Segher Boessenkool <segher@xxxxxxxxxxxxxxxxxxx> writes: > On Thu, Aug 06, 2020 at 12:03:33PM +1000, Michael Ellerman wrote: >> Segher Boessenkool <segher@xxxxxxxxxxxxxxxxxxx> writes: >> > On Wed, Aug 05, 2020 at 04:24:16PM +1000, Michael Ellerman wrote: >> >> Christophe Leroy <christophe.leroy@xxxxxxxxxx> writes: >> >> > Indeed, 32-bit doesn't have a redzone, so I believe it needs a stack >> >> > frame whenever it has anything to same. > > ^^^ > >> >> > fbb60: 94 21 ff e0 stwu r1,-32(r1) >> > >> > This is the *only* place where you can use a negative offset from r1: >> > in the stwu to extend the stack (set up a new stack frame, or make the >> > current one bigger). >> >> (You're talking about 32-bit code here right?) > > The "SYSV" ELF binding, yeah, which is used for 32-bit on Linux (give or > take, ho hum). > > The ABIs that have a red zone are much nicer here (but less simple) :-) Yep, just checking I wasn't misunderstanding your comment about negative offsets. >> >> At the same time it's much safer for us to just save/restore r2, and >> >> probably in the noise performance wise. >> > >> > If you want a function to be able to work with ABI-compliant code safely >> > (in all cases), you'll have to make it itself ABI-compliant as well, >> > yes :-) >> >> True. Except this is the VDSO which has previously been a bit wild west >> as far as ABI goes :) > > It could get away with many things because it was guaranteed to be a > leaf function. Some of those things even violate the ABIs, but you can > get away with it easily, much reduced scope. Now if this is generated > code, violating the rules will catch up with you sooner rather than > later ;-) Agreed. cheers
