Date: Sat, 6 Aug 2022 14:25:22 -0700
From: Nathan Bossart <nathandbossart@gmail.com>
To: Andres Freund <andres@anarazel.de>
Cc: John Naylor <john.naylor@enterprisedb.com>,
	John Naylor <jcnaylor@gmail.com>,
	pgsql-hackers <pgsql-hackers@postgresql.org>
Subject: Re: optimize lookups in snapshot [sub]xip arrays
Message-ID: <20220806212522.GA1221353@nathanxps13>
References: <20220802225539.nm7bgkwvkpzgib5x@awork3.anarazel.de>
 <20220802234357.GA802740@nathanxps13>
 <20220803180658.3ofq6v34wxchg4wc@awork3.anarazel.de>
 <20220803202540.GA1088968@nathanxps13>
 <CAFBsxsH51bBtyr3rf13x7YMNkFfuzZPSQxxbfUepBj8m4=ck=Q@mail.gmail.com>
 <20220804221551.GA1090071@nathanxps13>
 <CAFBsxsGmNefB-LAo8hNXLDAwWN0vHXc6H2A+NfFUQrj=rn50DQ@mail.gmail.com>
 <20220805202510.GA1208450@nathanxps13>
 <20220805220434.6h2aa7eww5rxird6@awork3.anarazel.de>
 <20220806181326.GA1216305@nathanxps13>
MIME-Version: 1.0
Content-Type: multipart/mixed; boundary="8t9RHnE3ZwKMSgU+"
Content-Disposition: inline
In-Reply-To: <20220806181326.GA1216305@nathanxps13>
Archived-At: 
 <https://www.postgresql.org/message-id/20220806212522.GA1221353%40nathanxps13>
Precedence: bulk


--8t9RHnE3ZwKMSgU+
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline

On Sat, Aug 06, 2022 at 11:13:26AM -0700, Nathan Bossart wrote:
> On Fri, Aug 05, 2022 at 03:04:34PM -0700, Andres Freund wrote:
>> But mainly I'd expect to find a difference if the SIMD code were optimized a
>> further on the basis of not needing to return the offset. E.g. by
>> replacing _mm_packs_epi32 with _mm_or_si128, that's cheaper.
> 
> I haven't been able to find a significant difference between the two.  If
> anything, the _mm_packs_epi* approach actually seems to be slightly faster
> in some cases.  For something marginally more concrete, I compared the two
> in perf-top and saw the following for the relevant instructions:

Nevermind, I'm wrong.  When compiled with -O2, it uses more than just the
xmm0 and xmm1 registers, and the _mm_or_si128 approach consistently shows a
speedup of slightly more than 5%.  Patches attached.

-- 
Nathan Bossart
Amazon Web Services: https://aws.amazon.com

--8t9RHnE3ZwKMSgU+
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0001-Introduce-optimized-routine-for-linear-searches-t.patch"

From bdfab91f1f2fa647d1b8a888dd6f8ad61ab80523 Mon Sep 17 00:00:00 2001
From: Nathan Bossart <nathandbossart@gmail.com>
Date: Wed, 3 Aug 2022 09:49:04 -0700
Subject: [PATCH v8 1/2] Introduce optimized routine for linear searches
 through an array of integers.

If SSE2 is available, this function uses it to speed up the search.  Otherwise,
it uses a simple 'for' loop.  This is a prerequisite for a follow-up commit
that will use this function to optimize [sub]xip lookups in
XidInMVCCSnapshot(), but it can be used anywhere that might benefit from such
an optimization.

It might be worthwhile to add an ARM-specific code path to this function in the
future.

Author: Nathan Bossart
Reviewed by: Andres Freund, John Naylor
Discussion: https://postgr.es/m/20220713170950.GA3116318%40nathanxps13
---
 src/include/port/pg_lfind.h | 70 +++++++++++++++++++++++++++++++++++++
 1 file changed, 70 insertions(+)
 create mode 100644 src/include/port/pg_lfind.h

diff --git a/src/include/port/pg_lfind.h b/src/include/port/pg_lfind.h
new file mode 100644
index 0000000000..8a212cc06b
--- /dev/null
+++ b/src/include/port/pg_lfind.h
@@ -0,0 +1,70 @@
+/*-------------------------------------------------------------------------
+ *
+ * pg_lfind.h
+ *	  Optimized linear search routines.
+ *
+ * Copyright (c) 2022, PostgreSQL Global Development Group
+ *
+ *
+ * IDENTIFICATION
+ *	  src/port/pg_lfind.h
+ *
+ *-------------------------------------------------------------------------
+ */
+#ifndef PG_LFIND_H
+#define PG_LFIND_H
+
+#include "port/simd.h"
+
+/*
+ * pg_lfind32
+ *
+ * Returns true if there is an element in 'base' that equals 'key'.  Otherwise,
+ * returns false.
+ */
+static inline bool
+pg_lfind32(uint32 key, uint32 *base, uint32 nelem)
+{
+	uint32		i = 0;
+
+	/* If possible, use SSE2 intrinsics to speed up the search. */
+#ifdef USE_SSE2
+	__m128i		keys = _mm_set1_epi32(key);	/* load 4 copies of key */
+	uint32		iterations = nelem & ~0xF;	/* round down to multiple of 16 */
+
+	for (; i < iterations; i += 16)
+	{
+		/* load the next 16 values into __m128i variables */
+		__m128i vals1 = _mm_loadu_si128((__m128i *) &base[i]);
+		__m128i vals2 = _mm_loadu_si128((__m128i *) &base[i + 4]);
+		__m128i vals3 = _mm_loadu_si128((__m128i *) &base[i + 8]);
+		__m128i vals4 = _mm_loadu_si128((__m128i *) &base[i + 12]);
+
+		/* perform the comparisons */
+		__m128i result1 = _mm_cmpeq_epi32(keys, vals1);
+		__m128i result2 = _mm_cmpeq_epi32(keys, vals2);
+		__m128i result3 = _mm_cmpeq_epi32(keys, vals3);
+		__m128i result4 = _mm_cmpeq_epi32(keys, vals4);
+
+		/* shrink the results into a single variable */
+		__m128i tmp1 = _mm_or_si128(result1, result2);
+		__m128i tmp2 = _mm_or_si128(result3, result4);
+		__m128i result = _mm_or_si128(tmp1, tmp2);
+
+		/* see if there was a match */
+		if (_mm_movemask_epi8(result) != 0)
+			return true;
+	}
+#endif
+
+	/* Process the remaining elements the slow way. */
+	for (; i < nelem; i++)
+	{
+		if (key == base[i])
+			return true;
+	}
+
+	return false;
+}
+
+#endif							/* PG_LFIND_H */
-- 
2.25.1


--8t9RHnE3ZwKMSgU+
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
	filename="v8-0002-Optimize-linear-searches-in-XidInMVCCSnapshot.patch"

From 826cc190c9ad76558adc41f1cbee76fc47e78c15 Mon Sep 17 00:00:00 2001
From: Nathan Bossart <nathandbossart@gmail.com>
Date: Wed, 3 Aug 2022 09:59:28 -0700
Subject: [PATCH v8 2/2] Optimize linear searches in XidInMVCCSnapshot().

This change makes use of the recently-introduced optimized linear search
routine to speed up searches through the [sub]xip arrays when possible, which
should improve performance significantly when the arrays are large.

Author: Nathan Bossart
Reviewed by: Andres Freund, John Naylor
Discussion: https://postgr.es/m/20220713170950.GA3116318%40nathanxps13
---
 src/backend/utils/time/snapmgr.c | 28 +++++++---------------------
 1 file changed, 7 insertions(+), 21 deletions(-)

diff --git a/src/backend/utils/time/snapmgr.c b/src/backend/utils/time/snapmgr.c
index 5bc2a15160..9b504c9745 100644
--- a/src/backend/utils/time/snapmgr.c
+++ b/src/backend/utils/time/snapmgr.c
@@ -56,6 +56,7 @@
 #include "datatype/timestamp.h"
 #include "lib/pairingheap.h"
 #include "miscadmin.h"
+#include "port/pg_lfind.h"
 #include "storage/predicate.h"
 #include "storage/proc.h"
 #include "storage/procarray.h"
@@ -2284,8 +2285,6 @@ RestoreTransactionSnapshot(Snapshot snapshot, void *source_pgproc)
 bool
 XidInMVCCSnapshot(TransactionId xid, Snapshot snapshot)
 {
-	uint32		i;
-
 	/*
 	 * Make a quick range check to eliminate most XIDs without looking at the
 	 * xip arrays.  Note that this is OK even if we convert a subxact XID to
@@ -2317,13 +2316,8 @@ XidInMVCCSnapshot(TransactionId xid, Snapshot snapshot)
 		if (!snapshot->suboverflowed)
 		{
 			/* we have full data, so search subxip */
-			int32		j;
-
-			for (j = 0; j < snapshot->subxcnt; j++)
-			{
-				if (TransactionIdEquals(xid, snapshot->subxip[j]))
-					return true;
-			}
+			if (pg_lfind32(xid, snapshot->subxip, snapshot->subxcnt))
+				return true;
 
 			/* not there, fall through to search xip[] */
 		}
@@ -2344,16 +2338,11 @@ XidInMVCCSnapshot(TransactionId xid, Snapshot snapshot)
 				return false;
 		}
 
-		for (i = 0; i < snapshot->xcnt; i++)
-		{
-			if (TransactionIdEquals(xid, snapshot->xip[i]))
-				return true;
-		}
+		if (pg_lfind32(xid, snapshot->xip, snapshot->xcnt))
+			return true;
 	}
 	else
 	{
-		int32		j;
-
 		/*
 		 * In recovery we store all xids in the subxact array because it is by
 		 * far the bigger array, and we mostly don't know which xids are
@@ -2383,11 +2372,8 @@ XidInMVCCSnapshot(TransactionId xid, Snapshot snapshot)
 		 * indeterminate xid. We don't know whether it's top level or subxact
 		 * but it doesn't matter. If it's present, the xid is visible.
 		 */
-		for (j = 0; j < snapshot->subxcnt; j++)
-		{
-			if (TransactionIdEquals(xid, snapshot->subxip[j]))
-				return true;
-		}
+		if (pg_lfind32(xid, snapshot->subxip, snapshot->subxcnt))
+			return true;
 	}
 
 	return false;
-- 
2.25.1


--8t9RHnE3ZwKMSgU+--