Re: [PATCH v2] gdb: add 'maintenance print record-instruction' command

[Bruno Larsen <blarsen@redhat.com>]

On 14/12/2022 01:46, Lancelot SIX wrote:
> Hi Bruno,
>
> I have a couple more comments and suggestions below.
>
> On Mon, Dec 12, 2022 at 11:44:17AM +0100, Bruno Larsen via Gdb-patches wrote:
>> While chasing some reverse debugging bugs, I found myself wondering what
>> was recorded by GDB to undo and redo a certain instruction. This commit
>> implements a simple way of printing that information.
>> ---
>>   gdb/NEWS            |  6 ++++
>>   gdb/doc/gdb.texinfo |  8 +++++
>>   gdb/record-full.c   | 81 +++++++++++++++++++++++++++++++++++++++++++++
>>   3 files changed, 95 insertions(+)
>>
>> diff --git a/gdb/NEWS b/gdb/NEWS
>> index c4ccfcc9e32..d6ce6bf86a0 100644
>> --- a/gdb/NEWS
>> +++ b/gdb/NEWS
>> @@ -103,6 +103,12 @@
>>   
>>   * New commands
>>   
>> +maintenance print record-instruction [ N ]
>> +  Print the recorded information for a given instruction. If N is not given
>> +  prints how GDB would undo the last instruction executed. If N is negative,
>> +  prints how GDB would undo the N-th previous instruction, and if N is
>> +  positive, it prints how GDB will redo the N-th following instruction.
>> +
>>   maintenance set ignore-prologue-end-flag on|off
>>   maintenance show ignore-prologue-end-flag
>>     This setting, which is off by default, controls whether GDB ignores the
>> diff --git a/gdb/doc/gdb.texinfo b/gdb/doc/gdb.texinfo
>> index 5b566669975..807af351e79 100644
>> --- a/gdb/doc/gdb.texinfo
>> +++ b/gdb/doc/gdb.texinfo
>> @@ -40531,6 +40531,14 @@ that symbol is described.  The type chain produced by this command is
>>   a recursive definition of the data type as stored in @value{GDBN}'s
>>   data structures, including its flags and contained types.
>>   
>> +@kindex maint print record-instruction
>> +@item maint print record-instruction
>> +@itemx maint print record-instruction @var{N}
>> +@cindex print how GDB recorded a given instruction.  If N is not positive
>> +number, it prints the values stored by the inferior before the N-th previous
>> +instruction was exectued.  If N is positive, print the values after the N-th
>> +following instruction is executed.  If N is not given, 0 is assumed.
>> +
>>   @kindex maint selftest
>>   @cindex self tests
>>   @item maint selftest @r{[}-verbose@r{]} @r{[}@var{filter}@r{]}
>> diff --git a/gdb/record-full.c b/gdb/record-full.c
>> index 48b92281fe6..25e1fd22c6f 100644
>> --- a/gdb/record-full.c
>> +++ b/gdb/record-full.c
>> @@ -2764,6 +2764,79 @@ set_record_full_insn_max_num (const char *args, int from_tty,
>>       }
>>   }
>>   
>> +/* Implement the 'maintenance print record-instruction' command.  */
>> +
>> +static void
>> +maintenance_print_record_instruction (const char *args, int from_tty)
>> +{
>> +  struct record_full_entry* to_print = record_full_list;
>> +
>> +  if (args != nullptr)
>> +    {
>> +      int offset = value_as_long (parse_and_eval (args));
>> +      if (offset > 0)
>> +	{
>> +	  /* Move forward OFFSET instructions.  We know we found the
>> +	     end of an instruction when to_print->type is 0.  */
> I think the literal 0 in the comment is a left-over from V1, right?
>
>> +	  while (to_print->next != nullptr && offset > 0)
>> +	    {
>> +	      to_print = to_print->next;
>> +	      if (to_print->type == record_full_end)
>> +		offset--;
>> +	    }
>> +	  if (offset != 0)
>> +	    error (_("Not enough recorded history"));
>> +	}
>> +      else
>> +	{
>> +	  while (to_print->prev != nullptr && offset < 0)
>> +	    {
>> +	      to_print = to_print->prev;
>> +	      if (to_print->type == record_full_end)
>> +		offset++;
>> +	    }
>> +	  if (offset != 0)
>> +	    error (_("Not enough recorded history"));
>> +	}
>> +    }
>> +  gdb_assert (to_print != nullptr);
>> +
>> +  /* Go back to the start of the instruction.  */
>> +  while (to_print->prev != nullptr && to_print->prev->type != record_full_end)
>> +    to_print = to_print->prev;
>> +
>> +  while (to_print->type != record_full_end)
>> +    {
>> +      switch (to_print->type)
>> +	{
>> +	  case record_full_reg:
>> +	    {
>> +	      gdb_byte* b = record_full_get_loc (to_print);
>                        ^^
>
> The space comes before the * here.
>
>> +	      gdb_printf ("Register %%%s changed:",
> The '%' prefix for register names is specific to the att assembler
> syntax.  Do we want this here?  I think in most places GDB uses the
> plain register name without such prefix.
>
>> +			  gdbarch_register_name (target_gdbarch (),
>> +						 to_print->u.reg.num));
>> +	      for (int i = 0; i < to_print->u.reg.len; i++)
>> +		gdb_printf (" %02x",b[i]);
>                                      ^
> Space after the ",".
>
>> +	      gdb_printf ("\n");
> Did you consider printing the register value instead of the bytes
> composing the value?  I think something like this could do:
>
>      auto mark = value_mark ();
>      type *regtype = gdbarch_register_type (target_gdbarch (),
>                                             to_print->u.reg.num);
>      value *val
>        = value_from_contents (regtype, record_full_get_loc (to_print));
>      
>      gdb_printf ("Register %s changed: %s\n",
>                  gdbarch_register_name (target_gdbarch (),
>                                         to_print->u.reg.num),
>                  hex_string (value_as_long (val)));
>      value_free_to_mark (mark);
>
> This would however not work for vector registers (anything which does
> not fit in a long long).  You might want to use something like
> `value_print` to handle all cases, unless its output is too fancy:
>
>      struct value_print_options opts;
>      get_user_print_options (&opts);
>      auto mark = value_mark ();
>      type *regtype = gdbarch_register_type (target_gdbarch (),
>                                             to_print->u.reg.num);
>      value *val
>        = value_from_contents (regtype, record_full_get_loc (to_print));
>
>      gdb_printf ("Register %s changed: %s\n",
>                  gdbarch_register_name (target_gdbarch (),
>                                         to_print->u.reg.num));
>      value_print (val, gdb_stdout, &opts);
>      gdb_printf ("\n");
>      value_free_to_mark (mark);
>
> Not sure if there is a cleaner way to achieve this, but it should do the
> trick.

Hi Lancelot,

Thanks for the review, I've fixed all the style nits.

I wasn't aware of how to do this, thanks for explaining! I was a bit 
reticent of using the value_as_long option specifically because I wanted 
to handle all types of registers, but if I can make value_print work it 
will definitely be better. However, when I tried this I got the 
following output:

(gdb) maint print record-instruction
8 bytes of memory at address 0x00007fffffffde48 changed from: e0 de ff 
ff ff 7f 00 00
(gdb) Register rip changed: (void (*)()) 0x40113e <main+8>

I see a few issues here:
* The casting at the front is pretty confusing. I want the printing to 
be raw so the user can see what is going on
* I'm not necessary fond of the <main+8> at the end, mostly because I am 
worried that it might be misplaced in other instances
* it is being printed after the gdb prompt

I'm saying this mostly because it will probably take me a while to make 
a new version, but I really appreciate your feedback!

-- 
Cheers,
Bruno

>
>> +	      break;
>> +	    }
>> +	  case record_full_mem:
>> +	    {
>> +	      gdb_byte* b = record_full_get_loc (to_print);
> Same here, the space shoud be before the "*".
>
>> +	      gdb_printf ("%d bytes of memory at address %s changed from:",
>> +			  to_print->u.mem.len,
>> +			  print_core_address (target_gdbarch (),
>> +					      to_print->u.mem.addr));
>> +	      for (int i = 0; i < to_print->u.mem.len; i++)
>> +		gdb_printf (" %02x",b[i]);
> Space after the ",".
>
> Best,
> Lancelot.
>
>> +	      gdb_printf ("\n");
>> +	      break;
>> +	    }
>> +	}
>> +      to_print = to_print->next;
>> +    }
>> +}
>> +
>>   void _initialize_record_full ();
>>   void
>>   _initialize_record_full ()
>> @@ -2868,4 +2941,12 @@ When ON, query if PREC cannot record memory change of next instruction."),
>>     c = add_alias_cmd ("memory-query", record_full_memory_query_cmds.show,
>>   		     no_class, 1,&show_record_cmdlist);
>>     deprecate_cmd (c, "show record full memory-query");
>> +
>> +  add_cmd ("record-instruction", class_maintenance,
>> +	   maintenance_print_record_instruction,
>> +	   _("\
>> +Print a recorded instruction.\nIf no argument is provided, print the last \
>> +instruction recorded.\nIf a negative argument is given, prints how the nth \
>> +previous instruction will be undone.\nIf a positive argument is given, prints \
>> +how the nth following instruction will be redone."), &maintenanceprintlist);
>>   }
>> -- 
>> 2.38.1
>>