On Wed, Sep 16, 2020 at 11:22:02AM -0700, Nick Desaulniers wrote: > I looked into this a bit, and IIRC, the issue was that compiler > generated functions aren't very good about keeping track of whether > they should or should not emit framepointer setup/teardown > prolog/epilogs. In LLVM's IR, -fno-omit-frame-pointer gets attached > to every function as a function level attribute. > https://godbolt.org/z/fcn9c6 ("frame-pointer"="all"). > > There were some recent LLVM patches for BTI (arm64) that made some BTI > related command line flags module level attributes, which I thought > was interesting; I was wondering last night if -fno-omit-frame-pointer > and maybe even the level of stack protector should be? I guess LTO > would complicate things; not sure it would be good to merge modules > with different attributes; I'm not sure how that's handled today in > LLVM. > > Basically, when the compiler is synthesizing a new function > definition, it should check whether a frame pointer should be emitted > or not. We could do that today by maybe scanning all other function > definitions for the presence of "frame-pointer"="all" fn attr, > breaking early if we find one, and emitting the frame pointer setup in > that case. Though I guess it's "frame-pointer"="none" otherwise, so > maybe checking any other fn def would be fine; I don't see any C fn > attr's that allow you to keep frame pointers or not. What's tricky is > that the front end flag was resolved much earlier than where this code > gets generated, so it would need to look for traces that the flag ever > existed, which sounds brittle on paper to me.
For code generated by the kernel at runtime, our current (x86) policy is "always use frame pointers for non-leaf functions". A lot of this compiler talk is over my head, but if *non-leaf* generated functions are rare enough then it might be worth considering to just always use frame pointers for them. -- Josh
