You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Jun 26, 2020. It is now read-only.
This is a cool project, and if we need software floating point at compile time, we should probably use this.
When compiling WebAssembly, the code generator is unlikely to discover opportunities for floating point constant folding that weren't already taken care of by LLVM when it generated the WASM. This makes FP constant folding a low priority at least until we start compiling WebAssembly that hasn't been properly optimized.
When JIT-compiling, we also need to make sure that we are bit-exact with the other tiers. Since WebAssembly floating point permits some non-determinism around NaN's, we should probably avoid constant-folding any FP expressions yielding NaN results.
Given this restriction, and given that we only need f32 and f64, it is possible that we can use the host's hardware FP for constant folding. That would be preferable in terms of compiler code size, performance, and dependency management.
See rust-lang/rust#43554 for more details. We of course have to publish that crate on crates.io first.
Would be glad to hear what specific needs Cretonne has from software floating-point.
The text was updated successfully, but these errors were encountered: