On 8/31/22 5:45 PM, Segher Boessenkool wrote: > On Tue, Aug 30, 2022 at 09:10:02AM +0000, Christophe Leroy wrote: >> Le 30/08/2022 à 11:01, Nicholas Piggin a écrit : >>> On Tue Aug 30, 2022 at 3:24 PM AEST, Christophe Leroy wrote: >>>>> This is still slightly concerning to me. Is there any guarantee that the >>>>> compiler would not use a different sequence for the address here? >>>>> >>>>> Maybe explicit r13 is required. >>>>> >>>> >>>> local_paca is defined as: >>>> >>>> register struct paca_struct *local_paca asm("r13"); > > And this is in global scope, making it a global register variable. > >>>> Why would the compiler use another register ? >>> >>> Hopefully it doesn't. Is it guaranteed that it won't? > > Yes, this is guaranteed.
Agree with Segher here. That said, there was a gcc bug a looooong time ago where gcc copied r13 into a temporary register and used it from there. That's ok (correctness wise, but not ideal) from user land standpoint, but we took a context switch after the reg copy and it was restarted on a different cpu, so differnt local_paca and r13 value. We went boom because the copy wasn't pointing to the correct local_paca anymore. So it is very important the compiler always use r13 when accessing the local_paca. Peter