https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80687
--- Comment #3 from Thomas Schwinge <tschwinge at gcc dot gnu.org> --- (In reply to Thomas Koenig from comment #2) > So we need to make this patch conditional on > the compiler supporting VLAs. ACK. > Otherwise, I think we should just go for malloc/free > (to reduce the stack usage anyway). Or would this also > be problematic on nvptx? As malloc/free are supported on nvptx, falling back to these will resolve the compilation problem. And it will then be our problem ;-) to get the performance right. (As necessary; PR65181 or something else -- curious if there are optimization possibilities, when the size of the memory region can actually be determined during compilation (?), can the compiler move this into an sufficiently-sized PTX ".local" region, avoiding the malloc/free calls?) > I'll look at this tonight. Thanks! I'll be available for testing any patches.