Re: inlining + optimization = nuisance bugs

As someone who has done a fair amount of numerics in my lifetime, and
knows others that have done orders of magnitude more, I have some
comments. The vast majority of code I write cares only about precision
in the crudest sense, and is more concerned with processing speed than
detailed numerics. However, compiler writers should recognize that the
presence of relational operators in the language implies that the
language permits reasoning about invariants in the code and should
preserve the pertinent invariants. I want to know that if the
arithmetic is guaranteed to be finite then I can do the following


IF (X > Y) THEN
do something with X > Y
ELSE
do something with X <= Y
ENDIF


I cannot at this time think of invariants which I want preserved in
all contexts, maybe someone from NAG or VISUAL NUMERICS can imagine
such contexts, but if I am dealing with a transcendental function, or
integrating a function with an integrable singularity, I want to be
able to deal with the tricky points in as clear a form as possible,
and I will identify those contexts through the explicit invocation of
relational expressions.


I will not expect magic from compiler writers, i.e., the recognition
of an expression in all contexts where it can be made to appear
through the mathematically proper invocation of commutativity,
associativity, and distributivity, but I do want simple invocations of
invariants between variables to be recognized.


William B. Clodius
--

Return to the comp.compilers page.
Search the comp.compilers archives again.