Re: 5.1 rc mod giving -0 (never just 0)

[Rici Lake <lua@...>]

On 22-Jan-06, at 4:25 PM, Asko Kauppi wrote:

> Not sure if this is worth bringing up, but...
>
> Lua 5.1-rc always gives "-0" for zero mod results. Since PIL2 says 
> "the result always having the same sign as the second argument", this 
> raises my eyebrowns.

I confirmed this on a similar powerpc.

The problem is that gcc (with -O2) produces a "fused negated multiply 
and subtract" opcode for the computation of luai_nummod: b - 
floor(b/c)*c. My guess is that fnmsub is implemented in hardware by 
simply negating the result of fmsub. The PPC numerics guide does warn 
than the four fused multiply/add instructions may violate ieee-754 
semantics.

In any event, if you compile lua (or at least lvm.c) with the option 
-mno-fused-madd, then you'll get 0 rather -0. (Or you could redefine 
luai_nummod in terms of fmod and copysign.)