[Bug middle-end/49373] [4.7 Regression] Many testcase failures

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49373

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #16 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-06-24 12:54:39 UTC ---
This is a bug in lto-streamer-out.c.  TYPE_MAXVAL of the ARRAY_TYPE's
TYPE_DOMAIN e.g. in f1 is a VAR_DECL (DECL_NAME is NULL, DECL_ARTIFICIAL, but
!DECL_IGNORED_P).  It is first encountered in a GIMPLE_DEBUG's first operand,
it hasn't been seen yet, so that VAR_DECL is output.  But, later on
write_global_stream calls lto_output_tree on the VLA ARRAY_TYPE, with ref_p set
to false, and this eventually calls lto_output_tree on that VAR_DECL again.
But this doesn't find it in ob->writer_cache, because the cache has been
destroyed in between, from output_function -> destroy_output_block ->
lto_streamer_cache_delete.
I guess VLA types need to be emitted immediately in the containing function,
rather than being deferred to be written much later on globally.