|Posted by Ivan Kazmenko||PermalinkReply|
While exploring quirks of floating-point values, as well as C/C++/D convenience with them, I stumbled on, in essence, the following (DMD32 on Windows):
void main ()
import std.stdio : writefln;
double x = 128.0; // same for real or float
writefln ("%.20a", x); // 0x1.00000000000000000000p+7, right
writefln ("%.20f", x); // 127.99999999999999999000, wrong
1. The internal representation is fine: the exponent (before shift) is 7, and the mantissa is all-zeroes (except the "1." part which is not stored).
2. Formatting to a decimal representation goes off in a bad way, giving the wrong third significant digit.
3. The trail of 9s is the same ~20 decimal digits for every floating-point type, which suggests that they are all converted to 80-bit real before formatting. This obscures their difference in width, which is bad from at least the learning standpoint.
Point 2 is sad. One would expect at least the exact powers of two to be stored exactly. And indeed they are. But trying to demonstrate it gives the wrong impression that they are not. This adds unnecessary confusion to the already complex subject of how floating-point values work, and learning the subject with D becomes much harder.
On the competition front, this already seems to be settled: at least with MinGW GCC, both C printf and C++ cout correctly display powers of two - and perhaps any small integers exactly stored as floating-point data, for that matter.
With DMD64, the issue vanishes. This strongly suggests that 32-bit druntime is the culprit. And indeed, the Phobos "formatValue ... if (is(FloatingPointTypeOf!T) && ...)" for floating-point calls snprintf from druntime, after which I couldn't easily track it to the source code. Should this be so? And regardless, perhaps snprintf (or Phobos or whatever will do the dirty work) can adapt a more modern approach so that integers in floating-point don't get corrupted when converted to string representation? Perhaps even without sacrificing much speed. The most obvious way seems to just make multiplications by 10, comparisons and subtractions in a loop, and is likely slow but at least correct for integers.
I'm going to make this a bugreport, but first wanted to explicitly point at this in the discussion group, since it's quirky, and I may have easily missed something important.