Is the order of procedural macro attribute execution guaranteed per item?

[azriel91]

Hiya, is there an intentional guarantee that procedural macro attributes are executed outward-in per item? To illustrate (repo):

// attr_* appends a new field to the struct

#[attr_0(u8)]
#[attr_1(u16)]
#[attr_2(u32)]
struct Hello;

#[attr_2(u32)]
#[attr_1(u16)]
#[attr_0(u8)]
struct Hello2;

generates:

struct Hello(u8, u16, u32);
struct Hello2(u32, u16, u8);

That's the behaviour, but I'd like to confirm if I can rely on it (and if it should be documented).

#56058 may be related, but I haven't checked the details whether the "each invocation" is per procedural macro (and potentially losing the order guarantee).

[Centril]

cc @eddyb @petrochenkov

-- To my knowledge we make no guarantees about the execution order of proc macro expansion.

[ehuss]

According to rust-lang/reference#578 (comment) it is not defined.

[Mark-Simulacrum]

That issue seems canonical enough and has a reply from @eddyb so I'm going to close this in favor of the reference (where such things should ultimately be documented anyway, I think).

[Nemo157]

This isn't really about the execution order of proc-macro-attributes, but rather the application order. Even if all the macros from the example were deterministic and side-effect free the order in which they are applied to the AST will affect the final result.

(But still, this might be better as a reference issue).