x/tools/internal/refactor/inline: a read from a non-address-taken variable should commute with global effects

[adonovan]

@lfolger reported a suboptimal inliner result of this form:

    var g = foo()

    func f(x, y any) { f2(x, y) }

    // before
    f(complicated(), g)

    // after
    var ( x = complicated(); y = g )
    f2(x, y)

In this case, the inliner's effects analysis assumed conservatively that the read from global var g would not commute with the call to complicated(). However, the variable g is not address-taken and is assigned only in its declaration.

The inliner's effects analysis should enumerate all non-address-taken variables (non-exported variables that appear in an lvalue position only in their own declaration) and allow reads of them to commute with global effects such as complicated(), resulting in this more optimal inlining: f2(g, complicated()).

[adonovan]

@lfolger raises the good point that effects analysis may be modeling reads too strictly in the first place. The spec guarantees only that calls, receives, and &&/|| operations are ordered, but not reads from memory, so if the program relies on g being read after the evaluation of complicated(), it already has a latent bug.

(The effects analysis still needs to model reads for other purposes: for example, if the callee function body repeatedly evaluates the parameter, and the argument is a read from a mutable variable, the read must not commute with some other side effect in the callee body.)