On Mon, Sep 10, 2018 at 7:07 PM, Julian Brown <jul...@codesourcery.com> wrote: > On Mon, 10 Sep 2018 10:52:47 -0700 > Cesar Philippidis <ce...@codesourcery.com> wrote: > >> On 09/10/2018 10:37 AM, Jason Merrill wrote: >> > On Mon, Sep 10, 2018 at 4:05 AM, Julian Brown >> > <jul...@codesourcery.com> wrote: >> >> This patch (by Cesar) changes the way C++ references are mapped in >> >> OpenACC regions, fixing an ICE in the non-scalar-data.C testcase. >> >> >> >> Post-patch, references are mapped like this (from the omplower >> >> dump): >> >> >> >> map(force_present:*x [len: 4]) map(firstprivate ref:x [pointer >> >> assign, bias: 0]) >> >> >> >> Tested with offloading to NVPTX and bootstrapped. OK for trunk? >> >> >> >> Thanks, >> >> >> >> Julian >> >> >> >> ChangeLog >> >> >> >> 2018-09-09 Cesar Philippidis <ce...@codesourcery.com> >> >> Julian Brown <jul...@codesourcery.com> >> >> >> >> PR middle-end/86336 >> >> >> >> (gimplify_adjust_omp_clauses_1): Update handling of >> >> mapping of C++ references. >> > >> > How is reference handling specified differently between OpenMP and >> > OpenACC? It seems strange for them to differ. >> >> Both OpenACC and OpenMP privatize mapped array pointers on the >> accelerator for subarrays in the same way. However, for pointers >> without subarrays, OpenMP treats them as zero-length arrays, whereas >> OpenACC treats them as ordinary scalars so that the pointer target >> will not get remapped on the accelerator (which is odd because >> there's a deviceptr clause for that). Scalars in C++ are special, >> because references must treated like an array of length one, for lack >> of a better terminology. > > I think it's more accurate to say that OpenACC says nothing about C++ > references at all, nor about how unadorned pointers are mapped in > copy/copyin/copyout clauses. So arguably we get to choose whatever we > want, preferably based on the principle of least surprise. (ICE'ing > definitely counts as a surprise!) > > As noted in a previous email, PGI seems to treat pointers to > aggregates specially, mapping them as ptr[0:1], but it's unclear if the > same is true for pointers to scalars with their compiler. Neither > behaviour seems to be standard-mandated, but this patch extends the > idea to references to scalars nonetheless.
That certainly seems like the most sensible way of handling references to non-arrays. And the 'this' pointer, incidentally. Should we not do the same for OpenMP? Jakub? Jason
