I reordered the MBA transformation passes in MBASub::RunOnBasicBlock so that runOnMul runs first, before runOnSub, runOnAdd, and runOnXor.

The problem

Some of the transformations (especially in xor_ops and sub_ops) generate new Mul instructions as part of their MBA formulas.
For example:

From xor_ops:
a ^ b = (a + b) - 2 * (a & b)

From sub_ops:
X - Y == (X ^ -Y) + 2*(X & -Y)

Before runOnMul would run after these passes, immediately picking up these newly generated Mul instructions and expanding them again using mul_ops (which itself expands 1 Mul into 2 Muls + overhead).

The issue is that mul_ops expands a multiplication into multiple new multiplications:
b * c = (((b | c) * (b & c)) + ((b & ~c) * (c & ~b)))

So what ended up happening was:

Sub / Xor generate Mul, Mul expands into more Mul, Those can generate even more Mul and Repeat next iteration xD
On math-heavy or crypto-heavy code (like ChaCha20), this cause the IR size to explode and hanging at runtime