Re: ARM stub sizing fix

[Daniel Jacobowitz <dan at codesourcery dot com>]

On Thu, Jan 14, 2010 at 10:25:28PM +1030, Alan Modra wrote:
> required on another branch to the same destination.  However, the code
> in arm_type_of_stub clearly returns stub types that depend on branch
> offset and reloc type.  You could easily have both a
> short_branch_v4t_thumb_arm and a long_branch_v4t_thumb_arm to the same
> destination, with sorry results if the short version happened to be
> created first.  Also, addition of stubs for one iteration of stub

I think this is a bug in arm_type_of_stub.  short_branch_v4t_thumb_arm
and long_branch_v4t_thumb_arm are the same type of stub, based on how
far the destination is from the stub.  So I think the branch_offset
check is the wrong thing.

We're picking the stub type before we know where the stub ends up.  I
think we have to give the two the same type, but distinguish in
find_stub_size_and_template.

> I also add an assert that will trigger if there is a dependence on
> reloc type.  I don't know enough about ARM code to know whether it's
> possible to have, say, one branch to a function using R_ARM_THM_CALL
> and another using R_ARM_THM_JUMP24 (possible tail call?).

You're right, that's just a normal and tail call to the same function.

-- 
Daniel Jacobowitz
CodeSourcery