Implement floordiv operator for gdb.Value

  This is my attempt to implement the // operator on gdb.Value objects.
There is already BINOP_INTDIV which works fine for integral types, but
for floats I use BINOP_DIV and then call floor() on the result. This
doesn't support decimal floats though.

Is this a reasonable solution? Is the test sufficient?

I have a follow-up patch which changes the meaning of the / operator
for gdb.Value when built against Python 3, to be consistent with
Python (see comment 1 in the Bugzilla PR) but I expect that to be more
controversial :-)


gdb/ChangeLog:
2016-09-20  Jonathan Wakely  <jwakely@redhat.com>

	PR python/20624
	* python/py-value.c (VALPY_FLOORDIV): Define new enumerator.
	(valpy_binop_throw): Handle VALPY_FLOORDIV.
	(valpy_floordiv): New function.
	(value_object_as_number): Set valpy_floordiv in relevant slot.

gdb/testsuite/ChangeLog:
2016-09-20  Jonathan Wakely  <jwakely@redhat.com>

	PR python/20624
	* gdb.python/py-value.exp: Test floor division.
commit 3a019f3ba61bde5320419f5782d8f2554ac55ace
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Tue Sep 20 10:38:35 2016 +0100

    gdb: Implement floordiv operator for gdb.Value
    
    gdb/ChangeLog:
    2016-09-20  Jonathan Wakely  <jwakely@redhat.com>
    
    	PR python/20624
    	* python/py-value.c (VALPY_FLOORDIV): Define new enumerator.
    	(valpy_binop_throw): Handle VALPY_FLOORDIV.
    	(valpy_floordiv): New function.
    	(value_object_as_number): Set valpy_floordiv in relevant slot.
    
    gdb/testsuite/ChangeLog:
    2016-09-20  Jonathan Wakely  <jwakely@redhat.com>
    
    	PR python/20624
    	* gdb.python/py-value.exp: Test floor division.
  

Hi there,

Thanks!

On 09/20/2016 02:26 PM, Jonathan Wakely wrote:
> This is my attempt to implement the // operator on gdb.Value objects.
> There is already BINOP_INTDIV which works fine for integral types, but
> for floats I use BINOP_DIV and then call floor() on the result. This
> doesn't support decimal floats though.
> 
> Is this a reasonable solution? Is the test sufficient?
> 

See below.

> @@ -1142,7 +1160,15 @@ valpy_binop_throw (enum valpy_opcode opcode, PyObject *self, PyObject *other)
>      }
>  
>    if (res_val)
> -    result = value_to_value_object (res_val);
> +    {
> +      if (floor_it)
> +        {
> +          double d = value_as_double (res_val);

Should be s/double/DOUBLEST, I suppose?

> +          d = floor (d);
> +          res_val = value_from_double (value_type (res_val), d);
> +        }
> +      result = value_to_value_object (res_val);
> +    }
>  
>    do_cleanups (cleanup);
>    return result;
> @@ -1200,6 +1226,12 @@ valpy_remainder (PyObject *self, PyObject *other)
>  }
>  
>  static PyObject *
> +valpy_floordiv (PyObject *self, PyObject *other)
> +{
> +  return valpy_binop (VALPY_FLOORDIV, self, other);
> +}
> +
> +static PyObject *
>  valpy_power (PyObject *self, PyObject *other, PyObject *unused)
>  {
>    /* We don't support the ternary form of pow.  I don't know how to express
> @@ -1837,7 +1869,7 @@ static PyNumberMethods value_object_as_number = {
>    NULL,                       /* nb_inplace_and */
>    NULL,                       /* nb_inplace_xor */
>    NULL,                       /* nb_inplace_or */
> -  NULL,                       /* nb_floor_divide */
> +  valpy_floordiv,             /* nb_floor_divide */
>    valpy_divide,               /* nb_true_divide */
>    NULL,			      /* nb_inplace_floor_divide */
>    NULL,			      /* nb_inplace_true_divide */
> diff --git a/gdb/testsuite/gdb.python/py-value.exp b/gdb/testsuite/gdb.python/py-value.exp
> index 57a9ba1..81837e9 100644
> --- a/gdb/testsuite/gdb.python/py-value.exp
> +++ b/gdb/testsuite/gdb.python/py-value.exp
> @@ -87,6 +87,8 @@ proc test_value_numeric_ops {} {
>    gdb_test "python print ('result = ' + str(f/g))" " = 0.5" "divide two double values"
>    gdb_test "python print ('result = ' + str(i%j))" " = 1" "take remainder of two integer values"
>    # Remainder of float is implemented in Python but not in GDB's value system.
> +  gdb_test "python print ('result = ' + str(i//j))" " = 2" "floor-divide two integer values"
> +  gdb_test "python print ('result = ' + str(f//g))" " = 0" "floor-divide two double values"

Is the "two double values" test returning an integer somehow?

I ask because IIUC, regardless of Python version, a floor-divide
involving a float should result in a float, while a floor-divide of
integers should result in an integer.  And that's what the patch looks
like should end up with.  So I was expecting to see "0.0" in
the "two double values" case:

 (gdb) python print (5.0//6.0)
 0.0
 (gdb) python print (5//6)
 0

I think it'd be good to test with negative numbers too, to make
sure that we round (and keep rounding) toward the same
direction Python rounds:

 (gdb) python print (8.0//-3)
 -3.0
 (gdb) python print (8//-3)
 -3
 (gdb) print 8/-3
 $1 = -2

Thanks,
Pedro Alves
  

On 20/09/16 16:33 +0100, Pedro Alves wrote:
>Hi there,
>
>Thanks!
>
>On 09/20/2016 02:26 PM, Jonathan Wakely wrote:
>> This is my attempt to implement the // operator on gdb.Value objects.
>> There is already BINOP_INTDIV which works fine for integral types, but
>> for floats I use BINOP_DIV and then call floor() on the result. This
>> doesn't support decimal floats though.
>>
>> Is this a reasonable solution? Is the test sufficient?
>>
>
>See below.
>
>> @@ -1142,7 +1160,15 @@ valpy_binop_throw (enum valpy_opcode opcode, PyObject *self, PyObject *other)
>>      }
>>
>>    if (res_val)
>> -    result = value_to_value_object (res_val);
>> +    {
>> +      if (floor_it)
>> +        {
>> +          double d = value_as_double (res_val);
>
>Should be s/double/DOUBLEST, I suppose?

OK - if I do that then floor(d) will convert it back to double,
unless you #include <cmath> and using std::floor, so that the overload
for long double is visible (in C++ <math.h> names like floor are
overloaded so you don't need to use floorf/floor/floorl according to
the type).

>> diff --git a/gdb/testsuite/gdb.python/py-value.exp b/gdb/testsuite/gdb.python/py-value.exp
>> index 57a9ba1..81837e9 100644
>> --- a/gdb/testsuite/gdb.python/py-value.exp
>> +++ b/gdb/testsuite/gdb.python/py-value.exp
>> @@ -87,6 +87,8 @@ proc test_value_numeric_ops {} {
>>    gdb_test "python print ('result = ' + str(f/g))" " = 0.5" "divide two double values"
>>    gdb_test "python print ('result = ' + str(i%j))" " = 1" "take remainder of two integer values"
>>    # Remainder of float is implemented in Python but not in GDB's value system.
>> +  gdb_test "python print ('result = ' + str(i//j))" " = 2" "floor-divide two integer values"
>> +  gdb_test "python print ('result = ' + str(f//g))" " = 0" "floor-divide two double values"
>
>Is the "two double values" test returning an integer somehow?
>
>I ask because IIUC, regardless of Python version, a floor-divide
>involving a float should result in a float, while a floor-divide of
>integers should result in an integer.  And that's what the patch looks
>like should end up with.  So I was expecting to see "0.0" in
>the "two double values" case:
>
> (gdb) python print (5.0//6.0)
> 0.0
> (gdb) python print (5//6)
> 0

This seems to be an existing property of gdb.Value, as even using the
normal division operator (and without my patch) I see floats printed
without a decimal part when they are an integer value:

(gdb) python print (gdb.Value(5.0)/5.0)
1
(gdb) python print (5.0/5.0)
1.0



>I think it'd be good to test with negative numbers too, to make
>sure that we round (and keep rounding) toward the same
>direction Python rounds:
>
> (gdb) python print (8.0//-3)
> -3.0
> (gdb) python print (8//-3)
> -3
> (gdb) print 8/-3
> $1 = -2

Good point, I'll do that.
  

On 20/09/16 17:35 +0100, Jonathan Wakely wrote:
>On 20/09/16 16:33 +0100, Pedro Alves wrote:
>>>@@ -1142,7 +1160,15 @@ valpy_binop_throw (enum valpy_opcode opcode, PyObject *self, PyObject *other)
>>>     }
>>>
>>>   if (res_val)
>>>-    result = value_to_value_object (res_val);
>>>+    {
>>>+      if (floor_it)
>>>+        {
>>>+          double d = value_as_double (res_val);
>>
>>Should be s/double/DOUBLEST, I suppose?
>
>OK - if I do that then floor(d) will convert it back to double,
>unless you #include <cmath> and using std::floor, so that the overload
>for long double is visible (in C++ <math.h> names like floor are
>overloaded so you don't need to use floorf/floor/floorl according to
>the type).

P.S. In theory it should work with <math.h> and without the
using-declaration, but with GCC that wasn't true until GCC 6.1 (see
http://developerblog.redhat.com/2016/02/29/why-cstdlib-is-more-complicated-than-you-might-think/
for the gory details).

So to get the right result for older versions of GCC you need <cmath>.
  

On 09/20/2016 05:35 PM, Jonathan Wakely wrote:
> On 20/09/16 16:33 +0100, Pedro Alves wrote:
>> Hi there,
>>
>> Thanks!
>>
>> On 09/20/2016 02:26 PM, Jonathan Wakely wrote:
>>> This is my attempt to implement the // operator on gdb.Value objects.
>>> There is already BINOP_INTDIV which works fine for integral types, but
>>> for floats I use BINOP_DIV and then call floor() on the result. This
>>> doesn't support decimal floats though.
>>>
>>> Is this a reasonable solution? Is the test sufficient?
>>>
>>
>> See below.
>>
>>> @@ -1142,7 +1160,15 @@ valpy_binop_throw (enum valpy_opcode opcode,
>>> PyObject *self, PyObject *other)
>>>      }
>>>
>>>    if (res_val)
>>> -    result = value_to_value_object (res_val);
>>> +    {
>>> +      if (floor_it)
>>> +        {
>>> +          double d = value_as_double (res_val);
>>
>> Should be s/double/DOUBLEST, I suppose?
> 
> OK - if I do that then floor(d) will convert it back to double,
> unless you #include <cmath> and using std::floor, so that the overload
> for long double is visible (in C++ <math.h> names like floor are
> overloaded so you don't need to use floorf/floor/floorl according to
> the type).

OK.  I remember reading your blog about this mess a while ago.

If easy to do, sounds like we should just do it.  OOC, would calling 
std::floor directly instead of using "using" work just as well?

(This kind of raises the question of which float type / format / representation
to use for arithmetic here -- host's or target's.  gdb currently always
uses host's, but that's a much larger issue that we can just
continue to ignore.)

>> Is the "two double values" test returning an integer somehow?
>>
>> I ask because IIUC, regardless of Python version, a floor-divide
>> involving a float should result in a float, while a floor-divide of
>> integers should result in an integer.  And that's what the patch looks
>> like should end up with.  So I was expecting to see "0.0" in
>> the "two double values" case:
>>
>> (gdb) python print (5.0//6.0)
>> 0.0
>> (gdb) python print (5//6)
>> 0
> 
> This seems to be an existing property of gdb.Value, as even using the
> normal division operator (and without my patch) I see floats printed
> without a decimal part when they are an integer value:
> 
> (gdb) python print (gdb.Value(5.0)/5.0)
> 1
> (gdb) python print (5.0/5.0)
> 1.0

Curious.  Off hand looks like a bug to me.  But since it's orthogonal
to your patch, let's leave it.

>> I think it'd be good to test with negative numbers too, to make
>> sure that we round (and keep rounding) toward the same
>> direction Python rounds:
>>
>> (gdb) python print (8.0//-3)
>> -3.0
>> (gdb) python print (8//-3)
>> -3
>> (gdb) print 8/-3
>> $1 = -2
> 
> Good point, I'll do that.

Thanks,
Pedro Alves
  

> On Sep 20, 2016, at 12:35 PM, Jonathan Wakely <jwakely@redhat.com> wrote:
> 
>> ...
> 
> This seems to be an existing property of gdb.Value, as even using the
> normal division operator (and without my patch) I see floats printed
> without a decimal part when they are an integer value:
> 
> (gdb) python print (gdb.Value(5.0)/5.0)
> 1
> (gdb) python print (5.0/5.0)
> 1.0

In all this, please keep in mind that this is one place where Python 2 and Python 3 differ:

$ python
Python 2.7.6 (v2.7.6:3a1db0d2747e, Nov 10 2013, 00:42:54) 
[GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> 1/2
0
>>> ^D
$ python3
Python 3.5.1 (v3.5.1:37a07cee5969, Dec  5 2015, 21:12:44) 
[GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> 1/2
0.5
>>> ^D

Python 2 can be told to do it the new way:

$ python2
Python 2.7.6 (v2.7.6:3a1db0d2747e, Nov 10 2013, 00:42:54) 
[GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> from __future__ import division
>>> 1/2
0.5
>>> 

	paul
  

On 20/09/16 17:57 +0100, Pedro Alves wrote:
>On 09/20/2016 05:35 PM, Jonathan Wakely wrote:
>> On 20/09/16 16:33 +0100, Pedro Alves wrote:
>>> Hi there,
>>>
>>> Thanks!
>>>
>>> On 09/20/2016 02:26 PM, Jonathan Wakely wrote:
>>>> This is my attempt to implement the // operator on gdb.Value objects.
>>>> There is already BINOP_INTDIV which works fine for integral types, but
>>>> for floats I use BINOP_DIV and then call floor() on the result. This
>>>> doesn't support decimal floats though.
>>>>
>>>> Is this a reasonable solution? Is the test sufficient?
>>>>
>>>
>>> See below.
>>>
>>>> @@ -1142,7 +1160,15 @@ valpy_binop_throw (enum valpy_opcode opcode,
>>>> PyObject *self, PyObject *other)
>>>>      }
>>>>
>>>>    if (res_val)
>>>> -    result = value_to_value_object (res_val);
>>>> +    {
>>>> +      if (floor_it)
>>>> +        {
>>>> +          double d = value_as_double (res_val);
>>>
>>> Should be s/double/DOUBLEST, I suppose?
>>
>> OK - if I do that then floor(d) will convert it back to double,
>> unless you #include <cmath> and using std::floor, so that the overload
>> for long double is visible (in C++ <math.h> names like floor are
>> overloaded so you don't need to use floorf/floor/floorl according to
>> the type).
>
>OK.  I remember reading your blog about this mess a while ago.
>
>If easy to do, sounds like we should just do it.  OOC, would calling
>std::floor directly instead of using "using" work just as well?

Yes, and that's exactly what I did locally a few minutes ago. Calling
std::floor explicitly probably makes it more obvious that it relies on
the C++ overloads to preserve the right type.

>(This kind of raises the question of which float type / format / representation
>to use for arithmetic here -- host's or target's.  gdb currently always
>uses host's, but that's a much larger issue that we can just
>continue to ignore.)

I'm happy to ignore it :-)

>>> Is the "two double values" test returning an integer somehow?
>>>
>>> I ask because IIUC, regardless of Python version, a floor-divide
>>> involving a float should result in a float, while a floor-divide of
>>> integers should result in an integer.  And that's what the patch looks
>>> like should end up with.  So I was expecting to see "0.0" in
>>> the "two double values" case:
>>>
>>> (gdb) python print (5.0//6.0)
>>> 0.0
>>> (gdb) python print (5//6)
>>> 0
>>
>> This seems to be an existing property of gdb.Value, as even using the
>> normal division operator (and without my patch) I see floats printed
>> without a decimal part when they are an integer value:
>>
>> (gdb) python print (gdb.Value(5.0)/5.0)
>> 1
>> (gdb) python print (5.0/5.0)
>> 1.0
>
>Curious.  Off hand looks like a bug to me.  But since it's orthogonal
>to your patch, let's leave it.
>
>>> I think it'd be good to test with negative numbers too, to make
>>> sure that we round (and keep rounding) toward the same
>>> direction Python rounds:
>>>
>>> (gdb) python print (8.0//-3)
>>> -3.0
>>> (gdb) python print (8//-3)
>>> -3
>>> (gdb) print 8/-3
>>> $1 = -2
>>
>> Good point, I'll do that.

Definitely worth testing because my patch gives the wrong result:

(gdb) python print(  gdb.Value(8) // gdb.Value(-3) )
-2

This is because I'm using the same BINOP_DIV operation as GDB's 8/-3
uses, but Python's // should round to -inf. I'll fix that.
  

On 20/09/16 17:00 +0000, Paul.Koning@dell.com wrote:
>
>> On Sep 20, 2016, at 12:35 PM, Jonathan Wakely <jwakely@redhat.com> wrote:
>>
>>> ...
>>
>> This seems to be an existing property of gdb.Value, as even using the
>> normal division operator (and without my patch) I see floats printed
>> without a decimal part when they are an integer value:
>>
>> (gdb) python print (gdb.Value(5.0)/5.0)
>> 1
>> (gdb) python print (5.0/5.0)
>> 1.0
>
>In all this, please keep in mind that this is one place where Python 2 and Python 3 differ:

Right, see https://sourceware.org/ml/gdb-patches/2016-09/msg00221.html
for a crazy idea that would make gdb.Value match the Python version
it's built against.

That wouldn't help for Python 2 using __future__ division though.
  

> On Sep 20, 2016, at 1:11 PM, Jonathan Wakely <jwakely@redhat.com> wrote:
> 
> On 20/09/16 17:00 +0000, Paul.Koning@dell.com wrote:
>> 
>>> On Sep 20, 2016, at 12:35 PM, Jonathan Wakely <jwakely@redhat.com> wrote:
>>> 
>>>> ...
>>> 
>>> This seems to be an existing property of gdb.Value, as even using the
>>> normal division operator (and without my patch) I see floats printed
>>> without a decimal part when they are an integer value:
>>> 
>>> (gdb) python print (gdb.Value(5.0)/5.0)
>>> 1
>>> (gdb) python print (5.0/5.0)
>>> 1.0
>> 
>> In all this, please keep in mind that this is one place where Python 2 and Python 3 differ:
> 
> Right, see https://sourceware.org/ml/gdb-patches/2016-09/msg00221.html
> for a crazy idea that would make gdb.Value match the Python version
> it's built against.
> 
> That wouldn't help for Python 2 using __future__ division though.

Yes.  You'd expect that this would be easy to find, but that isn't the case.  It may be accessible from C code, if you can find the parser flags from where you are.  Look for CO_FUTURE in the source tree.

	paul