My function has passed unit test, but it writes wrong results sometimes.

Question

Jiaxin Zhang 60

My function has passed unit test, but it writes wrong results sometimes. What’s wrong?

When I test the function, the actual result is equal to expected result. But it writes a different result in the release version of my project.

Strange!

Viorel 124.9K Reputation points

2025-10-11T09:23:54.0533333+00:00

Is it related to floating-point calculations?
Jiaxin Zhang 60 Reputation points

2025-10-11T09:38:20.9033333+00:00

@Viorel Yes, but I use QString other than float. I wrote a decimal class myself, and it can perform some simple calculations just like addition or subtraction.
RLWA32 50,826 Reputation points

2025-10-11T09:46:53.34+00:00

Have a look at some suggestions from Microsoft -- How to Debug a Release Build
Darran Rowe 2,451 Reputation points

2025-10-13T09:28:40.71+00:00

How far off is the "incorrect" result from the "correct" result.

The CRT doesn't guarantee completely accurate results when calling the CRT floating point functions. The documentation does state that it will be within 1 ULP of accuracy. In other words, if you are looking at the full float/double range, then the very last digit may be inaccurate. The other issue is that because of how many bits are used for the mantissa, one of the decimal digits of float and double isn't fully represented.

Your answer

Viorel 124.9K Reputation points

2025-10-11T09:23:54.0533333+00:00

Is it related to floating-point calculations?
Jiaxin Zhang 60 Reputation points

2025-10-11T09:38:20.9033333+00:00

@Viorel Yes, but I use QString other than float. I wrote a decimal class myself, and it can perform some simple calculations just like addition or subtraction.
RLWA32 50,826 Reputation points

2025-10-11T09:46:53.34+00:00

Have a look at some suggestions from Microsoft -- How to Debug a Release Build
Darran Rowe 2,451 Reputation points

2025-10-13T09:28:40.71+00:00

How far off is the "incorrect" result from the "correct" result.

The CRT doesn't guarantee completely accurate results when calling the CRT floating point functions. The documentation does state that it will be within 1 ULP of accuracy. In other words, if you are looking at the full float/double range, then the very last digit may be inaccurate. The other issue is that because of how many bits are used for the mantissa, one of the decimal digits of float and double isn't fully represented.