Re: [PATCH 1/3] Fix ppc64 single step over atomic sequence testcase

[Luis Gustavo <luis_gustavo@mentor.com>]

Hi Anton,

On 06/12/2012 09:45 AM, Anton Blanchard wrote:
>> It would be useful for you to say what exactly is broken, and on
>> which platform. At least it seems to have been working for some
>> people (at IBM).
>
> Sorry the explanation was a bit terse. The testcase assumes it only
> requires two "next"s to get through the test function:
>
> set bp1 [gdb_get_line_number "lwarx"]
> gdb_breakpoint "$bp1" "Breakpoint $decimal at $hex" \
>    "Set the breakpoint at the start of the sequence"
>
> gdb_test continue "Continuing.*Breakpoint $decimal.*" \
>    "Continue until breakpoint"
>
> gdb_test next ".*__asm __volatile.*" \
>    "Step through the lwarx/stwcx sequence"
>
> gdb_test next ".*return 0.*" \
>    "Step through the ldarx/stdcx sequence"
>
> If I run this testcase manually on Fedora 16 (gcc 4.6.2), it
> actually takes 7 steps to get through it:
>
> Breakpoint 2, main () at ./gdb.arch/ppc64-atomic-inst.c:27
> 27	  __asm __volatile ("1:     lwarx   %0,0,%2\n"              \
> (gdb) next
> 32	                    : "b" (word_addr), "m" (*word_addr)     \
> (gdb) next
> 27	  __asm __volatile ("1:     lwarx   %0,0,%2\n"              \
> (gdb) next
> 39	                    : "=&b" (dword), "=m" (*dword_addr)     \
> (gdb) next
> 35	  __asm __volatile ("1:     ldarx   %0,0,%2\n"              \
> (gdb) next
> 40	                    : "b" (dword_addr), "m" (*dword_addr)   \
> (gdb) next
> 35	  __asm __volatile ("1:     ldarx   %0,0,%2\n"              \
> (gdb) next
> 43	  return 0;
> (gdb)
>
> I'm not sure what is expected here, is "next" supposed to step all the
> way through inline assembly? Perhaps it is, but it seems fragile to
> depend on this high level behaviour.

I authored that testcase a while ago. For the most common case we didn't 
have multiple branch instructions inside such a sequence IIRC, but if we 
have that now, the testcase should definitely be updated.

GDB is supposed to skip eacho sequence (__asm__ block) in response to a 
"next", but it may have had its behavior changed from when i originally 
wrote this. It seems to take multiple next's to get to a different 
__asm__ block now, so it seems to be broken.

It may be a difference in GCC's debuginfo output or GDB just got smarter.

Regards,
Luis