Speed up tail processing when hashing aligned C strings
authorJohn Naylor <john.naylor@postgresql.org>
Sun, 31 Mar 2024 05:19:16 +0000 (12:19 +0700)
committerJohn Naylor <john.naylor@postgresql.org>
Sun, 31 Mar 2024 05:19:16 +0000 (12:19 +0700)
After encountering the NUL terminator, the word-at-a-time loop exits
and we must hash the remaining bytes. Previously we calculated the
terminator's position and re-loaded the remaining bytes from the input
string. We already have all the data we need in a register, so let's
just mask off the bytes we need and hash them immediately. The mask can
be cheaply computed without knowing the terminator's position. We still
need that position for the length calculation, but the CPU can now
do that in parallel with other work, shortening the dependency chain.

Ants Aasma and John Naylor

Discussion: https://postgr.es/m/CANwKhkP7pCiW_5fAswLhs71-JKGEz1c1%2BPC0a_w1fwY4iGMqUA%40mail.gmail.com

src/include/common/hashfn_unstable.h

index 791750d136c48b4818a1c3f41b4681fa52b45b14..bd7323fe052bd18f33c6ecadc4c5c618bfe9668e 100644 (file)
@@ -219,8 +219,9 @@ static inline size_t
 fasthash_accum_cstring_aligned(fasthash_state *hs, const char *str)
 {
        const char *const start = str;
-       size_t          remainder;
+       uint64          chunk;
        uint64          zero_byte_low;
+       uint64          mask;
 
        Assert(PointerIsAligned(start, uint64));
 
@@ -239,7 +240,7 @@ fasthash_accum_cstring_aligned(fasthash_state *hs, const char *str)
         */
        for (;;)
        {
-               uint64          chunk = *(uint64 *) str;
+               chunk = *(uint64 *) str;
 
 #ifdef WORDS_BIGENDIAN
                zero_byte_low = haszero64(pg_bswap64(chunk));
@@ -254,14 +255,37 @@ fasthash_accum_cstring_aligned(fasthash_state *hs, const char *str)
                str += FH_SIZEOF_ACCUM;
        }
 
-       /*
-        * The byte corresponding to the NUL will be 0x80, so the rightmost bit
-        * position will be in the range 7, 15, ..., 63. Turn this into byte
-        * position by dividing by 8.
-        */
-       remainder = pg_rightmost_one_pos64(zero_byte_low) / BITS_PER_BYTE;
-       fasthash_accum(hs, str, remainder);
-       str += remainder;
+       if (zero_byte_low & 0xFF)
+       {
+               /*
+                * The next byte in the input is the NUL terminator, so we have
+                * nothing to do.
+                */
+       }
+       else
+       {
+               /*
+                * Create a mask for the remaining bytes so we can combine them into
+                * the hash. The mask also covers the NUL terminator, but that's
+                * harmless. The mask could contain 0x80 in bytes corresponding to the
+                * input past the terminator, but only where the input byte is zero or
+                * one, so also harmless.
+                */
+               mask = zero_byte_low | (zero_byte_low - 1);
+#ifdef WORDS_BIGENDIAN
+               /* need to mask the upper bytes */
+               mask = pg_bswap64(mask);
+#endif
+               hs->accum = chunk & mask;
+               fasthash_combine(hs);
+
+               /*
+                * The byte corresponding to the NUL will be 0x80, so the rightmost
+                * bit position will be in the range 15, 23, ..., 63. Turn this into
+                * byte position by dividing by 8.
+                */
+               str += pg_rightmost_one_pos64(zero_byte_low) / BITS_PER_BYTE;
+       }
 
        return str - start;
 }