http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51782
Martin Jambor <jamborm at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rguenth at gcc dot gnu.org
--- Comment #16 from Martin Jambor <jamborm at gcc dot gnu.org> 2012-02-17
17:53:34 UTC ---
OK, I clearly misremembered where the address space information is
supposed to be stored. Richi's patch in comment #12 clearly shows
it's the address operand of MEM_REF should carry that information, so
please disregard my previous comment.
Nevertheless, during expansion of a COMPONENT_REF or any
handled_component we do rely on its type's TYPE_ADDR_SPACE because
expand_expr_real_1 calls set_mem_attributes with the whole reference
tree to deduce the attributes from and it uses the address space of
its type.
If only the base object or pointer of a reference is supposed to hold
the address space information, then this function needs to be changed
to look at the base object. I'll attach an untested patch straight
away.
Alternatively, we could alter types o references built by SRA at each
step (in build_ref_for_offset and build_ref_for_model) which also
works (I also have a patch) but would be rather inconsistent with how
we expand MEM_REFs (but front-end generated stuff looks like this).
