Performance differences of record creation methods

[JustinAzoff]

Given this record:

type Info: record {
    a: count;
    b: count;
    c: count;
};

And the following methods:

noop:

function foo() {
}

expanded:

function foo() {
    local tmp: Info;
    tmp$a=1;
    tmp$b=2;
    tmp$c=3;
}

list:

function foo() {
    local tmp: Info = [$a=1,$b=2,$c=3];
}

record:

function foo() {
    local tmp = Info($a=1,$b=2,$c=3);
}

record typed:

function foo(){
    local tmp: Info = Info($a=1,$b=2,$c=3);
}

and this driver script:

event zeek_init() {
    local x = 0;
    while ( ++x < 50000000 ) {
        foo();
    }
}

I see the following results:

##  record_inline/noop.bro
  50000000 took 12.07          4142427 it/sec
  50000000 took 11.83          4226811 it/sec
  50000000 took 11.99          4171400 it/sec
  50000000 took 11.90          4203419 it/sec
  50000000 took 11.86          4215261 it/sec
##  record_inline/expanded.bro
  50000000 took 25.20          1983888 it/sec
  50000000 took 25.45          1964645 it/sec
  50000000 took 24.95          2004054 it/sec
  50000000 took 25.22          1982687 it/sec
  50000000 took 24.96          2002827 it/sec
##  record_inline/list.bro
  50000000 took 32.99          1515566 it/sec
  50000000 took 30.32          1648815 it/sec
  50000000 took 29.68          1684754 it/sec
  50000000 took 29.94          1669781 it/sec
  50000000 took 30.67          1630465 it/sec
##  record_inline/record.bro
  50000000 took 38.16          1310325 it/sec
  50000000 took 42.22          1184264 it/sec
  50000000 took 44.39          1126294 it/sec
  50000000 took 37.63          1328741 it/sec
  50000000 took 38.24          1307655 it/sec
##  record_inline/record_typed.bro
  50000000 took 38.79          1288888 it/sec
  50000000 took 42.41          1179106 it/sec
  50000000 took 39.67          1260294 it/sec
  50000000 took 41.97          1191245 it/sec
  50000000 took 39.13          1277753 it/sec

It appears that once you account for the base function call overhead, creating records and then filling them in takes ~13s for 50000000, but using the typed record syntax takes 28s. I've seen this somewhat when looking at memory allocations as well, I believe the difference is because the 'nicer' syntax ends up allocating a temporary record that it has to merge into the final record.. so it takes 2x the time because it does 2x the allocations and 2x the work.