PR ipa/59775 (get_binfo_at_offset not handling virtual inheritance well)

Jan Hubicka hubicka@ucw.cz
Wed Jan 15 21:27:00 GMT 2014


> Hi,
> this patch fixes ICE in ipa-devirt that is caused by get_binfo_at_offset
> reporting NULL for a valid query.  This is because how virtual inheritance
> is represented.
> 
> Here we have chain A<-B<-C where A is a virtual base.  We look for A within
> C that is there at offset 64.  On this query get_binfo_at_offset walks
> fields of C and it finds A, it matches it and it looks for BINFO. The
> catch is that BINFO represents only direct bases and A is not direct base.
> Conseuqentely it returns NULL. In the past this only prevented us from
> devirtualizing here, now we ICE, since ipa-devirt depends on fact that
> it can resolve all possible queries. get_binfo_at_offset needs to make a hop
> through B.
> 
> This patch is kind of minimal change needed to get get_binfo_at_offset to
> look into B: when base is not found, it finds a base that contains A
> and dives into it.
> 
> I plan to rewrite the function for next stage1: the walk of fields is already
> done by ipa-devirt separately, so it really should do only BINFO walk.  For
> that we however need to retire other uses of get_binfo_at_offset that are still
> used by ipa-devirt.  That is on the TODO list to switch it to be basically a

	  ^^^ ipa-cp, that is the only place still ging through
	gimple_extract_devirt_binfo_from_cst that works in a way so
	get_binfo_at_offset needs to dive into fields.

> propagation engine for ipa-devirt's polymorphic_call_context structures.
> (basically I want to have one local pass doing non-trivial propagation and
> one IPA pass propagation across function boundaries, both sharing
> polymorphic_call_context structure and a lattice operations)



More information about the Gcc-patches mailing list