Re: Array mappings

[dewar@cs.nyu.edu (Robert Dewar)]

Mike wonders about the difficulty of applying traditional fortran optimization
techniques to Ada.

There is no conceptual problem in writing highly optimizing Ada backends.
I don't think the "highly structured for loop" of Ada is any particular
help, it is essentially identical to the do loop of Fortran in any case.

THe only conceptual problem arises as a result of the dynamic lower bounds,
which threaten to generate extra instructions in some cases, although
address reduction in loops will generally eliminate this extra work in
most loops. Another approach is to use virtual origins for arrays, which
is our eventual plan in GNAT.

The difficulty is that the amount of technical resources to do a highly
optimizing backend of this case is HUGE, and I can't believe it would have
been worth the investment of tens of person years to develop Ada optimization
to this level, especially early on. Furthermore, such an expenditure of
effort would have harmed stability and reliability.

Fortran users tend to be more concerned with speed than with reliability (the
infamous Cray division which gives wrong results very fast is an indicator of
this general attitude), and that of course is the wrong way round for most
Ada users.

Obviously the way to get high performance is to borrow an existing backend,
which of course is what we have done with GNAT. The performance of GNAT for
a simple arithmetic loop should be entirely comparable to that of 
GNU Fortran or GNU C, which is pretty darned good, but probably STILL not
as good as the really hot commercial Fortran compilers on which huge
optimziation efforts have been spent.