Re: non trivial constant folding

Anton Ertl <anton@mips.complang.tuwien.ac.at> wrote:
[...]


> Actually I was thinking more about evaluation order issues; the
> compiler does not need to detect that the program depends on the
> evaluation order in order to produce the same evaluation order for a
> given expression. However, if it wants to optimize by reordering
> the evaluation, and still wants to produce consistent behaviour, it
> probably needs to detect order dependences, and suppress the
> optimization (and issue a warning) if it cannot prove that there are
> no order dependences.


Which it'll practically never be able to do, in the case of floating
point arithmetics. Except for those expressions which can be
evaluated at compile time, already, just about *no* FP computation can
be known to be order-independant, before it's executed. Order
dependence is mainly determined by the actual operands, not by the
structure of the expression.


I.e. unless the compiler does a successful analysis of the "possible"
values of a numerical expression by interval arithmetics or similar
means, it cannot ever know whether it's allowed to optimize evaluation
order or not. In the end, the compiler would effectively be forbidden
to do any such optimization, in any programming language that has
doesn't regard ordering effects as free to decide by the implementor.
--
Hans-Bernhard Broeker (broeker@physik.rwth-aachen.de)
Even if all the snow were burnt, ashes would remain.

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