Re: [PING^3] Wconversion: fixes for C++ front-end

[Gabriel Dos Reis <gdr at cs dot tamu dot edu>]

On Wed, 7 Feb 2007, Manuel López-Ibáñez wrote:

| On 06 Feb 2007 18:46:44 -0600, Gabriel Dos Reis <gdr@cs.tamu.edu> wrote:
| > "Manuel López-Ibáñez" <lopezibanez@gmail.com> writes:
| >
| > [...]
| >
| > | So the basic differences are the call to fold_if_not_in_template and
| > | the use of LOOKUP_NO_CONVERSION when calling ocp_convert. I honestly
| > | cannot see how to relate cp_convert and convert.
| >
| > In that case, there does not seem reason to believe that the chain you
| > introduced is correct.  We must conservatitely regard it as incorrect
| > until proven otherwise.  the conversion code in cp/ has proven to be
| > tricky and sources of bugs, hence the caution that must be exercise.
|
| That is the reason for not using the return value of
| convert_and_check.

actually it is a reason to be careful about calling the conversion
functions.  I suspect I did not make my previous points clear enough:

   (1) you must understand why you call each of the functions
   (2) given that both convert() and cp_convert() convert, you must
       understand their relations.  Just quoting the definitions at me
       is NOT what I was trying to get at.
   (4) if you must call one function that does conversion, potentially
       issues diagnostic but throws its result away, just to call
       another conversion function that does similar computations, then
       something is wrong.
   (5) you must be able to explain your patch, what it does and why it
       does it.

The difference here is in the flag parameters. Before a new function
cp_convert_and_check is introduced, you must explain why the existing
ones are insufficient -- that brings you back to understanding
ocp_convert.


Conversions in C++ can involve not just standard conversions (implicit
conversions defined for built-in types), but also user-defined
conversions.  There is a well-defined ordering in which they must
happen -- see clause 4 of the C++ standard.

An implicit conversion sequence in C++ (they actions you need to take
to bring arguments to parameter types, etc.) in general is:

  (1) standard conversion sequence, followed by
  (2) a possible user-defined conversion, followed by
  (3) another standard conversion sequence.

In both (1) and (3) you want to catch non-value preserving
conversions.  And ocp_convert's job is to implement each step of the
conversion sequences.

Both convert() and cp_convert() essentially asks for doing the same
kind of conversions.  They differ in "who" is allowed to participating
in the conversion sequence.  More precisely, convert() cares only
about _standard conversion_ sequences.  cp_convert() tolerates
user-defined conversions.  It looks to me that you wan to have
ocp_convert() actually warns if necessary, instead of doing the
computations twice and throwing one results away.

-- Gaby