f32/64.neg/abs may be wrong about wasmi

[pepyakin]

It well might be! We didn't spent much time on float stuff and, I think, we even not passing testsuite for float.

[pepyakin]

It seems that there is no separate instruction in LLVM that acts as fneg: fsub 0, x is used to represent it. I guess we fine since both LLVM and wasm strives to be IEEE 754 compatible.

[pepyakin]

Also, just in case you interested, in #72 I've enabled fp tests and they should pass now.

[CryZe]

While there is no separate instruction for it in LLVM, the reference manual specifically calls out that subtracting from 0 is not correct:

This differs from subtracting the operand value from negative zero

I don't know enough about floating point to be able to judge whether there's any actual difference, but is there anything stopping you from just implementing it as:

f32::from_bits(val.to_bits() ^ 0x8000_0000)

[pepyakin]

Yeah, that's because the IEEE 754 tells the same!
I'm actually pretty not sure that fsub 0, x does something other than flipping the sign bit.

Personally, I'd perfer using f32::neg rather than bit-twiddling since it's cleaner (and maybe there is a chance that llvm might generate a better code?)

So before doing that change, I would like to be sure that this is actually a problem.

I think running the testsuite on a some exotic (fp-wise) arch such as MIPS would be enough to discover problems if they are actually present!

[pepyakin]

After all we ended up with raw bit manipulation, see #87 !