Visual studio compiler optimisation for type casting in C++

Question

I had the following statement in code:

int a = (int)( (float)(b * 1000) / (float)c + .5f );

where b is an int as well, and c is an unsigned intwith a constant value (of 15 in my test runs)

This statement was inside a while loop where b gets incremented by 1 on each iteration.

This code was working fine till I decided to switch on the optimisation flags (maximise speed) in Visual studio. After which, a would randomly run into an overflow (i.e it had a value of -2147483647). On investigating further this overflow used to happen at different values of b. The values of b where the overflow happened were between 9-12 in the test runs I observed.

What solved the problem was making the a small change as illustrated below:

int a = (int)( (b * 1000.f) / c + .5f );

Any ideas on how did this help? It is working fine but I can't figure out why?

Edit: Adding some more information based on comments:

While logging, I observed that :

(float)(b * 1000) / (float)c + .5f

was being evaluated to a really large value , which when type-casted to int was leading to the overflow. Individually, float(b*1000) and float(c) were being calculated correctly however.

Answer 1

b * 1000 is most likely overflowing the int type; the behaviour of doing this is undefined.

Drop all those obfuscating casts, and use

b * 1000.0 / c + 0.5

instead. 1000.0 is a double literal and causes the first term to be evaluated in floating point.

You ought to check the size of the expression before converting back to int, especially if c is small. std::numeric_limits is useful for that.