Re: and then

[sampson@cod.nosc.mil (Charles H. Sampson)]

In article <1993Apr7.173556.17184@sei.cmu.edu> ae@sei.cmu.edu (Arthur Evans) wr
ites:
>In connection with a discussion of Ada's short-circuit control forms,
>sampson@nosc.mil (Charles H. Sampson) says
>    Ada compiler writers appear to have abdicated their responsibility
>    and make no attempt to determine when boolean operations can be
>    optimized as short-circuit.
>
>Well, I'm not surprised that they make no such attempt, but it's not
>because they have abdicated any responsibility.  ARM 4.5(5) says of
>expressions involving 'and' and 'or':
>    The operands ... are evaluated in some order that is not defined by
>    the language (but before application of the corresponding operator).
>In other words, both operands MUST be evaluated.  That requirement is
>not unreasonable given the availability of the short-circuit forms.

     The nature of optimization is to generate code other than the obvious
_provided the semantics of the program are unchanged_.  (Ada is one of the
few languages to insist on that.)  Furthermore, a general rule in compiler
writing is "no harm, no foul".  If a programmer writes

                            if X > 1 and Y > 1

I find it hard to believe that any protest would be raised if it were dis-
covered (by reading the generated code) that the compiler had optimized
this as

                         if X > 1 and then Y > 1 .

     A large percentage of boolean expressions that contain a binary opera-
tor are of such simple forms and are worthy of optimization, in my opinion. 
When you look at something like

                          if X > 1 and X * Y < 4

it gets more complicated because of the possibility that NUMERIC_ERROR
might be raised.  It appears to me that the letter of 11.6(4) would allow
short-circuit in this case, but I'm not sure that that's the spirit of this
paragraph.  Whatever it might be, I would still initially write the above
form and even write such things as

                          if X > 1 and F(X) < 5 ,

which only a heroic optimizer could decide anything about, and change to
short-circuit only if execution was unacceptable and tuning analysis showed
that this statement was part of the problem.  The last form clearly indi-
cates to the maintainer that evaluation of F(X) is not dependent on X being
larger than 1.  If forced to change to short-circuit, I would add comments
stating this non-dependency.

                                   Charlie

P. S. I've just received information that the Verdix compilers now do such
transformations when they can verify that the semantics are unchanged.