Consider the following build options (v3):

-DMI_OPT_ARCH=ON (the default on arm64) enables architecture-specific optimizations: -march=haswell -mavx2 on x64, -march=armv8.1-a on arm64, and -march=rv64gcb_zacas on RISC-V. These improve bit operations and atomic operations, but unsupported hardware or emulators may fail with an illegal instruction. Apple arm64 defaults to -march=armv8.3-a, supported on M1 and later, for efficient load-acquire instructions. RISC-V hardware supporting zalasr can use -DMI_OPT_ARCH=rv64gcb_zacas_zalasr.
-DMI_ALLOW_THP=OFF can significantly reduce resident memory usage for some programs.
-DMI_ALLOW_THP=FULL can significantly improve performance for some programs, but still respects an OS THP setting of [never].
-DMI_FREE_IS_CHECKED=ON checks pointers passed to mi_free(), at a small performance cost. This can be useful when overriding malloc/free process-wide. Alternatively, mi_cfree() always checks pointer validity. Pointer checking is always enabled at secure level CHECKED and higher.

In source code:

Use mi_zalloc() for zero-initialized allocations.
Use typed macros such as mi_malloc_tp(): they are safer and use mi_malloc_csize() internally.
Use mi_malloc_csize() and mi_free_csize() when the size is a compile-time constant. These inline and select mi_malloc_small() or mi_free_small() where possible. The C++ mimalloc-new-delete.h sized overrides use these too. Use mi_free_csize_nonnull() when the pointer is known to be non-NULL.
Use mi_malloc_small() when the requested size is at most MI_SMALL_SIZE_MAX (128 machine words, or 1 KiB on a 64-bit system). Use mi_free_small() for pointers returned by mi_malloc_small(), and mi_free_small_nonnull() when the pointer is known to be non-NULL.

For runtimes and compilers, such as Koka, Lean, or Rust, prefer the constant-size or small-allocation APIs when sizes are known. If a runtime already carries thread-local state, it can cache mi_theap_get_default() for each thread and pass that thread's theap directly to mi_theap_malloc() or mi_theap_malloc_small(). This avoids repeatedly looking up the thread-local pointer and skips a NULL theap check. The benefit depends on the OS implementation of thread-local storage.