http://bugs.winehq.org/show_bug.cgi?id=14717
--- Comment #284 from Alexander E. Patrakov patrakov@gmail.com 2012-04-23 11:58:19 CDT --- (In reply to comment #282)
I did find one problem. In "put24" in patch 2,
t = lrintf(value * 0x800000);should be
t = lrintf(value * 0x80000000);to mirror the coefficient in "get24".
This indeed looks like a bug. However, the fix is wrong. First, 0x80000000U just to tell gcc that it is a large positive number, not the most negative one. Second, the overflow-case values above that line should be also modified.
Or take a simpler approach:
+ buf[0] = t & 0xFF; + buf[1] = (t >> 8) & 0xFF; + buf[2] = (t >> 16) & 0xFF;
This does make the numbers inconsistent, but note that the shift values are different, too.
Also I think this piece of code needs unit tests. Namely, that three bytes (0xff, 0xff, 0x7f) convert to something like 0.9999998807907104 and back, that 1.01 gets clipped and converted to (0xff, 0xff, 0x7f), and that (0x00, 0x00, 0x80) means -1.0