> The rest of the GCN hook is quite inconsistent, it says misalignment > == -1 is OK, is_packed > is not and then the above ... and gather-scatter is also always OK > (even if the scalar accesses > are 'packed' aka not naturally aligned?)
The gather_scater -> ok I added during the introduction of is_gather_scatter. That was to keep the old behavior where we didn't check the alignment at all. > Note this leaves unspecified whether a known misalignment means the > original scalar memory access _is_ naturally aligned? I think it needs > not, but I'd have to check alignment analysis if we force the result to -1 > in such case. We can just set is_packed every time? And adjust the targets that unconditionally return false when is_packed to misalignment == -1 && is_packed in order to keep the current behavior? > I think most is_packed uses think of alignment according to size. > > Note that we only communicate misalignment to the hook, not alignment > (that's implicit, alignment of the passed mode). But that means for the > x86 'double' the misalignment with respect to the vector alignment is unknown > but we might know the access is aligned to 4 bytes. So is_packed is basically > the poor-mans way to communicate this, and we should see to define that > with respect to alignment according to size? -- Regards Robin
