I don't know what ANSI says, but...
A cascade is a series of messages sent to the same receiver. The receiver that is used is the receiver of the (outermost) message that precedes the first semicolon.
Another way of thinking about it is... r m1; m2; m3; m4 is essentially equivalent to x := r. x m1. x m2. x m3 except that this is not an expression (it's a block body), and there's an extra temp needed here.
How about when non-cascaded messages are interspersed? I'd love to know a simple rule. Here are some example cases:
a x y; z 0 x y: 1 ; z 0 + 3 x y ; z 0 + 3 x y ; * 3 0 + 3 x y ; times: 3
Cascades are the only part of Smalltalk syntax that confused me when I wrote a Smalltalk parser a few years ago. In fact, I still don't understand it. Reading Squeak's parser, it appears that cascades happen at the location that is most convenient. :) Reading the BNF syntax on the Swiki, handling cascades seems to the main reason it is more than just a few lines long.
Okay, here's my shot. First, cascades cannot have a message sent to them (so, "a b; c + 3" is illegal). That is, the cascade messages are "top-level", in a certain sense (probably a sense that's not strictly correct, but this is English. I'm allowed.) Second, the cascade applies to the message of *lowest priority* just to the left of the cascade. That is, if there is a keyword message, then the cascade applies to the keyword message; otherwise, look for a binary message; otherwise, use a keyword message. (There has to be at least *one* message before you start cascading!)
So here are my guesses on the above:
(a x) y; z (0 x) y: 1; z 0 + (3 x y) ; z 0 + (3 x y) ; * 3 9 + (3 x y) ; times: 3
I guess I could generate some parse trees and check, but already I'm spending too long reading this exciting mailing list when I should be doing other things!
Incidentally, ANSI doesn't make this issue obvious, though it might be in there and I just didn't dig deeply enough.
-Lex