Constant Division in Compilers: The Granlund-Montgomery Method

A breakthrough in compiler optimization has nearly doubled the performance of division operations in processors from Apple and Intel, marking one of the most significant low-level speedups in modern computing in recent years.

The advancement centers on replacing the Granlund and Montgomery (GM) method, which has been the standard approach for division by constants in compilers since 1994, with a new algorithm that improves the efficiency of converting division into multiplication and bit-shifting operations. This technique, long used to avoid costly division instructions on CPUs, has now been refined to reduce computational overhead further, particularly in tight loops and performance-critical code.

According to researchers and engineers involved in the update, the new method improves the accuracy and range of the magic numbers and shift values used in the GM method, allowing compilers to generate faster code for a broader set of divisors without sacrificing correctness. The optimization is particularly effective for 32-bit and 64-bit integer division, which remains common in systems programming, game engines and real-time applications.

Independent benchmarks conducted by performance analysts show that the updated compiler passes in LLVM and GCC now produce code that runs up to 1.8 times faster on Apple’s M-series chips and Intel’s latest Core and Xeon processors when handling division-heavy workloads. The gains are most noticeable in scenarios involving hash tables, graphics processing, and numerical simulations, where division operations occur frequently in inner loops.

The improvement stems from collaborative work between compiler developers at Apple, Intel, and contributors to the LLVM project, who identified limitations in the original GM method’s handling of edge cases and signed integers. By refining the mathematical bounds used to compute the multiplicative inverse and shift amount, the new approach reduces the need for fallback correction steps, which previously eroded performance gains.

One engineer familiar with the changes, speaking on condition of anonymity due to corporate policy, noted that the update required no changes to hardware or instruction sets, making it a pure software win. “We’re getting nearly double the throughput on division-bound code just by improving how compilers translate high-level expressions into machine instructions,” the engineer said. “It’s rare to see such a broad impact from a compiler tweak alone.”

The optimization has been integrated into recent versions of LLVM (version 19 and later) and is being backported into GCC 14, with Microsoft’s MSVC also evaluating similar approaches for future releases. Apple has confirmed that the changes are active in its latest Xcode toolchain, benefiting apps compiled for iOS, macOS, and visionOS. Intel has validated the improvements in its oneAPI and Intel C++ Compiler toolchains.

While the GM method remains a foundational technique in compiler design, this refinement represents the first major improvement to its core logic in three decades. Experts say the update could influence energy efficiency in data centers and extend battery life in mobile devices by reducing the number of cycles spent on arithmetic operations.

As demand grows for efficient computation in AI inference, edge computing, and real-time systems, such low-level optimizations are becoming increasingly valuable. The change underscores how advances in compiler technology can deliver tangible performance benefits without requiring new hardware, offering a path forward for continued progress in computing performance beyond the limits of Moore’s Law.