When should the compiler NOT use conditional moves?

Whenever evaluating both branches would be wasteful, unsafe, or have side effects — because cmov always computes both results before selecting one.

The whole speed trick of cmov is that it skips branching by computing both values. That backfires in three situations:

1. Expensive computations — both run even though only one is kept:

val = Test(x) ? Hard1(x) : Hard2(x);   // both Hard1 and Hard2 evaluated

2. Unsafe computations — a path that would normally never execute now does:

val = p ? *p : 0;                      // *p dereferences even when p is NULL → crash

3. Side effects — both updates happen, corrupting state:

val = x > 0 ? (x *= 7) : (x += 3);     // both assignments execute

The rule: cmov is only safe when both candidate computations are cheap, side-effect-free, and cannot fault. The compiler weighs this with heuristics (and sometimes profiling) when deciding between cmov and a real branch — which is why you see both in optimized code.

Go deeper:

• Predication (Wikipedia) — predication trade-offs when both paths must execute.
• Felix Cloutier — CMOVcc — shows cmov continues unconditionally — both candidates computed.