On Mon, Mar 17, 2025 at 06:57:01PM -0700, Chris Torek wrote: > On Mon, Mar 17, 2025 at 11:38 AM Willy Tarreau <w@xxxxxx> wrote: > > OK thanks, but that remains quite strange to me. How can we end up > > here with such an unaligned stack ? At the very minimum I'd expect > > all offsets to be multiple of 8. > > It's a peculiar feature of the version 9 SPARC architecture and runtime. > This also ties into your window save area question. Let's start with these: > > * There are 16 save-able registers in a window. > * Before V9, registers were 32 bits wide. > * V9 and later, registers are 64 bits wide. > * Each stack frame must provide an area for register data. > > Now 32 bits = 4 bytes, times 16 regs = 64 bytes. So for V8 and lower, the > register save area is [%sp+0] through [%sp+63] inclusive. > > Now V9 comes along and we need 128 bytes. But we're going to > run old V8 code in compatibility mode! How will we tell that some > function f() is running in V8 mode instead of V9 mode? [footnote] > > Someone decided that the way to tell would be to use a deliberate > weird alignment of the stack pointer. If the stack pointer was 7 mod 8, > then we're in 64 bit V9 mode and [%sp+2047+0] through > [%sp+2047+127] inclusive are the register save area. If not, it > must be 0 mod 8 and we're in V8 mode and things are as before. > > Why 2047? Well, by observation, it's more common to need negative > offsets from the stack pointer (for a large stack-area array for instance) > than it is to need positive ones (register window save area and > overflow function argument area beyond that). But the instruction > set is more or less symmetric, with a 13-bit immediate constant > offset of -4096 to +4095. Solution: add some offset to the stack > pointer so that function-stack memory is [%sp-4096] through [%sp+2046], > a 6 kilobyte range instead of a 4k one. > > The stack offset therefore helps solve both problems: the offset > indicates whether to use V8 or V9 register dump conventions > and, at the same time, increases the amount of easily-accessed > stack memory. > > [footnote] This provides the ability to dynamically link V8 and V9 > code together. As far as I know this was never used, so that a per > process mode bit suffices just as well. Still, the offset went in. Super clear, thank you Chris! I think a short comment saying "2047: 7 mod 8 = v9 mode" or so would then help and remind that it's not a mistake. Cheers, Willy
