Improved architecture detection

3 files changed, 9 insertions, 11 deletions

diff --git a/configure.sh b/configure.sh
index 9a790ca..b9779c2 100755
--- a/configure.sh
+++ b/configure.sh
@@ -1,8 +1,13 @@
 #!/bin/sh
 
-cc -mavx2 utils/detect-immintrin.c && detected="AVX2" && rm a.out
+detect() { cc -march=native -dM -E - < /dev/null | grep "$1"; }
+
+[ -n "$(detect __AVX2__)" ] && detected="AVX2"
+
 TYPE=${TYPE-"$detected"}
 
+echo "Using CUBE_$TYPE"
+
 STD="-std=c99"
 WFLAGS="-pedantic -Wall -Wextra -Wno-unused-parameter -Wno-unused-function"
 [ "$TYPE" = "AVX2" ] && AVX="-mavx2"
diff --git a/cube.c b/cube.c
index b5a30a6..895c088 100644
--- a/cube.c
+++ b/cube.c
@@ -2,9 +2,7 @@
 #include <stdbool.h>
 #include <string.h>
 
-#ifdef CUBE_AVX2
-#include <immintrin.h>
-#endif
+#include "cube.h"
 
 #ifdef DEBUG
 #include <stdio.h>
@@ -25,8 +23,6 @@
 #define DBG_ASSERT(condition, retval, ...)
 #endif
 
-#include "cube.h"
-
 /******************************************************************************
 Section: constants, strings and other stuff
 ******************************************************************************/
@@ -495,6 +491,8 @@ Note: the #ifdef below is closed in the next section.
 
 #ifdef CUBE_AVX2
 
+#include <immintrin.h>
+
 typedef __m256i cube_fast_t;
 
 #define _co2_avx2 _mm256_set_epi64x(0, 0, 0, 0x6060606060606060)
diff --git a/utils/detect-immintrin.c b/utils/detect-immintrin.c
deleted file mode 100644
index 20da678..0000000
--- a/utils/detect-immintrin.c
+++ /dev/null
@@ -1,5 +0,0 @@
-#include <immintrin.h>
-
-int main() {
-        return 0;
-}