On Fri, 27 Oct 2017, Jakub Jelinek wrote: > On Fri, Oct 27, 2017 at 12:31:46PM +0200, Richard Biener wrote: > > I fear it doesn't work at all with LTO (you'll always get the old ABI > > if I read the patch correctly). This is because the function > > computing the size looks at flag_abi_version which isn't saved > > per function / TU. > > > > Similarly you'll never get the ABI warning with LTO (less of a big > > deal of course) because the langhook doesn't reflect things correctly > > either. > > > > So... can we instead compute whether a type is "empty" according > > to the ABI early and store the result in the type (thinking of > > doing this in layout_type?). Similarly set a flag whether to > > warn. Why do you warn from backends / code emission and not > > from the FEs? Is that to avoid warnings for calls that got inlined? > > Maybe the FE could set a flag on the call itself (ok, somewhat > > awkward to funnel through gimple). > > Warning in the FE is too early both because of the inlining, never > emitted functions and because whether an empty struct is passed differently > from the past matters on the backend (whether its psABI says it should be > not passed at all or not). > > Perhaps if empty types are rare enough it could be an artificial attribute > on the type if we can't get a spare bit for that. But computing in the FE > or before free_lang_data and saving on the type whether it is empty or not > seems reasonable to me.
There are 18 unused bits in tree_type_common if we don't want to re-use any. For the warning I first thought of setting TREE_NO_WARNING on it but that bit is used already. OTOH given the "fit" of TREE_NO_WARNING I'd move TYPE_ARTIFICIAL somewhere else. Richard.
