From 2fc73aca6a9e78bbb4750b8b3a985dfd628895b5 Mon Sep 17 00:00:00 2001
From: vasudeva8 <vasudeva.sarma@sanger.ac.uk>
Date: Thu, 5 Feb 2026 10:45:18 +0000
Subject: [PATCH 1/2] added prefetch and minor updates to speedup pileup

---
 sam.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/sam.c b/sam.c
index 0a386fe1d..69205ec73 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)
+                __builtin_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);

From 572ecc78e5fd498a814804653192ebae64030388 Mon Sep 17 00:00:00 2001
From: James Bonfield <jkb@sanger.ac.uk>
Date: Mon, 16 Feb 2026 14:29:03 +0000
Subject: [PATCH 2/2] Add autoconf check for __builtin_prefetch

This is to facilitate the use of prefetch on deep pileups to avoid
memory latency.

On a Covid sample (local 34005_1#18.fixmate.bam file) of 150-200k
depth, the perf report -n stats for bam_plp64_next are:

    prefetch  16523
    volatile  16839
    none      17895

These are the 5th best time of 10 runs, showing 8% gain for
prefetching and a 6% gain for the poor-man's alternative.
On shallow data it's not beneficial, but at a cost of ~2% instead.
(We may wish to consider removing the volatile case as it's a little
more costly on shallow data and a little less beneficial on deep, but
it's going to be rarely used regardless.)

NB: speed has not been checked on ARM CPUs.

Also added a check for __builtin_clz which we already use elsewhere
via compiler version checks, as we may as well add both.
---
 configure.ac      | 29 +++++++++++++++++++++++++++++
 htslib/hts_defs.h | 13 +++++++++++++
 sam.c             |  2 +-
 3 files changed, 43 insertions(+), 1 deletion(-)

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 69205ec73..11e3f3e69 100644
--- a/sam.c
+++ b/sam.c
@@ -6011,7 +6011,7 @@ const bam_pileup1_t *bam_plp64_next(bam_plp_t iter, int *_tid, hts_pos_t *_pos,
         lbnode_t **pptr = &iter->head;
         while (*pptr != iter->tail) {
             if ((*pptr)->next)
-                __builtin_prefetch((*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);