Re: PR28977 tc-i386.c internal error in parse_register

[Alan Modra <amodra@gmail.com>]

On Fri, Mar 18, 2022 at 10:39:42AM +0100, Jan Beulich wrote:
> On 18.03.2022 08:32, Alan Modra wrote:
> > My testcase was a little tidier but still gives:
> > 
> > mad.s: Error: invalid operands (*GAS `reg' section* and *GAS `reg' section* sections) for `%' when setting `s'
> 
> That's only with your change in place, I assume?

Yes.

> > The aim of the patch is to stop an abort *before* we decide the
> > expression is invalid.  i386 parse_register was being called via
> > md_parse_name in gas/expr.c:operand.
> 
> It still feels like your change is merely hiding a problem elsewhere.

If you think the patch is wrong, please feel free to revert it.  Since
we are getting into this discussion, I guess I should have asked
permission.

I don't doubt there are problems elsewhere, but I have some experience
tweaking the expression evaluation code (as do you) and don't want to
go down that rabbit hole for the sake of correcting the ssegment in
erroneous expressions.

I most definitely do not want to spend a large amount of my time
fixing things that never occur in real assembly source, as opposed to
the ridiculuous stuff fed to gas by fuzzers.

> Going from your example (and observing where the abort actually is
> reported) I added a 3rd instance of x=s. Then the abort continues to
> be reported on the 2nd instance. If things were working consistently,
> I would expect it to happen either on the first instance or at the
> end of the file (in this latter case the location reported would
> simply be bogus).

Oh yes, I noticed the same.  Just one instance of x=s doesn't hit the
problem either.  There were rather a lot more duplicated assignments
in the original 17k of fuzzer garbage, and the symbol name wasn't a
nice "x"!

> I think the original know(e->X_op == O_register) was actually quite
> appropriate when seeing a reg_section symbol come in. No reg_section
> symbols violating this should ever be constructed, at least not for
> x86.

Might be difficult to enforce if forward references are allowed.

Heh, if you really want to get your teeth into this, try this one:

b=a
a=%rax
 xor b,%rax

> There might be architectures where such makes sense, albeit
> code like this in expr.c:
> 
> 	  else if (mode != expr_defer && segment == reg_section)
> 	    {
> 	      expressionP->X_op = O_register;
> 	      expressionP->X_add_number = S_GET_VALUE (symbolP);
> 	    }
> 
> makes me think such should never be put into existence. I seem to
> have a vague recollection of, very long ago, having discussed with
> you already the question of too little use of expr_section in the
> course of expression evaluation (without any actual outcome as far
> as changes to the code).
> 
> Jan

-- 
Alan Modra
Australia Development Lab, IBM