alexcrichton opened PR #3195 from spec-nan-fuzz
to main
:
This should fix a recent fuzz "bug" where the spec interpreter had different results from Wasmtime when Wasmtime had NaN canonicalization enabled. Now wasm-smith generated modules we're using for differential fuzzing here should manually be canonicalizing NaN at the wasm level so Cranelift doesn't need to do it.
The initial, misguided, attempt for this is at https://github.com/bytecodealliance/wasmtime/pull/3190 but as pointed out there it's far more robust to do canonicalization somewhere, and for differential fuzzing it's easiest if it's done at the wasm level so we don't have to implement canonicalization differently in all engines.
pchickey submitted PR review.
alexcrichton merged PR #3195.
Last updated: Nov 22 2024 at 17:03 UTC