| Message ID | 20240328101356.300374-4-e@80x24.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | switch to tombstone-free khashl table | expand |
diff --git a/khashl.h b/khashl.h index 1e724bbf88..30f85dc5e2 100644 --- a/khashl.h +++ b/khashl.h @@ -265,7 +265,7 @@ typedef struct { low = hash & ((1U<<g->bits) - 1); \ h = &g->sub[low]; \ ret = prefix##_sub_getp_core(h, key, hash); \ - if (ret == 1U<<h->bits) r.sub = low, r.pos = (khint_t)-1; \ + if (ret >= kh_end(h)) r.sub = low, r.pos = (khint_t)-1; \ else r.sub = low, r.pos = ret; \ return r; \ } \
The ->bits field of regular khashl structs is invalid when the ->keys array is NULL. Thus the ensemble *_getp implementation must follow existing *_get and *_getp usage conventions and check the iterator against kh_end(). This fixes a fast-import crash on t3427-rebase-subtree.sh in an abandoned commit to use the ensemble implementation for oid_map and oid_pos. I've abandoned the aforementioned commit for now since it was more intrusive, more expensive for small tables, and realloc(3) on glibc is already optimized using mremap(2) for large hash resizes. Signed-off-by: Eric Wong <e@80x24.org> --- khashl.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)