[petsc-dev] configure: generating guesses from liblist and includes

[Jed Brown]

Satish Balay <balay at mcs.anl.gov> writes:

> On Thu, 16 May 2013, Jed Brown wrote:
>
>> Matthew Knepley <knepley at gmail.com> writes:
>> 
>> > I don't think this makes sense. A certain pair of library and include
>> > makes sense. Decoupling these invites all sorts of mismatches which
>> > would be hard to track down. I do not see what is wrong with the
>> > strategy above?
>> 
>> It causes confusing error conditions and in almost all cases, the same
>> includes are used for all library guesses.
>> 
>
> Even for just the libraries I think we needed:
> try (liba) if all true (test1,test2,test3) then accept (liba)
> else try (libb) if all true (test1,test2,test3) then   accept (libb)
>
> hence the current code. Perhaps we need multiple semantics for
> different situations.

The problem that after finding a _library_ that works, then finding that
the includes do not work, and we go on a tangent trying lots of
different libs with that same broken include.  The include does not
depend on the library so I think it should be tested independently.

Note that if multiple includes need to be tried, the current approach requires

  num_include_alternatives * num_lib_alternatives

but it could be tested in only

  num_include_alternatives + num_lib_alternatives

Use of multiple includes is rare now, so I don't know if this would
significantly affect speed.