On 08/07/12 19:30, Wolfgang Denk wrote:
Most architectures keep the global data pointer (gd) in a register.
This may, or may not be. You should not make any assumptions on how gd is implemented.
The comment did, the code didn't, as long as gd was a pointer (which should be a safe assumption). But that's irrelevant as I seem to have gotten the wrong end of the stick.
When using the external app API, because they are calling us rather than we calling them, this register can be corrupted.
How can this be? The caller should always use the same register convention as we do - otherwise we are in a much deeper trouble. It is up to the caller to make sure it uses the published API (resp. ABI).
Hmm, ok - this was not clear to me from the docs and example.
Indeed, the example does not reserve r8 (on ARM) or r2 (on PPC). Nor does it save/restore it on entry/syscall. I don't know the exact ABI semantics of r2 for PPC, but on ARM this is an error.
This could be worked around by something like:
diff --git a/examples/api/crt0.S b/examples/api/crt0.S index 6daf127..5f956e4 100644 --- a/examples/api/crt0.S +++ b/examples/api/crt0.S @@ -49,13 +49,21 @@ syscall: _start: ldr ip, =search_hint str sp, [ip] + ldr ip, =gd_backup + str r8, [ip] b main
.globl syscall syscall: + push {r6-r8, lr} + ldr r6, =gd_backup + ldr r8, [r6] ldr ip, =syscall_ptr + mov lr, pc ldr pc, [ip] + str r8, [r6] + pop {r6-r8, pc}
#else #error No support for this arch! @@ -69,3 +77,7 @@ syscall_ptr: .globl search_hint search_hint: .long 0 + + .globl gd_backup +gd_backup: + .long 0
An alternative would be to mandate DECLARE_GLOBAL_DATA_PTR for all API applications, but I'd frankly prefer not providing direct access to gd.
Best Regards,
Leif