On 15/12/14 11:44, Florian Weimer wrote: > On 12/15/2014 12:39 PM, Richard Earnshaw wrote: >> On 12/12/14 23:50, Florian Weimer wrote: >>> On 12/12/2014 08:02 PM, Jeff Law wrote: >>> >>>> It'd probably work for the specific cases Florian wants to tackle, but >>>> I'd consider it highly non-portable. >>> >>> I agree, that's why I asked for something “GCC-portable”. >>> >> >> Don't go this way. Seriously, you can't assume that a future platform >> ABI won't break if you do. ABI's at this level really aren't controlled >> by GCC, which just has to conform to whatever the platform ABI requires. >> We were very close in AArch64 to defining the ABI for variadics to push >> all variadic (from the ... onwards) arguments onto the stack - it leads >> to a very simple ABI from an implementation viewpoint and eliminates >> much of the grotty code around reconstructing the parameter list. > > This would still work except in a very pathological corner case which > cannot happen with glibc (the stack cannot be empty when the open > function is called because all user code runs with a non-empty stack). > > Did you shy away from the simple implementation because of too many > interop failures? > We certainly ran into some problems with non-conforming code. Not least of which was GCC itself :-( - though I think that particular issue has been fixed now. It certainly created enough uncertainty as to the depth of the rabbit hole that we decided not to pursue that avenue. R.
