Re: [PATCH v2 1/1] gdb/dwarf2: Fix 'rw_pieced_value' for values casted to different type.

[Simon Marchi <simark@simark.ca>]

On 10/25/22 09:42, Stephan Rohr via Gdb-patches wrote:
> From: "Rohr, Stephan" <stephan.rohr@intel.com>
> 
> The 'rw_pieced_value' function is executed when fetching a (lazy)
> variable described by 'DW_OP_piece' or 'DW_OP_bit_piece'.  The
> function checks the 'type' and 'enclosing_type' fields of the value
> for identity.
> 
>   * The 'type' field describes the type of a value.
>   * In most cases, the 'enclosing_type' field is identical to the
>     'type' field.
>   * Scenarios where the 'type' and 'enclosing_type' of an object
>     differ are described in 'gdb/value.c'.  Possible cases are:
>   * If a value represents a C++ object, then the 'type' field
>     gives the object's compile-time type.  If the object actually
>     belongs to some class derived from `type', perhaps with other
>     base classes and additional members, then `type' is just a
>     subobject of the real thing, and the full object is probably
>     larger than `type' would suggest.
>   * If 'type' is a dynamic class (i.e. one with a vtable), then GDB
>     can actually determine the object's run-time type by looking at
>     the run-time type information in the vtable.  GDB may then elect
>     to read the entire object.
>   * If the user casts a variable to a different type
>     (e.g. 'print (<type> []) <variable>'), the value's type is
>     updated before reading the value.

I think you should indent the last 3 bullets.

> 
> If a lazy value is fetched, GDB allocates space based on the enclosing
> type's length and typically reads the 'full' object.  This is not
> implemented for pieced values and causes an internal error if 'type'
> and 'enclosing_type' of a value are not identical.
> 
> However, GDB can read the value based on its type.  Thus, this patch
> fixes the previously mentioned cases by removing the check for identity.
> 
> Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=28605
> 
> gdb/ChangeLog:
> 2022-04-13  Stephan Rohr  <stephan.rohr@intel.com>
> 
> 	* dwarf2/loc.c (rw_pieced_value): Fix check on 'type' and
> 	'enlcosing_type' when reading pieced value 'v'.
> 
> gdb/testsuite/ChangeLog:
> 2022-04-13  Stephan Rohr  <stephan.rohr@intel.com>
> 
> 	* gdb.dwarf2/shortpiece.exp: Added test cases.
> ---
>  gdb/dwarf2/expr.c                       |  3 --
>  gdb/testsuite/gdb.dwarf2/shortpiece.exp | 50 ++++++++++++++++++++++++-
>  2 files changed, 49 insertions(+), 4 deletions(-)
> 
> diff --git a/gdb/dwarf2/expr.c b/gdb/dwarf2/expr.c
> index 73dfd4b4ffb..fa08ad89024 100644
> --- a/gdb/dwarf2/expr.c
> +++ b/gdb/dwarf2/expr.c
> @@ -161,9 +161,6 @@ rw_pieced_value (value *v, value *from, bool check_optimized)
>      }
>    else
>      {
> -      if (value_type (v) != value_enclosing_type (v))
> -	internal_error (_("Should not be able to create a lazy value with "
> -			  "an enclosing type"));

I'm still wondering about whether value_cast should change the enclosing
type.

When you do:

    p (int []) s1

The original value has:

 - type = struct S
 - enclosing_type = struct S
 - embedded_offset = 0

This means the object we're looking at is of type struct S, and it's at
offset 0 of a top-level containing object of type struct S (itself, in
this trivial case).  After the case, we have:

 - type = int[2]
 - enclosing_type = struct S
 - embedded_offset = 0

This means the object we're looking at is of type int[2], and it,s at
offset 0 of a top-level containing object of type struct S.  I don't
think that's a wrong way of looking at it.  But it's perhaps not that
useful either.


I'm wondering what happens if the value we cast has a different type and
enclosing_type to begin with.  I imagine the following test case:

  struct Base1
  {
    virtual ~Base1() = default;
    int a = 1;
    int b = 2;
  };

  struct Base2
  {
    virtual ~Base2() = default;
    int c = 3;
    int d = 4;
  };

  struct Derived : public Base1, Base2
  {
    virtual ~Derived() = default;
    int x = 98;
    int z = 99;
  };

  int main()
  {
    Derived d;
    Base1 *b1 = &d;
    Base2 *b2 = &d;
    return 0;
  }

In GDB:

  (gdb) print (int []) *b2

Before the cast we have:

 - type = Base2
 - enclosing_type = Derived
 - embedded_offset = 16

After the cast:

 - type = int[4] (4, not 2, because of the vtable)
 - enclosing_type = Derived
 - embedded_offset = 16

This tells us we have an object of type int[4], at offset 16 within a
containing object of type Derived.  Again, not a wrong way to look at
it.  It's just that the type -> enclosing_type relationship seems
useless now that the value is cast to a type that does not involve
inheritance anymore.

But this shows that if we make value_cast set the enclosing_type, it
should also reset the embedded_offset to 0.

Anyway, it's still not clear to me what the right answers are.  If Tom
is happy with the patch, I am happy with it too.

> @@ -98,3 +129,20 @@ if { [prepare_for_testing "failed to prepare" ${testfile} \
>  gdb_test "p s1" " = {a = 1, b = 0}"
>  gdb_test "p s2" \
>      "access outside bounds of object referenced via synthetic pointer"
> +
> +# When fetching a lazy value, GDB typically tries to fetch the 'full'
> +# object based on the enclosing type.  GDB does not support the reading
> +# of a pieced value with a (possibly larger) enclosing type.  However,
> +# the user may want to print a value casted to a different type,
> +# e.g. print (<type> []) <variable>.  This cast causes an update of the
> +# value's type.  In case of a pieced value, GDB failed to fetch the
> +# value's content.
> +# This test verifies that GDB can print a pieced value casted to a
> +# different type.
> +gdb_test "p (int \[\]) s1" " = \\{1\\, 0\\}"
> +gdb_test "p (short \[\]) s1" " = \\{1\\, 0\\, 0\\, <synthetic pointer>\\}"
> +
> +# Test for correct output if the size of the original object is not a
> +# multiple of the array's element size.
> +gdb_test "p s3" " = {a = 0, b = 1}"
> +gdb_test "p (int \[\]) s3" " = \\{0\\}"

I think you should match the warning here.  Since it's the behavior we
expect from GDB, we want to make sure there's no regression that causes
this warning to not be printed.  It's not tested at all elsewhere from
what I can see.

Simon