Use SIMD-specific compiler flags only for files needing them

configure.ac

@@ -271,7 +271,8 @@ AS_IF([test "x$EMSCRIPTEN" = "x"],[
 #include <mmintrin.h>
 ]], [[ __m64 x = _mm_setzero_si64(); ]])],
     [AC_MSG_RESULT(yes)
-     AC_DEFINE([HAVE_MMINTRIN_H], [1], [mmx is available])],
+     AC_DEFINE([HAVE_MMINTRIN_H], [1], [mmx is available])
+     AX_CHECK_COMPILE_FLAG([-mmmx], [CFLAGS_MMX="-mmmx"])],
     [AC_MSG_RESULT(no)])
 
   AC_MSG_CHECKING(for SSE2 instructions set)
@@ -283,7 +284,8 @@ AS_IF([test "x$EMSCRIPTEN" = "x"],[
 #include <emmintrin.h>
 ]], [[ __m128d x = _mm_setzero_pd(); ]])],
     [AC_MSG_RESULT(yes)
-     AC_DEFINE([HAVE_EMMINTRIN_H], [1], [sse2 is available])],
+     AC_DEFINE([HAVE_EMMINTRIN_H], [1], [sse2 is available])
+     AX_CHECK_COMPILE_FLAG([-msse2], [CFLAGS_SSE2="-msse2"])],
     [AC_MSG_RESULT(no)])
 
   AC_MSG_CHECKING(for SSE3 instructions set)
@@ -296,7 +298,8 @@ AS_IF([test "x$EMSCRIPTEN" = "x"],[
 ]], [[ __m128 x = _mm_addsub_ps(_mm_cvtpd_ps(_mm_setzero_pd()),
                                 _mm_cvtpd_ps(_mm_setzero_pd())); ]])],
     [AC_MSG_RESULT(yes)
-     AC_DEFINE([HAVE_PMMINTRIN_H], [1], [sse3 is available])],
+     AC_DEFINE([HAVE_PMMINTRIN_H], [1], [sse3 is available])
+     AX_CHECK_COMPILE_FLAG([-msse3], [CFLAGS_SSE3="-msse3"])],
     [AC_MSG_RESULT(no)])
 
   AC_MSG_CHECKING(for SSSE3 instructions set)
@@ -308,7 +311,21 @@ AS_IF([test "x$EMSCRIPTEN" = "x"],[
 #include <tmmintrin.h>
 ]], [[ __m64 x = _mm_abs_pi32(_m_from_int(0)); ]])],
     [AC_MSG_RESULT(yes)
-     AC_DEFINE([HAVE_TMMINTRIN_H], [1], [ssse3 is available])],
+     AC_DEFINE([HAVE_TMMINTRIN_H], [1], [ssse3 is available])
+     AX_CHECK_COMPILE_FLAG([-mssse3], [CFLAGS_SSSE3="-mssse3"])],
+    [AC_MSG_RESULT(no)])
+
+  AC_MSG_CHECKING(for SSE4.1 instructions set)
+  AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
+#pragma GCC target("sse4.1")
+#ifndef __SSE4_1__
+# define __SSE4_1__
+#endif
+#include <smmintrin.h>
+]], [[ __m128i x = _mm_minpos_epu16(_mm_setzero_si128()); ]])],
+    [AC_MSG_RESULT(yes)
+     AC_DEFINE([HAVE_SMMINTRIN_H], [1], [sse4.1 is available])
+     AX_CHECK_COMPILE_FLAG([-msse4.1], [CFLAGS_SSE4_1="-msse4.1"])],
     [AC_MSG_RESULT(no)])
 
   AC_MSG_CHECKING(for AESNI instructions set and PCLMULQDQ)
@@ -326,12 +343,20 @@ AS_IF([test "x$EMSCRIPTEN" = "x"],[
        __m128i y = _mm_clmulepi64_si128(_mm_setzero_si128(), _mm_setzero_si128(), 0);]])],
     [AC_MSG_RESULT(yes)
      AC_DEFINE([HAVE_WMMINTRIN_H], [1], [aesni is available])
-     AX_CHECK_COMPILE_FLAG([-maes], [CFLAGS="$CFLAGS -maes"])
-     AX_CHECK_COMPILE_FLAG([-mpclmul], [CFLAGS="$CFLAGS -mpclmul"])
+     AX_CHECK_COMPILE_FLAG([-maes], [CFLAGS_AESNI="-maes"])
+     AX_CHECK_COMPILE_FLAG([-mpclmul], [CFLAGS_PCLMUL="-mpclmul"])
      ],
     [AC_MSG_RESULT(no)])
 ])
 
+AC_SUBST(CFLAGS_MMX)
+AC_SUBST(CFLAGS_SSE2)
+AC_SUBST(CFLAGS_SSE3)
+AC_SUBST(CFLAGS_SSSE3)
+AC_SUBST(CFLAGS_SSE4_1)
+AC_SUBST(CFLAGS_AESNI)
+AC_SUBST(CFLAGS_PCLMUL)
+
 AC_CHECK_HEADERS([sys/mman.h])
 
 dnl Checks for typedefs, structures, and compiler characteristics.

src/libsodium/Makefile.am

@@ -2,7 +2,6 @@ lib_LTLIBRARIES = \
 	libsodium.la
 
 libsodium_la_SOURCES = \
-	crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c \
 	crypto_aead/chacha20poly1305/sodium/aead_chacha20poly1305.c \
 	crypto_auth/crypto_auth.c \
 	crypto_auth/hmacsha256/auth_hmacsha256_api.c \
@@ -63,7 +62,6 @@ libsodium_la_SOURCES = \
 	crypto_pwhash/scryptsalsa208sha256/pwhash_scryptsalsa208sha256.c \
 	crypto_pwhash/scryptsalsa208sha256/sysendian.h \
 	crypto_pwhash/scryptsalsa208sha256/nosse/pwhash_scryptsalsa208sha256_nosse.c \
-	crypto_pwhash/scryptsalsa208sha256/sse/pwhash_scryptsalsa208sha256_sse.c \
 	crypto_scalarmult/crypto_scalarmult.c \
 	crypto_scalarmult/curve25519/scalarmult_curve25519_api.c \
 	crypto_secretbox/crypto_secretbox.c \
@@ -269,3 +267,20 @@ endif
 
 SUBDIRS = \
 	include
+
+libsodium_la_LIBADD = libaesni.la libsse2.la
+noinst_LTLIBRARIES =  libaesni.la libsse2.la
+
+libaesni_la_LDFLAGS = $(libsodium_la_LDFLAGS)
+libaesni_la_CPPFLAGS = $(libsodium_la_CPPFLAGS) \
+	@CFLAGS_SSE4_1@ @CFLAGS_AESNI@ @CFLAGS_PCLMUL@
+
+libaesni_la_SOURCES = \
+	crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c
+
+libsse2_la_LDFLAGS = $(libsodium_la_LDFLAGS)
+libsse2_la_CPPFLAGS = $(libsodium_la_CPPFLAGS) \
+	@CFLAGS_SSE2@
+
+libsse2_la_SOURCES = \
+	crypto_pwhash/scryptsalsa208sha256/sse/pwhash_scryptsalsa208sha256_sse.c

src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c

@@ -13,25 +13,10 @@
 
 #ifdef HAVE_WMMINTRIN_H
 
-#pragma GCC target("sse")
-#pragma GCC target("sse2")
-#pragma GCC target("ssse3")
 #pragma GCC target("sse4.1")
 #pragma GCC target("aes")
 #pragma GCC target("pclmul")
 
-#ifndef __SSSE3__
-# define __SSSE3__
-#endif
-#ifndef __SSE4_1__
-# define __SSE4_1__
-#endif
-#ifndef __AES__
-# define __AES__
-#endif
-#ifndef __PCLMUL__
-# define __PCLMUL__
-#endif
 #include <immintrin.h>
 
 #if defined(__INTEL_COMPILER) || defined(_bswap64)