Initial Commit for CRC32_vpmsum integration and WeakHash Failure

cmake/cpu_features.cmake

@@ -65,9 +65,9 @@ elseif (ARCH_PPC64LE)
     # Note that gcc and clang have support for x86 SSE2 intrinsics when building for PowerPC
     option (POWER9 "Build for Power 9 CPU and above" 0)
     if(POWER9)
-        set (COMPILER_FLAGS "${COMPILER_FLAGS} -maltivec -mcpu=power9 -D__SSE2__=1 -DNO_WARN_X86_INTRINSICS")
+        set (COMPILER_FLAGS "${COMPILER_FLAGS} -maltivec -mcpu=power9 -D__SSE2__=1 -DNO_WARN_X86_INTRINSICS -DHAVE_POWER8=1 -DHAS_ALTIVEC=1")
     else ()
-        set (COMPILER_FLAGS "${COMPILER_FLAGS} -maltivec -mcpu=power8 -D__SSE2__=1 -DNO_WARN_X86_INTRINSICS")
+        set (COMPILER_FLAGS "${COMPILER_FLAGS} -maltivec -mcpu=power8 -D__SSE2__=1 -DNO_WARN_X86_INTRINSICS -DHAVE_POWER8=1 -DHAS_ALTIVEC=1")
     endif ()
 
 elseif (ARCH_AMD64)

src/Common/HashTable/Hash.h

@@ -44,10 +44,16 @@ inline DB::UInt64 intHash64(DB::UInt64 x)
 #include <nmmintrin.h>
 #endif
 
+#if (defined(__PPC64__) || defined(__powerpc64__)) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 
+#include "Common/HashTable/crc32_constants.h"
+# include "Common/HashTable/vec_crc32.h"
+#endif
+
+
+
 #if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32)
 #include <arm_acle.h>
 #endif
-
 /// NOTE: Intel intrinsic can be confusing.
 /// - https://code.google.com/archive/p/sse-intrinsics/wikis/PmovIntrinsicBug.wiki
 /// - https://stackoverflow.com/questions/15752770/mm-crc32-u64-poorly-defined
@@ -57,6 +63,8 @@ inline DB::UInt64 intHashCRC32(DB::UInt64 x)
     return _mm_crc32_u64(-1ULL, x);
 #elif defined(__aarch64__) && defined(__ARM_FEATURE_CRC32)
     return __crc32cd(-1U, x);
+#elif (defined(__PPC64__) || defined(__powerpc64__)) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    return crc32_ppc(-1U, reinterpret_cast<const unsigned char *>(&x), sizeof(x));
 #else
     /// On other platforms we do not have CRC32. NOTE This can be confusing.
     /// NOTE: consider using intHash32()
@@ -69,6 +77,8 @@ inline DB::UInt64 intHashCRC32(DB::UInt64 x, DB::UInt64 updated_value)
     return _mm_crc32_u64(updated_value, x);
 #elif defined(__aarch64__) && defined(__ARM_FEATURE_CRC32)
     return __crc32cd(static_cast<UInt32>(updated_value), x);
+#elif (defined(__PPC64__) || defined(__powerpc64__)) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    return crc32_ppc(updated_value, reinterpret_cast<const unsigned char *>(&x), sizeof(x));
 #else
     /// On other platforms we do not have CRC32. NOTE This can be confusing.
     return intHash64(x) ^ updated_value;

src/Common/HashTable/clang_workaround.h

@@ -0,0 +1,87 @@
+#ifndef CLANG_WORKAROUND_H
+#define CLANG_WORKAROUND_H
+
+/*
+ * These stubs fix clang incompatibilities with GCC builtins.
+ */
+
+#ifndef __builtin_crypto_vpmsumw
+#define __builtin_crypto_vpmsumw __builtin_crypto_vpmsumb
+#endif
+#ifndef __builtin_crypto_vpmsumd
+#define __builtin_crypto_vpmsumd __builtin_crypto_vpmsumb
+#endif
+
+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);
+}
+
+/*
+ * GCC __builtin_pack_vector_int128 returns a vector __int128_t but Clang
+ * does not recognize this type. On GCC this builtin is translated to a
+ * xxpermdi instruction that only moves the registers __a, __b instead generates
+ * a load.
+ *
+ * Clang has vec_xxpermdi intrinsics. It was implemented in 4.0.0.
+ */
+static inline
+__vector unsigned long long  __builtin_pack_vector (unsigned long __a,
+						    unsigned long __b)
+{
+	#if defined(__BIG_ENDIAN__)
+	__vector unsigned long long __v = {__a, __b};
+	#else
+	__vector unsigned long long __v = {__b, __a};
+	#endif
+	return __v;
+}
+
+/*
+ * Clang 7 changed the behavior of vec_xxpermdi in order to provide the same
+ * behavior of GCC. That means code adapted to Clang >= 7 does not work on
+ * Clang <= 6.  So, fallback to __builtin_unpack_vector() on Clang <= 6.
+ */
+#if !defined vec_xxpermdi || __clang_major__ <= 6
+
+static inline
+unsigned long __builtin_unpack_vector (__vector unsigned long long __v,
+				       int __o)
+{
+	return __v[__o];
+}
+
+#if defined(__BIG_ENDIAN__)
+#define __builtin_unpack_vector_0(a) __builtin_unpack_vector ((a), 0)
+#define __builtin_unpack_vector_1(a) __builtin_unpack_vector ((a), 1)
+#else
+#define __builtin_unpack_vector_0(a) __builtin_unpack_vector ((a), 1)
+#define __builtin_unpack_vector_1(a) __builtin_unpack_vector ((a), 0)
+#endif
+
+#else
+
+static inline
+unsigned long __builtin_unpack_vector_0 (__vector unsigned long long __v)
+{
+	#if defined(__BIG_ENDIAN__)
+	return vec_xxpermdi(__v, __v, 0x0)[0];
+	#else
+	return vec_xxpermdi(__v, __v, 0x3)[0];
+	#endif
+}
+
+static inline
+unsigned long __builtin_unpack_vector_1 (__vector unsigned long long __v)
+{
+	#if defined(__BIG_ENDIAN__)
+	return vec_xxpermdi(__v, __v, 0x3)[0];
+	#else
+	return vec_xxpermdi(__v, __v, 0x0)[0];
+	#endif
+}
+#endif /* vec_xxpermdi */
+
+#endif