Conversions from BigInteger to Double are not always as precise as from Long

[SingleAccretion]

Description

The gist of the issue is that some precision is apparently lost when using the explicit conversion operator operator double(BigInteger integer), as compared to a direct cast from long. Here's a very simple reproduction:

using System;
using System.Numerics;

BigInteger a = long.MaxValue / 2;

Console.WriteLine($"Source: {a}");
Console.WriteLine($"BigInt: {(double)a:F0}");
Console.WriteLine($"Long:   {(double)(long)a:F0}");

This will print:

Source: 4611686018427387903
BigInt: 4611686018427387392
Long:   4611686018427387904

Configuration

Tested with the latest bits from dotnet/installer - 6.0.100-preview.3.21164.5 on Windows 10 x64. I do not believe this is an architecture-specific issue as all the code that does it is fully managed and uses simple arithmetic.

Regression?

Not a regression - the reproduction works on net6.0, net5.0 and netcoreapp3.1. It appears that the code responsible for the conversion hasn't been touched in a long time.

[ghost]

Tagging subscribers to this area: @tannergooding, @pgovind
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Description

The gist of the issue is that some precision is apparently lost when using the explicit conversion operator double(BigInteger integer), as compared to a direct cast from long. Here's a very simple reproduction:

 BigInteger a = long.MaxValue / 2;

 Console.WriteLine($"Source: {a}");
 Console.WriteLine($"BigInt: {(double)a:F0}");
 Console.WriteLine($"Long:   {(double)(long)a:F0}");

This will print:

Source: 4611686018427387903
BigInt: 4611686018427387392
Long:   4611686018427387904

Configuration

Tested with the latest bits from dotnet/installer - 6.0.100-preview.3.21164.5 on Windows 10 x64. I do not believe this is an architecture-specific issue as all the code that does it is fully managed and uses simple arithmetic.

Regression?

Not a regression - the reproduction works on net6.0, net5.0 and netcoreapp3.1. It appears that the code responsible for the conversion hasn't been touched in a long time.

Author:	SingleAccretion
Assignees:	-
Labels:	area-System.Numerics, untriaged
Milestone:	-

[tannergooding]

This looks like a rounding bug. Putting into 6.0 so we can do an investigation to determine if this is a simple fix or not.

[Maximys]

@tannergooding , currently I trying to fix this bug, but I have little question linked with RunSingleExplicitCastFromBigIntegerTests. Can you explain me this lines of test:

bigInteger = new BigInteger(float.MinValue);
bigInteger -= BigInteger.One;
VerifySingleExplicitCastFromBigInteger(float.MinValue, bigInteger);

This lines makes next:

1. create BigInteger value by float.MinValue;
2. decreasing this value;
3. casting result value to float;
4. checking equation of (float.MinValue - 1) and float.MinValue.

Is it normall? I think, it should return false but I'm not sure.
I see the same problem on next lines
Can you explain me?

[tannergooding]

Is it normall? I think, it should return false but I'm not sure.

Yes. IEEE 754 binary floating-point numbers are non-exact and have a sliding epsilon (delta/difference) between values. This epsilon starts at float.Epsilon at 0 and doubles ever power of two. So, for example, the delta between 2^22 (4194304) and the next largest representable value is 0.5. That means 4194304 + any value that is less than 0.5 will result in the same value (4194304). This delta doubles at the next power of two (2^23, which is 8388608) and becomes 1.0, then again at the next (2^24, which is 16777216) to become 2.0, etc.

float.MaxValue is approx. 3.4028235E+38 which means the delta between it and the next largest number is very big (approx. 2.028241E+31) and anything smaller will result in float.MaxValue being returned.