Bit-twiddling vs. Logic in Modern C++

Why if (a & b) is not a free lunch

The folklore

“Branching is slow, so replace && with & (or || with |) and the optimizer will thank you.”

That advice shows up in code reviews, on Reddit threads, and even in seasoned code bases. The intuition is simple:

• && may short-circuit, so the compiler must generate a conditional branch.
• & always evaluates both operands, so the resulting instruction stream can be branch-free.

Therefore, the bitwise form “must” be faster—right?

We are going to stress-test that claim from four angles:

1. Language rules (what the Standard guarantees).
2. What compilers actually emit in 2025.
3. Micro-architecture (branch prediction, cache, ILP).
4. Human factors—maintainability, bugs, and intent.

The result is a nuanced picture: sometimes & wins, often it makes no difference, and in many real programs it is a liability.

Ground truth: what the Standard says¹

• Built-in && must evaluate left → right and must not evaluate the right operand when the left is false (or when the result is already known for ||). This is sequenced, not negotiable.²
• Overloaded operator&&/|| are just function calls—no short-circuiting.
• Bitwise &/| always evaluate both operands and produce an integral value.

Key takeaway: Behavioural semantics are fixed. The optimizer can re-arrange side-effect-free instructions, but it may not “speculatively” evaluate an operand if doing so would create an observable difference in the abstract machine.

Cppreference summarises the rule crisply: “Built-in && and || perform short-circuit evaluation; bitwise logic operators do not.”³

“But my disassembler shows identical code!”

That is not a myth. For side-effect-free operands the optimiser is allowed to collapse:

bool fast = (x & y);      // bitwise
bool safe = (x && y);     // logical

into the same branch-free sequence because it can prove that:

• x and y are scalars already materialised in registers,
• reading them twice is free,
• the observable result (fast/safe) is identical.

A nice real-world benchmark shows exactly that: Clang 15 with -O3 produced nearly identical inner loops for & and && scans over a 160 M-element array.⁴

What happened to short-circuiting? It is still honoured semantically, but the optimiser recognised that evaluating the second operand eagerly does not change program behaviour, so it emitted branch-free code that happens to read both variables. The standard’s “as-if” rule is satisfied.

The hidden foot-guns of if (a & b)

Performance anatomy

1. Branch prediction: Modern CPUs predict the jne generated by && with >95 % accuracy for stable data patterns. Misprediction costs ~14 cycles, but only when the data is adversarial.

2. Instruction-level parallelism: Eager evaluation can sometimes reduce ILP because both operands must be ready before the fused compare can retire.

3. Cache traffic: If b is a pointer dereference, forcing the read every time may hurt cache residency. Short-circuiting can be a win.

4. Speculative loads: An out-of-order core may fetch b early anyway; if it turns out the load was unnecessary the cost is often hidden.

5. GPU / SIMD kernels: Here branch divergence is lethal; replacing && with & (or *) makes sense—but you should express it explicitly in a data-parallel style (e.g. select(mask, …)), not with C++ control flow.

Measuring instead of guessing

Below is a minimal benchmark you can paste into Compiler Explorer or run with perf stat—modify the MODE macro to flip operators.

#include <vector>
#include <random>
#include <chrono>
#include <iostream>

#ifndef MODE               // 0 = bitwise &, 1 = logical &&
#define MODE 1
#endif

int main() {
    constexpr size_t N = 128 * 1024 * 1024;
    std::vector<uint8_t> a(N), b(N);
    std::mt19937 rng(0);
    std::uniform_int_distribution<int> dist(0, 255);
    for (size_t i = 0; i < N; ++i) {
        a[i] = dist(rng);
        b[i] = dist(rng);
    }

    size_t hits = 0;
    auto t0 = std::chrono::steady_clock::now();
    for (size_t i = 0; i < N; ++i) {
#if MODE
        if ((a[i] > 200) && (b[i] > 200))  // logical
#else
        if ((a[i] > 200) &  (b[i] > 200))  // bitwise
#endif
            ++hits;
    }
    auto t1 = std::chrono::steady_clock::now();
    std::cout << "hits=" << hits << "  "
              << std::chrono::duration<double>(t1 - t0).count()
              << " s\n";
}

On an Ice Lake i7 at -O3 the difference is typically < 1 %, completely drowned by measurement noise unless you pin the benchmark and flush caches between runs.

When does bitwise pay off?

• Inside tight, branch-averse GPU shaders or SIMD loops, where every warp must execute the same instruction stream.

• When computing a compound predicate that you will reuse later:

  uint32_t m = flags & READY & ENABLED & !ERROR;
  if (m) …
  

• For bit-mask idioms (if (flags & FLAG_WRITE))—but that is not a boolean logic replacement; the intent is genuinely “test bit k”.

Even in those cases document your intent with a comment; the form is unusual enough that reviewers will ask.

Debunking common myths

Concurrency and atomics

For std::atomic<bool> the value read is a side effect visible to other threads (§6.9.2.1 ¶7). A conforming compiler must not merge or speculate away the short-circuit. Rely on && to avoid spurious loads/stores across threads; replacing it with & can lengthen the critical section or even change lock-free algorithms.

Style, intent, and code-review heuristics

1. Default to &&/|| unless you have a measured reason.
2. Document any deliberate use of & on booleans (// branch-free predicate for GPU).
3. Isolate tricks behind helper functions or constexpr lambdas so the intent is explicit.
4. Write tests that capture side-effect expectations. A refactor that changes ++b to b can silently break logic when & is used.

A quick guideline table

Conclusion

The performance delta between && and & has mostly evaporated on today’s optimising compilers and superscalar CPUs. What remains is a semantic delta that can—and regularly does—bite maintainers.

Unless you have a profiler trace, an ISA you know intimately, and operands guaranteed side-effect-free, treat if (a & b) for what it is: a micro-optimisation gamble that trades clarity and correctness for at best marginal speed gains.

Opt for the code that communicates intent first. Let the optimiser prove you wrong later.

TL;DR – Use && unless you can prove that & is faster and harmless. The compiler usually does the proving for you.