On Thu, Aug 10, 2023 at 02:34:04PM +0100, Szabolcs Nagy wrote: > The 08/10/2023 12:41, Mark Brown wrote: > > I agree that it's going to be excessive for pretty much all > > applications, I adjusted it to match x86 as part of the general effort > > to avoid divergence and because I was a bit concerned about non-PCS > > cases (eg, JITed code) potentially running into trouble, especially with > is that even possible? > 16byte alignment is not a convention but architectural: > access via unaligned sp traps (at least in userspace). > it is possible to use bl such that the stack is not involved > e.g. if there is no bl/ret pairing, but if we base the gcs > size on the stack size then i'd expect one stack frame per > bl/ret pair with 16byte alignment, or is there a programming > model possible that uses 8byte stack per bl? That's definitely what I'd expect most of the time. You'd need to be tracking what needs pushing in some other register but it's possible. Quite why you'd do this is a separate question, I think I'm being overly cautious worrying about anyone actually having done it but it wouldn't be the first time I was surprised by someone doing something unexpected. Like I say I think it's excessive and was erring on the side of being conservative.
Attachment:
signature.asc
Description: PGP signature
