On Tue, Aug 28, 2018 at 7:59 AM Alexander Monakov <amona...@ispras.ru> wrote: > > > > > So - how difficult is it to fix BRIG to not use MULT_HIGHPART_EXPR if > > > > not supported? > > Richard, how should we proceed from here? Do you like the solution in the > initial mail, or would you prefer something else? FWIW I agree with Pekka, > no need to burden BRIG FE with expanding mul-highpart.
I think that if we want to support MULT_HIGHPART generally then we need to support it for all modes - what's your plan for 32bit targets here? This means providing libgcc fallback implementations for modes we cannot directly expand to. I'd also like to see better support for MULT_HIGHPART then, for example verify_gimple_assign_binary does nothing (TODO). As for the expansion code I wonder if handling it in expand_binop would be better? Do we want to expose highpart multiply to users somehow? Thanks, Richard. > > > Pekka, can you comment? I think you have fallback paths for vector types > > > only at the moment? > > > > I think it involves pretty much moving the code of your patch to the > > BRIG frontend. > > To me it'd look a bit wrong in terms of "division of responsibilities" > > as this is not really > > frontend-specific as far as I understand (even if BRIG/HSAIL happened > > to be the only language > > supporting them currently). > > > > > Does BRIG have mult-highpart for 64-bit integers? On 32-bit targets we > > > won't be able to easily expand them (but on 64-bit targets it is fine). > > > > Yes it does. 32b targets are not high priority for BRIG FE at this > > point, so I wouldn't > > worry about this as we assume HSA's "large" model is used, so this is > > likely not > > the only problem when trying to generate for 32b machines. > > > > > For scalar types I think we should prefer to implement a generic expansion > > > rather than have the frontend query the backend. For vector types I am not > > > sure. > > > > In my relatively gcc-uneducated humble opinion these both belong more > > naturally to a > > target-specific expansion or "legalization" pass, which tries to > > convert tree nodes to what the target > > can support. > > > > BR, > > Pekka > >
