Avoid memcpy(dst, NULL, 0) in forder.c (#7055)

* flush(): avoid memcpy(<dst>, NULL, 0)

During thread contention, it's possible for a thread to not call push()
even once, leaving gs_thread[me] unallocated. The resulting call to
memcpy() was noticed by a sanitizer during additional CRAN checks.

* NEWS item

* Add nocov to the hard-to-test branch

Currently, the condition is only encountered during thread contention,
which is prohibitively expensive to test or at least requires
non-portable tricks.

* Explain why the branch is #nocov

Author: aitap

NEWS.md

@@ -38,6 +38,8 @@
 
 9. Joins to extended data.frames, e.g. `x[i, col := x.col1 + i.col2]` where `i` is a `tbl`, can use the `x.` and `i.` prefix forms, [#6998](https://github.com/Rdatatable/data.table/issues/6998). Thanks @MichaelChirico for the bug and PR.
 
+10. On a heavily loaded machine, a `forder` thread could try to perform a zero-length copy from a null pointer, which was de-facto harmless but is against the C standard and was caught by additional CRAN checks, [#7051](https://github.com/Rdatatable/data.table/issues/7051). Thanks to @helske for the report and @aitap for the PR.
+
 ### NOTES
 
 1. Continued work to remove non-API C functions, [#6180](https://github.com/Rdatatable/data.table/issues/6180). Thanks Ivan Krylov for the PRs and for writing a clear and concise guide about the R API: https://aitap.codeberg.page/R-api/.

src/forder.c

@@ -128,6 +128,8 @@ static void flush(void) {
   if (!retgrp) return;
   int me = omp_get_thread_num();
   int n = gs_thread_n[me];
+  // normally doesn't happen, can be encountered under heavy load, #7051
+  if (!n) return; // # nocov
   int newn = gs_n + n;
   if (gs_alloc < newn) {
     gs_alloc = (newn < nrow/3) ? (1+(newn*2)/4096)*4096 : nrow;