low level optimization

[Jim Giles]

In article <1991Apr18.233807.19552 at zoo.toronto.edu>, henry at zoo.toronto.edu (Henry Spencer) writes:
|> In article <21846 at lanl.gov> jlg at lanl.gov (Jim Giles) writes:
|> >if the user chooses only to recompile one file and not any of the others
|> >(which is allowed by the "may" clause), the implementation _must_ allow
|> >this to be linked with the previously translated files to produce a
|> >program.
|> 
|> Of course, it may not run as fast as if he had compiled them all together.

I see that I'm going to have to give a specific example.  Suppose
that functions A and B are in two separate files.  Suppose that
I compile them 'together' in some way (that is, the compiler makes
some use of internal knowledge of B while translating A and vice-versa).
Now, suppose I _change_ A and retranslate just A (I do _not_ retranslate
B).  The standard requires that I still be able to link both into a
single program.  If the translation of B made any _profitable_ use
of internal information about A, then that information is obsolete
(and probably downright _wrong_) after A has been changed.  Yet, the
standard requires that the new version of A still be usable with the
previously translated B.  Either the use the translator made of the
interprocedural analysis was trivial, or B will contain unsafe
optimizations - the only third way is if 'translation' does not
include code generation (which will subsequently be done by some
post-translation tool: like the loader - just like I keep saying).

I don't understand why you continue to hang on like grim death
to the fiction that dependent compilation satisfies the requirements
of the standard.

J. Giles