On 05/01/2018 04:13 AM, Alex Bennée wrote: > If we are going to the trouble of adding a read_fp_hreg() we might as > well do the same for the write case.
The write case is much harder. Almost all of the time we have a helper function that has already zero-extended the result from 16 to 32-bits, and therefore write_fp_sreg is perfectly fine. In more than a few cases, this writeback is shared by both single and half precision paths and so adjusting the writeback is not appropriate. There are in fact zero instances at this point in the tree that do require extra zero-extension. So I'm going to leave this alone. r~