PARQUET-488: Add SSE cmake toggle, fix build on systems without SSE

I added an option to make SSE strictly opt-in for now. As a side effect of this, parquet-cpp now builds and the test suite passes out of the box on 32-bit ARMv7 (I tried it on my RaspberryPi Model B 2).

Author: Wes McKinney <wesm@apache.org>

Closes apache#74 from wesm/PARQUET-488 and squashes the following commits:

61225e9 [Wes McKinney] Use -march=native
3833efd [Wes McKinney] Remove stale cmake comment
70fcf65 [Wes McKinney] Add cmake PARQUET_USE_SSE option
775c72d [Wes McKinney] Fix compilation on arm7/raspberrypi

Change-Id: If8e4e7e1b7fc64df952cb8b82662bb017ca56f72

cpp/src/parquet/util/bit-util.h

@@ -151,11 +151,15 @@ class BitUtil {
 
   /// Returns the number of set bits in x
   static inline int Popcount(uint64_t x) {
+#ifdef PARQUET_USE_SSE
     if (LIKELY(CpuInfo::IsSupported(CpuInfo::POPCNT))) {
       return POPCNT_popcnt_u64(x);
     } else {
       return PopcountNoHw(x);
     }
+#else
+    return PopcountNoHw(x);
+#endif
   }
 
   // Compute correct population count for various-width signed integers

cpp/src/parquet/util/cpu-info.cc

@@ -24,7 +24,6 @@
 #include <sys/sysctl.h>
 #endif
 
-#include <mmintrin.h>
 #include <stdlib.h>
 #include <string.h>
 #include <unistd.h>

cpp/src/parquet/util/hash-util.h

@@ -202,11 +202,15 @@ class HashUtil {
   /// Seed values for different steps of the query execution should use different seeds
   /// to prevent accidental key collisions. (See IMPALA-219 for more details).
   static uint32_t Hash(const void* data, int32_t bytes, uint32_t seed) {
+#ifdef PARQUET_USE_SSE
     if (LIKELY(CpuInfo::IsSupported(CpuInfo::SSE4_2))) {
       return CrcHash(data, bytes, seed);
     } else {
       return MurmurHash2_64(data, bytes, seed);
     }
+#else
+    return MurmurHash2_64(data, bytes, seed);
+#endif
   }
 
   /// The magic number (used in hash_combine()) 0x9e3779b9 = 2^32 / (golden ratio).

cpp/src/parquet/util/sse-util.h

@@ -21,7 +21,9 @@
 #ifndef PARQUET_UTIL_SSE_UTIL_H
 #define PARQUET_UTIL_SSE_UTIL_H
 
+#ifdef PARQUET_USE_SSE
 #include <emmintrin.h>
+#endif
 
 namespace parquet_cpp {
 
@@ -71,6 +73,8 @@ namespace SSEUtil {
   };
 } // namespace SSEUtil
 
+#ifdef PARQUET_USE_SSE
+
 /// Define the SSE 4.2 intrinsics.  The caller must first verify at runtime (or codegen
 /// IR load time) that the processor supports SSE 4.2 before calling these.  These are
 /// defined outside the namespace because the IR w/ SSE 4.2 case needs to use macros.
@@ -81,12 +85,6 @@ namespace SSEUtil {
 /// (IMPALA-1399/1646).  The compiler intrinsics cannot be used without -msse4.2, so we
 /// define our own implementations of the intrinsics instead.
 
-#if defined(__SSE4_1__) || defined(__POPCNT__)
-/// Impala native code should not be compiled with -msse4.1 or higher until the minimum
-/// CPU requirement is raised to at least the targeted instruction set.
-#error "Do not compile with -msse4.1 or higher."
-#endif
-
 /// The PCMPxSTRy instructions require that the control byte 'mode' be encoded as an
 /// immediate.  So, those need to be always inlined in order to always propagate the
 /// mode constant into the inline asm.
@@ -214,7 +212,36 @@ static inline int64_t POPCNT_popcnt_u64(uint64_t a) {
   return 0;
 }
 
-#endif
+#endif // IR_COMPILE
+
+#else
+
+static inline uint32_t SSE4_crc32_u8(uint32_t crc, uint8_t v) {
+  DCHECK(false) << "SSE support is not enabled";
+  return 0;
+}
+
+static inline uint32_t SSE4_crc32_u16(uint32_t crc, uint16_t v) {
+  DCHECK(false) << "SSE support is not enabled";
+  return 0;
+}
+
+static inline uint32_t SSE4_crc32_u32(uint32_t crc, uint32_t v) {
+  DCHECK(false) << "SSE support is not enabled";
+  return 0;
+}
+
+static inline uint32_t SSE4_crc32_u64(uint32_t crc, uint64_t v) {
+  DCHECK(false) << "SSE support is not enabled";
+  return 0;
+}
+
+static inline int64_t POPCNT_popcnt_u64(uint64_t a) {
+  DCHECK(false) << "SSE support is not enabled";
+  return 0;
+}
+
+#endif // PARQUET_USE_SSE
 
 } // namespace parquet_cpp