Add an altivec variant of "count_lengths" in inftrees

This accounts for a small bump in performance

1 files changed, 25 insertions, 0 deletions

diff --git a/arch/power/power_intrins.h b/arch/power/power_intrins.h
index 965387c9e9..3efcfb9722 100644
--- a/arch/power/power_intrins.h
+++ b/arch/power/power_intrins.h
@@ -26,11 +26,36 @@
 #define __builtin_crypto_vpmsumd __builtin_crypto_vpmsumb
 #endif
 
+#ifdef __VSX__
 static inline __vector unsigned long long __attribute__((overloadable))
 vec_ld(int __a, const __vector unsigned long long* __b) {
     return (__vector unsigned long long)__builtin_altivec_lvx(__a, __b);
 }
+#endif
 
 #endif
 
+/* There's no version of this that operates over unsigned and if casted, it does
+ * sign extension. Let's write an endian independent version and hope the compiler
+ * eliminates creating another zero idiom for the zero value if one exists locally */
+static inline vector unsigned short vec_unpackl(vector unsigned char a) {
+    vector unsigned char zero = vec_splat_u8(0);
+
+#if BYTE_ORDER == BIG_ENDIAN
+    return (vector unsigned short)vec_mergel(zero, a);
+#else
+    return (vector unsigned short)vec_mergel(a, zero);
+#endif
+}
+
+static inline vector unsigned short vec_unpackh(vector unsigned char a) {
+    vector unsigned char zero = vec_splat_u8(0);
+
+#if BYTE_ORDER == BIG_ENDIAN
+    return (vector unsigned short)vec_mergeh(zero, a);
+#else
+    return (vector unsigned short)vec_mergeh(a, zero);
+#endif
+}
+
 #endif