[AIE] add StickyWAW mutator to postscheduler/postpipeliner

[martien-de-jong]

Not checked on QoR yet, but I do see a significant improvement in one of the postpipeliner unit tests

[andcarminati]

Hi @martien-de-jong, in the last update of this mutator, we left an orphan command line option:

static cl::opt<unsigned> WAWStickyRegistersMemOpsThreshold(
    "aie-waw-sticky-register-mem-threshold", cl::Hidden, cl::init(4),
    cl::desc("Number of memory instructions to enable the register exclusion "
             "heuristic in WAW sticky registers dep. removal"));

Is it possible to include this removal in this PR?

[andcarminati]

LGTM.

[gbossu]

I'd be happy to see QoR numbers, especially for Neg :)

[krishnamtibrewala]

CHECK : Why do we need the call to "postProcessDAG" ?

[gbossu]

Some of the DAG mutations are required for correctness, e.g. giving a correct latency to memory edges

[gbossu]

I'd rather not merge this now, running the DAG mutators for pipelining seems to have a nasty side effect on the final loop schedule when pipelining fails. In the post-processing loop of softmax, which essentially does VLD -> VMUL -> VST.CONV, we now have two NOPs at the end, probably because we see a memory dep between VST.CONV and the VLD of the next iteration.

[gbossu]

Marking as "request changes" to make sure we don't inadvertently merge this before looking into #222 (comment)

[andcarminati]

Now that this change was integrated through other PR, we can close this one!

Change already integrated.