Only use x86 intrinsics for endian swapping on x86

src/util/common.c

@@ -15,43 +15,71 @@
 #include <stdio.h>
 #include <string.h>
 
+#if defined(__x86_64__) || defined(_M_AMD64)
 #ifdef _WIN32
 #include <intrin.h>
 #else
 #include <immintrin.h>
 #include <sys/mman.h>
 #endif
+#endif
 
 // This function actually works for both copying LE file data to BE system
 // memory, and vice versa. all we care about is if the source and target
 // endianness mismatch (which will only happen if the system is BE since the
 // file data is guaranteed to be LE), so simply flipping the bytes in this case
 // will work regardless of which direction we're copying.
+#if defined(__x86_64__) || defined(_M_AMD64)
+void copy_int_as_le(void *dst, void *src, size_t len) {
+    memcpy(dst, src, len);
+
+    #ifdef _WIN32
+    if (len == 2) {
+        *((uint16_t*) dst) = _byteswap_ushort(*((uint16_t*) dst));
+    } else if (len == 3 || len == 4) {
+        *((uint32_t*) dst) = _byteswap_ulong(*((uint32_t*) dst));
+    } else if (len == 8) {
+        *((uint64_t*) dst) = _byteswap_uint64(*((uint64_t*) dst));
+    }
+    #else
+    if (len == 2) {
+        *((uint16_t*) dst) = __builtin_bswap16(*((uint16_t*) dst));
+        ((uint8_t*) dst)[0] ^= ((uint8_t*) dst)[1];
+    } else if (len == 3 || len == 4) {
+        *((uint32_t*) dst) = __builtin_bswap32(*((uint32_t*) dst));
+    } else if (len == 8) {
+        *((uint64_t*) dst) = __builtin_bswap64(*((uint64_t*) dst));
+    }
+    #endif
+}
+#else
+inline static void _swap_bytes(void *bytes, uint8_t off_a, uint8_t off_b) {
+    unsigned char *byte_arr = (unsigned char *) bytes;
+    byte_arr[off_a] ^= byte_arr[off_b];
+    byte_arr[off_b] ^= byte_arr[off_a];
+    byte_arr[off_a] ^= byte_arr[off_b];
+}
+
 void copy_int_as_le(void *dst, void *src, size_t len) {
     memcpy(dst, src, len);
 
     int x = 1;
     if (((unsigned char*) &x)[0] == 0) {
         // system is big-endian, so we need to convert to little
-        #ifdef _WIN32
         if (len == 2) {
-            *((uint16_t*) dst) = _byteswap_ushort(*((uint16_t*) dst));
+            _swap_bytes(&dst, 0, 1);
         } else if (len == 3 || len == 4) {
-            *((uint32_t*) dst) = _byteswap_ulong(*((uint32_t*) dst));
+            _swap_bytes(&dst, 0, 3);
+            _swap_bytes(&dst, 1, 2);
         } else if (len == 8) {
-            *((uint64_t*) dst) = _byteswap_uint64(*((uint64_t*) dst));
+            _swap_bytes(&dst, 0, 7);
+            _swap_bytes(&dst, 1, 6);
+            _swap_bytes(&dst, 2, 5);
+            _swap_bytes(&dst, 3, 4);
         }
-        #else
-        if (len == 2) {
-            *((uint16_t*) dst) = __builtin_bswap16(*((uint16_t*) dst));
-        } else if (len == 3 || len == 4) {
-            *((uint32_t*) dst) = __builtin_bswap32(*((uint32_t*) dst));
-        } else if (len == 8) {
-            *((uint64_t*) dst) = __builtin_bswap64(*((uint64_t*) dst));
-        }
-        #endif
     }
 }
+#endif
 
 int validate_path_component(const char *cmpnt, uint8_t len_s) {
     for (uint8_t i = 0; i < len_s; i++) {