diff --git a/configure.ac b/configure.ac
index 095dce23..d25f2d22 100644
--- a/configure.ac
+++ b/configure.ac
@@ -244,6 +244,7 @@ dnl Checks for headers
 
 AS_IF([test "x$EMSCRIPTEN" = "x"],[
   AC_CHECK_HEADERS([emmintrin.h], [], [], [#pragma GCC target("sse2")])
+  AC_CHECK_HEADERS([pmmintrin.h], [], [], [#pragma GCC target("sse3")])
   AC_CHECK_HEADERS([tmmintrin.h], [], [], [#pragma GCC target("ssse3")])
   AC_CHECK_HEADERS([smmintrin.h], [], [], [#pragma GCC target("sse4.1")])
   AC_CHECK_HEADERS([immintrin.h], [], [], [#pragma GCC target("avx")])
diff --git a/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/crypto_scrypt-common.c b/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/crypto_scrypt-common.c
index 9791863d..9d5de17a 100644
--- a/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/crypto_scrypt-common.c
+++ b/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/crypto_scrypt-common.c
@@ -22,86 +22,85 @@
 #include <string.h>
 
 #include "crypto_scrypt.h"
-
-#define escrypt_kdf escrypt_kdf_nosse
+#include "runtime.h"
 
 #define BYTES2CHARS(bytes) \
-	((((bytes) * 8) + 5) / 6)
+        ((((bytes) * 8) + 5) / 6)
 
 #define HASH_SIZE 32 /* bytes */
 #define HASH_LEN BYTES2CHARS(HASH_SIZE) /* base-64 chars */
 
 static const char * const itoa64 =
-	"./0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+        "./0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
 
 static uint8_t * encode64_uint32(uint8_t * dst, size_t dstlen,
     uint32_t src, uint32_t srcbits)
 {
-	uint32_t bit;
+        uint32_t bit;
 
-	for (bit = 0; bit < srcbits; bit += 6) {
-		if (dstlen < 1)
-			return NULL;
-		*dst++ = itoa64[src & 0x3f];
-		dstlen--;
-		src >>= 6;
-	}
+        for (bit = 0; bit < srcbits; bit += 6) {
+                if (dstlen < 1)
+                        return NULL;
+                *dst++ = itoa64[src & 0x3f];
+                dstlen--;
+                src >>= 6;
+        }
 
-	return dst;
+        return dst;
 }
 
 static uint8_t * encode64(uint8_t * dst, size_t dstlen,
     const uint8_t * src, size_t srclen)
 {
-	size_t i;
+        size_t i;
 
-	for (i = 0; i < srclen; ) {
-		uint8_t * dnext;
-		uint32_t value = 0, bits = 0;
-		do {
-			value |= (uint32_t)src[i++] << bits;
-			bits += 8;
-		} while (bits < 24 && i < srclen);
-		dnext = encode64_uint32(dst, dstlen, value, bits);
-		if (!dnext)
-			return NULL;
-		dstlen -= dnext - dst;
-		dst = dnext;
-	}
+        for (i = 0; i < srclen; ) {
+                uint8_t * dnext;
+                uint32_t value = 0, bits = 0;
+                do {
+                        value |= (uint32_t)src[i++] << bits;
+                        bits += 8;
+                } while (bits < 24 && i < srclen);
+                dnext = encode64_uint32(dst, dstlen, value, bits);
+                if (!dnext)
+                        return NULL;
+                dstlen -= dnext - dst;
+                dst = dnext;
+        }
 
-	return dst;
+        return dst;
 }
 
 static int decode64_one(uint32_t * dst, uint8_t src)
 {
-	const char * ptr = strchr(itoa64, src);
-	if (ptr) {
-		*dst = ptr - itoa64;
-		return 0;
-	}
-	*dst = 0;
-	return -1;
+        const char * ptr = strchr(itoa64, src);
+        if (ptr) {
+                *dst = ptr - itoa64;
+                return 0;
+        }
+        *dst = 0;
+        return -1;
 }
 
 static const uint8_t * decode64_uint32(uint32_t * dst, uint32_t dstbits,
     const uint8_t * src)
 {
-	uint32_t bit;
-	uint32_t value;
+        uint32_t bit;
+        uint32_t value;
 
-	value = 0;
-	for (bit = 0; bit < dstbits; bit += 6) {
-		uint32_t one;
-		if (decode64_one(&one, *src)) {
-			*dst = 0;
-			return NULL;
-		}
-		src++;
-		value |= one << bit;
-	}
+        value = 0;
+        for (bit = 0; bit < dstbits; bit += 6) {
+                uint32_t one;
+                if (decode64_one(&one, *src)) {
+                        *dst = 0;
+                        return NULL;
+                }
+                src++;
+                value |= one << bit;
+        }
 
-	*dst = value;
-	return src;
+        *dst = value;
+        return src;
 }
 
 uint8_t *
@@ -110,64 +109,68 @@ escrypt_r(escrypt_local_t * local,
     const uint8_t * setting,
     uint8_t * buf, size_t buflen)
 {
-	uint8_t hash[HASH_SIZE];
-	const uint8_t * src, * salt;
-	uint8_t * dst;
-	size_t prefixlen, saltlen, need;
-	uint64_t N;
-	uint32_t r, p;
+        uint8_t hash[HASH_SIZE];
+    escrypt_kdf_t escrypt_kdf;
+        const uint8_t * src, * salt;
+        uint8_t * dst;
+        size_t prefixlen, saltlen, need;
+        uint64_t N;
+        uint32_t r, p;
 
-	if (setting[0] != '$' || setting[1] != '7' || setting[2] != '$')
-		return NULL;
-	src = setting + 3;
+        if (setting[0] != '$' || setting[1] != '7' || setting[2] != '$')
+                return NULL;
+        src = setting + 3;
 
-	{
-		uint32_t N_log2;
-		if (decode64_one(&N_log2, *src))
-			return NULL;
-		src++;
-		N = (uint64_t)1 << N_log2;
-	}
+        {
+                uint32_t N_log2;
+                if (decode64_one(&N_log2, *src))
+                        return NULL;
+                src++;
+                N = (uint64_t)1 << N_log2;
+        }
 
-	src = decode64_uint32(&r, 30, src);
-	if (!src)
-		return NULL;
+        src = decode64_uint32(&r, 30, src);
+        if (!src)
+                return NULL;
 
-	src = decode64_uint32(&p, 30, src);
-	if (!src)
-		return NULL;
+        src = decode64_uint32(&p, 30, src);
+        if (!src)
+                return NULL;
 
-	prefixlen = src - setting;
+        prefixlen = src - setting;
 
-	salt = src;
-	src = (uint8_t *)strrchr((char *)salt, '$');
-	if (src)
-		saltlen = src - salt;
-	else
-		saltlen = strlen((char *)salt);
+        salt = src;
+        src = (uint8_t *)strrchr((char *)salt, '$');
+        if (src)
+                saltlen = src - salt;
+        else
+                saltlen = strlen((char *)salt);
 
-	need = prefixlen + saltlen + 1 + HASH_LEN + 1;
-	if (need > buflen || need < saltlen)
-		return NULL;
+        need = prefixlen + saltlen + 1 + HASH_LEN + 1;
+        if (need > buflen || need < saltlen)
+                return NULL;
 
-	if (escrypt_kdf(local, passwd, passwdlen, salt, saltlen,
-	    N, r, p, hash, sizeof(hash)))
-		return NULL;
+    escrypt_kdf =
+        sodium_runtime_has_sse2() ? escrypt_kdf_sse : escrypt_kdf_nosse;
+        if (escrypt_kdf(local, passwd, passwdlen, salt, saltlen,
+            N, r, p, hash, sizeof(hash))) {
+                return NULL;
+    }
 
-	dst = buf;
-	memcpy(dst, setting, prefixlen + saltlen);
-	dst += prefixlen + saltlen;
-	*dst++ = '$';
+        dst = buf;
+        memcpy(dst, setting, prefixlen + saltlen);
+        dst += prefixlen + saltlen;
+        *dst++ = '$';
 
-	dst = encode64(dst, buflen - (dst - buf), hash, sizeof(hash));
-	/* Could zeroize hash[] here, but escrypt_kdf() doesn't zeroize its
-	 * memory allocations yet anyway. */
-	if (!dst || dst >= buf + buflen) /* Can't happen */
-		return NULL;
+        dst = encode64(dst, buflen - (dst - buf), hash, sizeof(hash));
+        /* Could zeroize hash[] here, but escrypt_kdf() doesn't zeroize its
+         * memory allocations yet anyway. */
+        if (!dst || dst >= buf + buflen) /* Can't happen */
+                return NULL;
 
-	*dst = 0; /* NUL termination */
+        *dst = 0; /* NUL termination */
 
-	return buf;
+        return buf;
 }
 
 uint8_t *
@@ -175,40 +178,40 @@ escrypt_gensalt_r(uint32_t N_log2, uint32_t r, uint32_t p,
     const uint8_t * src, size_t srclen,
     uint8_t * buf, size_t buflen)
 {
-	uint8_t * dst;
-	size_t prefixlen = 3 + 1 + 5 + 5;
-	size_t saltlen = BYTES2CHARS(srclen);
-	size_t need;
+        uint8_t * dst;
+        size_t prefixlen = 3 + 1 + 5 + 5;
+        size_t saltlen = BYTES2CHARS(srclen);
+        size_t need;
 
-	need = prefixlen + saltlen + 1;
-	if (need > buflen || need < saltlen || saltlen < srclen)
-		return NULL;
+        need = prefixlen + saltlen + 1;
+        if (need > buflen || need < saltlen || saltlen < srclen)
+                return NULL;
 
-	if (N_log2 > 63 || ((uint64_t)r * (uint64_t)p >= (1U << 30)))
-		return NULL;
+        if (N_log2 > 63 || ((uint64_t)r * (uint64_t)p >= (1U << 30)))
+                return NULL;
 
-	dst = buf;
-	*dst++ = '$';
-	*dst++ = '7';
-	*dst++ = '$';
+        dst = buf;
+        *dst++ = '$';
+        *dst++ = '7';
+        *dst++ = '$';
 
-	*dst++ = itoa64[N_log2];
+        *dst++ = itoa64[N_log2];
 
-	dst = encode64_uint32(dst, buflen - (dst - buf), r, 30);
-	if (!dst) /* Can't happen */
-		return NULL;
+        dst = encode64_uint32(dst, buflen - (dst - buf), r, 30);
+        if (!dst) /* Can't happen */
+                return NULL;
 
-	dst = encode64_uint32(dst, buflen - (dst - buf), p, 30);
-	if (!dst) /* Can't happen */
-		return NULL;
+        dst = encode64_uint32(dst, buflen - (dst - buf), p, 30);
+        if (!dst) /* Can't happen */
+                return NULL;
 
-	dst = encode64(dst, buflen - (dst - buf), src, srclen);
-	if (!dst || dst >= buf + buflen) /* Can't happen */
-		return NULL;
+        dst = encode64(dst, buflen - (dst - buf), src, srclen);
+        if (!dst || dst >= buf + buflen) /* Can't happen */
+                return NULL;
 
-	*dst = 0; /* NUL termination */
+        *dst = 0; /* NUL termination */
 
-	return buf;
+        return buf;
 }
 
 int
@@ -216,14 +219,17 @@ crypto_scrypt(const uint8_t * passwd, size_t passwdlen,
     const uint8_t * salt, size_t saltlen, uint64_t N, uint32_t r, uint32_t p,
     uint8_t * buf, size_t buflen)
 {
-	escrypt_local_t local;
-	int retval;
+    escrypt_kdf_t escrypt_kdf;
+        escrypt_local_t local;
+        int retval;
 
-	if (escrypt_init_local(&local))
-		return -1;
-	retval = escrypt_kdf(&local,
-	    passwd, passwdlen, salt, saltlen, N, r, p, buf, buflen);
-	if (escrypt_free_local(&local))
-		return -1;
-	return retval;
+        if (escrypt_init_local(&local))
+                return -1;
+    escrypt_kdf =
+        sodium_runtime_has_sse2() ? escrypt_kdf_sse : escrypt_kdf_nosse;
+        retval = escrypt_kdf(&local,
+            passwd, passwdlen, salt, saltlen, N, r, p, buf, buflen);
+        if (escrypt_free_local(&local))
+                return -1;
+        return retval;
 }
diff --git a/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/crypto_scrypt.h b/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/crypto_scrypt.h
index 2a1d8d18..1ffc7830 100644
--- a/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/crypto_scrypt.h
+++ b/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/crypto_scrypt.h
@@ -60,6 +60,12 @@ extern int escrypt_free_local(escrypt_local_t * __local);
 extern void *alloc_region(escrypt_region_t * region, size_t size);
 extern int free_region(escrypt_region_t * region);
 
+typedef int (*escrypt_kdf_t)(escrypt_local_t * __local,
+                             const uint8_t * __passwd, size_t __passwdlen,
+                             const uint8_t * __salt, size_t __saltlen,
+                             uint64_t __N, uint32_t __r, uint32_t __p,
+                             uint8_t * __buf, size_t __buflen);
+
 extern int escrypt_kdf_nosse(escrypt_local_t * __local,
     const uint8_t * __passwd, size_t __passwdlen,
     const uint8_t * __salt, size_t __saltlen,
diff --git a/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/scrypt_platform.c b/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/scrypt_platform.c
index 6dc26c57..f8d8e8bf 100644
--- a/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/scrypt_platform.c
+++ b/src/libsodium/crypto_pwhash/scryptxsalsa208sha256/scrypt_platform.c
@@ -25,6 +25,7 @@
 #include <stdlib.h>
 
 #include "crypto_scrypt.h"
+#include "runtime.h"
 
 void *
 alloc_region(escrypt_region_t * region, size_t size)
diff --git a/src/libsodium/sodium/runtime.c b/src/libsodium/sodium/runtime.c
index 2e4ae191..4bc200bb 100644
--- a/src/libsodium/sodium/runtime.c
+++ b/src/libsodium/sodium/runtime.c
@@ -66,8 +66,18 @@ _sodium_runtime_intel_cpu_features(CPUFeatures * const cpu_features)
         return -1;
     }
     _cpuid(cpu_info, 0x00000001);
+
+#ifndef HAVE_EMMINTRIN_H
     cpu_features->has_sse2 = ((cpu_info[3] & CPUID_SSE2) != 0x0);
+#else
+    cpu_features->has_sse2 = 0;
+#endif
+
+#ifndef HAVE_PMMINTRIN_H
     cpu_features->has_sse3 = ((cpu_info[2] & CPUIDECX_SSE3) != 0x0);
+#else
+    cpu_features->has_sse3 = 0;
+#endif
 
     return 0;
 }