Use 64 bit little endian on ppc64le.

This has tangible performance benefits.

This lands https://github.com/google/snappy/pull/27

snappy-internal.h

@@ -80,9 +80,8 @@ char* CompressFragment(const char* input,
 // Does not read *(s1 + (s2_limit - s2)) or beyond.
 // Requires that s2_limit >= s2.
 //
-// Separate implementation for x86_64, for speed.  Uses the fact that
-// x86_64 is little endian.
-#if defined(ARCH_K8)
+// Separate implementation for 64-bit, little-endian cpus.
+#if defined(ARCH_K8) || (defined(ARCH_PPC) && !defined(WORDS_BIGENDIAN))
 static inline std::pair<size_t, bool> FindMatchLength(const char* s1,
                                                       const char* s2,
                                                       const char* s2_limit) {

snappy-stubs-internal.h

@@ -52,6 +52,10 @@
 // Enable 64-bit optimized versions of some routines.
 #define ARCH_K8 1
 
+#elif defined(__ppc64__)
+
+#define ARCH_PPC 1
+
 #endif
 
 // Needed by OS X, among others.