Dirty Optimization Secrets (C for Playdate)

Practical, low-level optimization techniques for C on Playdate aimed at extreme workloads like emulators, sims, or renderers. Key observation: the CPU is fast but memory (especially non-cached accesses) is slow, so code that stays in registers and minimizes loads runs much faster. Instruction cache is tiny (Rev A ~4 KB, Rev B ~16 KB), so smaller, denser core loops can beat -O3-produced larger code; using -Os and compressing a 20 KB emulator core down to ~2 KB produced large speedups. Keep hot core paths contiguous in memory (use attribute ((section(".text.foo"))) on core functions) and move rare-edge handlers out of the hot section to preserve icache residency and branch locality.

Concrete tooling and practices: use a custom linker script (copy link_map.ld and set override LDSCRIPT in your makefile/CMake) to place object files and align sections precisely, and inspect symbol addresses/sizes with nm Source/pdex.elf | sort to verify icache fit. Real-world results include a reported 20-30% slowdown from adding unused C functions (icache/size effects) and a surprising speed gain when replacing integer divisions in a tight loop with float additions; that latter case likely benefitted from freeing integer registers (float registers reduced register pressure) rather than raw division cycle counts.