Canonicalize NaNs in the interpreter (#1046)

Even though we store f32 and f64 values as their representation, we
still use the underlying system's implementations for float operations.
These have non-deterministic behavior w.r.t. NaNs.

This change canoncalizes all NaNs where it is allowed. This excludes
`fxx.abs`, `fxx.neg`, `fxx.copysign` and `fxx.reinterpret*`, which
always work on the floating-point representation directly, and the
`fxx.convert*` instructions, which cannot be NaN because the input is an
integer.

0 files changed, 0 insertions, 0 deletions