Re: Inferior function call command set

[Andrew Cagney <ac131313 at redhat dot com>]

> Andrew Cagney writes:
>  > > Out of curiousity, is there any need to have a runtime choice?
>  > 
>  > Entry point in ROM, non 1:1 code/stack, ...
>
> Apologies, still confused.
> [having spent the last few days buried in the guts of
> hand-called-function support such things are very much on my
> mind these days]
>
> How does having an entry point in ROM affect things?
> It appears to me that all AT_ENTRY_POINT does is use the entry point
> address as a magic number that will "never appear" in user code.
> [thus if the callee is returning to it you know you're back in the "stub"]

GDB tries to insert a breakpoint at that address (that doesn't work when 
it is in a ROM :-) Looking at the MIPS, it turned out to also not work 
very well when the entry point couldn't be found.

> In my port I added the ability for the user to override
> CALL_DUMMY_ADDRESS since the entry point is ambiguous/unspecified.
> [THAT would be a very welcome addition to the mainline code. :-)]
> Pproviding both AT_ENTRY_POINT and ON_STACK is _far_ more effort than
> providing the ability to override what gdb uses for CALL_DUMMY_ADDRESS.
> Perhaps what I should have done is just hardwire it to 42.  1/2 :-).

That isn't true.

For an up-to-date architecture, assuming the stack is executable(1), 
there should be zero change.  In fact, given problems with finding the 
entry point address, on-stack is technically a better choice (although a 
quick reality-check will stop that change dead).  The d10v's tweak and 
the retention of entry_point_address, stem from it being a true harvard 
architecture.

A survey of architectures explicitly setting call_dummy_address reveals:

alpha: does something slightly weird (that may now be redundant)
avr: wrapped entry_point_address
m68hc11: wrapped entry_point_address
mips: wrapped entry_point_address, but with a hack for ROMs made 
redundant by ``set call breakpoint-location on-stack''
sparc: depending on a compile time option, uses either 
entry_point_address or the stack

so CALL_DUMMY_ADDRESS can/should probably be wired to entry_point_address.

> No claim is made that there isn't a need for the runtime
> stack/entry-point choice.  But I still don't understand the need for it.
> [Not that anyone has to spend time clearing up my understanding of course;
> but if it's not that much effort, or if other people are also curious ...]
>
>  > An addition to the testsuite is implicit.

Andrew

(1) Someone needs to get this working when the stack isn't executable. 
The inferior would stop with a sigseg instead of sigtrap.