Many pipelined architectures do not define in datasheets exactly what can cause a pipeline stall - usually they provide rough guidance, but not information about corner cases. Branch prediction likewise is rarely fully defined in the datasheets.
Some architectures have some operations which are data dependant - for example, "division takes 6-8 cycles". The software developer can't be sure it will be 6, 7, or 8, and if they were to measure performance, they'd be relying on undefined performance behaviour, since a new revision of the same model of chip could change performance.
Also, most architectures with DMA aren't cycle accurate because the SRAM is single-ported, so you get a CPU stall if you happen to read memory in the same cycle as a hardware device needs more data.