On 13/8/5 下午10:24, Mike Stump wrote: > On Aug 5, 2013, at 7:15 AM, Chung-Lin Tang <clt...@codesourcery.com> wrote: >> On 13/8/5 10:06 PM, Mike Stump wrote: >>> On Aug 4, 2013, at 8:14 AM, Chung-Lin Tang <clt...@codesourcery.com> wrote: >>>> On 13/7/15 1:43 AM, Diego Novillo wrote: >>>>> Could you please repost the patch with its description? This thread >>>>> is sufficiently old and noisy that I'm not even sure what the patch >>>>> does nor why. >>>> >>>> Taking the same example in my first post: >>>> >>>> Under -fPIC, the code in rtlanal.c:nonzero_address_p() does not properly >>>> recognize the "PIC-reg + <constant>" form of load as a weak symbol; it >>>> returns 'true' immediately after seeing the pic-reg indexing, and does >>>> not test the wrapped symbol for DECL_WEAK. >>> >>> So, I can't help but think that others would say that looking into an >>> unspec is by nature, the wrong way to do it, unless that code is in the >>> port. >>> >>> I think the followup from Bernhard points to a better solution, though the >>> wording in the comment was objectionable. Merely say that the symbol, if >>> weak and not defined, is then not local. >> >> When I last tested that patch which moves the DECL_WEAK check, the >> testcases for C++ TLS wrappers fail. I don't remember the fine details, >> but effectively it filters out the TLS wrappers from being treated >> locally, causing them to be called through @PLT, and regressing on some >> tests specifically checking for that… > > Hum… I wonder if there is a TLS predicate one can mix in to the check that > is appropriate. > >> The UNSPEC interpretation here is fairly restricted, FWIW. Earlier talk >> on this thread also mentioned that maybe specific RTL constructs for >> reasoning about PIC addresses should be introduced, rather than common >> idiomatic pattern, though that may be a long shot for now. > > specifying the unspecified, make is specified, and calling it unspec, would > seem to be wrong. > > The right approach, long term, is to have address forms all specified and > documented and a port merely can use the forms they are interested in. pic > is one of those things that should be bumped up, unspec is kinda silly.
Yes, that's what I meant. e.g. instead of using (const (unspec...)), new RTL operators for specifying the common forms of relocations (including those used PIC) should be defined; this will involve changing lots of backends, so should be aimed in the long term. Chung-Lin
