Anyone using efficient algorithm or HW support for NEXT/STEP?

Anyone using efficient algorithm or HW support for NEXT/STEP?

[Wendy Peikes]

Here at Cisco, we've gotten both the host (started with gnu gdb 5.0)
and target gdb working for arm-elf.

(Note:  I'd be glad to answer any questions about the arm gdb host
side and help with target-side questions)

We're (my target-side colleague and myself) are wondering if someone
is using a better algorithm for the NEXT and STEP commands:

Our gdb's single stepping is very slow. Current implementation (A)
(Note: applies, in general, to all platforms, not just to arm):
1- restore  user installed breakpoints
2- command to single step.
3- ack and set temp breakpoint(s).  
For arm, sets two temp breakpoints on branches and function calls; 
one at the destination and one at next instruction.  These cover 
conditional execution either way it may go.
4- exception signal
5- remove breakpoints
6- collect regs to see where we are.
7- if still between start addr and stop addr, repeat,
 if == stop addr: stop
8 - if in another function, determine return addr and
 set breakpoint there
9 - continue
10 - go to 4

When it could be more efficient, simpler, and more robust to do B:
1- restore user installed breakpoints
2- set a temp breakpoint at stop addr
3 - continue
4- exception signal
5- collect regs and display where we are

Is anyone out there using a more efficient algorithm similar to B? 
Also, is anyone using hardware single-stepping or hardware-assisted
stepping? For which platforms?

Thank you much for your help,

Wendy Peikes
Cisco Systems
Software Tools Group
GDB Support

Re: Anyone using efficient algorithm or HW support for NEXT/STEP?

[Andrew Cagney]

Some thoughts:

- The latest GDB supports ``set trust-readonly-sections'' on.  Which 
stops GDB going to the target for prologe analysis.  This is a big win.

- enhance the ``T'' packet to return registers such as PC, FP, ARG0, ... 
that can help avoid the need for GDB to fetch all registers

- investigate a thread aware step-out-of-range packet.

Andrew



> Here at Cisco, we've gotten both the host (started with gnu gdb 5.0)
> and target gdb working for arm-elf.
> 
> (Note:  I'd be glad to answer any questions about the arm gdb host
> side and help with target-side questions)
> 
> We're (my target-side colleague and myself) are wondering if someone
> is using a better algorithm for the NEXT and STEP commands:
> 
> Our gdb's single stepping is very slow. Current implementation (A)
> (Note: applies, in general, to all platforms, not just to arm):
> 1- restore  user installed breakpoints
> 2- command to single step.
> 3- ack and set temp breakpoint(s).  
> For arm, sets two temp breakpoints on branches and function calls; 
> one at the destination and one at next instruction.  These cover 
> conditional execution either way it may go.
> 4- exception signal
> 5- remove breakpoints
> 6- collect regs to see where we are.
> 7- if still between start addr and stop addr, repeat,
>  if == stop addr: stop
> 8 - if in another function, determine return addr and
>  set breakpoint there
> 9 - continue
> 10 - go to 4
> 
> When it could be more efficient, simpler, and more robust to do B:
> 1- restore user installed breakpoints
> 2- set a temp breakpoint at stop addr
> 3 - continue
> 4- exception signal
> 5- collect regs and display where we are
> 
> Is anyone out there using a more efficient algorithm similar to B? 
> Also, is anyone using hardware single-stepping or hardware-assisted
> stepping? For which platforms?
> 
> Thank you much for your help,
> 
> Wendy Peikes
> Cisco Systems
> Software Tools Group
> GDB Support
>