[Question] fflags for vector floating point non-arithmetic instructions

[liangkaiwang]

Hi,

In section 13.1, I see

13.1. Vector Floating-Point Exception Flags A vector floating-point exception at any active floating-point element sets the standard FP exception flags in the fflags register. Inactive elements do not set FP exception flags.

But does this apply to the instruction that only moves the data from one element to another like vfmerge, vfmv, vfslide1[up/down], etc? As an example, if the input is Signaling NaN, do we raise the invalid flag for the instruction?

thanks, Liang-Kai

[aswaterman]

Moves don't give rise to floating-point operations, in the IEEE 754 sense of the term, and so they don't set any FP exception flags.

(And yes, this no-exception property holds even if NaN boxing causes an operand to be transformed into canonical NaN.)

[liangkaiwang]

Thanks for your quick response. Really appreciate it.

Given that there are some non-intuitive behavior surrounding NaN Boxing and NaN canonicalization, I feel it may be worthwhile to specify them in the spec. But I will leave it up to the committee to decide if it is useful.

If I want to summarize the behavior, for data-move instructions like vfmerge, vfmv, vfslide1up, and vfslide1down, we have

• NaN boxing check needs to be applied when the input is a floating point scalar
• After NaN Boxing check is applied, the data is treated as a bag of bits and data is populated to the vector elements based on the instructions behavior and the elements' status without any modification
• For those data-move instructions, exception flags are never raised under any condition

thanks, Liang-Kai

[aswaterman]

Your summary sounds right. I'll leave it to the authors of the relevant specs to decide if clarification is needed. (IIRC, it's sufficiently specified, but the information is scattered around.)