Improvements for intel -m32 builds.

On this platform _mm256_extract_epi64 isn't defined, but the rest of
AVX2 is.  It needs to fail auto-detection.

Also we get unaligned accesses in the SSE4 code with tbuf due to
differing data alignment caused by 32-bit pointers instead of 64-bit.
This exposes an underlying problem of using aligned SIMD writes on
tbuf without explicitly asking for alignment. (The new code is also
sometimes a little faster.)

See also samtools/htslib#1500

configure.ac

@@ -165,7 +165,8 @@ AX_CHECK_COMPILE_FLAG([-mavx2], [
 	]],[[
 	  __m256i a = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8);
 	  __m256i b = _mm256_add_epi32(a, a);
-	  return *((char *) &b);
+	  long long c = _mm256_extract_epi64(b, 0);
+	  return (int) c;
 	]])
 ])
 AM_CONDITIONAL([RANS_32x16_AVX2],[test "x$MAVX2" != "x"])

htscodecs/rANS_static32x16pr_avx2.c

@@ -1123,7 +1123,7 @@ unsigned char *rans_uncompress_O1_32x16_avx2(unsigned char *in,
     union {
         unsigned char tbuf[32][32];
         uint64_t tbuf64[32][4];
-    } u;
+    } u  __attribute__((aligned(32)));
     unsigned int tidx = 0;
 
     if (0) {

htscodecs/rANS_static32x16pr_avx512.c

@@ -735,7 +735,7 @@ unsigned char *rans_uncompress_O1_32x16_avx512(unsigned char *in,
     union {
         unsigned char tbuf[32][32];
         uint64_t tbuf64[32][4];
-    } u;
+    } u __attribute__((aligned(32)));
 #else
     uint32_t tbuf[32][32];
 #endif

htscodecs/rANS_static32x16pr_sse4.c

@@ -1423,7 +1423,7 @@ unsigned char *rans_uncompress_O1_32x16_sse4(unsigned char *in,
         uint16_t *sp = (uint16_t *)ptr;
         const uint32_t mask = ((1u << TF_SHIFT_O1_FAST)-1);
         __m128i maskv  = _mm_set1_epi32(mask); // set mask in all lanes
-        uint8_t tbuf[32][32];
+        uint8_t tbuf[32][32] __attribute__((aligned(32)));
         int tidx = 0;
         LOAD128(Rv, R);
         LOAD128(Lv, l);

m4/ax_check_compile_flag.m4

@@ -36,13 +36,17 @@
 
 #serial 6
 
+# LOCAL modification; change AC_COMPILE_IFELSE to AC_LINK_IFELSE so
+# _mm256_extract_epi64 tests on a x86_64 running under -m32 still
+# fails. (Otherwise it compiles, but fails to link.)
+
 AC_DEFUN([AX_CHECK_COMPILE_FLAG],
 [AC_PREREQ(2.64)dnl for _AC_LANG_PREFIX and AS_VAR_IF
 AS_VAR_PUSHDEF([CACHEVAR],[ax_cv_check_[]_AC_LANG_ABBREV[]flags_$4_$1])dnl
 AC_CACHE_CHECK([whether _AC_LANG compiler accepts $1], CACHEVAR, [
   ax_check_save_flags=$[]_AC_LANG_PREFIX[]FLAGS
   _AC_LANG_PREFIX[]FLAGS="$[]_AC_LANG_PREFIX[]FLAGS $4 $1"
-  AC_COMPILE_IFELSE([m4_default([$5],[AC_LANG_PROGRAM()])],
+  AC_LINK_IFELSE([m4_default([$5],[AC_LANG_PROGRAM()])],
     [AS_VAR_SET(CACHEVAR,[yes])],
     [AS_VAR_SET(CACHEVAR,[no])])
   _AC_LANG_PREFIX[]FLAGS=$ax_check_save_flags])