eval_trick.md

The eval Trick

Each direct eval creates a separate browser execution context, so the idea is to use it to break down scripts without the need of including additional scripts. The high-level idea is to split each script and subsequently each block recursively into blocks of eval calls smaller than 1kB (an arbitrary number), and somehow have the whole script behave as before.

• Test run in the browser.
• Crawl top 100 websites and analyze the logs.

Why it should work

Direct eval calls are executed in the same scope as the caller, with read-write access to all of the surrounding variables.

Challenges and workarounds

In reality, the eval trick is a giant hack due to the quirks of eval.

• Hoisting: some constructs behave as if they are evaluated first before the script is run.

  • Top-level functions.

    We simply eval them first.

  • Variable declarations. Functions cannot capture variables declared in adjacent evals, and would capture the variable in the outer scope with the same name instead, causing errors. Variables created in eval do not leak out unless they are declared with var in non-strict mode.

    We identify all variable declarations in the current scope, declare all of them first using var or let, then make sure they are not declared again in the eval calls on the same level.

    • (Deferred) Some assignments do not have a single variable identifier on their left-hand side (LHS), such as destructuring. The LHS can contain arbitrary expressions.

      We currently stick var to their beginning so most of them work, but the assigned variables do not leak out. This is usually fine because most such assignments are for local variables anyway.

• return statements cannot return from inside eval.

  For each eval block that contains return statements not in nested functions or classes, we wrap it in an immediately invoked function expression (IIFE) so that the returns are valid. We call the IIFE with .call(this) to preserve this. We then check the return value of eval and return early it if it is not undefined.

  • Functions declared in IIFEs do not leak out.

    We declare the function identifiers as variables first, then convert all function declarations to assignments of function expressions to those variables.

    • (Deferred) This seems to break a few scripts, especially when they call functions defined in other scripts.

      No idea for solutions yet because the error messages did not reveal the cause.

• import statements are not allowed inside eval.

  We put them at the top of the script.

• export statements are not allowed inside eval.

  We keep these scripts as is.

• break, continue and yield statements may not be able to reach the correct outer scopes inside eval.

  We do not rewrite with eval in loops or generators until we hit a function or class boundary.

• await does not work inside eval.

  We await on the eval and use an async IIFE if an IIFE is used.

Inherent limitations

• Performance: eval disables the JIT and forces slow variable lookups.
• Bloat: deeply nested evals cause mountains of backslashes.
• Integrity: e.g., if the website checks the checksum of the script, it will fail.