Add fast path for validating UTF-8 text

Our previous validator used a traditional algorithm that performed
comparison and branching one byte at a time. It's useful in that
we always know exactly how many bytes we have validated, but that
precision comes at a cost. Input validation can show up prominently
in profiles of COPY FROM, and future improvements to COPY FROM such
as parallelism or faster line parsing will put more pressure on input
validation. Hence, add fast paths for both ASCII and multibyte UTF-8:

Use bitwise operations to check 16 bytes at a time for ASCII. If
that fails, use a "shift-based" DFA on those bytes to handle the
general case, including multibyte. These paths are relatively free
of branches and thus robust against all kinds of byte patterns. With
these algorithms, UTF-8 validation is several times faster, depending
on platform and the input byte distribution.

The previous coding in pg_utf8_verifystr() is retained for short
strings and for when the fast path returns an error.

Review, performance testing, and additional hacking by: Heikki
Linakangas, Vladimir Sitnikov, Amit Khandekar, Thomas Munro, and
Greg Stark

Discussion:
https://www.postgresql.org/message-id/CAFBsxsEV_SzH%2BOLyCiyon%3DiwggSyMh_eF6A3LU2tiWf3Cy2ZQg%40mail.gmail.com

1 files changed, 215 insertions, 0 deletions

diff --git a/src/common/wchar.c b/src/common/wchar.c
index a6bffd06428..be931c5e92a 100644
--- a/src/common/wchar.c
+++ b/src/common/wchar.c
@@ -1750,11 +1750,226 @@ pg_utf8_verifychar(const unsigned char *s, int len)
 	return l;
 }
 
+/*
+ * The fast path of the UTF-8 verifier uses a deterministic finite automaton
+ * (DFA) for multibyte characters. In a traditional table-driven DFA, the
+ * input byte and current state are used to compute an index into an array of
+ * state transitions. Since the address of the next transition is dependent
+ * on this computation, there is latency in executing the load instruction,
+ * and the CPU is not kept busy.
+ *
+ * Instead, we use a "shift-based" DFA as described by Per Vognsen:
+ *
+ * https://gist.github.com/pervognsen/218ea17743e1442e59bb60d29b1aa725
+ *
+ * In a shift-based DFA, the input byte is an index into array of integers
+ * whose bit pattern encodes the state transitions. To compute the next
+ * state, we simply right-shift the integer by the current state and apply a
+ * mask. In this scheme, the address of the transition only depends on the
+ * input byte, so there is better pipelining.
+ *
+ * The naming convention for states and transitions was adopted from a UTF-8
+ * to UTF-16/32 transcoder, whose table is reproduced below:
+ *
+ * https://github.com/BobSteagall/utf_utils/blob/6b7a465265de2f5fa6133d653df0c9bdd73bbcf8/src/utf_utils.cpp
+ *
+ * ILL  ASC  CR1  CR2  CR3  L2A  L3A  L3B  L3C  L4A  L4B  L4C CLASS / STATE
+ * ==========================================================================
+ * err, END, err, err, err, CS1, P3A, CS2, P3B, P4A, CS3, P4B,      | BGN/END
+ * err, err, err, err, err, err, err, err, err, err, err, err,      | ERR
+ *                                                                  |
+ * err, err, END, END, END, err, err, err, err, err, err, err,      | CS1
+ * err, err, CS1, CS1, CS1, err, err, err, err, err, err, err,      | CS2
+ * err, err, CS2, CS2, CS2, err, err, err, err, err, err, err,      | CS3
+ *                                                                  |
+ * err, err, err, err, CS1, err, err, err, err, err, err, err,      | P3A
+ * err, err, CS1, CS1, err, err, err, err, err, err, err, err,      | P3B
+ *                                                                  |
+ * err, err, err, CS2, CS2, err, err, err, err, err, err, err,      | P4A
+ * err, err, CS2, err, err, err, err, err, err, err, err, err,      | P4B
+ *
+ * In the most straightforward implementation, a shift-based DFA for UTF-8
+ * requires 64-bit integers to encode the transitions, but with an SMT solver
+ * it's possible to find state numbers such that the transitions fit within
+ * 32-bit integers, as Dougall Johnson demonstrated:
+ *
+ * https://gist.github.com/dougallj/166e326de6ad4cf2c94be97a204c025f
+ *
+ * This packed representation is the reason for the seemingly odd choice of
+ * state values below.
+ */
+
+/* Error */
+#define	ERR  0
+/* Begin */
+#define	BGN 11
+/* Continuation states, expect 1/2/3 continuation bytes */
+#define	CS1 16
+#define	CS2  1
+#define	CS3  5
+/* Leading byte was E0/ED, expect 1 more continuation byte */
+#define	P3A  6
+#define	P3B 20
+/* Leading byte was F0/F4, expect 2 more continuation bytes */
+#define	P4A 25
+#define	P4B 30
+/* Begin and End are the same state */
+#define	END BGN
+
+/* the encoded state transitions for the lookup table */
+
+/* ASCII */
+#define ASC (END << BGN)
+/* 2-byte lead */
+#define L2A (CS1 << BGN)
+/* 3-byte lead */
+#define L3A (P3A << BGN)
+#define L3B (CS2 << BGN)
+#define L3C (P3B << BGN)
+/* 4-byte lead */
+#define L4A (P4A << BGN)
+#define L4B (CS3 << BGN)
+#define L4C (P4B << BGN)
+/* continuation byte */
+#define CR1 (END << CS1) | (CS1 << CS2) | (CS2 << CS3) | (CS1 << P3B) | (CS2 << P4B)
+#define CR2 (END << CS1) | (CS1 << CS2) | (CS2 << CS3) | (CS1 << P3B) | (CS2 << P4A)
+#define CR3 (END << CS1) | (CS1 << CS2) | (CS2 << CS3) | (CS1 << P3A) | (CS2 << P4A)
+/* invalid byte */
+#define ILL ERR
+
+static const uint32 Utf8Transition[256] =
+{
+	/* ASCII */
+
+	ILL, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+	ASC, ASC, ASC, ASC, ASC, ASC, ASC, ASC,
+
+	/* continuation bytes */
+
+	/* 80..8F */
+	CR1, CR1, CR1, CR1, CR1, CR1, CR1, CR1,
+	CR1, CR1, CR1, CR1, CR1, CR1, CR1, CR1,
+
+	/* 90..9F */
+	CR2, CR2, CR2, CR2, CR2, CR2, CR2, CR2,
+	CR2, CR2, CR2, CR2, CR2, CR2, CR2, CR2,
+
+	/* A0..BF */
+	CR3, CR3, CR3, CR3, CR3, CR3, CR3, CR3,
+	CR3, CR3, CR3, CR3, CR3, CR3, CR3, CR3,
+	CR3, CR3, CR3, CR3, CR3, CR3, CR3, CR3,
+	CR3, CR3, CR3, CR3, CR3, CR3, CR3, CR3,
+
+	/* leading bytes */
+
+	/* C0..DF */
+	ILL, ILL, L2A, L2A, L2A, L2A, L2A, L2A,
+	L2A, L2A, L2A, L2A, L2A, L2A, L2A, L2A,
+	L2A, L2A, L2A, L2A, L2A, L2A, L2A, L2A,
+	L2A, L2A, L2A, L2A, L2A, L2A, L2A, L2A,
+
+	/* E0..EF */
+	L3A, L3B, L3B, L3B, L3B, L3B, L3B, L3B,
+	L3B, L3B, L3B, L3B, L3B, L3C, L3B, L3B,
+
+	/* F0..FF */
+	L4A, L4B, L4B, L4B, L4C, ILL, ILL, ILL,
+	ILL, ILL, ILL, ILL, ILL, ILL, ILL, ILL
+};
+
+static void
+utf8_advance(const unsigned char *s, uint32 *state, int len)
+{
+	/* Note: We deliberately don't check the state's value here. */
+	while (len > 0)
+	{
+		/*
+		 * It's important that the mask value is 31: In most instruction sets,
+		 * a shift by a 32-bit operand is understood to be a shift by its mod
+		 * 32, so the compiler should elide the mask operation.
+		 */
+		*state = Utf8Transition[*s++] >> (*state & 31);
+		len--;
+	}
+
+	*state &= 31;
+}
+
 static int
 pg_utf8_verifystr(const unsigned char *s, int len)
 {
 	const unsigned char *start = s;
+	const int	orig_len = len;
+	uint32		state = BGN;
+
+/*
+ * Sixteen seems to give the best balance of performance across different
+ * byte distributions.
+ */
+#define STRIDE_LENGTH 16
+
+	if (len >= STRIDE_LENGTH)
+	{
+		while (len >= STRIDE_LENGTH)
+		{
+			/*
+			 * If the chunk is all ASCII, we can skip the full UTF-8 check,
+			 * but we must first check for a non-END state, which means the
+			 * previous chunk ended in the middle of a multibyte sequence.
+			 */
+			if (state != END || !is_valid_ascii(s, STRIDE_LENGTH))
+				utf8_advance(s, &state, STRIDE_LENGTH);
+
+			s += STRIDE_LENGTH;
+			len -= STRIDE_LENGTH;
+		}
+
+		/*
+		 * The error state persists, so we only need to check for it here. In
+		 * case of error we start over from the beginning with the slow path
+		 * so we can count the valid bytes.
+		 */
+		if (state == ERR)
+		{
+			len = orig_len;
+			s = start;
+		}
+
+		/*
+		 * We treat all other states as success, but it's possible the fast
+		 * path exited in the middle of a multibyte sequence, since that
+		 * wouldn't have caused an error. Before checking the remaining bytes,
+		 * walk backwards to find the last byte that could have been the start
+		 * of a valid sequence.
+		 */
+		while (s > start)
+		{
+			s--;
+			len++;
+
+			if (!IS_HIGHBIT_SET(*s) || pg_utf_mblen(s) > 1)
+				break;
+		}
+	}
 
+	/* check remaining bytes */
 	while (len > 0)
 	{
 		int			l;