Re: [PATCH 2/2] libopcodes: extend the styling within the i386 disassembler

[Andrew Burgess <aburgess@redhat.com>]

Jan Beulich via Binutils <binutils@sourceware.org> writes:

> On 29.04.2022 15:42, Andrew Burgess via Binutils wrote:
>> The i386 disassembler is pretty complex.  Most disassembly is done
>> indirectly; operands are built into buffers within a struct instr_info
>> instance, before finally being printed later in the disassembly
>> process.
>> 
>> Sometimes the operand buffers are built in a different order to the
>> order in which they will eventually be printed.
>> 
>> Each operand can contain multiple components, e.g. multiple registers,
>> immediates, other textual elements (commas, brackets, etc).
>> 
>> When looking for how to apply styling I guess the ideal solution would
>> be to move away from the operands being a single string that is built
>> up, and instead have each operand be a list of "parts", where each
>> part is some text and a style.  Then, when we eventually print the
>> operand we would loop over the parts and print each part with the
>> correct style.
>> 
>> But it feels like a huge amount of work to move from where we are
>> now to that potentially ideal solution.  Plus, the above solution
>> would be pretty complex.
>> 
>> So, instead I propose a .... different solution here, one that works
>> with the existing infrastructure.
>> 
>> As each operand is built up, piece be piece, we pass through style
>> information.  This style information is then encoded into the operand
>> buffer (see below for details).  After this the code can continue to
>> operate as it does right now in order to manage the set of operand
>> buffers.
>> 
>> Then, as each operand is printed we can split the operand buffer into
>> chunks at the style marker boundaries, with each chunk being printed
>> in the correct style.
>> 
>> For encoding the style information I use the format "~%x~".  As far as
>> I can tell the '~' is not otherwise used in the i386 disassembler, so
>> this should serve as a unique marker.  To speed up writing and then
>> reading the style markers, I take advantage of the fact that there are
>> less than 16 styles so I know the '%x' will only ever be a single hex
>> character.
>
> Like H.J. I'd like to ask that you avoid ~ here (I actually have plans
> to use it to make at least some 64-bit constants better recognizable);
> I'm not sure about using non-ASCII though, as that may cause issues with
> compilers treating non-ASCII wrong. I'd soften this to non-alnum, non-
> operator characters (perhaps more generally non-printable). Otoh I guess
> about _any_ character could be used in symbol names, so I'm not
> convinced such an escaping model can be generally conflict free.

Hi Jan,

I've addressed all the simple feedback from H.J. and Vladimir, and I
just need to figure out something for the escaping mechanism.

I'm still keen to try and go with an escaping based solution, my
reasoning is that I think that this is the solution least likely to
introduce latent disassembler bugs.

However, that position is based on my belief that there's no exhaustive
test for the i386 based disassembler, i.e. one that tests every single
valid instruction disassembles correctly.  If there was such a test then
I might be more tempted to try something more radical...

That said, if I was going to stick with an escaping scheme, then I have
some ideas for moving forward.

The current scheme relies on the fact that symbols are not printed
directly from the i386 disassembler, instead the i386 disassembler calls
back into the driver application (objdump, gdb) to print the symbol.  As
a result, symbols don't go through the instr_info::obuf buffer.  This
means that we never try to interpret a symbol name for escape
characters.

This means we avoid one of the issues that you raised, what if the
escape character appears in a symbol name; the answer is, I just don't
need to worry about this!

So, I only need to ensure that the escape character is:

  (a) not a character that the disassembler currently tries to directly
  print itself, and

  (b) not something that will ever be printed as part of an immediate.

Clearly my choice passes both right now, but looks like it will not pass
(b) forever.

One possible solution would be to replace all the remaining places where
we directly write to instr_info::obuf with calls to oappend_char.  I
could then extend the oappend API such that we do "real" escaping, that
is (assuming the continued use of '~' for now): '~X' would indicate a
style marker, with X being the style number, and '~~' would indicate a
literal '~' character.  In this was we really wouldn't care which
character we used (though we'd probably pick one that didn't crop up too
ofter just for ease of parsing the buffers).

An alternative solution would be to pick a non-printable character,
e.g. \001, and use this as the escape character in place of the current
'~'.  This seems to pass the (a) and (b) tests above, and if such a
character does ever appear in a symbol name, then, as I've said above, I
don't believe this would cause us any problems.

Here's a session that demonstrates a symbol containing '~' with the
current patch (obviously the final disassembler call actually has
colour in the output, which all looks correct to me):

  $ cat /tmp/weird.s
          .text
          .global "foo~bar"
          "foo~bar":
          nop
          nop
          nop
          call       "foo~bar"
  $ as -o /tmp/weird.o /tmp/weird.s
  $ ./binutils/objdump --disassembler-color=extended-color -d /tmp/weird.o 
  
  /tmp/weird.o:     file format elf64-x86-64
  
  
  Disassembly of section .text:
  
  0000000000000000 <foo~bar>:
     0:	90                   	nop
     1:	90                   	nop
     2:	90                   	nop
     3:	e8 00 00 00 00       	call   8 <foo~bar+0x8>


Thanks,
Andrew