffixes:
for sub_suffix in sub_suffixes:
isa_propagate_pair("rv32i" + sub_suffix, "rv32i" + suffix)
isa_propagate_pair("rv64i" + sub_suffix, "rv64i" + suffix)
def propagate_isas(parts=isa_suffixes, suffix=""):
if not parts:
isa_propagate(suffix)
else:
propagate_isas(parts[1:], suffix)
propagate_isas(parts[1:], suffix + parts[0][0])
for part in parts[0][1]:
propagate_isas(parts[1:], suffix + part)
propagate_isas()There was a problem hiding this comment.
Yeah I was thinking about this too. I think for a generic "any combination of extensions" it would be better to generate them during the core testing setup, rather than producing a file for each and every combination.
There was a problem hiding this comment.
I guess the advantage of having each variation is that it makes it easier to use them in a testbench outside of the genchecks.py setup, like https://github.com/YosysHQ/riscv-formal/blob/7793823fe3c281205093b1c5ac7a3aa772f5e223/cores/picorv32/complete.sby
There was a problem hiding this comment.
Spun this out as a new issue #35 so I can resolve (read: ignore) it here :)
Includes updated NERV demonstrator.