# ARM Timing ## Architectural versus implementation timing ARM instruction timing varies strongly across ARM generations and implementations. SLeeLa records: - instruction fetch; - decode; - issue; - execution unit; - dependency/stall; - branch prediction/result; - load/store; - cache hit/miss; - TLB/translation walk; - exception; - retirement. A Cortex-M profile can use a relatively direct pipeline model, while a Cortex-A or Neoverse profile can supply wider out-of-order and speculative timing. ## AArch64 instruction size A64 instructions are fixed 32-bit words. citeturn0search20 ## AArch32/Thumb A32 and T32 have different instruction encodings and timing characteristics. The decoder and timing model therefore track instruction-state explicitly. ## Expert timing, voltages, and variable clocks (shared across all CPU models) The per-CPU timing above is the architectural/behavioral view for this part. The adjustable, quality-first timing system is shared by every CPU model in `lib/cpu` and is documented once, in the package guide: - [`lib/cpu/TIMING.md`](../../TIMING.md) — Voltages, Variable Clocks, and Expert Timings: `SLVoltageRail` (actual voltages in mV), `SLCpuTiming`, `SLBusTiming`, `SLRamTiming`, `SLIOTiming`, the Novice/Basic/Advanced/Expert `SLTimingClass` tiers, and the `SLTimingProfile` -> `SLTimingManifest` a compiled `.sleela` image carries so the compiler/VM honor the **declared** operating point. - [`lib/cpu/components/README.md`](../../components/README.md) — running a CPU at spec and **overclocking** it (`SLClockPolicy` / `setSpecClock` / `overclock`), plus the RAM / storage / optical / system-bus parts bin and the DMA / GPU / native-RAM options. Every CPU model (this one included) inherits the clock/overclock surface from `SLCPURuntime` and can be assembled with those shared expert-timing, voltage, and component options; they are a build-layer concern, not re-declared per CPU.