Re: Government Policy on Ada Acquisitions

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

It is true that GNAT is not finished, but it is using the GCC backend which is
(a) quite mature at this point and
(b) highly optimizing

If you write a chunk of Ada code that is essentially equivalent to some C code,
and of course you turn on the optimizer (-O2), you get code that is equivalent
to what you would get from C, which means in practice very good, comparable to
the best C compilers around. Someone at the London SigAda remarked to me that
he was amazed at the Whetstone performance of GNAT, "it was essentially the
same as C". I would have been amazed if this were NOT the case.

Now Ada 9X specific stuff, such as aggregates, pointers to unconstrained arrays,etc. is still quite unoptimizied. The performance of GNAT is by no means terrible
for such things, but it can be greatly improved. Right now, that kind of
optimization is of course low priority for us.

Similarly, the library packages can stand a lot of improvement. Until the most
recent release, Text_IO was checking (with a system call) for the console case
on every character. That was really a bug, but not one that showed up as a
functionality problem in our tests. That bug iis now fixed, but there may
well be other similar performance problems.

In any case, running GCC unoptimized and reporting its performance is quite
ludicrous and gives no useful hints at all about the real performance of
GNAT code. The compiler itself is always generated using -O2. It would not
occur to us to leave optimization off. Similarly all our test programs are
typically run using -O2. Indeed it is often the case that the compiler is
more reliable with optimization turned on, simply because that's the way
people run it for any real code.