Originally Gitea issue #81 by @wschmidt-ibm on 2021-10-19
We need to properly document that the semantics for ties are different for vector float and vector double arguments. For vector float, the semantics are round-to-nearest, ties-to-even. For vector double, the semantics are round-to-nearest, ties-away-from-zero. The reasons these semantics differ are somewhat lost in the mists of time, but they need to be documented.
Changing the semantics so they are consistent would be possible, but might break working code in the field, and is likely to perform poorly on several processor implementations. So the compiler leadership team has determined to simply make sure the documentation is correct.
We need to properly document that the semantics for ties are different for vector float and vector double arguments. For vector float, the semantics are round-to-nearest, ties-to-even. For vector double, the semantics are round-to-nearest, ties-away-from-zero. The reasons these semantics differ are somewhat lost in the mists of time, but they need to be documented.
Changing the semantics so they are consistent would be possible, but might break working code in the field, and is likely to perform poorly on several processor implementations. So the compiler leadership team has determined to simply make sure the documentation is correct.