Issue 5908 - Optimizer generates wrong value with divide-by-zero.

[yebblies]

Reinstate optimizer error for integer division by zero

https://issues.dlang.org/show_bug.cgi?id=5908

[dnadlinger]

Huh – shouldn't this produce a runtime FP exception?

[yebblies]

It does if it can't catch it at compile time.

[WalterBright]

Not necessary since OPremquo is only for integral operands. You can document this by inserting assert(!tyfloating(tym)) instead.

[yebblies]

That makes sense. Updated.

[WalterBright]

Integer divide by zero should not raise a FP exception.

There are also issues with the following code:

int b = 0;
...
if (condition that is false when b is 0 but the optimizer doesn't know that)
    x = a / b;

The optimizer will, because of constant propagation, see a/0 and issue a bogus error message.

[yebblies]

There are also issues with the following code:

I'm not so sure about that example. b is a constant, so I think the division by zero is always a bug, whether it's guarded by an if or not.

We have the same issue with null dereference. It can easily be worked around by adding && b to the if condition.

[WalterBright]

I'm not so sure about that example. b is a constant, so I think the division by zero is always a bug, whether it's guarded by an if or not. We have the same issue with null dereference. It can easily be worked around by adding && b to the if condition.

That's right, we do have the same issue with null dereference being detected by the optimizer using data flow analysis, and I had to disable the error message because real code would trip the error even though the code logic ensured it would never happen. Here's another example:

if (condition)
    b = 0;
else
    b = 1;
...
if (!condition)
    a = c / b;   // uh-oh! b could be 0!

The optimizer sees a path where b could be 0, and would issue the error. But that path would never happen.

[yebblies]

The optimizer sees a path where b could be 0, and would issue the error. But that path would never happen.

This error doesn't appear when b could be 0, only when it must be zero. My understanding is it's conservative.

[ghost]

@WalterBright: Feedback on the previous comment?

[WalterBright]

Auto-merge toggled on

[WalterBright]

Thought we'd give it a try and see what happens.

[yebblies]

Fingers crossed!