Re: Parentheses expressions as actual arguments of elemental procedures

[Daniel Kraft <d at domob dot eu>]

Hi all,

Paul Richard Thomas wrote:
> Maybe you have done this already on #gfortran - could you get together
> to discuss how to attack the problem in the title line.  Mikael, you
> obviously are well into the scalarizer and would be able to help here
> - these actual arguments need to be written to a temporary array.  At
> present, the scalarizer treats them to a temporary scalar, element by
> element.  Daniel, I don't know if you have noted this but the problem
> is not restricted to subroutines and is, in fact, a scalarizer bug.

no, so far we've not discussed this between ourselves...  And shame on 
me it took me so long to post my promised comments on this problem :$

As far as I understand the scalarizer (and that's not very far) there's 
some "context" to hold "scalarized expressions" for base-expressions; 
when looking into trans-stmt.c:gfc_conv_elemental_dependencies, for each 
expression that should get a temporary there, it is looked up in this 
linked list.  So I guess when the scalarizer translates an expression, 
it will see if there's already a temporary or such for it in this linked 
list (ss->expr or so?).  So far as I've seen, this is only done for 
EXPR_VARIABLE expressions, though.  Would it be possible to do the same 
thing for other expressions as well?  I tried this with 
gfc_conv_elemental_dependencies, but those expressions were not in this 
linked list and could not be found.  Is it possible to simply add them 
there (when the ss/se/context/whatever is created?) or would this break 
something?

Another idea would be to take up my ideas for front-end temporaries 
again (that would be useful for finalization, too), see 
http://gcc.gnu.org/ml/fortran/2008-08/msg00026.html.  Then we could use 
this feature to generate all temporaries for actual argument expressions 
in ELEMENTAL calls on the higher level.

That's my ideas, and I would like to hear your opinions on them and 
comments from someone better into the scalarizer about point 1...

Yours,
Daniel

--
Done:  Arc-Bar-Cav-Rog-Sam-Val-Wiz
To go: Hea-Kni-Mon-Pri-Ran-Tou