arm64 crc prefetch optimise (#5773)

Summary:
prefetch data for following block，avoid cache miss when doing crc caculate

I do performance test at kunpeng-920 server(arm-v8, 64core@2.6GHz)
./db_bench --benchmarks=crc32c --block_size=500000000
before optimise : 587313.500 micros/op 1 ops/sec;  811.9 MB/s (500000000 per op)
after optimise  : 289248.500 micros/op 3 ops/sec; 1648.5 MB/s (500000000 per op)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5773

Differential Revision: D17347339

fbshipit-source-id: bfcd74f0f0eb4b322b959be68019ddcaae1e3341

util/crc32c_arm64.cc

@@ -64,7 +64,10 @@ uint32_t crc32c_arm64(uint32_t crc, unsigned char const *data,
      */
     uint32_t k0 = 0xe417f38a, k1 = 0x8f158014;
 
-    /* First 8 bytei for better pipelining */
+    /* Prefetch data for following block to avoid cache miss */
+    PREF1KL1((uint8_t *)buf64, 1024);
+
+    /* First 8 byte for better pipelining */
     crc0 = crc32c_u64(crc, *buf64++);
 
     /* 3 blocks crc32c parallel computation

util/crc32c_arm64.h

@@ -17,6 +17,17 @@
 #define crc32c_u16(crc, v) __crc32ch(crc, v)
 #define crc32c_u32(crc, v) __crc32cw(crc, v)
 #define crc32c_u64(crc, v) __crc32cd(crc, v)
+#define PREF4X64L1(buffer,PREF_OFFSET, ITR) \
+        __asm__("PRFM PLDL1KEEP, [%x[v],%[c]]"::[v]"r"(buffer), [c]"I"((PREF_OFFSET) + ((ITR) + 0)*64));\
+        __asm__("PRFM PLDL1KEEP, [%x[v],%[c]]"::[v]"r"(buffer), [c]"I"((PREF_OFFSET) + ((ITR) + 1)*64));\
+        __asm__("PRFM PLDL1KEEP, [%x[v],%[c]]"::[v]"r"(buffer), [c]"I"((PREF_OFFSET) + ((ITR) + 2)*64));\
+        __asm__("PRFM PLDL1KEEP, [%x[v],%[c]]"::[v]"r"(buffer), [c]"I"((PREF_OFFSET) + ((ITR) + 3)*64));
+
+#define PREF1KL1(buffer,PREF_OFFSET) \
+        PREF4X64L1(buffer,(PREF_OFFSET), 0) \
+        PREF4X64L1(buffer,(PREF_OFFSET), 4) \
+        PREF4X64L1(buffer,(PREF_OFFSET), 8) \
+        PREF4X64L1(buffer,(PREF_OFFSET), 12)
 
 extern uint32_t crc32c_arm64(uint32_t crc, unsigned char const *data, unsigned len);
 extern uint32_t crc32c_runtime_check(void);