How to test if float is NaN?

[Don Clugston]

Lionello Lunesu wrote:
> xs0 wrote:
>> I think this should work as well:
>>
>> if (num == num) {
>>     // not NaN
>> } else {
>>     // NaN
>> }
> 
> Indeed! This works just fine in VC on x86, but fails miserably on x64...
> 
> In a release build the VC8 x64 compiler optimized that "if" to "if(1)" :D
> 
> Here's the assembly from a debug build:
> 
> inline bool _is_nan( double d ) { return !(d==d); }
> 0000000000404000  movsd       mmword ptr [rsp+8],xmm0
> 0000000000404006  push        rdi
> 0000000000404007  sub         rsp,10h
> 000000000040400B  mov         rdi,rsp
> 000000000040400E  mov         rcx,4
> 0000000000404018  mov         eax,0CCCCCCCCh
> 000000000040401D  rep stos    dword ptr [rdi]
> 000000000040401F  movsd       xmm0,mmword ptr [d]
> 0000000000404025  ucomisd     xmm0,mmword ptr [d]
> 000000000040402B  je          citkTypes::_is_nan+36h (404036h)
> 000000000040402D  mov         dword ptr [rsp],1
> 0000000000404034  jmp         citkTypes::_is_nan+3Dh (40403Dh)
> 0000000000404036  mov         dword ptr [rsp],0
> 000000000040403D  mov         al,byte ptr [rsp]
> 0000000000404040  add         rsp,10h
> 0000000000404044  pop         rdi
> 0000000000404045  ret
> 
> According to
> http://www.ews.uiuc.edu/~cjiang/reference/vc314.htm
> they should not have been checking for ZF,PF,CF==1,0,0, but the code 
> only checks ZF?
> 
> Since by definition a NaN is not equal to anything, including itself, I 
> suppose it's a compiler bug?

Definitely. I've seen that sort of behaviour before in VC++.
<cynic> Does anyone at Microsoft understand floating point arithmetic? 
</cynic>