regex_macros: Fix valgrind errors

[alexcrichton]

While speedier to not perform initialization, these calls to mem::init() are
causing valgrind errors when using compiled macros.

Closes #13969

[sfackler]

Is Valgrind just not smart enough to figure out that the fields have actually been initialized by the time they're read, or is there actually a logic problem here?

[alexcrichton]

I suspect there is an actual logic problem (I'd be surprised if valgrind had a false positive here), but I also have not taken the time to investigate the errors to figure out why it may be a logic problem.

[sfackler]

cc @BurntSushi

[gereeter]

It is actually reading uninitialised data, but this is deliberate and not a logic error. See the cited source for more details.

Interestingly, RE2, which uses the same trick, has the following code:

// Don't need to zero the memory, but do so anyway
// to appease Valgrind.
if (valgrind_) {
  for (int i = 0; i < max_size; i++) {
    dense_[i] = 0xababababU;
    sparse_to_dense_[i] = 0xababababU;
  }
}

Additionally, it should be more efficient if you initialize sparse to !0 instead of 0, as this will make the membership test fail on the first condition instead of the second.

[alexcrichton]

Sounds like this needs some more investigation.

[thestinger]

Reading uninitialized data is valid behaviour as long as the code is able to handle the results of reads varying each time.