diff --git a/src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c b/src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c
index 8e4b7f2b..fbc66230 100644
--- a/src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c
+++ b/src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c
@@ -14,21 +14,21 @@
 #include "crypto_aead_aes256gcm.h"
 #include "export.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 #include "randombytes.h"
 #include "runtime.h"
 #include "utils.h"
 
 #if defined(HAVE_TMMINTRIN_H) && defined(HAVE_WMMINTRIN_H)
 
-# ifdef __GNUC__
-#  pragma GCC target("ssse3")
-#  pragma GCC target("aes")
-#  pragma GCC target("pclmul")
-# endif
+#ifdef __GNUC__
+# pragma GCC target("ssse3")
+# pragma GCC target("aes")
+# pragma GCC target("pclmul")
+#endif
 
 #include <tmmintrin.h>
 #include <wmmintrin.h>
+#include "private/sse2_64_32.h"
 
 #ifndef ENOSYS
 # define ENOSYS ENXIO
diff --git a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-avx2.c b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-avx2.c
index 7cb41fb6..4945462d 100644
--- a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-avx2.c
+++ b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-avx2.c
@@ -8,7 +8,6 @@
 
 #include "blake2.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 
 #if defined(HAVE_AVX2INTRIN_H) && defined(HAVE_EMMINTRIN_H) && \
     defined(HAVE_TMMINTRIN_H) && defined(HAVE_SMMINTRIN_H)
@@ -24,6 +23,7 @@
 # include <immintrin.h>
 # include <smmintrin.h>
 # include <tmmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "blake2b-compress-avx2.h"
 
diff --git a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-sse41.c b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-sse41.c
index 9e5c0c50..f085c615 100644
--- a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-sse41.c
+++ b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-sse41.c
@@ -7,7 +7,6 @@
 
 #include "blake2.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 
 #if defined(HAVE_EMMINTRIN_H) && defined(HAVE_TMMINTRIN_H) && \
     defined(HAVE_SMMINTRIN_H)
@@ -21,6 +20,7 @@
 # include <emmintrin.h>
 # include <smmintrin.h>
 # include <tmmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "blake2b-compress-sse41.h"
 
diff --git a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-ssse3.c b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-ssse3.c
index 5684bc88..6372da03 100644
--- a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-ssse3.c
+++ b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-ssse3.c
@@ -4,7 +4,6 @@
 
 #include "blake2.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 
 #if defined(HAVE_EMMINTRIN_H) && defined(HAVE_TMMINTRIN_H)
 
@@ -15,6 +14,7 @@
 
 # include <emmintrin.h>
 # include <tmmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "blake2b-compress-ssse3.h"
 
diff --git a/src/libsodium/crypto_onetimeauth/poly1305/sse2/poly1305_sse2.c b/src/libsodium/crypto_onetimeauth/poly1305/sse2/poly1305_sse2.c
index e80c961d..aa81777e 100644
--- a/src/libsodium/crypto_onetimeauth/poly1305/sse2/poly1305_sse2.c
+++ b/src/libsodium/crypto_onetimeauth/poly1305/sse2/poly1305_sse2.c
@@ -6,7 +6,6 @@
 #include "crypto_verify_16.h"
 #include "poly1305_sse2.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 #include "utils.h"
 
 #if defined(HAVE_TI_MODE) && defined(HAVE_EMMINTRIN_H)
@@ -16,6 +15,7 @@
 # endif
 
 # include <emmintrin.h>
+# include "private/sse2_64_32.h"
 
 typedef __m128i xmmi;
 
diff --git a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx2.c b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx2.c
index ab400ded..a35e1f91 100644
--- a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx2.c
+++ b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx2.c
@@ -18,7 +18,6 @@
 #include "argon2-core.h"
 #include "argon2.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 
 #if defined(HAVE_AVX2INTRIN_H) && defined(HAVE_EMMINTRIN_H) && \
     defined(HAVE_TMMINTRIN_H) && defined(HAVE_SMMINTRIN_H)
@@ -33,10 +32,11 @@
 # ifdef _MSC_VER
 #  include <intrin.h> /* for _mm_set_epi64x */
 # endif
-#include <emmintrin.h>
-#include <immintrin.h>
-#include <smmintrin.h>
-#include <tmmintrin.h>
+# include <emmintrin.h>
+# include <immintrin.h>
+# include <smmintrin.h>
+# include <tmmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "blamka-round-avx2.h"
 
diff --git a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx512f.c b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx512f.c
index d0d7c8c8..6566804e 100644
--- a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx512f.c
+++ b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx512f.c
@@ -18,7 +18,6 @@
 #include "argon2-core.h"
 #include "argon2.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 
 #if defined(HAVE_AVX512FINTRIN_H) && defined(HAVE_AVX2INTRIN_H) && \
     defined(HAVE_EMMINTRIN_H) &&  defined(HAVE_TMMINTRIN_H) && defined(HAVE_SMMINTRIN_H)
@@ -34,10 +33,11 @@
 # ifdef _MSC_VER
 #  include <intrin.h> /* for _mm_set_epi64x */
 # endif
-#include <emmintrin.h>
-#include <immintrin.h>
-#include <smmintrin.h>
-#include <tmmintrin.h>
+# include <emmintrin.h>
+# include <immintrin.h>
+# include <smmintrin.h>
+# include <tmmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "blamka-round-avx512f.h"
 
diff --git a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-ssse3.c b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-ssse3.c
index 36eac24f..fc85a478 100644
--- a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-ssse3.c
+++ b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-ssse3.c
@@ -18,7 +18,6 @@
 #include "argon2-core.h"
 #include "argon2.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 
 #if defined(HAVE_EMMINTRIN_H) && defined(HAVE_TMMINTRIN_H)
 
@@ -32,6 +31,7 @@
 # endif
 # include <emmintrin.h>
 # include <tmmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "blamka-round-ssse3.h"
 
diff --git a/src/libsodium/crypto_pwhash/scryptsalsa208sha256/sse/pwhash_scryptsalsa208sha256_sse.c b/src/libsodium/crypto_pwhash/scryptsalsa208sha256/sse/pwhash_scryptsalsa208sha256_sse.c
index ee2aa511..73ee4025 100644
--- a/src/libsodium/crypto_pwhash/scryptsalsa208sha256/sse/pwhash_scryptsalsa208sha256_sse.c
+++ b/src/libsodium/crypto_pwhash/scryptsalsa208sha256/sse/pwhash_scryptsalsa208sha256_sse.c
@@ -35,7 +35,6 @@
 #include <string.h>
 
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 
 #ifdef HAVE_EMMINTRIN_H
 
@@ -46,6 +45,7 @@
 # if defined(__XOP__) && defined(DISABLED)
 #  include <x86intrin.h>
 # endif
+# include "private/sse2_64_32.h"
 
 # include "../crypto_scrypt.h"
 # include "../pbkdf2-sha256.h"
diff --git a/src/libsodium/crypto_stream/chacha20/dolbeau/chacha20_dolbeau-avx2.c b/src/libsodium/crypto_stream/chacha20/dolbeau/chacha20_dolbeau-avx2.c
index f63e0552..61000500 100644
--- a/src/libsodium/crypto_stream/chacha20/dolbeau/chacha20_dolbeau-avx2.c
+++ b/src/libsodium/crypto_stream/chacha20/dolbeau/chacha20_dolbeau-avx2.c
@@ -6,7 +6,6 @@
 #include "core.h"
 #include "crypto_stream_chacha20.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 #include "utils.h"
 
 #if defined(HAVE_AVX2INTRIN_H) && defined(HAVE_EMMINTRIN_H) && \
@@ -23,6 +22,7 @@
 # include <immintrin.h>
 # include <smmintrin.h>
 # include <tmmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "../stream_chacha20.h"
 # include "chacha20_dolbeau-avx2.h"
diff --git a/src/libsodium/crypto_stream/chacha20/dolbeau/chacha20_dolbeau-ssse3.c b/src/libsodium/crypto_stream/chacha20/dolbeau/chacha20_dolbeau-ssse3.c
index 6f5d3851..ad13c3af 100644
--- a/src/libsodium/crypto_stream/chacha20/dolbeau/chacha20_dolbeau-ssse3.c
+++ b/src/libsodium/crypto_stream/chacha20/dolbeau/chacha20_dolbeau-ssse3.c
@@ -6,7 +6,6 @@
 #include "core.h"
 #include "crypto_stream_chacha20.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 #include "utils.h"
 
 #if defined(HAVE_EMMINTRIN_H) && defined(HAVE_TMMINTRIN_H)
@@ -18,6 +17,7 @@
 
 # include <emmintrin.h>
 # include <tmmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "../stream_chacha20.h"
 # include "chacha20_dolbeau-ssse3.h"
diff --git a/src/libsodium/crypto_stream/salsa20/xmm6int/salsa20_xmm6int-avx2.c b/src/libsodium/crypto_stream/salsa20/xmm6int/salsa20_xmm6int-avx2.c
index 18d4773e..507d7fed 100644
--- a/src/libsodium/crypto_stream/salsa20/xmm6int/salsa20_xmm6int-avx2.c
+++ b/src/libsodium/crypto_stream/salsa20/xmm6int/salsa20_xmm6int-avx2.c
@@ -5,7 +5,6 @@
 
 #include "crypto_stream_salsa20.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 #include "utils.h"
 
 #if defined(HAVE_AVX2INTRIN_H) && defined(HAVE_EMMINTRIN_H) && \
@@ -22,6 +21,7 @@
 #include <immintrin.h>
 #include <smmintrin.h>
 #include <tmmintrin.h>
+#include "private/sse2_64_32.h"
 
 # include "../stream_salsa20.h"
 # include "salsa20_xmm6int-avx2.h"
diff --git a/src/libsodium/crypto_stream/salsa20/xmm6int/salsa20_xmm6int-sse2.c b/src/libsodium/crypto_stream/salsa20/xmm6int/salsa20_xmm6int-sse2.c
index d8e53a65..16cca0da 100644
--- a/src/libsodium/crypto_stream/salsa20/xmm6int/salsa20_xmm6int-sse2.c
+++ b/src/libsodium/crypto_stream/salsa20/xmm6int/salsa20_xmm6int-sse2.c
@@ -5,7 +5,6 @@
 
 #include "crypto_stream_salsa20.h"
 #include "private/common.h"
-#include "private/sse2_64_32.h"
 #include "utils.h"
 
 #ifdef HAVE_EMMINTRIN_H
@@ -14,6 +13,7 @@
 #  pragma GCC target("sse2")
 # endif
 # include <emmintrin.h>
+# include "private/sse2_64_32.h"
 
 # include "../stream_salsa20.h"
 # include "salsa20_xmm6int-sse2.h"