Re: [RFA] Function return type checking

[Klee Dienes <kdienes@apple.com>]

[-- Attachment #1: Type: text/plain, Size: 3091 bytes --]

On Saturday, March 9, 2002, at 08:30 PM, Andrew Cagney wrote:

> I've personally got reservations over introducing a change that, given 
> a file like:
>
> void
> b(void)
> {
>   double d;
>   d = (double) bar ((float) 3);
> }
>
> radically alters gdb's behavour given:
>
> 	print (double) bar((float) 3)
>
> and rejects (?)
>
> 	print bar((float) 3)

I think that's a very valid reservation.  But let me expand the example 
a bit:

$ cat a.c
double foo (float f)
{
   return (f + 1.0);
}

double bar (float f)
{
   return (f + 1.0);
}

$ cat b.c
double foo (float);

int main ()
{
   double d1, d2;

   d1 = (double) foo ((float) 3);
   d2 = (double) bar ((float) 3);
}

$ cc -c a.c
$ cc -g -c b.c
$ cc a.o b.o -o a

In "standard" GDB, you get the following results:

(gdb) print d1
$1 = 4
(gdb) print d2
$2 = 1 							/* generic bogus value */
(gdb) print (double) foo ((float) 3)
$3 = 1 							/* generic bogus value */
(gdb) print (double) bar ((float) 3)
$4 = 1 							/* generic bogus value */

In GDB as patched, you get:

(gdb) print d1
$1 = 4
(gdb) print d2
$2 = 1074266112 					/* generic bogus value */
(gdb) print (double) foo ((float) 3)
$3 = 4
(gdb) print (double) bar ((float) 3)
$4 = 4

I'd argue that the real problem here is that the calling conventions for 
a function call are determined for GCC and for GDB in two very different 
ways.  In the case of GCC, it's based on what signature GCC has seen for 
the function.  In the case of GDB, it's based on what debugging 
information is available for the function being called.  This means that 
it's always going to be possible for there to be a mismatch between 
behavior of GCC and GDB when evaluating expressions, and my patch does 
not try to address that.  My patch only changes the default behavior 
from two "incorrect" results (only one of which matches the behavior of 
GCC), to two "correct" results (only one of which matches the behavior 
of GCC).  I'd argue that it's better to err on the side of returning the 
correct result, if you have to err at all, and that in the case where 
you do want the exact behavior of GCC,

print (float) (int) bar ((float) 3)

is much more intuitive and easy to type than

print ((float (*) ()) bar) ((float) 3)

> and rejects (?)
>
> 	print bar((float) 3)

Correct; this would print:

Unable to call function at 0x1df8: no return type information available.
To call this function anyway, you can cast the return type explicitly 
(e.g. 'print (float) fabs (3.0)')

> Is this feature intended for C or ObjectiveC developers?

I'd intend for this to be used by everyone. We specifically added it in 
response to bug reports from people making heavy use of the system math 
libraries; as well as from Cocoa (Objective-C) developers making heavy 
use of functions returning NSRect objects.  The reason it's of interest 
to me in preparing the Objective-C patches is that much of the 
Objective-C GDB code makes use of  being able to pass 'expected_type' 
arguments to the modified functions, and I'd rather not have to 
re-architect all those calls before submitting the patches.

[-- Attachment #2: Type: text/enriched, Size: 3211 bytes --]

On Saturday, March 9, 2002, at 08:30 PM, Andrew Cagney wrote:


<excerpt>I've personally got reservations over introducing a change
that, given a file like:


void

b(void)

{

  double d;

  d = (double) bar ((float) 3);

}


radically alters gdb's behavour given:


	print (double) bar((float) 3)


and rejects (?)


	print bar((float) 3)

</excerpt>

I think that's a very valid reservation.  But let me expand the
example a bit:


$ cat a.c

double foo (float f)

{

  return (f + 1.0);

}


double bar (float f)

{

  return (f + 1.0);

}


$ cat b.c

double foo (float);


int main ()

{

  double d1, d2;


  d1 = (double) foo ((float) 3);

  d2 = (double) bar ((float) 3);

}


$ cc -c a.c

$ cc -g -c b.c

$ cc a.o b.o -o a


In "standard" GDB, you get the following results:


(gdb) print d1

$1 = 4

(gdb) print d2

$2 = 1 							/* generic bogus value */

(gdb) print (double) foo ((float) 3)

$3 = 1 							/* generic bogus value */

(gdb) print (double) bar ((float) 3)

$4 = 1 							/* generic bogus value */


In GDB as patched, you get:


(gdb) print d1

$1 = 4

(gdb) print d2

$2 = 1074266112 					/* generic bogus value */

(gdb) print (double) foo ((float) 3)

$3 = 4

(gdb) print (double) bar ((float) 3)

$4 = 4


I'd argue that the real problem here is that the calling conventions
for a function call are determined for GCC and for GDB in two very
different ways.  In the case of GCC, it's based on what signature GCC
has seen for the function.  In the case of GDB, it's based on what
debugging information is available for the function being called. 
This means that it's always going to be possible for there to be a
mismatch between behavior of GCC and GDB when evaluating expressions,
and my patch does not try to address that.  My patch only changes the
default behavior from two "incorrect" results (only one of which
matches the behavior of GCC), to two "correct" results (only one of
which matches the behavior of GCC).  I'd argue that it's better to err
on the side of returning the correct result, if you have to err at
all, and that in the case where you <italic>do</italic> want the exact
behavior of GCC,


print (float) (int) bar ((float) 3)


is much more intuitive and easy to type than


print ((float (*) ()) bar) ((float) 3)


<excerpt><color><param>0000,0000,DEDE</param>and rejects (?)


	print bar((float) 3)</color>

</excerpt>

Correct; this would print:


Unable to call function at 0x1df8: no return type information
available.

To call this function anyway, you can cast the return type explicitly
(e.g. 'print (float) fabs (3.0)')


<excerpt>Is this feature intended for C or ObjectiveC developers?

</excerpt>

I'd intend for this to be used by everyone. We specifically added it
in response to bug reports from people making heavy use of the system
math libraries; as well as from Cocoa (Objective-C) developers making
heavy use of functions returning NSRect objects.  The reason it's of
interest to me in preparing the Objective-C patches is that much of
the Objective-C GDB code makes use of  being able to pass
'expected_type' arguments to the modified functions, and I'd rather
not have to re-architect all those calls before submitting the patches.