exp10(Dec128(0)) produces incorrect result

[jmkuhn]

julia> exp10(Dec32(0))
+1E+0

julia> exp10(Dec64(0))
+1E+0

julia> exp10(Dec128(0))
+0E+0

[stevengj]

Oh weird, I see the same thing! This looks like it could be a bug in the underlying library.

[stevengj]

Yup, the source code bid128_exp10.c contains the lines:

    // x is 0, return 1.0
	 res.w[BID_HIGH_128W] = 0x3040000000000000ull;
	 res.w[BID_LOW_128W] = 0;
     BID_RETURN (res);

whereas I think it should be res.w[BID_LOW_128W] = 1 since reinterpret(UInt128, Dec128(1)) is 0x30400000000000000000000000000001

I submitted a bug report to [email protected].

[stevengj]

Even simpler, the code for bid128_exp2.c and bid128_exp.c uses res = BID128_1 for the same condition, since BID128_1 is apparently a predefined constant for 1.0.

[stevengj]

Intel responded to confirm the problem:

Thank you for pointing this out. Indeed, this is a corner case that requires a correction…
I believe your change is correct (I will double-check and will confirm). We will also post an updated library in the coming days.

[quinnj]

@stevengj, do you know if the library was updated? Do we need to do anything to build the new version?

[stevengj]

I don't see any updates on Intel's website; I pinged them again via email to see what the timeframe is. Once we have an updated library (or we patch it ourselves if they don't update soon), we'll have to build new Mac and Windows binaries and post them on bintray.

[stevengj]

I just got an email from them that they plan a new release "within days."

[quinnj]

@stevengj, do you have an official diff/patch for the fix? I have the BinaryBuilder/BinaryProvider working for all platforms now, so we could apply the fix ourselves until Intel gets around to releasing the new version.

[stevengj]

I have an unofficial copy of version 20U2, but I'm not sure if I'm supposed to distribute it. They are allegedly going to post it within a week. In the meantime, you can just use my one-line patch res.w[BID_LOW_128W] = 1 from above.

[stevengj]

Version 20U2 was just officially released: https://www.netlib.org/misc/intel/IntelRDFPMathLib20U2.tar.gz

[jmkuhn]

In April I submitted a PR to DecFPBuilder quinnj/DecFPBuilder#6 for updates to BinaryBuilder. I'm not sure if there have been any significant changes to BinaryBuilder since then.

[jmkuhn]

Fixed with #106