[PATCH 4/5] Remove predictors that are unrealiable.

[Martin Liška]

On 01/19/2018 12:57 PM, Martin Liška wrote:
> Yes, there's a huge difference in between CPU 2006 and 2017. Former has 63% w/ dominant edges,
> and later one only 11%. It's caused by these 2 benchmarks with a high coverage:
> 

Hi.

I'm sending details about the 2 edges that influence the statistics significantly:

> 500.perlbench_r: regexec.c.065i.profile:
>   negative return heuristics of edge 1368->1370: 2.0%  exec 2477714850 hit 2429863555 (98.1%)

2477714850: 3484:S_regtry(pTHX_ regmatch_info *reginfo, char **startposp)
        -: 3485:{
2477714850: 3486:    CHECKPOINT lastcp;
2477714850: 3487:    REGEXP *const rx = reginfo->prog;
2477714850: 3488:    regexp *const prog = ReANY(rx);
2477714850: 3489:    SSize_t result;
[snip]
2477714850: 8046:    assert(!result ||  locinput - reginfo->strbeg >= 0);
2477714850: 8047:    return result ?  locinput - reginfo->strbeg : -1;
        -:  8048:}

As seen it return -1 if a regex is not found, which is in case of perlbench very likely branch.

> 
> and 523.xalancbmk_r:
> build/build_peak_gcc7-m64.0000/NameDatatypeValidator.cpp.065i.profile:  negative return heuristics of edge 3->4: 2.0%  exec 1221735072 hit 1221522453 (100.0%)

1221735072:   74:int NameDatatypeValidator::compare(const XMLCh* const lValue
        -:   75:                                   , const XMLCh* const rValue
        -:   76:                                   ,       MemoryManager*     const)
        -:   77:{
1221735072:   78:    return ( XMLString::equals(lValue, rValue)? 0 : -1);
        -:   79:}
        -:   80:

IP profile dump file:

xercesc_2_7::NameDatatypeValidator::compare (struct NameDatatypeValidator * const this, const XMLCh * const lValue, const XMLCh * const rValue, struct MemoryManager * const D.17157)
{
  bool _1;
  int iftmp.0_2;

  <bb 2> [count: 1221735072]:
  # DEBUG BEGIN_STMT
  _1 = xercesc_2_7::XMLString::equals (lValue_4(D), rValue_5(D));
  if (_1 != 0)
    goto <bb 4>; [0.02%]
  else
    goto <bb 3>; [99.98%]

  <bb 3> [count: 1221522453]:

  <bb 4> [count: 1221735072]:
  # iftmp.0_2 = PHI <0(2), -1(3)>
  return iftmp.0_2;

}

Likewise, XML values are more commonly non-equal.
Ok, so may I mark also negative return with PROB_EVEN to track it?

Thanks,
Martin

> 
> Ideas what to do with the predictor for GCC 8 release?
> Martin