> On Sat, 23 Jun 2012, Marc Glisse wrote:
>> Actually, what would you think of s/flag_trapping_math/flag_signaling_nans/
>> in invert_tree_comparison? Are there other possible ways a<b can trap than
>> having a sNaN for a or b?
> Ordered comparisons raise exceptions for both quiet and signaling NaNs.
> It's generally operations with floating-point results that raise
> exceptions only for signaling NaN operands.
Thank you for this explanation, it makes more sense now. And it seems to
mean that writing a non-intrusive patch won't help.