optimizations

optimizations

[Karl A. Nyberg]

I presume you had intended this for info-ada, and not info-ada-request?  If
so, please mail it again.  Thanks.

-- Karl --

Received: by grebyn.com (1.2/smail2.3/07-01-87)
	id AA19926; Wed, 10 Feb 88 13:17:47 est
Received: by grebyn.com (1.2/smail2.3/07-01-87)
	id AA19917; Wed, 10 Feb 88 13:17:40 est
Received: from AJPO.SEI.CMU.EDU by umd5.UMD.EDU (5.54/umd.04)
        for grebyn.COM!karl id AA25998; Wed, 10 Feb 88 13:07:06 EST
Message-Id: <8802101756.AA06942@ajpo.sei.cmu.edu>
Received: from TL-20B.ARPA by SIMTEL20.ARPA with TCP; Wed, 10 Feb 88 10:56:30 MST
Received: ID <PLOEDEREDER@TL-20B.ARPA.#Internet>; Wed 10 Feb 88 12:55:27-EST
Date: Wed 10 Feb 88 12:55:24-EST
From: umd5!TL-20B.ARPA!PLOEDEREDER
Subject: optimizations
To: info-ada-request@simtel20.arpa
Cc: ploedEREDER@tl-20b.arpa


I have followed the discussion on optimization with considerable interest
and was truly astonished about the (mis)conceptions that seem to exist.

There is a first law of code optimization (and code generation in general)
that must never be violated:
THE OBSERVABLE SEMANTICS OF EXECUTING CODE MUST NOT BE IN VIOLATION OF
THE LANGUAGE DEFINITION.

"Observable" in the statement above means a difference in input/output or
execution behavior of the program (it does not mean "observable in the assembly
code").  Optimization is not a carte blanche for the compiler implementer to
"optimize" into code that is 99% correct and fails to conform to the language
definition in 1% of all possible cases.

The example in case is definitely a bug in the compiler. The semantics of
assignment are copy semantics by language definition. Thus
   X: STRING := Y.all;
must have the observable semantics of copying the string contents of Y (and not
just copying the pointer value of Y). A compiler may optimize this assignment
into a pointer assignment ONLY IF the observable semantics are as if the string
were copied. This is not the case in the example, since, after deallocation, it
is indeed observable that the semantics of the program have been altered in
violation of the prescribed language definition of assignment. The situation
would be entirely different if the Ada source had copied the pointer value,
i.e., "X := Y"; in this case, subsequent deallocation applied to Y does indeed
make any reference to the contents of X erroneous by language definition.

As a final note to comments that recommend supressing optimization to ensure
"proper" observation of semantics: You will note that the "law" that I cited
above applies to non-optimizing code generation as well. If your Ada code
relies on semantics not prescribed by the language definition, its portability
(to another system or to another compiler on the same system, or even to the
next release of the same compiler) will be severely endangered, since compilers
WILL differ in their choice of implementation-defined semantics.

I refer you to publications by John Nissen et al. or by the ARTEWG that
delineate many of the areas of implementation-defined semantics, so you can
write code that does not rely on unjustified assumptions about the prescribed
semantics of Ada.

Erhard Ploedereder
Tartan Laboratories
PLOEDEREDER@TL-20B.ARPA
-------