Unfortunately, some assemblers still don't know about AVX opcodes

configure.ac

@@ -399,7 +399,7 @@ AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
 
 HAVE_AMD64_ASM_V=0
 AS_IF([test "$enable_asm" != "no"],[
-  AC_MSG_CHECKING(whether we should use x86_64 asm code)
+  AC_MSG_CHECKING(whether we can use x86_64 asm code)
   AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
   ]], [[
 #if defined(__amd64) || defined(__amd64__) || defined(__x86_64__)
@@ -413,13 +413,36 @@ AS_IF([test "$enable_asm" != "no"],[
 __asm__("pxor %xmm12,%xmm6");
 ]])],
   [AC_MSG_RESULT(yes)
-   AC_DEFINE([HAVE_AMD64_ASM], [1], [x86_64 asm code should be used])
+   AC_DEFINE([HAVE_AMD64_ASM], [1], [x86_64 asm code can be used])
    HAVE_AMD64_ASM_V=1],
   [AC_MSG_RESULT(no)])
 ])
 AM_CONDITIONAL([HAVE_AMD64_ASM], [test $HAVE_AMD64_ASM_V = 1])
 AC_SUBST(HAVE_AMD64_ASM_V)
 
+HAVE_AVX_ASM_V=0
+AS_IF([test "$enable_asm" != "no"],[
+  AC_MSG_CHECKING(whether we can assemble AVX opcodes)
+  AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
+  ]], [[
+#if defined(__amd64) || defined(__amd64__) || defined(__x86_64__)
+# if defined(__CYGWIN__) || defined(__MINGW32__) || defined(__MINGW64__) || defined(_WIN32) || defined(_WIN64)
+#  error Windows x86_64 calling conventions are not supported yet
+# endif
+/* neat */
+#else
+# error !x86_64
+#endif
+__asm__("vpunpcklqdq %xmm0,%xmm13,%xmm0");
+]])],
+  [AC_MSG_RESULT(yes)
+   AC_DEFINE([HAVE_AVX_ASM], [1], [AVX opcodes are supported])
+   HAVE_AVX_ASM_V=1],
+  [AC_MSG_RESULT(no)])
+])
+AM_CONDITIONAL([HAVE_AVX_ASM], [test $HAVE_AVX_ASM_V = 1])
+AC_SUBST(HAVE_AVX_ASM_V)
+
 AC_MSG_CHECKING(for 128-bit arithmetic)
 HAVE_TI_MODE_V=0
 AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[

src/libsodium/Makefile.am

@@ -204,10 +204,8 @@ EXTRA_DIST = \
 	crypto_scalarmult/curve25519/sandy2x/ladder.S \
 	crypto_scalarmult/curve25519/sandy2x/ladder_base.S
 
-if HAVE_AMD64_ASM
+if HAVE_AVX_ASM
 libsodium_la_SOURCES += \
-	crypto_stream/salsa20/amd64_xmm6/api.h \
-	crypto_stream/salsa20/amd64_xmm6/stream_salsa20_amd64_xmm6.S \
 	crypto_scalarmult/curve25519/sandy2x/consts_namespace.h \
 	crypto_scalarmult/curve25519/sandy2x/curve25519_sandy2x.c \
 	crypto_scalarmult/curve25519/sandy2x/curve25519_sandy2x.h \
@@ -221,6 +219,12 @@ libsodium_la_SOURCES += \
 	crypto_scalarmult/curve25519/sandy2x/ladder_base_namespace.h \
 	crypto_scalarmult/curve25519/sandy2x/ladder_namespace.h \
 	crypto_scalarmult/curve25519/sandy2x/sandy2x.S
+endif
+
+if HAVE_AMD64_ASM
+libsodium_la_SOURCES += \
+	crypto_stream/salsa20/amd64_xmm6/api.h \
+	crypto_stream/salsa20/amd64_xmm6/stream_salsa20_amd64_xmm6.S
 else
 libsodium_la_SOURCES += \
 	crypto_stream/salsa20/ref/api.h \

src/libsodium/crypto_scalarmult/curve25519/sandy2x/curve25519_sandy2x.c

@@ -7,7 +7,7 @@
 
 #include <stddef.h>
 
-#ifdef HAVE_AMD64_ASM
+#ifdef HAVE_AVX_ASM
 
 #include "utils.h"
 #include "curve25519_sandy2x.h"

src/libsodium/crypto_scalarmult/curve25519/sandy2x/fe51_invert.c

@@ -5,7 +5,7 @@
 
 #include "fe51.h"
 
-#ifdef HAVE_AMD64_ASM
+#ifdef HAVE_AVX_ASM
 
 #define fe51_square(x, y) fe51_nsquare(x, y, 1)
 

src/libsodium/crypto_scalarmult/curve25519/sandy2x/fe_frombytes_sandy2x.c

@@ -5,7 +5,7 @@
 #include "fe.h"
 #include "crypto_uint64.h"
 
-#ifdef HAVE_AMD64_ASM
+#ifdef HAVE_AVX_ASM
 
 static crypto_uint64 load_3(const unsigned char *in)
 {

src/libsodium/crypto_scalarmult/curve25519/sandy2x/sandy2x.S

@@ -1,5 +1,5 @@
 
-#ifdef HAVE_AMD64_ASM
+#ifdef HAVE_AVX_ASM
 
 #define IN_SANDY2X
 

src/libsodium/crypto_scalarmult/curve25519/scalarmult_curve25519.c

@@ -3,7 +3,7 @@
 #include "scalarmult_curve25519.h"
 #include "runtime.h"
 
-#ifdef HAVE_AMD64_ASM
+#ifdef HAVE_AVX_ASM
 # include "sandy2x/curve25519_sandy2x.h"
 #endif
 #ifdef HAVE_TI_MODE
@@ -49,7 +49,7 @@ _crypto_scalarmult_curve25519_pick_best_implementation(void)
 #else
     implementation = &crypto_scalarmult_curve25519_ref10_implementation;
 #endif
-#ifdef HAVE_AMD64_ASM
+#ifdef HAVE_AVX_ASM
     if (sodium_runtime_has_avx()) {
         implementation = &crypto_scalarmult_curve25519_sandy2x_implementation;
     }