[llvm-dev] RFC: Remove inaccessiblememonly from 3.8 branch

Philip Reames via llvm-dev llvm-dev at lists.llvm.org
Wed Feb 10 13:12:57 PST 2016


A while back, we introduced new attributes to model functions which read 
and wrote from locations not otherwise visible in the IR.  The 
motivation was to enable much more aggressive aliasing around such calls.

Since that time, we've reverted the original patch and AFAIK we have no 
transformation or analysis passes in tree which use this information.  
Given this, I don't think it's really fair to say the attribute has 
baked at all.  We have no reason to believe that we picked the right 
semantics, and last I remember, there was some legitimate question about 
that.  Worse, AFAIK, all active development has stopped.  I don't know 
of any active reviews for this attribute.

These attributes have not yet shipped in any LLVM release, but this is 
about to change in 3.8.  Once that changes, we will have accepted a 
backwards compatibility guarantee with bitcode which is potentially 
concerning for a new attribute in such a half developed state.

Given the preceding, I'd like to propose that we remove the 
inaccessiblememonly and inaccessiblemem_or_argmemonly attributes from 
the 3.8 release branch.  This could be done by either a) actually 
removing the code, or b) simply disabling it.  If we remove the 
documentation from the LangRef, disable the LL and bitcode parsing for 
the attribute, and explicitly document that we are not reserving the 
attribute index numbers, I think that prevents any assumption of forward 
compatibility.  Another option would be to mark them experimental, which 
would have effectively the same effect.

Philip

The original GlobalsAA:
http://reviews.llvm.org/D15605

The revert:
http://reviews.llvm.org/D15919

The last active revision I know of: (Abandoned Jan 6)
http://reviews.llvm.org/D15665


More information about the llvm-dev mailing list