https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93385
--- Comment #13 from Martin Jambor <jamborm at gcc dot gnu.org> --- The assumption is that DCE will eventually remove all gimple assignments with LHS which has no uses if gimple_has_volatile_ops returns false for them. I have had only a brief look at DCE today and it seems to use gimple_has_side_effects instead, which for assignments looks equivalent but perhaps IPA-SRA should use that for the sake of consistency. I thought this assumption was conservative enough, but perhaps I am too optimistic. Scheduling a DCE before IPA-SRA summary building would not help in the general case because IPA-SRA is optimistic about call argument uses. If all calculations end up being used only as a call argument and the pass finds out in the interprocedural phase that these arguments are going to be removed, it would then also remove the parameter in question - and without subsequent DCE the calculations would not be removed again. Adding this kind of DCE to the modification phase of IPA-SRA is of course an option - we'll have to plug it into tree-inline.c somewhere, making it a tiny bit more complicated again though.
