Issue in dwarf2/expr.c

Issue in dwarf2/expr.c

[Rohr, Stephan]

Hi all,

I came across a bug in dwarf2/expr.c:177 in function  rw_pieced_value. when reading a pieced value described by DW_OP_piece or DW_OP_bit_piece.

      if (value_type (v) != value_enclosing_type (v))
                internal_error (__FILE__, __LINE__,
                                                _("Should not be able to create a lazy value with "
                                                  "an enclosing type"));

This check can easily be triggered when casting a variable with location described by DW_OP_piece or DW_OP_bit_piece.

A minimal reproducer can be found in gdb/testsuite/gdb.dwarf2/shortpiece.exp, e.g. casting s1 to an array type triggers the error

p (short []) s1

I currently struggle to understand the reasoning for this check. I could think of changing this to

if (TYPE_LENGTH(value_type (v)) > TYPE_LENGTH( value_enclosing_type (v)))
to support the cast operation.

But in general, I'm not sure why this check is needed at all.

I appreciate your feedback.


Best
stephan

Re: Issue in dwarf2/expr.c

[Andreas Arnez]

On Wed, Apr 13 2022, Rohr, Stephan via Gdb wrote:

> Hi all,
>
> I came across a bug in dwarf2/expr.c:177 in function  rw_pieced_value. when reading a pieced value described by DW_OP_piece or DW_OP_bit_piece.
>
>       if (value_type (v) != value_enclosing_type (v))
>                 internal_error (__FILE__, __LINE__,
>                                                 _("Should not be able to create a lazy value with "
>                                                   "an enclosing type"));

Since I did some work on this function in the past, I just looked where
this check comes from, and it turns out that it was introduced in 2010
by Tom Tromey with commit afd74c5ff76010405caddd2834be4a0178fa93dd --

gdb
	* dwarf2loc.c (read_pieced_value): Work properly when 'v' has an
	offset.
	(write_pieced_value): Likewise.

Perhaps Tom still remembers the rationale?

>
> This check can easily be triggered when casting a variable with location described by DW_OP_piece or DW_OP_bit_piece.
>
> A minimal reproducer can be found in gdb/testsuite/gdb.dwarf2/shortpiece.exp, e.g. casting s1 to an array type triggers the error
>
> p (short []) s1
>
> I currently struggle to understand the reasoning for this check. I could think of changing this to
>
> if (TYPE_LENGTH(value_type (v)) > TYPE_LENGTH( value_enclosing_type (v)))
> to support the cast operation.
>
> But in general, I'm not sure why this check is needed at all.
>
> I appreciate your feedback.
>
>
> Best
> stephan

--
Andreas

Re: Issue in dwarf2/expr.c

[Tom Tromey]

>> I came across a bug in dwarf2/expr.c:177 in function  rw_pieced_value. when reading a pieced value described by DW_OP_piece or DW_OP_bit_piece.
>> 
>> if (value_type (v) != value_enclosing_type (v))
>> internal_error (__FILE__, __LINE__,
>> _("Should not be able to create a lazy value with "
>> "an enclosing type"));

> Since I did some work on this function in the past, I just looked where
> this check comes from, and it turns out that it was introduced in 2010
> by Tom Tromey with commit afd74c5ff76010405caddd2834be4a0178fa93dd --

> gdb
> 	* dwarf2loc.c (read_pieced_value): Work properly when 'v' has an
> 	offset.
> 	(write_pieced_value): Likewise.

> Perhaps Tom still remembers the rationale?

I don't remember, sorry.
It's possible I assumed that it should be impossible to create a value
with an enclosing type without un-lazying it.

Removing the assertion seems fine but it's important to make sure the
result works correctly.  An enclosing type normally means that the data
representing the 'type' is at some offset in the contents, but it seems
to me that rw_pieced_value may not handle this case.

Tom

RE: Issue in dwarf2/expr.c

[Rohr, Stephan]

Hi Tom,

Thanks for your feedback. 

The 'embedded_offset' is considered in the ' value_contents_raw (v)' function call whereas the 'offset' is handled in 'rw_pieced_value'. So from my understanding it should work as expected. I will check for regressions and submit a patch.

Best
stephan

-----Original Message-----
From: Tom Tromey <tom@tromey.com> 
Sent: Friday, April 15, 2022 10:25 PM
To: Andreas Arnez <arnez@linux.ibm.com>
Cc: Rohr, Stephan via Gdb <gdb@sourceware.org>; Rohr, Stephan <stephan.rohr@intel.com>; Tom Tromey <tom@tromey.com>
Subject: Re: Issue in dwarf2/expr.c

>> I came across a bug in dwarf2/expr.c:177 in function  rw_pieced_value. when reading a pieced value described by DW_OP_piece or DW_OP_bit_piece.
>> 
>> if (value_type (v) != value_enclosing_type (v)) internal_error 
>> (__FILE__, __LINE__, _("Should not be able to create a lazy value 
>> with "
>> "an enclosing type"));

> Since I did some work on this function in the past, I just looked 
> where this check comes from, and it turns out that it was introduced 
> in 2010 by Tom Tromey with commit 
> afd74c5ff76010405caddd2834be4a0178fa93dd --

> gdb
> 	* dwarf2loc.c (read_pieced_value): Work properly when 'v' has an
> 	offset.
> 	(write_pieced_value): Likewise.

> Perhaps Tom still remembers the rationale?

I don't remember, sorry.
It's possible I assumed that it should be impossible to create a value with an enclosing type without un-lazying it.

Removing the assertion seems fine but it's important to make sure the result works correctly.  An enclosing type normally means that the data representing the 'type' is at some offset in the contents, but it seems to me that rw_pieced_value may not handle this case.

Tom