Re: ARM thumb interworking hole?

[Ian Lance Taylor <ian at airs dot com>]

Charles Manning <manningc2@actrix.gen.nz> writes:

> From where I stand (ie. almost totally ignorant of gcc's guts), the way code 
> should be generated for handling a function pointer under vanilla 
> interworking is something along the lines of
> 
> 
>         mov   rx,....  @ load rx with pointer
>         bl      call_rx_trampoline 
>         .......
> 
> call_rx_trampoline:
>          bx  rx
> 
> The rx_trampoline is very cheap (2 bytes) but needs to be close so it matters 
> little if this is emitted for every section that uses it. I'm sure the 
> compiler could emit this along the same lines as used to manage literal pools 
> etc.

That is pretty much how it works today, except that there is only copy
of the _call_rx_trampoline code, and it is in libgcc.

It's true that gcc could be changed to put that in the literal pool.
But most people who use Thumb code are interested in minimal code
size, so having one copy of the trampoline per program, rather than
one per function which needs it, seems best in the normal case.  You
have certainly identified a bug in the -mlong-calls case.  As Richard
says, that bug is probably best fixed in the linker, which can insert
the trampoline only when it is required.

Anyhow, for ARMv5 code these issues do not arise.  Are you still using
an ARMv4 processor?

Ian