IntegerOverflowChecker generates redundant instructions for unsigned intrinsics

[keram88]

Currently, instructions like @llvm.uadd.with.overflow.i32 are translated to approximately:

// %res = call {i32, i1} @llvm.uadd.with.overflow.i32(%a, %b)
...
$a1 = $zext.i32.i64($a);
$a2 = $zext.i32.i64($b);
$res = $add.i64($a1, $b1);
$over = $ugt.i64($res, UINT_MAX); // 1 if overflow
$under = $ult.i64($res, 0); // 1 if underflow
$flag = $or($over, $under); // 1 if overflow

Both arguments are unsigned, so the result will always be greater than or equal to zero, making $under redundant.

[keram88]

This was fixed quite some time ago. I'm not sure why it didn't get closed automatically.

[keram88]

Apparently it never got merged in...

[zvonimir]

@keram88 : so what is happening with this in the end? I just noticed you closed the pull request.

[keram88]

I couldn't remember why I never made a PR for that branch, which is from June 2018. I realized that it's not quite right to completely remove the lower bound check for unsigned math since subtraction can overflow in integer mode, e.g., 2-3. I'll reopen the PR since I think I have a clean-ish way to deal with it.