On Tue, 2022-03-29 at 13:39 +0200, Janosch Frank wrote: > > I think that won't work because that function might clean its float > > registers in the epilogue and hence destroy the contents. Except if > > you > > have an idea on how to avoid that? I missed that this is a theoretical problem because the function will be inlined anyway. > About that: > > Well, who guarantees you that the compiler won't change a fpr (and > thereby the overlapped vrs) between the vlms here and your infinite > loop > at the end of the function? :-) gcc uses fprs and acrs in the most > interesting places and I've just been hit by that again a few hours > ago. > > I.e. to be safe we'll need to implement the next few lines in > assembly > as well, no? I guess so. I will go ahead and reimplement everything in assembly?
