search for: alias_set_subset_of_p

...;t needed is because I thought the >> relation `alias` was symmetric: if x aliases y then y alias x in all cases. >> Is this the case for tbaa? > > It should not be, but LLVM may have done this :) > See the difference in GCC's alias.c between alias_sets_conflict_p and > alias_set_subset_of_p. > > You can see where the subset (one-way checking) is used in tree-ssa-alias.c > > If LLVM does always use this rule, and plans on keeping it that way, > you are correct that there is no point in having a directed graph at > all, since you aren't using the direction. That...

Hi all, A while ago there was a discussion on changing the current "set of trees" representation of TBAA metadata to be more expressive, prompted by the need to support C structs. Dan Gohman also talked about the issue here: http://llvm.org/devmtg/2012-11/Gohman-AliasAnalysis.pdf. It was suggested that the trees be replaced by a type DAG then. While working on this compiler