http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46815
--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-06 15:39:51 UTC --- Created attachment 22660 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22660 gcc46-pr46815.patch Ugh. So what happens is that we set DECL_VALUE_EXPR of the NRV optimized var to RESULT_DECL, but then in cp_genericize as it is addressable change the RESULT_DECL into DECL_BY_REFERENCE decl and convert_from_reference whenever it is seen in the stmts (but don't adjust the DECL_VALUE_EXPR of the var). Which means we end up using DW_OP_reg3 instead of DW_OP_breg3 0. Either we can do something like this patch, i.e. replace DECL_VALUE_EXPR during genericization, or we could do something similar, but faster by saving pointer to the NRV optimized VAR_DECL somewhere in the lang specific stuff for the FUNCTION_DECL, or, if we can be 100% sure by finalize_nrv time that non-addressable RESULT_DECL isn't ever going to turn into TREE_ADDRESSABLE RESULT_DECL, we could assume early what cp_genericize will do and set DECL_VALUE_EXPR in finalize_nrv to INDIRECT_REF of the RESULT_DECL (which would mean somewhat invalid tree until cp_genericize is run, but perhaps nothing would notice), or we could teach dwarf2out.c if it uses DECL_VALUE_EXPR and it points to DECL_BY_REFERENCE decl that it should add indirection automatically (but then, the question is what all will actually break by that). Jason, any preferences?
