Re: plt for arm gnu thumb

[Daniel Jacobowitz <drow@false.org>]

On Tue, Apr 18, 2006 at 04:13:47PM +0100, Richard Earnshaw wrote:
> There's only ever going to be a problem if you try to run on an
> Architecture v4T device.  These chips cannot transition to Thumb mode
> when loading a PC value from memory, but need instead to use a BX
> instruction.  The problem is that for a shared library trampoline we
> need to:
> 
>  - preserve all of r0-r11, r13 and r14
>  - have the 'index' of the called function in ip
> 
> There's simply no way to achieve all that efficiently and support proper
> interworking on these devices.  (v5 and later chips can swap from ARM to
> Thumb on a load into the PC, so there is no problem there).
> 
> We investigated this quite thoroughly as part of the EABI development. 
> We concluded that given that the standard trampoline would work on all
> v4 and earlier chips and all v5 or later chips (ie everything but v4T),
> and that on v4T the sequence works if the target of the call is entered
> in ARM state, the best solution overall was to mandate that on a v4T
> device, the called address must be an ARM instruction (it can transition
> to Thumb state immediately thereafter).
> 
> In theory this could all be handled by the linker when constructing the
> library, but currently the GNU linker does not do this.

Oh, right; I forgot about this issue.

I did once implement alternate PLT sequences which could target Thumb,
even on v4t.  However, they are mighty awkward, so I would strongly
prefer not to submit the patch.  John, are you really using v4t?

-- 
Daniel Jacobowitz
CodeSourcery