Closed
Description
Tracking issue for the conversion float_bits_conv
feature, added by PR #39271.
The main topic during the PR's thread has been about how to handle signaling NaNs, and the final decision was to silently mask them, especially as LLVM behaviour regarding them is not well specified.
Another issue raised by @dwrensha was that sNaNs are not encoded consistently. IEEE 754-2008 (section 8.2.1 in the latest draft) specifies quiet NaNs to have their first bit of the trailing significand field set to 1, for signaling NaNs it should be set to 0. Some MIPS based platforms have it reversed. Should they be cfg'd out?