On Thu, Apr 12, 2012 at 4:24 PM, Richard Guenther <richard.guent...@gmail.com> wrote: > On Thu, Apr 12, 2012 at 2:00 PM, Alexander Monakov <amona...@ispras.ru> wrote: >> >>> Can atom execute two IMUL in parallel? Or what exactly is the pipeline >>> behavior? >> >> As I understand from Intel's optimization reference manual, the behavior is >> as >> follows: if the instruction immediately following IMUL has shorter latency, >> execution is stalled for 4 cycles (which is IMUL's latency); otherwise, a >> 4-or-more cycles latency instruction can be issued after IMUL without a >> stall. >> In other words, IMUL is pipelined with respect to other long-latency >> instructions, but not to short-latency instructions. > > It seems to be modeled in the pipeline description though: > > ;;; imul insn has 5 cycles latency > (define_reservation "atom-imul-32" > "atom-imul-1, atom-imul-2, atom-imul-3, atom-imul-4, > atom-port-0") > > ;;; imul instruction excludes other non-FP instructions. > (exclusion_set "atom-eu-0, atom-eu-1" > "atom-imul-1, atom-imul-2, atom-imul-3, atom-imul-4") >
The main idea is quite simple: If we are going to schedule IMUL instruction (it is on the top of ready list) we try to find out producer of other (independent) IMUL instruction that is in ready list too. The goal is try to schedule such a producer to get another IMUL in ready list and get scheduling of 2 successive IMUL instructions. And MD allows us only prefer scheduling of successive IMUL instructions, i.e. If IMUL was just scheduled and ready list contains another IMUL instruction then it will be chosen as the best candidate for scheduling. > at least from my very limited guessing of what the above does. So, did you > analyze why the scheduler does not properly schedule things for you? > > Richard. > >> >> From reading the patch, I could not understand the link between pipeline >> behavior and what the patch appears to do. >> >> Hope that helps. >> >> Alexander
