OK. It have been changed to sign-extended in the "processing of Narrower
Values" section of the new spec(version 1.0.0.rc) . I'll fix this and
update other nan-boxing processing in current implementation.
在 2021/12/26 上午9:42, liweiwei 写道:
Sorry. In the old spec(version 0.41), nanboxing is not totally
disabled, but "NaN-boxing is limited to |XLEN| bits, not |FLEN| bits".
Taking misa.mxl into acount, if misa.mxl is RV32, and maximum is RV64,
this should be sign-extended. Is there any other new update for
nanboxing to the spec?
在 2021/12/26 上午6:00, Richard Henderson 写道:
On 12/24/21 7:13 PM, liweiwei wrote:
In RV64 case, this should be nan-boxing value( upper bits are all
ones). However, zfinx will not check nan-boxing of source, the
upper 32 bits have no effect on the final result. So I think both
zero-extended or sign-extended are OK.
There is no nanboxing in zfinx at all -- values are sign-extended.
r~
