On 10/2/25 12:16 PM, Robin Dapp wrote:
To summarize what we discussed at the Cauldron:
Even though I continue to be unhappy with the fact that we split at LMUL8
regardless of -mrvv-max-lmul, there is no way around this given the ABI.
As the ABI is opt-in we can still change our minds in the future by adding
yet another ABI variant ;) For the time being and with LLVM having implemented
the variant discussed here for over a year, Kito and I agreed to move forward
with the patch as-is. Implementation details (like removing vls_mode_valid_p
from riscv-vector-switch.def) can be changed later if needed but let's rather
have the ABI now than later.
Given it's ABI, opt-in and matches LLVM's behavior I agree that we
really don't have a choice here.
We probably want to add a nugget to the -mrvv-max-lmul documentation
that touches on this subject so that this quirk is documented.
jeff
