range_str: propagate rehashed marks to caller

Author: aitap

inst/tests/tests.Rraw

@@ -21876,3 +21876,10 @@ test(2347, DT[i, .(result = all(is.na(grp) == is.na(a))), by = grp][,all(result)
 DT = data.table(a = as.Date("2010-01-01"), b = 1L)
 test(2348.1, tryCatch(DT[a == as.Date("20100101")], error=conditionCall)[[1L]], quote(charToDate))
 test(2348.2, tryCatch(DT[a == as.Date("20100101") | b == 2L], error=conditionCall)[[1L]], quote(charToDate))
+
+# exercise rehashing during forder, #6694
+strings = as.character(6145:1)
+DT = data.table(x = strings)
+setorder(DT, x)
+test(2349, DT[["x"]], sort.int(strings, method='radix'))
+rm(DT, strings)

src/forder.c

@@ -291,14 +291,15 @@ static void cradix(SEXP *x, int n)
   free(cradix_xtmp);   cradix_xtmp=NULL;
 }
 
-static void range_str(const SEXP *x, int n, uint64_t *out_min, uint64_t *out_max, int *out_na_count, bool *out_anynotascii, bool *out_anynotutf8, hashtab * marks)
+static void range_str(const SEXP *x, int n, uint64_t *out_min, uint64_t *out_max, int *out_na_count, bool *out_anynotascii, bool *out_anynotutf8, hashtab **marks_)
 // group numbers are left in truelength to be fetched by WRITE_KEY
 {
   int na_count=0;
   bool anynotascii=false, anynotutf8=false;
   if (ustr_n!=0) internal_error_with_cleanup(__func__, "ustr isn't empty when starting range_str: ustr_n=%d, ustr_alloc=%d", ustr_n, ustr_alloc);  // # nocov
   if (ustr_maxlen!=0) internal_error_with_cleanup(__func__, "ustr_maxlen isn't 0 when starting range_str");  // # nocov
   bool fail = false;
+  hashtab *marks = *marks_;
   #pragma omp parallel for num_threads(getDTthreads(n, true)) shared(marks, fail)
   for(int i=0; i<n; i++) {
     SEXP s = x[i];
@@ -338,6 +339,8 @@ static void range_str(const SEXP *x, int n, uint64_t *out_min, uint64_t *out_max
       }
     }
   }
+  // if the hash table grew, propagate the changes to the caller
+  *marks_ = marks;
   if (fail) internal_error_with_cleanup(__func__, "failed to grow the 'marks' hash table");
   *out_na_count = na_count;
   *out_anynotascii = anynotascii;
@@ -613,7 +616,7 @@ SEXP forder(SEXP DT, SEXP by, SEXP retGrpArg, SEXP retStatsArg, SEXP sortGroupsA
       // need2utf8 now happens inside range_str on the uniques
       marks = hash_create(4096); // relatively small to allocate, can grow exponentially later
       PROTECT(marks->prot); n_protect++;
-      range_str(STRING_PTR_RO(x), nrow, &min, &max, &na_count, &anynotascii, &anynotutf8, marks);
+      range_str(STRING_PTR_RO(x), nrow, &min, &max, &na_count, &anynotascii, &anynotutf8, &marks);
       break;
     default:
       STOP(_("Column %d passed to [f]order is type '%s', not yet supported."), col+1, type2char(TYPEOF(x)));

src/hash.c

@@ -149,7 +149,7 @@ void hash_set(hashtab *self, SEXP key, R_xlen_t value) {
     if (!new_h) internal_error( // # nocov
       __func__, "hash table full at n_full=%zu and failed to rehash", h->size
     );
-    // overwrite the existing table, keeping the external pointer
+    // overwrite the existing table, keeping the EXTPTR -> (next ->)* h chain intact
     free(h->table);
     *h = *new_h;
     free(new_h);