*PING* On 6/24/20 7:32 PM, Tobias Burnus wrote:
(OpenMP 5 extends this a lot, but this is about OpenMP 4.5. It touches code which is also used by OpenACC's attach/detach.)@OpenACC/Julian: I think the character attach/detach for deferred-length strings does not work properly with OpenACC; I did not touch this code – but I think it needs some love. This code adds support for map(dt%comp, dt%comp2) where "comp" can be either a nonpointer, nonallocatable element scalar, array or array section. Or it can be a pointer - where character strings are one complication as for deferred-length ones, the length is stored in an extra DT component. While testing, I encountered two bugs, one relating to kind=4 character string (patch pending review; PR95837) not part of testcase) and one related to deferred-length character strings (commented in the test case; larger issue; PR95868). Like always, some more tests/testcase probably would not harm. Regarding the patch: (a) openmp.c: This enabled component matching for 'map(' and piggybacks on the OpenACC code for the checks. I think that some additional checks might be useful – and I hope that no check is too strict. The "depend" clause was excluded as one otherwise gets a testsuite fails due to the is-contiguous check. (b) trans-openmp.c: - gfc_trans_omp_clauses now has a "bool openacc". - GOMP_MAP_ATTACH_DETACH is replaced by GOMP_MAP_ALWAYS_POINTER - For arrays, the mapping of the descriptor is squeezed before "node" which contains the data transfer (var.desc.data mapping followed by the always_pointer for the mapping). In this array case, the latter gets a pointless cast in order to prevent that for both var.desc and var.desc.data memory gets allocated in the struct. → That's also the reason the big switch table is moved up. - For deferred-length strings, the string-length is in an extra struct element (derived-type component) and will be mapped in addition. - Bugs in the previous version: * gfc_trans_omp_array_section for "element == true", the size of a pointer instead of the size of the element was mapped. * For string variables (with constant length) the kind=4 was not properly handled. * Allocatable scalars were not handled – missing second clause for the always_pointer (and attach_detach, I assume) Comments, remarks, suggestions? Otherwise: OK for the trunk? Tobias
----------------- Mentor Graphics (Deutschland) GmbH, Arnulfstraße 201, 80634 München / Germany Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Alexander Walter
