[PATCH] D31885: Remove TBAA information from LValues representing union members
John McCall via Phabricator via cfe-commits
cfe-commits at lists.llvm.org
Mon May 15 14:01:17 PDT 2017
rjmccall added a comment.
In https://reviews.llvm.org/D31885#755456, @kparzysz wrote:
> In https://reviews.llvm.org/D31885#751164, @rjmccall wrote:
>
> > The right fix is probably just to make sure that EmitLValueForField doesn't add TBAA information when the base LValue doesn't have it. That will also fix problems with recursive member access where the base is may_alias.
>
>
> That won't work for arrays that are member of unions. We'd need to change EmitArraySubscriptExpr, which at some point calls MakeAddrLValue for the type being loaded/stored. MakeAddrLValue adds TBAAInfo for that type (unsigned short in case of the C testcase) and it does so regardless of any struct paths. This TBAA info must be removed for the testcase to be fixed. At that point we would have to examine the base expression to see if it refers to a union member, which is not different from the approach in this review. The LValue construction is done in a piecewise manner where some elements of the previously (recursively) created LValues may or may not be preserved. This makes me concerned that any more detailed attempt at fixing this may accidentally omit some case.
That is a reasonable concern.
Hmm. It seems to me that we do already propagate an important semantic property from the base l-value in every important case: the alignment source. That audit has already been done, and all you really need to do do is generalize it: look at all the places that generically propagate an AlignmentSource, e.g. EmitPointerWithAlignment and almost everywhere that calls LValue::getAlignmentSource(), and make them propagate an LValueBaseInfo (or whatever) instead. You can then make sure that may_alias-ness is stored in an LValueBaseInfo along with the alignment source, and then suppress TBAA when a base info is known to be may_alias, and then your temporary fix is just to make all union accesses unconditionally set may_alias to true.
Repository:
rL LLVM
https://reviews.llvm.org/D31885
More information about the cfe-commits
mailing list