https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83590
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Created attachment 43098 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=43098&action=edit gcc8-pr83590.patch Actually, the problem is much more severe. The thing is, neither the TYPE_SIZE_UNIT nor DECL_SIZE_UNIT is INTEGER_CST from the FE, it becomes INTEGER_CST when gimplify_one_sizepos is called on these. First we gimplify the TYPE_SIZE*, then check DECL_SIZE* for not being INTEGER_CST and then we gimplify DECL_SIZE*, at which point it becomes also INTEGER_CST. But we are already determined to treat it as VLA and create DECL_VALUE_EXPR for it. Of course all the OpenMP/OpenACC lowering/expansion code checks for special treatment of VLA by checking either DECL_SIZE* or TYPE_SIZE* or variably_modified*, but none of these are true. So, either we can do what this patch does, or we'd need to introduce some flag whether gimplify_vla_decl has been done on it (parsing DECL_VALUE_EXPR and trying to understand if it is due to being VLA or myriads of other reasons why it has DECL_VALUE_EXPR is too hard). Can you give this a spin?
