Path: utzoo!attcan!uunet!tank!ncar!mailrus!cornell!uw-beaver!uw-june!david
From: david@june.cs.washington.edu (David Callahan)
Newsgroups: comp.lang.fortran
Subject: Re: FORTRAN 8x parser/lexer .NE. compiler
Keywords: FORTRAN 8x
Message-ID: <6148@june.cs.washington.edu>
Date: 20 Oct 88 19:18:50 GMT
References: <656@convex.UUCP>
Reply-To: david@uw-june.UUCP (David Callahan)
Distribution: comp.lang.fortran
Organization: U of Washington, Computer Science, Seattle
Lines: 110


The following is a reply to Steve Rown (rowan@convex.COM) who used the
Convex Ada implementation to measure the difficulties in implementing
Fortran 8x.  I am neither an expert on Fortran 8X nor on Ada so please bare
with my mistakes and questions.

#Granted, an Ada compiler is simpler than a FORTRAN 8x compiler but this
#should provide some idea of the magnitude of the problem.

Would someone please motivate this statement? Ada requires some
type-inference, generic packages, and handling of multi-tasking and
execptions. What are the features (aside from array-valued expressions)in
Fortran 8X that require similar amounts if implementation effort?

#The front-end of the CONVEX Ada compiler, the portion that performs
#lexing, parsing, and semantic analysis is 100,000 lines of C code.
# ...  The Ada parser and lexer come to less than 5,000 lines of the
#front-end total.  

It does not follow that a Fortran 8X front end will have similar 
characteristics.

#Now lets add the vectorized and code generator.  Let's
#say another 80,000 lines of C again taking the CONVEX Ada compiler as an
#example.  
#Before it's said that we don't need a vectorized let me point
#out that one has to do exactly the same data dependency analysis to
#determine when the extra temps introduced by FORTRAN 8x's copy semantics
#can be removed. 

(I don't want to argue against using vectorization technology for
optimizing Fortran 8X, but the author implies that it is necessary
for essentially any efficient implementation).

Dependence analysis is only part of a vectorizer. In particular, to optimize
array operations, you don't have to recognize induction variables and loop
invariant expressions. You don't have to pattern-match for things like sum
reductions and min/max reductions. Hence, the 80k figure may be very high.
The necessity of removing the implied temporary is unclear since its
lifetime is short. Of coarse, when adding one to every element of a
gig-word array you don't want to copy it, but it seems reasonable that for a
compiler to handle the easy cases well and the let the programmer help
(eliminate copies by adding loops) when things get out of hand. Explicit
array syntax is intended to improve the expressiveness of the language not
allow programmers to be oblivous to the efficiency of processing huge
amounts of data.

#Now lets add say 20,000 lines of code for the library managers necessary
#for dependent compilation.  

What are the aspects of Fortran8x that require more library management than
"incude" files? The UNIX world learned how to handle "compilation
dependencies" outside of the compiler. What about Fortran 8x invalidates
those techniques?.

# ... the compiler Callahan, Dongarra and Levine found to be
#the best vectorizing FORTRAN compiler in the world (See "Vectorizing
#Compilers: A Test Suite and Results, Technical Memorandum No. 109,
#Mathematics and Computer Science Division, Argonne National Laboratory).

Not to disagree with the statement, but this is not a conclusion the
authors make in the cited tech report.

#To develop a full FORTRAN 8x compiler which will optimize and do
#parallelization (without vectorization) will take many person-years to
#complete.  I think one can see how the numbers add up.

Why is parallelization necessary? You work for a company that markets
a parallel computer and I work for one that intends to, but that is 
certainly not a requirement of most implementations of Fortran 8X.

#One other Ada example and I will stop beating this dead horse.  Ada
#lexers and parsers were available in 1979.  The first production quality
#Ada compiler was not validated until 1985.  I choose the DEC compiler as
#the first real Ada compiler.  The academic Ada-ED compiler just does not
#count in my book.

Ada was developed from scatch without existing experience with a subset and
includes a very complex runtime environment to support multitasking and
execption handling. Aside from recursion and dynamic allocation, does
Fortran 8x have any feature that requies similar support?

Won't any substantive change to the language (like user defined types,
recursion, bit types, and dynamic memory allocation) require several years
before compilers are available and reliable? Aside from the array syntax,
what features can be deleted that will significantly reduce the development
time?

#If I have missed something here I would be interested in hearing the
#counter argument.  I have only been writing vectorizing compilers,
#FORTRAN, C and, Ada for the past ten years.
#
#I hope this did not come off as too much of a flame but as my public
#review comment stated, I feel we are doing the FORTRAN community a
#disservice by proposing a language with features the average FORTRAN user
#does not want or need.  

A couple of  final questions in reply: is it necessary that the "average"
Fortran user (if such a beast exists) want or need a feature for that
feature to be added to the language? Shouldn't it be sufficient that some
substantial portion of the community would find it useful, particularly
given that Fortran 77 is a subset so that the only noticable change would be
in compile speed? 

#Steve Rowan
#rowan@convex.COM

David Callahan
Tera Computer Company
(david@june.cs.washington.edu)