(j3.2006) Irksome not to be able to invoke type-bound function on an expression

[Clune, Thomas L. GSFC-6101]

Unfortunately, I keep forgetting some implications with regard to limits on compilers identifying sub-expressions.    But even then, the expressiveness could be useful in many cases where the optimization is minimal or irrelevant.   We?re not going to outlaw user-defined temporaries, so when it matters, we keep doing what we are doing.



> On Sep 27, 2016, at 12:06 AM, Rafik Zurob <rzurob at ca.ibm.com> wrote:
> 
> One thing to watch out for with v%geod()%h_geod is performance if you have 
> multiple instances of it.  C++ gets away with it because you typically 
> #include your inlinable member function definitions into the same 
> translation unit and the compiler frontend or low-opt backend inlines 
> them.  Fortran type-bound procedures will typically be in a separate 
> compilation unit (e.g. a module) and typically will not be inlined unless 
> you enable expensive whole program optimization / link time optimization. 
> If you're only accessing one component of v%geod(), and accessing it only 
> once, you probably don't care.  But if you're accessing multiple 
> components or bindings of it, the explicit temp is a much better choice 
> that's not dependent on compiler optimization.  Also, note that if you're 
> using virtual or non-const (*) functions, the optimizer might not be able 
> to inline, and you might call v%geod() every time it appears in the 
> source.  The temp is an even better approach if v%geod() is expensive.
> 
> Regards
> 
> Rafik
> (*) const in the C++ sense.  I think we had a proposal for adding a 
> "virtuous" procedure prefix with the same meaning.  Pure functions can 
> still access global data.  So you can't replace
> x = pure_foo()
> y = pure_foo()
> with
> x = pure_foo()
> y = x
> without seeing the definition of pure_foo.  Similarly, if you have:
> x = v%geod()%h_geod
> y = v%geod()%h_geod
> and v%geod() is non-virtuous, you'd still execute the code in v%geod() 
> twice.
> 
> j3-bounces at mailman.j3-fortran.org wrote on 26/09/2016 09:31:53 PM:
> 
>> From: Van Snyder <Van.Snyder at jpl.nasa.gov>
>> To: j3 <j3 at j3-fortran.org>
>> Date: 26/09/2016 09:32 PM
>> Subject: (j3.2006) Irksome not to be able to invoke type-bound 
>> function on an expression
>> Sent by: j3-bounces at mailman.j3-fortran.org
>> 
>> I have a type ECR_t that represents Earth-centered-rotating Cartesian
>> coordinates as 3-vectors.  It has type-bound functions and operators to
>> add, subtract, scale, compute 2-norm, and other stuff.
>> 
>> I'd like to compute
>> 
>>  s = ( ( myP + myH * myP%grad_geoid() - line(1) )%norm2()
>> 
>> but this is prohibited, so I need either to export norm2() as an
>> ordinary (not type-bound) function and use
>> 
>>  s = norm2( myP + myH * myP%grad_geoid() - line(1) )
>> 
>> or
>> 
>>  temp = ( ( myP + myH * myP%grad_geoid() - line(1) )
>>  s = temp%norm2()
>> 
>> One particularly important case is function composition.  I'd like to
>> use something like
>> 
>>  s = a%f1()%f2()
>> 
>> Is there a good reason we can't eventually allow to invoke a type-bound
>> function using an expression, and in particular using the result of
>> another type-bound function?
>> 
>> A related question is whether we can eventually select components from a
>> function result (or more general expression).  I have a function Geod()
>> bound to ECR_t that computes 3-dimensional geodetic coordinates
>> (longitude, geodetic latitude, and geodetic height), from
>> Earth-centered-rotating Cartesian coordinates.  The iteration (Either
>> Bowring's or Fukushima's) that does this necessarily computes both
>> geodetic latitude and geodetic height.  But sometimes all I want is
>> geodetic latitude.  My type H_V_Geod that represents 3-dimensional
>> geodetic coordinates is an extension of the one H_Geod that represents
>> only longitude and geodetic latitude.
>> 
>> If I have
>> 
>>  type(ECR_t) :: V
>>  type(H_Geod) :: Geod
>> 
>> I can't get the geodetic surface components (longitude and geodetic
>> latitude) of V using
>> 
>>  geod = v%geod()%h_geod
>> 
>> I need something like
>> 
>>  type(H_V_Geod) :: Temp
>>  temp = v%geod()
>>  geod = temp%h_geod
>> 
>> Is there a good reason we can't eventually allow to select a component
>> of an expression?
>> 
>> 
>> _______________________________________________
>> J3 mailing list
>> J3 at mailman.j3-fortran.org
>> http://mailman.j3-fortran.org/mailman/listinfo/j3
>> 
> 
> 
> _______________________________________________
> J3 mailing list
> J3 at mailman.j3-fortran.org
> http://mailman.j3-fortran.org/mailman/listinfo/j3