Path: utzoo!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!yale!cmcl2!lanl!lambda!jlg
From: jlg@lambda.UUCP (Jim Giles)
Newsgroups: comp.lang.misc
Subject: Re: NOT Educating FORTRAN programmers to use C
Message-ID: <14199@lambda.UUCP>
Date: 19 Jan 90 23:26:30 GMT
References: <12950@cbnewsc.ATT.COM>
Lines: 84

From article <12950@cbnewsc.ATT.COM>, by res@cbnewsc.ATT.COM (Rich Strebendt):
> [...]
> It seems that this discussion has gotten to the "my optimizer is
> bigger than your optimizer" stage.  [...]

Not at all.  So far, the discussion has been restricted to things that
could be optimized in one language and not in the other.  Fortran permits
optimization of procedure arguments that C must assume are aliased - no
optimizer, no matter how sophisticated, can eliminate this difference.
Fortran allows intrinsic functions to be optimized, C hasn't any intrinsics
(I know, the new standard permits such - show me an available implementation
before you mention it again).

On the other hand, C can optimize procedure calls (or even 'inline' them)
provided that the procedure and the call are in the same scope (file).
I don't know of any C implementation that actually does this, but it _is_
permitted by the existing (de facto) language specification.  (Note that
Fortran 90 also allows this for internal procedures and for procedure
calls within MODULES.  Just like C and intrinsics, this is not a fair
comparison until such time as implementations exist which actually perform
the optimization.)

> [... C is better because it ...] allows me to better describe my INTENT
> to the compiler and let it generate code that needs LESS optimization.

Actually, C is worse at this in some ways and only average in others.
For example, When I wish to pass an array to a procedure, Fortran (and
Pascal, Modula2, ADA, ...) lets me do so.  In C, the procedure always
sees a pointer instead of an array.  Another example: when writing
numerical programs you often need to force the order of evaluation
of an expression to avoid over/underflow or to eliminate cancellation
in subtractions.  Fortran (pascal, Modula2, ADA, ...) lets me do so
with parenthesis - C forces me to introduce temporary variables and
_hope_ that the optimizer doesn't actually do a store/load.

> [...]                                                   Similarly,
> using the x += y construct instead of X=X+Y will be a winner for the
> compiler (especially if x is a multidimensional array) and will
> require no pattern matching abilities in the optimizer to detect --

This is a particularly bad example for you to use.  The optimization
that you want the compiler to do without is called "common subexpression
elimination".  It is important that your compiler know how to do this
optimization even if you do have a '+=' operator.  I wouldn't ever buy
a compiler which didn't have this capability.  A compiler without this
capability is somewhat behind the state-of-the-art anyway (by about 30
years).

Another reason this is a bad example is the nature of the operator itself.
It is obviously very easy to implement.  This kind of thing was known
before C was invented.  Certainly language designers have known about it
in C for the last 15-18 years.  YET: there has been no stampede among
language designers to put these operators into their languages.  Nor has
there been a groundswell of user support (neither to the standards
committees nor to the vendors) for these operators to be added to existing
languages.  In fact, aside from C users, there seems to be no interest
in these operators at all.

One last reason that this is a bad example is that these operators work
on the expression level and not on the statement level.  This is known
(from several language design experiments) to be harmful to user's
productivity.

> [...]
> (such as incrementing a pointer rather than incrementing an index for
> an array then recomputing the address of the array element) [...]

Once again, the optimization you refer to has a name: "strength reduction".
Once again, I wouldn't _have_ a compiler which didn't do this.  A compiler
which fails to do this is, once again, slightly behind the times (with
30 years again being a fair estimate).  In fact, the use of arrays (which
are _NOT_ aliased to each other) is a direct win over using pointers
(which must be _assumed_ aliased even when they're not).

A further problem with this is that pointer incrementing on multi-dimensional
arrays is particularly error prone.  Unless your inner loop is the one which
steps over the first/last (depending on language) index of the arrays, your
stride for the pointer may not be one (1).  If not (as is often the case
in numerical programs), _you_ are responsible for using the correct strides,
bounds, and corrections in each level of loop nesting.  This is further
aggravated if you are only computing something for a corner or region of
the array.

J. Giles