Check for SSE2/SSE3/NEON support at runtime.

configure.ac

@@ -323,18 +323,19 @@ void fcontract(uint128_t *t) {
 AM_CONDITIONAL([HAVE_TI_MODE], [test $HAVE_TI_MODE_V = 1])
 AC_SUBST(HAVE_TI_MODE_V)
 
-AC_MSG_CHECKING(for cpuid instruction)
 HAVE_CPUID_V=0
+AS_IF([test "$enable_asm" != "no"],[
+  AC_MSG_CHECKING(for cpuid instruction)
   AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[ ]], [[
-__asm__("movl %ebx,%esi\n"
+unsigned int a;
-        "cpuid\n"
+__asm__ __volatile__("cpuid" : "=a" (a) : "a" (0U), "c" (0U));
-        "movl %esi,%ebx");
   ]])],
   [AC_MSG_RESULT(yes)
    AC_DEFINE([HAVE_CPUID], [1], [cpuid instruction is available])
    HAVE_CPUID_V=1],
   [AC_MSG_RESULT(no)])
   AM_CONDITIONAL([HAVE_CPUID], [test $HAVE_CPUID_V = 1])
+  ])
 AC_SUBST(HAVE_CPUID_V)
 
 AS_CASE([$host_cpu],

src/libsodium/Makefile.am

@@ -193,6 +193,7 @@ libsodium_la_SOURCES = \
 	randombytes/sysrandom/randombytes_sysrandom.c \
 	sodium/compat.c \
 	sodium/core.c \
+	sodium/runtime.c \
 	sodium/utils.c \
 	sodium/version.c
 

src/libsodium/include/Makefile.am

@@ -49,6 +49,7 @@ SODIUM_EXPORT = \
 	sodium/randombytes.h \
 	sodium/randombytes_salsa20_random.h \
 	sodium/randombytes_sysrandom.h \
+	sodium/runtime.h
 	sodium/utils.h
 
 EXTRA_SRC = $(SODIUM_EXPORT) \

src/libsodium/include/sodium.h

@@ -41,6 +41,7 @@
 #include <sodium/randombytes.h>
 #include <sodium/randombytes_salsa20_random.h>
 #include <sodium/randombytes_sysrandom.h>
+#include <sodium/runtime.h>
 #include <sodium/utils.h>
 #include <sodium/version.h>
 

src/libsodium/include/sodium/runtime.h

@@ -0,0 +1,19 @@
+
+#ifndef __SODIUM_RUNTIME_H__
+#define __SODIUM_RUNTIME_H__ 1
+
+#include "export.h"
+
+SODIUM_EXPORT
+int sodium_runtime_get_cpu_features(void);
+
+SODIUM_EXPORT
+int sodium_runtime_have_neon(void);
+
+SODIUM_EXPORT
+int sodium_runtime_have_sse2(void);
+
+SODIUM_EXPORT
+int sodium_runtime_have_sse3(void);
+
+#endif

src/libsodium/sodium/core.c

@@ -2,6 +2,7 @@
 #include "core.h"
 #include "crypto_onetimeauth.h"
 #include "randombytes.h"
+#include "runtime.h"
 
 static int initialized;
 
@@ -11,6 +12,7 @@ sodium_init(void)
     if (initialized != 0) {
         return 1;
     }
+    sodium_runtime_get_cpu_features();
     if (crypto_onetimeauth_pick_best_implementation() == NULL) {
         return -1;
     }

src/libsodium/sodium/runtime.c

@@ -0,0 +1,100 @@
+
+#ifdef HAVE_ANDROID_GETCPUFEATURES
+# include <cpu-features.h>
+#endif
+
+#include "runtime.h"
+
+typedef struct CPUFeatures_ {
+    int initialized;
+    int have_neon;
+    int have_sse2;
+    int have_sse3;
+} CPUFeatures;
+
+static CPUFeatures cpu_features;
+
+#define CPUID_SSE2     0x04000000
+#define CPUIDECX_SSE3  0x00000001
+
+static int
+_sodium_runtime_arm_cpu_features(CPUFeatures * const cpu_features)
+{
+#ifndef __arm__
+    cpu_features->have_neon = 0;
+    return -1;
+#else
+# ifdef __APPLE__
+#  ifdef __ARM_NEON__
+    cpu_features->have_neon = 1;
+#  else
+    cpu_features->have_neon = 0;
+#  endif
+# elif defined(HAVE_ANDROID_GETCPUFEATURES) && defined(ANDROID_CPU_ARM_FEATURE_NEON)
+    cpu_features->have_neon =
+        (android_getCpuFeatures() & ANDROID_CPU_ARM_FEATURE_NEON) != 0x0;
+# else
+    cpu_features->have_neon = 0;
+# endif
+    return 0;
+#endif
+}
+
+static void
+_cpuid(unsigned int cpu_info[4U], const unsigned int cpu_info_type)
+{
+#ifdef _MSC_VER
+    __cpuidex((int *) cpu_info, cpu_info_type, 0);
+#elif defined(HAVE_CPUID)
+    __asm__ __volatile__ ("cpuid" :
+                          "=a" (cpu_info[0]), "=b" (cpu_info[1]),
+                          "=c" (cpu_info[2]), "=d" (cpu_info[3]) :
+                          "a" (cpu_info_type), "c" (0U));
+#else
+    cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
+#endif
+}
+
+static int
+_sodium_runtime_intel_cpu_features(CPUFeatures * const cpu_features)
+{
+    unsigned int cpu_info[4];
+    unsigned int id;
+
+    _cpuid(cpu_info, 0x0);
+    if ((id = cpu_info[0]) == 0U) {
+        return -1;
+    }
+    _cpuid(cpu_info, 0x00000001);
+    cpu_features->have_sse2 = ((cpu_info[3] & CPUID_SSE2) != 0x0);
+    cpu_features->have_sse3 = ((cpu_info[2] & CPUIDECX_SSE3) != 0x0);
+
+    return 0;
+}
+
+int
+sodium_runtime_get_cpu_features(void)
+{
+    int ret = -1;
+
+    ret &= _sodium_runtime_arm_cpu_features(&cpu_features);
+    ret &= _sodium_runtime_intel_cpu_features(&cpu_features);
+    cpu_features.initialized = 1;
+
+    return ret;
+}
+
+int
+sodium_runtime_have_neon(void) {
+    return cpu_features.have_neon;
+}
+
+int
+sodium_runtime_have_sse2(void) {
+    return cpu_features.have_sse2;
+}
+
+int
+sodium_runtime_have_sse3(void) {
+    return cpu_features.have_sse3;
+}