FunC optimizer incorrectly removes division-by-zero checks in muldiv, muldivc, and muldivr

[Gusarich]

FunC’s optimizer improperly constant-folds expressions like muldiv(x, y, z), muldivc(x, y, z), and muldivr(x, y, z) to zero when either x or y is known to be zero, entirely ignoring the denominator z. According to TVM semantics, these instructions must fail with a division-by-zero error at runtime if z equals zero. The optimizer currently bypasses this check, causing incorrect behavior.

Minimal Example (muldiv):

Input FunC code:

int test(int z) {
    return muldiv(0, 1, z);
}

Incorrectly optimized output:

DECLPROC test
test PROC:<{
  DROP
  0 PUSHINT
}>

The same incorrect behavior occurs with muldivc and muldivr, since all three instructions share the same optimizer logic.

Expected Behavior:
The optimizer should not bypass runtime division-by-zero checks, allowing the TVM to correctly raise errors when the denominator (z) is zero.

Impact:
This bug silently prevents crucial runtime errors, potentially leading to incorrect or unintended behavior in deployed smart contracts.

---

LLM Fuzzing discovery (see tact-lang/tact#3123)