Re: GDB returns wrong type when traversing optimized-out Fields

On 2018-02-04 11:06 PM, Roman Popov wrote:
> I apologize for code typo in previous email. Here is correct code sample:
> 
> template <unsigned v1, unsigned v2>
> struct TRAITS {
>     static const unsigned val1 = v1;
>     static const unsigned val2 = v2;
> };
> template < class TRAITS >
> struct foo {
>     static const unsigned x1 = TRAITS::val1;
>     static const unsigned x2 = TRAITS::val2;
> };
> 
> int main () {
>     foo<TRAITS<1,2>> f1;
>     // SET BREAKPOINT HERE
>     return 0;
> }
> 
> -Roman
> 
> 
> 2018-02-04 20:02 GMT-08:00 Roman Popov <[email protected]>:
> 
>> Hi all,
>> I've encountered strange GDB behavior when requesting a value of
>> optimized-away field.
>> Instead of returning None or raising exception, GDB returns an
>> optimizied-out value of wrong type.
>>
>> Here is a small reproducer *optimize_out.cpp*:
>>
>> template <unsigned v1, unsigned v2>
>> struct TRAITS {
>>     static const unsigned val1 = v1;
>>     static const unsigned val2 = v2;
>> };
>> template < class TRAITS >
>> struct foo {
>>     static const unsigned x1 = TRAITS::v1;
>>     static const unsigned x2 = TRAITS::v2;
>> };
>>
>> int main () {
>>     foo<TRAITS<1,2>> f1;
>>     // SET BREAKPOINT HERE
>>     return 0;
>> }
>>
>> # Using g++ 7.3
>> $ g++ -g optimize_out.cpp
>>
>> # Using gdb 8.1
>> $ gdb a.out
>>
>> (gdb) break optimize_out.cpp:14
>> (gdb) r
>> (gdb) p f1
>> $1 = {static x1 = <optimized out>, static x2 = <optimized out>}
>>
>> Ok, looks good.  Now traverse fields:
>>
>> (gdb) python
>>> f1 = gdb.parse_and_eval("f1")
>>> for field in f1.type.fields():
>>>    print ("field name: ", field.name, "field type: ", field.type)
>>>    field_val = f1[field]
>>>    print ("optout?: ",field_val.is_optimized_out, "type:
>> ",field_val.type)
>>> end
>> field name:  x1 field type:  const unsigned int
>> optout?:  True type:  foo<TRAITS<1, 2> >
>> field name:  x2 field type:  const unsigned int
>> optout?:  True type:  foo<TRAITS<1, 2> >
>>
>>
>>
>> So type we get is foo<TRAITS<1, 2> >, not unsigned int.
>>
>> Looks like GDB-MI has same behavior. At least this code sample totatlly
>> confuses GDB GUI I use.
>>
>> Thanks,
>> Roman
>>

Hi Roman,

It just seems to be that we return a value with the type of the structure
instead of the type of the field in an error path somewhere.   It causes
these weird behaviors in the CLI too:

(gdb) p f1
$1 = {static x1 = <optimized out>, static x2 = <optimized out>}
(gdb) p f1.x1
$2 = <optimized out>
(gdb) ptype f1.x1
type = struct foo<TRAITS<1, 2> > [with TRAITS = TRAITS<1, 2>] {
    static const unsigned int x1;
    static const unsigned int x2;
}
(gdb) ptype f1.x1.x1
type = struct foo<TRAITS<1, 2> > [with TRAITS = TRAITS<1, 2>] {
    static const unsigned int x1;
    static const unsigned int x2;
}
(gdb) p f1.x1.x1.x1.x1
$3 = <optimized out>

Can you try the patch below?  I did not run the testsuite on it.


From ef7f73557e291ed2d8bc7f175765bdbb91e2c817 Mon Sep 17 00:00:00 2001
From: Simon Marchi <[email protected]>
Date: Mon, 5 Feb 2018 00:34:08 -0500
Subject: [PATCH] Return right type in value_static_field

---
 gdb/value.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/gdb/value.c b/gdb/value.c
index 9a144fb7fb..063f57129a 100644
--- a/gdb/value.c
+++ b/gdb/value.c
@@ -2978,7 +2978,10 @@ value_static_field (struct type *type, int fieldno)
 	    = lookup_minimal_symbol (phys_name, NULL, NULL);

 	  if (!msym.minsym)
-	    return allocate_optimized_out_value (type);
+	    {
+	      struct type *field_type = TYPE_FIELD_TYPE (type, fieldno);
+	      return allocate_optimized_out_value (field_type);
+	    }
 	  else
 	    {
 	      retval = value_at_lazy (TYPE_FIELD_TYPE (type, fieldno),
-- 
2.16.1