function f() { print(Math.round(Math.fround(Math.expm1(Math.log2(4294967297)) | 0))) }; f(); f(); $ ./js-dbg-opt-64-dm-nsprBuild-darwin-89a190592267 --fuzzing-safe --no-threads --ion-eager a2cjs.js -13527757 -13527756 $ ./js-dbg-opt-64-dm-nsprBuild-darwin-89a190592267 --fuzzing-safe --no-threads --baseline-eager a2cjs.js -13527757 -13527757 Tested this on m-c rev 89a190592267. My configure flags are: CC="clang -Qunused-arguments" CXX="clang++ -Qunused-arguments" AR=ar AUTOCONF=/usr/local/Cellar/autoconf213/2.13/bin/autoconf213 sh /Users/skywalker/trees/mozilla-central/js/src/configure --target=x86_64-apple-darwin12.5.0 --enable-debug --enable-optimize --enable-nspr-build --enable-more-deterministic --with-ccache --enable-gczeal --enable-debug-symbols --disable-tests autoBisect is running.

autoBisect shows this is probably related to the following changeset: The first bad revision is: changeset: https://hg.mozilla.org/mozilla-central/rev/a1b396e1f1dd user: Benjamin Bouvier date: Fri Feb 28 12:07:05 2014 +0100 summary: Bug 930477: Specialize Round for Float32; r=jandem,mjrosenb Benjamin, is bug 930477 a possible regressor?

So that's the case from bug 1073910, showing up on x86. It's already fixed on ARM, but with some different implementation. My new theory is that numbers in [0.5, -0] should be added 0.5, and all other numbers 0.499999... (These rounding errors are frustrating, but I cannot craft a nice single test case file that would try all numbers. I think the functions Math.round and Math.round(Math.fround()) would need to get cloned at call-site in such a perfect test case file, and each value should be tested against its own clone. Tried a few things, but there's always sometimes that gets wrong at some point...)