Unittests fail on division, on x86_64 architecture. about libfixmath HOT 10 CLOSED

For now try reverting to the r52 division code which I believe is tested, this 
is a critical issue and I'll look into it as soon as I get time.

Also fails with gcc 4.6.3 on x86_64.

Hum, I tried
svn update -r52

and the tests still fail at that revision.


FAILED: fix16_unittests.c:194 failures == 0                                     

make: *** [run_fix16_unittests] Error 1                                         

12:06 louiz@abricot /tmp/libfixmath-read-only/unittests % svn log               

------------------------------------------------------------------------        

r52 | [email protected] | 2012-01-26 16:43:30 +0100 (Thu, 26 Jan 2012) 
| 9 lines

I've not really had much todo with the testing so I'm not sure how it judges 
success, but you should know that division in fixed point (like integer 
division) has limited accuracy since the result has dynamic range.
If you divide by a small/large or small/large number then you'll likely 
under/overflow. I know the old division code worked because it's been used in a 
number of my projects so it may be the case that the tests just expect a higher 
accuracy than is possible but I'll have a dig and see what I can find.
For cases where guaranteed accuracy isn't important it may be better to use 
saturated divides (e.g. graphics/games), otherwise you must handle the under or 
overflow cases.

This issue was closed by revision r83.

This was a real bug in my code; I had not realized that the return value of GCC 
builtin clzl (count leading zeros) depends on the sizeof(long). Revision 83 
fixes this.

I need to have deterministic results (always the same result for a given 
operation, whatever the arch, compiler, etc) for my non-integer operations. 
That means that I need the guaranteed accuracy. I’ll handle the overflow 
cases then.

I discovered this issue when some operation like 5.1/1.2 failed. I’ll try my 
own tests with the r52 division code and tell you if I have more success with 
that.

(And thank you very much for your fast and useful answers!)

You’re too fast for me, you fix the issue before I can even respond to your 
previous message!
And you’re awesome, the tests successfully pass! Thank you again! :)

Wow, nice one Petteri :)

You mind telling us what you're using libfixmath for louizatakk? You don't have 
to of-course but it's always interesting to see what people are using the code 
for.

Oh, I didn’t pay attention, you’re two different persons. So thanks to both 
of you :)

I’m using it in an online RTS game, where only “commands” are passed from 
the server to the client (for example “unit X move there”) and then the 
whole game simulation runs in parallel on all machines. That’s why, for that 
part of the game, I cannot use floats (result could be different on each 
machine, and the games would be desynchronized).