Trapping quiet NaN

Question

I have an application where some component occasionally inserts a qNaN in a large data stream, which then invalidates the entire processing (FFT on a vector containing a single qNaN results in an all-qNaN output). Now I'd like to catch that component in the act and find out why it's doing that.

For this, I'd need to somehow make all NaNs signalling during debug. Is there a way to do so, with a x64 CPU executing 32 bit code?

Not exactly a duplicate, but lots of details http://stackoverflow.com/questions/570669/checking-if-a-double-or-float-is-nan-in-c — user7116, Mar 08 '12 at 14:28
Another question with good details http://stackoverflow.com/questions/2247447/usefulness-of-signaling-nan — user7116, Mar 08 '12 at 15:19
Related: [Can I make gcc tell me when a calculation results in NaN or inf at runtime?](http://stackoverflow.com/questions/2941611/can-i-make-gcc-tell-me-when-a-calculation-results-in-nan-or-inf-at-runtime/20973509) — legends2k, May 02 '14 at 09:40
possible duplicate of [How to trace a NaN in C++](http://stackoverflow.com/questions/3615724/how-to-trace-a-nan-in-c) — legends2k, May 02 '14 at 09:45

Evgeny Kluev · Accepted Answer · 2012-03-08T17:10:54.730

23

If you want to make all NaNs, overflows, and zerodivides signaling during debug, it is possible.

For gcc:

#include <fenv.h>

#ifndef NDEBUG
feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);
#endif

For Visual Studio (not tested):

#include <float.h>

#ifndef NDEBUG
_clearfp();
_controlfp(_controlfp(0, 0) & ~(_EM_INVALID | _EM_ZERODIVIDE | _EM_OVERFLOW),
           _MCW_EM);
#endif

References: Microsoft, gcc.

These functions allow to catch either NaNs, produced by floating point operations (overflows, zerodivides, invalid operations), or sNaNs, used as input to some floating point operation. They do not allow to catch qNaNs, used as input to floating point operation. For such qNaNs, the only way to find them is to check each value individually (see Luchian Grigore's answer).

So if the component, that inserts a qNaN is in the same program, were you want to catch it, or if this component is in separate program, but you have its source codes, just enable FP exceptions with feenableexcept()/_controlfp(). Otherwise, check every value in the incoming data stream with isnan() (C++11) or with x != x.

edited Mar 08 '12 at 17:10

answered Mar 08 '12 at 15:17

Evgeny Kluev

24,287
7
55
98

Do you put this in the implementation file? – Luchian Grigore Mar 08 '12 at 16:01
@LuchianGrigore, yes, just call this from main(). – Evgeny Kluev Mar 08 '12 at 16:03
Hm, still didn't get an exception, but accepting as it put me on the right track with debugging. – Simon Richter Mar 09 '12 at 07:09
If you really need an exception, there is a way. But very slow and complicated. You can write driver (kernel module), set "EM" flag in "CR0" control register, and get an exception on every FP operation. Then check for qNaN and perform operation in kernel mode. – Evgeny Kluev Mar 09 '12 at 07:39
Or you can try to patch 387-emulation code to simplify this task. – Evgeny Kluev Mar 09 '12 at 08:25

score 7 · Answer 2 · answered Mar 08 '12 at 14:20

7

I doubt there's a way to put a data breakpoint all over the memory to catch all NaNs.

I'd look for the code that inserts into the vector and test for NaN there:

if ( x != x )
    assert(!"NaN detected");

answered Mar 08 '12 at 14:20

Luchian Grigore

253,575
64
457
625

4

Is there a reason you didn't put the condition in the `assert` itself? – Lightness Races in Orbit Mar 08 '12 at 14:28
@LightnessRacesinOrbit I did think about it, but it's more readable to the untrained eye like this. – Luchian Grigore Mar 08 '12 at 14:30
This is a valuable opportunity to train that eye. – John McFarlane Feb 22 '22 at 10:26

Trapping quiet NaN

2 Answers2

Linked