Backport htslib fix: configure checks for broken SIMD intrinsics

Apply PR samtools/htslib#1886. GCC 4.8.5 and earlier accept SSSE3 intrinsics with -mssse3; these compiler versions also accept __attribute__((target("ssse3"))) but the attribute fails to enable compilation of the intrinsics.
pysam-developers · Feb 10, 2025 · 3e3c8b0 · 3e3c8b0
1 parent c52203a
commit 3e3c8b0
Show file tree

Hide file tree

Showing 5 changed files with 33 additions and 16 deletions.
diff --git a/htslib/Makefile b/htslib/Makefile
@@ -302,7 +302,7 @@ config.h:
 	echo '#define HAVE_ATTRIBUTE_CONSTRUCTOR 1' >> $@
 	echo '#endif' >> $@
 	echo '#if (defined(__x86_64__) || defined(_M_X64))' >> $@
-	echo '#define HAVE_ATTRIBUTE_TARGET 1' >> $@
+	echo '#define HAVE_ATTRIBUTE_TARGET_SSSE3 1' >> $@
 	echo '#define HAVE_BUILTIN_CPU_SUPPORT_SSSE3 1' >> $@
 	echo '#endif' >> $@
 

diff --git a/htslib/config.h.in b/htslib/config.h.in
@@ -22,8 +22,8 @@
 /* Define if __attribute__((constructor)) is available. */
 #undef HAVE_ATTRIBUTE_CONSTRUCTOR
 
-/* Define if __attribute__((target(...))) is available. */
-#undef HAVE_ATTRIBUTE_TARGET
+/* Define if __attribute__((target("ssse3"))) works. */
+#undef HAVE_ATTRIBUTE_TARGET_SSSE3
 
 /* Defined to 1 if rANS source using AVX2 can be compiled. */
 #undef HAVE_AVX2

diff --git a/htslib/configure b/htslib/configure
@@ -4686,20 +4686,28 @@ fi
 rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 
-{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for __attribute__((target))" >&5
-printf %s "checking for __attribute__((target))... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for working __attribute__((target(\"ssse3\")))" >&5
+printf %s "checking for working __attribute__((target(\"ssse3\")))... " >&6; }
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
+  #ifdef __x86_64__
+  #include "x86intrin.h"
+
   __attribute__((target("ssse3")))
-  int zero(void) {
-    return 0;
+  void shuffle(char *aptr, char *bptr) {
+    __m128i a = _mm_lddqu_si128((__m128i *)aptr);
+    __m128i b = _mm_shuffle_epi8(a, a);
+    _mm_storeu_si128((__m128i *)bptr, b);
   }
+  #else
+  void shuffle(char *aptr, char *bptr) { }
+  #endif
 
 int
 main (void)
 {
-zero();
+shuffle(0, 0);
   ;
   return 0;
 }
@@ -4710,7 +4718,7 @@ then :
   { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: yes" >&5
 printf "%s\n" "yes" >&6; }
 
-printf "%s\n" "#define HAVE_ATTRIBUTE_TARGET 1" >>confdefs.h
+printf "%s\n" "#define HAVE_ATTRIBUTE_TARGET_SSSE3 1" >>confdefs.h
 
 
 else case e in #(

diff --git a/htslib/configure.ac b/htslib/configure.ac
@@ -174,16 +174,25 @@ AC_LINK_IFELSE([AC_LANG_PROGRAM([],[
 ])
 
 dnl Check for function attribute used in conjunction with __builtin_cpu_supports
-AC_MSG_CHECKING([for __attribute__((target))])
+dnl and that it does enable the corresponding intrinsics (which is broken on ancient GCCs)
+AC_MSG_CHECKING([for working __attribute__((target("ssse3")))])
 AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
+  #ifdef __x86_64__
+  #include "x86intrin.h"
+
   __attribute__((target("ssse3")))
-  int zero(void) {
-    return 0;
+  void shuffle(char *aptr, char *bptr) {
+    __m128i a = _mm_lddqu_si128((__m128i *)aptr);
+    __m128i b = _mm_shuffle_epi8(a, a);
+    _mm_storeu_si128((__m128i *)bptr, b);
   }
-]], [[zero();]])], [
+  #else
+  void shuffle(char *aptr, char *bptr) { }
+  #endif
+]], [[shuffle(0, 0);]])], [
   AC_MSG_RESULT([yes])
-  AC_DEFINE([HAVE_ATTRIBUTE_TARGET], 1,
-            [Define if __attribute__((target(...))) is available.])
+  AC_DEFINE([HAVE_ATTRIBUTE_TARGET_SSSE3], 1,
+            [Define if __attribute__((target("ssse3"))) works.])
 ], [
   AC_MSG_RESULT([no])
 ])

diff --git a/htslib/sam_internal.h b/htslib/sam_internal.h
@@ -100,7 +100,7 @@ static inline void nibble2base_default(uint8_t *nib, char *seq, int len) {
 }
 
 #if defined HAVE_ATTRIBUTE_CONSTRUCTOR && \
-    ((defined __x86_64__ && defined HAVE_ATTRIBUTE_TARGET && defined HAVE_BUILTIN_CPU_SUPPORT_SSSE3) || \
+    ((defined __x86_64__ && defined HAVE_ATTRIBUTE_TARGET_SSSE3 && defined HAVE_BUILTIN_CPU_SUPPORT_SSSE3) || \
      (defined __ARM_NEON))
 #define BUILDING_SIMD_NIBBLE2BASE
 #endif