[PowerPC] PR23774 stack backchain broken saga

[Segher Boessenkool]

>> Especially as the ABI states that the write of the backlink
>> and the stack pointer update _have_ to be done in one insn.
>
> That's on allocation.  Deallocation isn't so critical.  You just need 
> to
> ensure the backchain is written before updating sp.

Yes, but your example generated code showed allocation with
two insns.  Of course it works, but it's not allowed by the
ABI.

> Hmm, on powerpc64-linux you could even alloc up to 288 bytes without
> an atomic update, since 288 bytes below the current sp is available for
> use.

Yes.  Question is, can GCC support this (easily) -- it would
have to keep track how much space it already alloca()'ed
before doing _this_ alloca().


Segher