diff --git a/configure.ac b/configure.ac
index 25032ac7f..8e39bc5e3 100644
--- a/configure.ac
+++ b/configure.ac
@@ -162,6 +162,35 @@ HTS_CHECK_COMPILE_FLAGS_NEEDED([avx512f], [-mavx512f -mpopcnt],
   AC_DEFINE([HAVE_AVX512],1,[Defined to 1 if rANS source using AVX512F can be compiled.])
 ])
 
+dnl Check for presence of builtin_clz
+AC_MSG_CHECKING([for working __builtin_clz])
+AC_LINK_IFELSE([AC_LANG_PROGRAM([],[
+  int z = 10;
+  if (__builtin_clz(z)) {
+    return 0;
+  }
+])], [
+  AC_MSG_RESULT([yes])
+  AC_DEFINE([HAVE___BUILTIN_CLZ], 1,
+            [Defined to 1 if __builtin_clz works])
+], [
+  AC_MSG_RESULT([no])
+])
+
+dnl Check for presence of builtin_prefetch
+AC_MSG_CHECKING([for working __builtin_prefetch])
+AC_LINK_IFELSE([AC_LANG_PROGRAM([],[
+  int z = 10;
+  __builtin_prefetch(&z);
+  return 0;
+])], [
+  AC_MSG_RESULT([yes])
+  AC_DEFINE([HAVE___BUILTIN_PREFETCH], 1,
+            [Defined to 1 if __builtin_prefetch works])
+], [
+  AC_MSG_RESULT([no])
+])
+
 dnl Check for working __builtin_cpu_supports (ssse3 is broken on some clangs)
 AC_MSG_CHECKING([for working __builtin_cpu_supports("ssse3")])
 AC_LINK_IFELSE([AC_LANG_PROGRAM([],[
diff --git a/htslib/hts_defs.h b/htslib/hts_defs.h
index 71eb88d46..d7942f2d9 100644
--- a/htslib/hts_defs.h
+++ b/htslib/hts_defs.h
@@ -138,4 +138,17 @@ DEALINGS IN THE SOFTWARE.  */
 #define HTSLIB_EXPORT
 #endif
 
+// Prefetch implementations.
+// We only support a basic implementation here
+#ifdef HAVE___BUILTIN_PREFETCH
+static inline void hts_prefetch(void *p) {
+    __builtin_prefetch(p);
+}
+#else
+static inline void hts_prefetch(void *p) {
+    // Fetch and discard is quite close to a genuine prefetch
+    *(volatile char *)p;
+}
+#endif
+
 #endif
diff --git a/sam.c b/sam.c
index 0a386fe1d..11e3f3e69 100644
--- a/sam.c
+++ b/sam.c
@@ -5977,6 +5977,9 @@ static void overlap_remove(bam_plp_t iter, const bam1_t *b)
 {
     if ( !iter->overlaps ) return;
 
+    if ( b->core.flag&BAM_FUNMAP || !(b->core.flag&BAM_FPROPER_PAIR) ) //no need
+        return;
+
     khiter_t kitr;
     if ( b )
     {
@@ -6007,6 +6010,8 @@ const bam_pileup1_t *bam_plp64_next(bam_plp_t iter, int *_tid, hts_pos_t *_pos,
         // write iter->plp at iter->pos
         lbnode_t **pptr = &iter->head;
         while (*pptr != iter->tail) {
+            if ((*pptr)->next)
+                hts_prefetch((*pptr)->next);
             lbnode_t *p = *pptr;
             if (p->b.core.tid < iter->tid || (p->b.core.tid == iter->tid && p->end <= iter->pos)) { // then remove
                 overlap_remove(iter, &p->b);