Post

I wonder if there should be a test mode that randomize let's say the lowest 3 bits after any float operation. FMA is the visible tip of the iceberg, there are edge cases where some CPUs just implement IEEE754 at partial precision on some costly operations.

Sadly for perfect portability the only solution is fixed point math or software floats. Or maybe types like archfloat.x86, archfloat.riscv, .. that make sure to portably implement floats according to one exact ISA.

People have been pretty good at IEEE math for a few decades now. In scalar code at least, the widely-understood default mode for doing FP looks like Java's strictfp, with the only glitches at (1) multiply-and-add and (2) conversions. We owe a lot to Java for this.

I assume you mean go.dev/cl/799984 and go.dev/cl/801620; these kinds of changes I think have always been allowed go.dev/ref/spec#:~:... You can use pkg.go.dev/golang.org/x... with -compile=fma to sus out issues, I think, or GOCOMPILEDEBUG=fmahash=...

This made me wonder if any code I maintain has FMA bugs and yep... So now I'm making a lint rule for TS's repo that bans it entirely