Overflow behaviour relies on Optimization

[vaivaswatha]

While implementing #5385, I found that when optimizations are disabled, (in particular, constcombine and/or mem2reg for this bug), the IR we're generating for addition doesn't check for overflows correctly.

This is the IR for should_fail/arith_overflow/u8_add_overflow (note that there's no mem2reg or constcombine):

script {
    entry fn main() -> bool, !1 {
        local u8 a
        local u8 b
        local u8 res
        local u8 result

        entry():
        v0 = get_local ptr u8, a, !2
        v1 = const u8 255, !4
        store v1 to v0, !2
        v2 = get_local ptr u8, b, !5
        v3 = const u8 1, !6
        store v3 to v2, !5
        v4 = get_local ptr u8, a, !7
        v5 = load v4
        v6 = get_local ptr u8, b, !8
        v7 = load v6
        v8 = add v5, v7, !9
        v9 = get_local ptr u8, res, !12
        store v8 to v9, !13
        v10 = get_local ptr u8, res, !15
        v11 = load v10, !9
        v12 = const u8 255, !4
        v13 = cmp gt v11 v12, !9
        cbr v13, add_1_block0(), add_1_block1(), !17

        add_1_block0():
        v14 = const u64 0, !18
        revert v14, !20

        add_1_block1():
        ...
    }
   ...

Because we're using byte stores and loads for u8, we end up ignoring the overflowed part in the u64 register, thus not triggering the revert.

When this issue is fixed, we should be enabling the tests should_pass/arith_overflow/u8_add_overflow and should_pass/arith_overflow/u8_mul_overflow which were both disabled in #5385.

I suspect the bug got introduced with #4929.

[vaivaswatha]

Easy repro without dependence on #5385

script;

fn main() {

}

fn add(a: u8, b: u8) -> u8 {
   a + b
}

#[test]
fn should_overflow_revert() {
   let a: u8 = u8::max();
   let b: u8 = 1;
   add(a, b);
}

On master, forc test hits a revert, as it should, but with mem2reg (by adding Return Ok(false) at the beginning of promote_to_registers), forc test just passes (but it shouldn't have).