FP comparisons, C, X86, and NaN's

[Chris Lattner]

On 18 Jun 2004, Gabriel Dos Reis wrote:

>
> | > > 2. C does not define whether or not standard relational operations trap.
> | > >    If this is the case, there is no bug in GCC.
>
> | So you are saying that this is the case: it is undefined whether an
> | implementation traps on a comparison with nans?  This is my interpretation
> | of this passage:
>
> No!  That section explicitly says that isxxx type generic macros
> are _quiet_ in the sense that they *do not raise* floating-point
> exception or traps or whatever.  It is not implementation defined.  It
> is prescribed.  Please consider the whole section I quoted.

I'm sorry, I meant standard relational operators like < and >.  The
behavior of the isxxx macros are quite clear.  Is it undefined whether or
not the <,> operators trap? (I'm guessing no from your later response)

> | >         --   is  true.   Relational   operators   may   raise   the
> | >        ``invalid''  floating-point  exception  when argument values
> | >        are NaNs.
> |
> | I just want to make sure that I'm intepreting things correctly.
>
> "Relation operators" in the above quote means the usual <, >, <=, >=,
> ==, !=.  Not the isgreater, isxxxx... type generic macros.
>
> Said differently, you've discovered a bug in the code GCC generates.

>From this, I take it to mean that < is defined to trap if either operand
is a QNaN or SNaN.  Is this what you mean?

-Chris

-- 
http://llvm.cs.uiuc.edu/
http://www.nondot.org/~sabre/Projects/