[cfp-interest 4019] Re: WG14 meeting next week - New CFP potential items and my proposed responses
Joshua Cranmer
joshua.cranmer at intel.com
Fri Nov 13 14:04:30 PST 2026
On 11/13/2026 16:48, RAJAN BHAKTA wrote:
> *1062*: Intent is to allow it if the implementation wants
> (efficiency). (SNaNs cause exceptions on assignment?)
> Question 1: Implementation defined should still take precedence.
> This is the most likely case due to efficiency. Most implementations
> likely do byte-wise or chunk-wise copy on hardware so it would be a
> burden to require tracking contents and triggering the signal. Matches
> existing behaviour as well for implementations that I know of.
> Question 2: Like question 1. Most likely will be like a trap
> representation. Signals should only occur only on a read, not assignment.
> Question 3: No standard library function should produce a SNaN
> other than ones that directly modify it (like setting payloads).
> Uninitialized objects being a SNaN (Ex. malloc'd storage) and then
> assigned should signal if a read is done just like trap representations.
The notes I wrote down for this case summarizes as follows:
This basically asks if struct/union copies are |memcpy| or assign of
individual struct fields according to the abstract machine. I think the
abstract machine leans to the latter, in which case yeah, you can get FP
exceptions. Note that you definitely can get this on bare structs in
function calls/returns. Standard library functions probably can't
produce sNaN except under specific conditions. varargs can signal FP
exceptions for float args, although the exact place it gets signaled is
:shrug:
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20261113/cde8d876/attachment-0001.htm>
More information about the cfp-interest
mailing list