On 30 March 2013 21:19, Richard Henderson <[email protected]> wrote: > On 2013-03-30 14:14, Peter Maydell wrote: >> >> If you're doing a sub sp, sp, #n anyway, you might as well >> add 4 to that and drop the save/load of r12 from the stmdb/ldmia. >> It was only in there to save having an extra sub/add insn. > > Fair enough, not that it ought to make a difference to the cycle > count.
Yeah, the TRMs suggest the ARM9 was the last to do ldm/stm at 32 bits per cycle, and I'm guessing the prologue isn't actually a sufficiently hot path to be worth worrying about individual cycles anyway. I just felt it would be a bit tidier. -- PMM
