Perl_hv_common - don't call share_hek_flags when (keysv_hek != NULL)

[richardleach]

When creating a new HE for insertion into a hash that supports shared keys, Perl_hv_common always achieves this via S_share_hek_flags.

S_share_hek_flags searches the shared string table - PL_strtab - for a matching entry and inserts one if needs be, returning the relevant HEK*.

However, if Perl_hv_common's keysv is a "HEK in disguise" to begin with, keysv_hek already contains the resulting HEK* - and we are able to bump the refcount on it. So this commit checks for that case and branches accordingly.

gcov counts from a perl build and test_harness run suggest that having a valid keysv_hek is already very common.

129097943:  996:    if (LIKELY(HvSHAREKEYS(hv))) {
129097941:  997:        entry = new_HE();
129097941:  998:        if (keysv_hek) {
 50255991:  999:            HeKEY_hek(entry) = keysv_hek;
 50255991: 1000:            share_hek_hek(keysv_hek);
        -: 1001:        } else {
 78841950: 1002:            HeKEY_hek(entry) = share_hek_flags(key, klen, hash, flags);
        -: 1003:        }
        -: 1004:    }

If greater use is made of PL_strtab in the future, such as for OP_CONST SVs, the keysv_hek is likely to become correspondingly hotter.

(If someone points out a compelling reason why this isn't already done, I'll add comments to that effect instead.)

---

• I haven't benchmarked yet to see if this change is likely to be noticeable. It may warrant a perldelta entry.

[richardleach]

A rough comparison:

my %h1 = map { $_ => 0} qw(one two three four five);
for (1 .. 100_000_000) {
    my %h2 = %h1;
}

blead:

         26,730.35 msec task-clock                       #    1.000 CPUs utilized             
               127      context-switches                 #    4.751 /sec                      
                 0      cpu-migrations                   #    0.000 /sec                      
               197      page-faults                      #    7.370 /sec                      
   124,235,900,859      cycles                           #    4.648 GHz                       
     3,323,830,317      stalled-cycles-frontend          #    2.68% frontend cycles idle      
   503,542,464,696      instructions                     #    4.05  insn per cycle            
   106,009,840,217      branches                         #    3.966 G/sec                     
       101,478,688      branch-misses                    #    0.10% of all branches  

patched:

         25,003.63 msec task-clock                       #    1.000 CPUs utilized             
               119      context-switches                 #    4.759 /sec                      
                 0      cpu-migrations                   #    0.000 /sec                      
               194      page-faults                      #    7.759 /sec                      
   116,791,723,673      cycles                           #    4.671 GHz                       
     3,307,581,712      stalled-cycles-frontend          #    2.83% frontend cycles idle      
   477,589,786,476      instructions                     #    4.09  insn per cycle            
   101,859,216,594      branches                         #    4.074 G/sec                     
        51,308,472      branch-misses                    #    0.05% of all branches

Worth doing, but I'm not sure it warrants a perldelta entry.

[jkeenan]

Need to get make test_porting working again for this p.r.

[mauke]

I believe the porting tests are only failing because there are some tags missing from this fork. The tests would probably succeed in the main repository.

[richardleach]

I've rebased and pushed.