NaN vs -NaN (negative NaN) on macOS/arm64
Opened this issue · 2 comments
Bug description
Strange issue found while working on something unrelated. It seems that in certain cases, the compiler creates NaN on some platforms, and -NaN on others. This can lead to test failures.
I encountered it first here:
#2908 (comment)
But a search found that we've seen it before.
https://github.com/AssemblyScript/assemblyscript/pull/2352/files/403cae846a804c829b70bd2de388c3286cfdb899?diff=unified&w=0#r931976127
(wait about 5-10 seconds for GitHub to jump to the comment thread on that last link)
I'm not sure if this an macOS issue or an arm64 issue, but in both cases that was the platform. It does not occur on Linux or Windows x64.
Writing this up mostly so it doesn't get lost. I haven't found any practical implications yet. Some opinions about NaN vs -NaN seems to think it's not very important: https://stackoverflow.com/questions/21349847/positive-vs-negative-nans
Perhaps already fixed by a newer Binaryen implementation? 🤷
Steps to reproduce
On a macOS/arm64 machine:
npm run test:compiler -- std/math --create
git diff
Observe a bunch of changes like this:
- f64.const -nan:0x8000000000000
+ f64.const nan:0x8000000000000AssemblyScript version
v0.27.32
I can reproduce this on my Android phone via Termux (I knew I've experienced this issue before!).
It's definitely an arm64 thing, not macOS.
Some hint, it happened during binaryen's optimization pass. It may cause by nodejs wasm / js engine.