[erlang-questions] illegal guard expression for IF illegal guard expression for "if"

[Richard O'Keefe]

The essential point is that guards are part of pattern matching,
and that to make pattern matching efficient you want the compiler
to be able to move stuff around a fair bit.  Now consider

f(X, Y) when g(X), X > 0 -> ...

This has the *same* meaning and effect as

f(X, Y) when X > 0, g(X) -> ...

but it wouldn't have if g(_) could be an arbitrary function.

Back when I was still at RMIT I proposed stealing an idea from
NU Prolog, which had ":- pure <predicate name>" declarations to
inform the compiler that <predicate name> had no side effects &c.
That's not quite strong enough:  for a guard expression we want
to know that it is pure and *bounded*.  In the presence of bignum
arithmetic, "bounded" is a little hard to define, but
"time and temporary workspace polynominal in the inputs" gets close.

Abstract patterns turn out to provide *exactly* the right mixture of
expressiveness to handle most I-wish-I-could-call-this-in-a-guard
needs and limitation to make it safe.