Re: [PATCH 1/2] This patch fixes GDBServer's run control for single stepping

[Yao Qi <qiyaoltc at gmail dot com>]

Antoine Tremblay <antoine.tremblay@ericsson.com> writes:

> You have to add if the current instruction is an IT instruction in wich
> case the next instruction will be in an IT block.
>

Oh, you are right.

> Also if you have a conditional instruction that would evalutate to
> true and is not the last one, get_next_pcs may return an instruction
> after the IT block, arm_breakpoint_kind_from_current_state will be
> called from the IT block with that PC and return a THUMB2_KIND when it
> should not. See the else case in arm-get-next-pcs.c:~351

With the current PC and CPSR, it is not difficult to know whether
next_pc is still within IT block nor not, because all instructions in IT
block should be sequentially executed or skipped.

>
> My point was that we should use get_next_pc directly since it's the best
> place to detect if the next_pc is in the IT block. And the intent would
> be clear.

Yeah, we can record the information of breakpoint type in the return
value of get_next_pc, ...

>
> It would give something like the patch below. (Note the GDB part of this
> is missing but it works with GDBServer)
>

... but using extra bit in CORE_ADDR is not a good idea to me.

>> The problem of this patch is that we end up inserting different
>> kinds of breakpoints on the same instruction.  For a given 32-bit thumb
>> instruction, GDB and GDBserver knows 32-bit thumb breakpoint instruction
>> is used for GDB breakpoint, but only GDBserver knows 16-bit thumb
>> breakpoint is used for GDBserver single-step breakpoint, so GDB will be
>> confused on this.  I stopped here, and start to do something else.
>
> Humm but how will the GDBServer 16-bit breakpoint be reported to GDB ?
> Won't it always be hit and handled by GDBServer ?
>
> And if you have a GDB breakpoint on an instruction and GDBServer puts
> a single step breakpoint on that GDB breakpoint instruction, GDBServer
> still knows of the GDB and GDBServer breakpoint types.
>
> So how does GDB get confused ?

That was my conclusion at that point.  I got some regressions in
gdb.threads/*.exp when I tested my patch (gdb running is on
x86_64-linux), but I can't remember more details.

I am also wondering that we can use some code in
arm_adjust_breakpoint_address about detecting BPADDR is within IT block
or not by scanning instructions backward, if none of two bytes (can be
16-bit thumb instruction or the 2nd half of 32-bit thumb instruction)
matches IT instruction, the PC is not within IT block.

-- 
Yao (齐尧)