On Wed, 12 Dec 2018, Wilco Dijkstra wrote: > I've not seen such an alternative implementation (-fno-trampolines is > ignored on all targets I tried), but it wouldn't affect the ABI since you can > only take the address of a nested function when you're the parent function.
Regarding the trampoline-code-on-executable-stack variant of nested functions, don't miss one fine detail: While the choice of static-chain register does not affect the ABI, it's the other way round: the choice of static-chain register matters, specifically it's call-clobberedness. The sublime port (to invent a soundalike term for port maintainership alluding to unused terminology) is recommended to pick a call-clobbered register that isn't used to carry any odd special part of a function argument (including but not restricted to regular function arguments and the struct-value-regnum). It looks like the current aarch64 static-chain register R18 is call-saved but without special provisions to save and restore the static chain register, i.e. the port is broken wrt. trampolines but may appear to work (likely as-if you got the call-clobberedness wrong for a special case; I haven't investigated). I understand the i386 port gets this right. The CRIS port does not, but attempts and adds another bug (you can't use the trampoline as a register-save area on return). So, changing from R18 to R11 for aarch64 seems right, as the latter is call-clobbered and the former is call-saved IIUC. Unless of course I and grep both miss something or other. brgds, H-P
