diff --git a/htslib/hts_defs.h b/htslib/hts_defs.h
index 7719215c1..357684012 100644
--- a/htslib/hts_defs.h
+++ b/htslib/hts_defs.h
@@ -58,6 +58,15 @@ DEALINGS IN THE SOFTWARE.  */
 #define HTS_NORETURN
 #endif
 
+// Enable optimisation level 3, especially for gcc.  To be used
+// where we want to force vectorisation in hot loops and the default -O2
+// just doesn't cut it.
+#if HTS_COMPILER_HAS(optimize) || HTS_GCC_AT_LEAST(4,4)
+#define HTS_OPT3 __attribute__((optimize("O3")))
+#else
+#define HTS_OPT3
+#endif
+
 // GCC introduced warn_unused_result in 3.4 but added -Wno-unused-result later
 #if HTS_COMPILER_HAS(__warn_unused_result__) || HTS_GCC_AT_LEAST(4,5)
 #define HTS_RESULT_USED __attribute__ ((__warn_unused_result__))
diff --git a/sam.c b/sam.c
index 80c391939..1f2ecb7ba 100644
--- a/sam.c
+++ b/sam.c
@@ -4217,6 +4217,15 @@ int sam_read1(htsFile *fp, sam_hdr_t *h, bam1_t *b)
     return pass_filter < 0 ? -2 : ret;
 }
 
+// With gcc, -O3 or -ftree-loop-vectorize is really key here as otherwise
+// this code isn't vectorised and runs far slower than is necessary (even
+// with the restrict keyword being used).
+static inline void HTS_OPT3
+add33(uint8_t *a, const uint8_t * b, int32_t len) {
+    uint32_t i;
+    for (i = 0; i < len; i++)
+        a[i] = b[i]+33;
+}
 
 static int sam_format1_append(const bam_hdr_t *h, const bam1_t *b, kstring_t *str)
 {
@@ -4267,10 +4276,8 @@ static int sam_format1_append(const bam_hdr_t *h, const bam1_t *b, kstring_t *st
         if (s[0] == 0xff) {
             cp[i++] = '*';
         } else {
-            // local copy of c->l_qseq to aid unrolling
-            uint32_t lqseq = c->l_qseq;
-            for (i = 0; i < lqseq; ++i)
-                cp[i]=s[i]+33;
+            add33((uint8_t *)cp, s, c->l_qseq); // cp[i] = s[i]+33;
+            i = c->l_qseq;
         }
         cp[i] = 0;
         cp += i;