CData allocation recording

[sergos]

ffinew.lua.txt

Trace recording is aborted due to NYI for the case of a struct with aggregate allocation.
Ironically, it works for the array of the same structs - see the test in attach.

The VLA/VLS machinery was added in f1f7e40 on the v2.1 branch, closing NYI for array of aggregates around line 1013
I believe it could be also applied for the case of struct with aggregates at line 1042 since it handles zero init, failing with initializers list anyways.

The test attached shows the following 500x uplift with initial state

$ ./luajit ../../test/ffinew.lua 
Median time for struct allocation: 0.021003 sec
Median time for array allocation: 4.6000000000213e-05 sec

turned into

$ ./luajit ../../test/ffinew.lua 
Median time for struct allocation: 4.8e-05 sec
Median time for array allocation: 4.8000000000001e-05 sec

with

--- a/src/lj_crecord.c
+++ b/src/lj_crecord.c
@@ -1039,7 +1039,7 @@ static void crec_alloc(jit_State *J, RecordFFData *rd, CTypeID id)
          dc = ctype_rawchild(cts, df);  /* Field type. */
          if (!(ctype_isnum(dc->info) || ctype_isptr(dc->info) ||
                ctype_isenum(dc->info)))
-           lj_trace_err(J, LJ_TRERR_NYICONV);  /* NYI: init aggregates. */
+            goto special;
          if (J->base[i]) {
            sp = J->base[i];
            sval = &rd->argv[i];

[MikePall]

The fix is not that simple, since some stores may already have been emitted.

I've added a proper workaround that helps the zero-fill struct-of-NYI case.