BREAKING CHANGE: fix: detect the premature usage of @global variant

[HerrCai0907]

fix #2622

• I've read the contributing guidelines
• I've added my name and email to the NOTICE file

[HerrCai0907]

The bug will cause compile result is incorrect.

// index.ts
import {} from "./a";
import {} from "./b";

// a.ts
class T {
  a: i32;
}
let a = t;

// b.ts
class T {}
@global export const t = new T();
const b = new T();

the compile result is

 (func $start:assembly/a
  global.get $~lib/memory/__heap_base
  i32.const 4
  i32.add
  i32.const 15
  i32.add
  i32.const 15
  i32.const -1
  i32.xor
  i32.and
  i32.const 4
  i32.sub
  global.set $~lib/rt/stub/startOffset
  global.get $~lib/rt/stub/startOffset
  global.set $~lib/rt/stub/offset
  i32.const 0
  call $assembly/a/T#constructor
  global.set $assembly/b/t
  global.get $assembly/b/t
  global.set $assembly/a/a
 )

It use assign a/a and b/t with call $assembly/a/T#constructor but actually it should be call $assembly/b/T#constructor

[dcodeIO]

Interesting find, hmm. Appears that the issue arises because, in this particular situation with a custom @global in an imported file, using a @global variable before its source file executes will compile it out-of-order, here before any other top-level statements in b.ts, which then is similar in nature to @lazy, sharing its drawbacks. One could almost say that this sort of usage is not intended, as it modifies compilation order as a side-effect which can, even with the fix, lead to subtle issues.

Hence I wonder if it would be preferable to detect this situation and raise an error that t is used too early, say before its declaration? The fix would then be to import b.ts before a.ts, instead of permitting potentially undesirable side-effects?

Example of such a side-effect:

// index.ts
import {} from "./a";
import {} from "./b";

// a.ts
t += 41;

// b.ts
@global export let t = 1;
const b = t; // on a glimpse one would expect b = 1, but is 42

Similarly:

// index.ts
import {} from "./a";
import {} from "./b";

// a.ts
console.log(t.toString()); // one could expect "42", but is "1"

// b.ts
@global export let t = 1;
t = 42;

[HerrCai0907]

Hence I wonder if it would be preferable to detect this situation and raise an error that t is used too early, say before its declaration? The fix would then be to import b.ts before a.ts, instead of permitting potentially undesirable side-effects?

I am wondering how to handle @lazy @global. I have no idea to detect this case because the @lazy decorator make the variant not compile until used. So maybe we can make an exception for this case.

[dcodeIO]

If both @lazy and @global are present together, it seems OK to permit potential side-effects for now, given that @lazy permits it already. In practice, if both are present, the error would then not be diagnosed.

[dcodeIO]

Does this cover all possible cases? For example, if the global is not read (a = t1), but assigned (t1 = something), it looks like this branch alone would not suffice.

[HerrCai0907]

replaced by #2632