test: failing tests for constant folding and TDZ issues

[jluxenberg]

RE: #3125

Spent a bit of time trying to fix these in js_parser.go but couldn't solve it. This PR has a few failing tests which illustrate the problem.

There are two issues:

(1) Constant propagation not working properly when the constant is declared after it is used

Consider this program:

{ 
  const f = () => x; 
  const x = 0; 
  f();
}

The definition of x should be hoisted to the top of the enclosing block. Since it is constant, it is eligible for constant folding/propagation. The propagation should respect the Temporal Dead Zone; accessing x before it is initialized should raise a ReferenceError: Cannot access 'x' before initialization.

esbuild minifies this program to (() => x)(). I believe this is because x is correctly identified as a constant, but then isn't able to do the substitution into the function body because it is defined after the function that uses it.

(2) Dead statement removal doesn't respect the Temporal Dead Zone

Consider this program:

{
  number; 
  const number=10;
}

This should crash with an error ReferenceError: Cannot access 'number' before initialization.

However, esbuild minifies it to something like {const n=10}, omitting the number; statement.

[evanw]

Regarding (1), the issue #3125 was fixed and the fix already includes a test.

Regarding (2), this is because bound identifiers are considered to be side-effect free to allow for important optimizations. It's true that this means esbuild can incorrectly remove TDZ checks such as in the case you presented. However, code that triggers TDZ checks is inherently broken anyway, and disabling common and important tree-shaking optimizations for a rare edge case that only appears in broken code doesn't seem like a good trade-off for a production-oriented bundler like esbuild. This is why I have deliberately made the decision to not respect the TDZ check here.

[jluxenberg]

Thanks for the detailed explanation and fix for #3125!