matching: minor simplification in do_for_multiple_match

[gasche]

(This is a small refactoring commit that should be obviously correct.)

do_for_multiple_match is used to optimize code of the form match e1, e2 with ..., to avoid allocating a tuple (e1, e2) when possible (when the pattern does not require binding the scrutinee). (If the clause is | (x, y) as p -> ... for example, a binding let p = (e1, e2) in ... will still be emitted.)

At first sight, its implementation is wrong. It precompiles a set of pattern-matching clauses whose scrutinee is a tuple of arbitrary expressions, and then it binds each of those expressions to a fresh variable and transforms the result of the precompilation to use these per-element variables instead of the tuple, to save an allocation. This is wrong because the arbitrary expressions may thus occur twice in the generated code, once in a let p = (e1, e2) binding generated by the precompilation, and another to bind the per-element variables.

In fact this is okay because there is only one caller for do_for_multiple_match, and it always passes a tuple of expressions that are in fact variables, by performing an "outer binding" step that binds all the tuple elements to and calls do_for_multiple_match on a list of expressions-that-are-variables. So in fact there is no bug, and the fresh per-element variables in do_for_multiple_match are never useful.

This redundant binding of variables was introduced in 2009 in f8107f9, to fix the bug described above (duplication of effectful expressions in the generated code), see #4828.

The present commit simplifies and clarifies the code by changing do_for_multiple_match to take a list of variables instead of a list of expressions, and removing the redundant creation of per-element variables.

(cc @ncik-roberts, @trefis, @Octachron, @lthls)

[lthls]

This code looked familiar, and apparently I already spotted that while working on #12440. Thanks for actually taking care of it.

[gasche]

Thanks both for the prompt reviews, merging.