On 2017-07-11 09:21, Eric Botcazou wrote:
Applied without the undocumented tweaks to the divdf3_fix and sqrtdf2_fix patterns. Why are 2 nops necessary here? The stored value doesn't matter. And the length attribute should be adjusted if nops are added to the pattern.
The first nop was added to prevent sequence A from appearing (store -> fdivd -> std). But as you say, it is not needed as we do not read the value written by the std. The second nop was added to prevent sequence B (std -> store) and this one seems necessary as the value written by the store might be used later. OK to submit a new patch with only the second nop and a correct length attribute?
-- Daniel Cederman Cobham Gaisler