Re: Bug in avr-gdb? was:Re: [avr-gcc-list] AVR simulator and interrupts

[Theodore A. Roth]

On Fri, 16 Jul 2004, Klaus Rudolph wrote:

> Hi James,
>
> I could rproduce the problem also with actual simulavr.
> After traceing through the gdb output I could found
> a senseless continue command which is send from gdb to simulator.
> The result is endless running of simulavr which is exact what the
> gdb wants but what is not what the user want from gdb :-)
>
>
> I think we have a bug in gdb for avr.

I don't think this is a bug in gdb and it's not something that can be
easily worked around in gdb.

The problem is gdb doesn't have enough information to step through the
"__vectors" function. In reality, __vectors is not a function at all,
but a jump table. The assembler marks it as a function because of this
(from gcrt1.S):

        .section .vectors,"ax",@progbits
        .global __vectors
        .func   __vectors
__vectors:

I think it's done that way to allow locating the .vectors section at
flash addr 0x00 (someone correct me if I'm wrong). This confuses gdb
since it thinks it's looking at a function expecting it to return at
some point. Instead, we jmp/rjmp out of __vectors directly in the
interrupt handler.

I just hacked simulavr-0.1.2.1 to automatically perform an extra step
when it vectors into the __vectors table (see attached patch). This
makes the 'step' command step right into the interrupt when it fires off
which is what I would expect the user to expect. This is similar in idea
to avarice's --ignore-intr option.

There's still a problem with the 'next' command. In my modified example
program (see test.c attached), if I tell gdb to 'next' when it's at line
28 in this code,

     25     while (1)
     26     {
     27         /* Chew up some MCU cycles. */
     28         clk += TCNT0;
     29         //dummy -= TCNT0;
     30     }

it goes into an infinite loop of stepping (but not jumping into
__vectors). This falls under the "don't expect gdb to do the right thing
if you compile with optimizations"  category. A simple work around in
this case is to just set a breakpoint at line 28.

If I uncomment line 29, then 'next' works fine and I never jump int
__vectors or the interrupt.

---
Ted Roth
PGP Key ID: 0x18F846E9
Jabber ID: address@hidden

test.c
Description: Text document

simulavr-step-vect.diff
Description: Text document