String escaping fuzz bug with wasm-metadce

[tlively]

This is a fuzz bug where the fuzzer runs the CtorEval handler with test/lit/basic/name-high-bytes.wast as its initial contents.

test.wast (reduced):

(module
 (type $0 (func))
 (export "test\\c3\\a9_invoker" (func $0))
 (func $0 (type $0)
  (unreachable)
 )
)

graph.json produced by fuzz_opt.py's filter_exports (reduced):

[
  {
    "name": "outside",
    "reaches": ["export-test\\\\c3\\\\a9_invoker"],
    "root": true
  },
  {
    "name": "export-test\\\\c3\\\\a9_invoker",
    "export": "test\\\\c3\\\\a9_invoker"
  }
]

Here we have an export name containing backslashes. Note that they are escaped in the Wasm text format, so the actual unescaped byte content of the export name is test\c3\a9_invoker. But fuzz_opt.py is not unescaping the string it reads from the disassembly, and then it is JSON-encoding the escaped name, so graph.json ends up with doubly escaped backslashes.

To make matters worse, wasm-metadce is parsing the input JSON in "ASCII" mode, which does not do any unescaping either. So wasm-metadce is rooting an export named test\\\\c3\\\\a9_invoker but in fact the export's name is test\c3\a9_invoker, so the export is removed and the filtered module is empty.

This causes the fuzzer to fail when it later runs wasm-ctor-eval and passes it --kept-exports=test\c3\a9_invoker (note that it has unescaped the string for this step). This errors out because that export no longer exists.

IIUC, the proper fix would be to 1) perform Wasm text format unescaping immediately when extracting export names in get_exports, and 2) perform JSON unescaping when parsing the JSON in wasm-metadce.

[tlively]

(seed 12496077044581880887 at commit 5894db496d08 for future reference if we need to reproduce the original error)

[kripken]

Worth fixing but this also seems low priority, and I believe this error happens due to including test/lit/basic/name-high-bytes.wast as initial content. I think maybe we can just ignore that file for now? #8483