v16-0001-Improve-CRC32C-performance-on-recent-x86_64-plat.patch

From b3af802cf28cdc0937e163dbba005f823d74e0d0 Mon Sep 17 00:00:00 2001
From: John Naylor <john.naylor@postgresql.org>
Date: Tue, 25 Mar 2025 19:22:32 +0700
Subject: [PATCH v16 1/2] Improve CRC32C performance on recent x86_64 platforms

The previous implementation of CRC32C on x86 relied on the native
CRC32 instruction from the SSE 4.2 extension, which operates on
up to 8 bytes at a time. We can get a substantial speedup by using
carryless multiplication on SIMD registers, processing 64 bytes per
loop iteration. Shorter inputs fall back to ordinary CRC instructions.

The VPCLMULQDQ instruction on 512-bit registers has been available
on Intel hardware since 2019 and AMD since 2022. There is an older
variant for 128-bit registers, but at least on Zen 2 it performs worse
than normal CRC instructions for short inputs. (Thanks to David Rowley
for testing on that platform.)

We must now do a runtime check, even for builds that target SSE
4.2. This doesn't matter in practice for WAL (arguably the most
critical case), because since commit af0c24855 the final computation
with the 20-byte WAL header is inlined and unrolled. Compared with
two direct function calls, testing showed equal or slightly faster
performance in performing an indirect function call on several dozen
bytes followed by inlined instructions on constant input of 20 bytes.

The MIT-licensed implementation was generated with the "generate"
program from

https://github.com/corsix/fast-crc32/

Based on: "Fast CRC Computation for Generic Polynomials Using PCLMULQDQ
Instruction" V. Gopal, E. Ozturk, et al., 2009

Co-authored-by: Raghuveer Devulapalli <raghuveer.devulapalli@intel.com>
Co-authored-by: Paul Amonson <paul.d.amonson@intel.com>
Reviewed-by: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de> (earlier version)
Tested-by: Raghuveer Devulapalli <raghuveer.devulapalli@intel.com>
Tested-by: Matthew Sterrett <matthewsterrett2@gmail.com> (earlier version)
Discussion: https://postgr.es/m/BL1PR11MB530401FA7E9B1CA432CF9DC3DC192@BL1PR11MB5304.namprd11.prod.outlook.com
Discussion: https://postgr.es/m/PH8PR11MB82869FF741DFA4E9A029FF13FBF72@PH8PR11MB8286.namprd11.prod.outlook.com
---
 config/c-compiler.m4                  | 37 +++++++++++
 configure                             | 91 ++++++++++++++++++++++----
 configure.ac                          | 41 ++++++++----
 meson.build                           | 58 +++++++++++++----
 src/include/pg_config.h.in            |  3 +
 src/include/port/pg_crc32c.h          | 39 ++++++++---
 src/port/meson.build                  |  1 +
 src/port/pg_crc32c_sse42.c            | 94 ++++++++++++++++++++++++++-
 src/port/pg_crc32c_sse42_choose.c     | 75 ++++++++++++++++-----
 src/test/regress/expected/strings.out | 24 +++++++
 src/test/regress/sql/strings.sql      |  5 ++
 11 files changed, 408 insertions(+), 60 deletions(-)

diff --git a/config/c-compiler.m4 b/config/c-compiler.m4
index e9e54470e66..4c4863f6de0 100644
--- a/config/c-compiler.m4
+++ b/config/c-compiler.m4
@@ -581,6 +581,43 @@ fi
 undefine([Ac_cachevar])dnl
 ])# PGAC_SSE42_CRC32_INTRINSICS
 
+# PGAC_AVX512_PCLMUL_INTRINSICS
+# ---------------------------
+# Check if the compiler supports AVX-512 carryless multiplication
+# and AVX-512VL instructions used for computing CRC. AVX-512F is
+# assumed to be supported if the above are.
+#
+# If the intrinsics are supported, sets pgac_avx512_pclmul_intrinsics.
+AC_DEFUN([PGAC_AVX512_PCLMUL_INTRINSICS],
+[define([Ac_cachevar], [AS_TR_SH([pgac_cv_avx512_pclmul_intrinsics])])dnl
+AC_CACHE_CHECK([for _mm512_clmulepi64_epi128], [Ac_cachevar],
+[AC_LINK_IFELSE([AC_LANG_PROGRAM([#include <immintrin.h>
+    __m512i x;
+    __m512i y;
+
+    #if defined(__has_attribute) && __has_attribute (target)
+    __attribute__((target("vpclmulqdq,avx512vl")))
+    #endif
+    static int avx512_pclmul_test(void)
+    {
+      __m128i z;
+
+      y = _mm512_clmulepi64_epi128(x, y, 0);
+      z = _mm_ternarylogic_epi64(
+                _mm512_castsi512_si128(y),
+                _mm512_extracti32x4_epi32(y, 1),
+                _mm512_extracti32x4_epi32(y, 2),
+                0x96);
+      return _mm_crc32_u64(0, _mm_extract_epi64(z, 0));
+    }],
+  [return avx512_pclmul_test();])],
+  [Ac_cachevar=yes],
+  [Ac_cachevar=no])])
+if test x"$Ac_cachevar" = x"yes"; then
+  pgac_avx512_pclmul_intrinsics=yes
+fi
+undefine([Ac_cachevar])dnl
+])# PGAC_AVX512_PCLMUL_INTRINSICS
 
 # PGAC_ARMV8_CRC32C_INTRINSICS
 # ----------------------------
diff --git a/configure b/configure
index 30d949c3c46..03cdae7423c 100755
--- a/configure
+++ b/configure
@@ -17829,17 +17829,21 @@ fi
 
 # Select CRC-32C implementation.
 #
-# If we are targeting a processor that has Intel SSE 4.2 instructions, we can
-# use the special CRC instructions for calculating CRC-32C. If we're not
-# targeting such a processor, but we can nevertheless produce code that uses
-# the SSE intrinsics, compile both implementations and select which one to use
-# at runtime, depending on whether SSE 4.2 is supported by the processor we're
-# running on.
+# There are three methods of calculating CRC, in order of increasing
+# performance:
 #
-# Similarly, if we are targeting an ARM processor that has the CRC
-# instructions that are part of the ARMv8 CRC Extension, use them. And if
-# we're not targeting such a processor, but can nevertheless produce code that
-# uses the CRC instructions, compile both, and select at runtime.
+# 1. The fallback using a lookup table, called slicing-by-8
+# 2. CRC-32C instructions (found in e.g. Intel SSE 4.2 and ARMv8 CRC Extension)
+# 3. Algorithms using carryless multiplication instructions
+#    (e.g. Intel PCLMUL and Arm PMULL)
+#
+# If we can produce code (via function attributes or additional compiler
+# flags) that uses #2 (and possibly #3), we compile all implementations
+# and select which one to use at runtime, depending on what is supported
+# by the processor we're running on.
+#
+# If we are targeting a processor that has #2, we can use that without
+# runtime selection.
 #
 # Note that we do not use __attribute__((target("..."))) for the ARM CRC
 # instructions because until clang 16, using the ARM intrinsics still requires
@@ -17890,7 +17894,7 @@ if test x"$USE_SSE42_CRC32C" = x"1"; then
 
 $as_echo "#define USE_SSE42_CRC32C 1" >>confdefs.h
 
-  PG_CRC32C_OBJS="pg_crc32c_sse42.o"
+  PG_CRC32C_OBJS="pg_crc32c_sse42.o pg_crc32c_sse42_choose.o"
   { $as_echo "$as_me:${as_lineno-$LINENO}: result: SSE 4.2" >&5
 $as_echo "SSE 4.2" >&6; }
 else
@@ -17939,6 +17943,71 @@ $as_echo "slicing-by-8" >&6; }
 fi
 
 
+# Check for carryless multiplication intrinsics to do vectorized CRC calculations.
+#
+if test x"$host_cpu" = x"x86_64"; then
+  { $as_echo "$as_me:${as_lineno-$LINENO}: checking for _mm512_clmulepi64_epi128" >&5
+$as_echo_n "checking for _mm512_clmulepi64_epi128... " >&6; }
+if ${pgac_cv_avx512_pclmul_intrinsics+:} false; then :
+  $as_echo_n "(cached) " >&6
+else
+  cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+#include <immintrin.h>
+    __m512i x;
+    __m512i y;
+
+    #if defined(__has_attribute) && __has_attribute (target)
+    __attribute__((target("vpclmulqdq,avx512vl")))
+    #endif
+    static int avx512_pclmul_test(void)
+    {
+      __m128i z;
+
+      y = _mm512_clmulepi64_epi128(x, y, 0);
+      z = _mm_ternarylogic_epi64(
+                _mm512_castsi512_si128(y),
+                _mm512_extracti32x4_epi32(y, 1),
+                _mm512_extracti32x4_epi32(y, 2),
+                0x96);
+      return _mm_crc32_u64(0, _mm_extract_epi64(z, 0));
+    }
+int
+main ()
+{
+return avx512_pclmul_test();
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_link "$LINENO"; then :
+  pgac_cv_avx512_pclmul_intrinsics=yes
+else
+  pgac_cv_avx512_pclmul_intrinsics=no
+fi
+rm -f core conftest.err conftest.$ac_objext \
+    conftest$ac_exeext conftest.$ac_ext
+fi
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $pgac_cv_avx512_pclmul_intrinsics" >&5
+$as_echo "$pgac_cv_avx512_pclmul_intrinsics" >&6; }
+if test x"$pgac_cv_avx512_pclmul_intrinsics" = x"yes"; then
+  pgac_avx512_pclmul_intrinsics=yes
+fi
+
+fi
+
+{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for vectorized CRC-32C" >&5
+$as_echo_n "checking for vectorized CRC-32C... " >&6; }
+if test x"$pgac_avx512_pclmul_intrinsics" = x"yes"; then
+
+$as_echo "#define USE_AVX512_CRC_WITH_RUNTIME_CHECK 1" >>confdefs.h
+
+  { $as_echo "$as_me:${as_lineno-$LINENO}: result: AVX-512 with runtime check" >&5
+$as_echo "AVX-512 with runtime check" >&6; }
+else
+  { $as_echo "$as_me:${as_lineno-$LINENO}: result: none" >&5
+$as_echo "none" >&6; }
+fi
 
 # Select semaphore implementation type.
 if test "$PORTNAME" != "win32"; then
diff --git a/configure.ac b/configure.ac
index 25cdfcf65af..0bce658f296 100644
--- a/configure.ac
+++ b/configure.ac
@@ -2115,17 +2115,21 @@ AC_SUBST(CFLAGS_CRC)
 
 # Select CRC-32C implementation.
 #
-# If we are targeting a processor that has Intel SSE 4.2 instructions, we can
-# use the special CRC instructions for calculating CRC-32C. If we're not
-# targeting such a processor, but we can nevertheless produce code that uses
-# the SSE intrinsics, compile both implementations and select which one to use
-# at runtime, depending on whether SSE 4.2 is supported by the processor we're
-# running on.
-#
-# Similarly, if we are targeting an ARM processor that has the CRC
-# instructions that are part of the ARMv8 CRC Extension, use them. And if
-# we're not targeting such a processor, but can nevertheless produce code that
-# uses the CRC instructions, compile both, and select at runtime.
+# There are three methods of calculating CRC, in order of increasing
+# performance:
+#
+# 1. The fallback using a lookup table, called slicing-by-8
+# 2. CRC-32C instructions (found in e.g. Intel SSE 4.2 and ARMv8 CRC Extension)
+# 3. Algorithms using carryless multiplication instructions
+#    (e.g. Intel PCLMUL and Arm PMULL)
+#
+# If we can produce code (via function attributes or additional compiler
+# flags) that uses #2 (and possibly #3), we compile all implementations
+# and select which one to use at runtime, depending on what is supported
+# by the processor we're running on.
+#
+# If we are targeting a processor that has #2, we can use that without
+# runtime selection.
 #
 # Note that we do not use __attribute__((target("..."))) for the ARM CRC
 # instructions because until clang 16, using the ARM intrinsics still requires
@@ -2173,7 +2177,7 @@ fi
 AC_MSG_CHECKING([which CRC-32C implementation to use])
 if test x"$USE_SSE42_CRC32C" = x"1"; then
   AC_DEFINE(USE_SSE42_CRC32C, 1, [Define to 1 use Intel SSE 4.2 CRC instructions.])
-  PG_CRC32C_OBJS="pg_crc32c_sse42.o"
+  PG_CRC32C_OBJS="pg_crc32c_sse42.o pg_crc32c_sse42_choose.o"
   AC_MSG_RESULT(SSE 4.2)
 else
   if test x"$USE_SSE42_CRC32C_WITH_RUNTIME_CHECK" = x"1"; then
@@ -2206,6 +2210,19 @@ else
 fi
 AC_SUBST(PG_CRC32C_OBJS)
 
+# Check for carryless multiplication intrinsics to do vectorized CRC calculations.
+#
+if test x"$host_cpu" = x"x86_64"; then
+  PGAC_AVX512_PCLMUL_INTRINSICS()
+fi
+
+AC_MSG_CHECKING([for vectorized CRC-32C])
+if test x"$pgac_avx512_pclmul_intrinsics" = x"yes"; then
+  AC_DEFINE(USE_AVX512_CRC_WITH_RUNTIME_CHECK, 1, [Define to 1 to use AVX-512 CRC algorithms with a runtime check.])
+  AC_MSG_RESULT(AVX-512 with runtime check)
+else
+  AC_MSG_RESULT(none)
+fi
 
 # Select semaphore implementation type.
 if test "$PORTNAME" != "win32"; then
diff --git a/meson.build b/meson.build
index b8da4966297..a42f54bf493 100644
--- a/meson.build
+++ b/meson.build
@@ -2348,17 +2348,21 @@ endif
 ###############################################################
 # Select CRC-32C implementation.
 #
-# If we are targeting a processor that has Intel SSE 4.2 instructions, we can
-# use the special CRC instructions for calculating CRC-32C. If we're not
-# targeting such a processor, but we can nevertheless produce code that uses
-# the SSE intrinsics, compile both implementations and select which one to use
-# at runtime, depending on whether SSE 4.2 is supported by the processor we're
-# running on.
+# There are three methods of calculating CRC, in order of increasing
+# performance:
 #
-# Similarly, if we are targeting an ARM processor that has the CRC
-# instructions that are part of the ARMv8 CRC Extension, use them. And if
-# we're not targeting such a processor, but can nevertheless produce code that
-# uses the CRC instructions, compile both, and select at runtime.
+# 1. The fallback using a lookup table, called slicing-by-8
+# 2. CRC-32C instructions (found in e.g. Intel SSE 4.2 and ARMv8 CRC Extension)
+# 3. Algorithms using carryless multiplication instructions
+#    (e.g. Intel PCLMUL and Arm PMULL)
+#
+# If we can produce code (via function attributes or additional compiler
+# flags) that uses #2 (and possibly #3), we compile all implementations
+# and select which one to use at runtime, depending on what is supported
+# by the processor we're running on.
+#
+# If we are targeting a processor that has #2, we can use that without
+# runtime selection.
 #
 # Note that we do not use __attribute__((target("..."))) for the ARM CRC
 # instructions because until clang 16, using the ARM intrinsics still requires
@@ -2392,7 +2396,7 @@ int main(void)
 }
 '''
 
-    if not cc.links(prog, name: '_mm_crc32_u8 and _mm_crc32_u32',
+    if not cc.links(prog, name: 'SSE 4.2 CRC32C',
           args: test_c_args)
       # Do not use Intel SSE 4.2
     elif (cc.get_define('__SSE4_2__') != '')
@@ -2407,6 +2411,38 @@ int main(void)
       have_optimized_crc = true
     endif
 
+    # Check if the compiler supports AVX-512 carryless multiplication
+    # and AVX-512VL instructions used for computing CRC. AVX-512F is
+    # assumed to be supported if the above are.
+    prog = '''
+#include <immintrin.h>
+__m512i x;
+__m512i y;
+
+#if defined(__has_attribute) && __has_attribute (target)
+__attribute__((target("vpclmulqdq,avx512vl")))
+#endif
+int main(void)
+{
+     __m128i z;
+
+    y = _mm512_clmulepi64_epi128(x, y, 0);
+    z = _mm_ternarylogic_epi64(
+            _mm512_castsi512_si128(y),
+            _mm512_extracti32x4_epi32(y, 1),
+            _mm512_extracti32x4_epi32(y, 2),
+            0x96);
+    /* return computed value, to prevent the above being optimized away */
+    return _mm_crc32_u64(0, _mm_extract_epi64(z, 0));
+}
+'''
+
+    if cc.links(prog,
+        name: 'AVX-512 CRC32C',
+        args: test_c_args)
+      cdata.set('USE_AVX512_CRC32C_WITH_RUNTIME_CHECK', 1)
+    endif
+
   endif
 
 elif host_cpu == 'arm' or host_cpu == 'aarch64'
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 92f0616c400..97f2f8c19cc 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -654,6 +654,9 @@
 /* Define to 1 to build with assertion checks. (--enable-cassert) */
 #undef USE_ASSERT_CHECKING
 
+/* Define to 1 to use AVX-512 CRC algorithms with a runtime check. */
+#undef USE_AVX512_CRC_WITH_RUNTIME_CHECK
+
 /* Define to 1 to use AVX-512 popcount instructions with a runtime check. */
 #undef USE_AVX512_POPCNT_WITH_RUNTIME_CHECK
 
diff --git a/src/include/port/pg_crc32c.h b/src/include/port/pg_crc32c.h
index 9376d223fef..82313bb7fcf 100644
--- a/src/include/port/pg_crc32c.h
+++ b/src/include/port/pg_crc32c.h
@@ -42,7 +42,10 @@ typedef uint32 pg_crc32c;
 #define EQ_CRC32C(c1, c2) ((c1) == (c2))
 
 #if defined(USE_SSE42_CRC32C)
-/* Use Intel SSE4.2 instructions. */
+/*
+ * Use either Intel SSE 4.2 or AVX-512 instructions. We don't need a runtime check
+ * for SSE 4.2, so we can inline those in some cases.
+ */
 
 #include <nmmintrin.h>
 
@@ -50,7 +53,11 @@ typedef uint32 pg_crc32c;
 	((crc) = pg_comp_crc32c_dispatch((crc), (data), (len)))
 #define FIN_CRC32C(crc) ((crc) ^= 0xFFFFFFFF)
 
+extern pg_crc32c (*pg_comp_crc32c) (pg_crc32c crc, const void *data, size_t len);
 extern pg_crc32c pg_comp_crc32c_sse42(pg_crc32c crc, const void *data, size_t len);
+#ifdef USE_AVX512_CRC32C_WITH_RUNTIME_CHECK
+extern pg_crc32c pg_comp_crc32c_avx512(pg_crc32c crc, const void *data, size_t len);
+#endif
 
 /*
  * We can only get here if the host compiler targets SSE 4.2, but on some
@@ -82,9 +89,27 @@ pg_comp_crc32c_dispatch(pg_crc32c crc, const void *data, size_t len)
 		return crc;
 	}
 	else
-		return pg_comp_crc32c_sse42(crc, data, len);
+		/* Otherwise, use a runtime check for AVX-512 instructions. */
+		return pg_comp_crc32c(crc, data, len);
 }
 
+#elif defined(USE_SSE42_CRC32C_WITH_RUNTIME_CHECK)
+
+/*
+ * Use Intel SSE 4.2 or AVX-512 instructions, but perform a runtime check first
+ * to check that they are available.
+ */
+#define COMP_CRC32C(crc, data, len) \
+	((crc) = pg_comp_crc32c((crc), (data), (len)))
+#define FIN_CRC32C(crc) ((crc) ^= 0xFFFFFFFF)
+
+extern pg_crc32c pg_comp_crc32c_sb8(pg_crc32c crc, const void *data, size_t len);
+extern PGDLLIMPORT pg_crc32c (*pg_comp_crc32c) (pg_crc32c crc, const void *data, size_t len);
+extern pg_crc32c pg_comp_crc32c_sse42(pg_crc32c crc, const void *data, size_t len);
+#ifdef USE_AVX512_CRC32C_WITH_RUNTIME_CHECK
+extern pg_crc32c pg_comp_crc32c_avx512(pg_crc32c crc, const void *data, size_t len);
+#endif
+
 #elif defined(USE_ARMV8_CRC32C)
 /* Use ARMv8 CRC Extension instructions. */
 
@@ -103,10 +128,10 @@ extern pg_crc32c pg_comp_crc32c_armv8(pg_crc32c crc, const void *data, size_t le
 
 extern pg_crc32c pg_comp_crc32c_loongarch(pg_crc32c crc, const void *data, size_t len);
 
-#elif defined(USE_SSE42_CRC32C_WITH_RUNTIME_CHECK) || defined(USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK)
+#elif defined(USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK)
 
 /*
- * Use Intel SSE 4.2 or ARMv8 instructions, but perform a runtime check first
+ * Use ARMv8 instructions, but perform a runtime check first
  * to check that they are available.
  */
 #define COMP_CRC32C(crc, data, len) \
@@ -115,13 +140,7 @@ extern pg_crc32c pg_comp_crc32c_loongarch(pg_crc32c crc, const void *data, size_
 
 extern pg_crc32c pg_comp_crc32c_sb8(pg_crc32c crc, const void *data, size_t len);
 extern pg_crc32c (*pg_comp_crc32c) (pg_crc32c crc, const void *data, size_t len);
-
-#ifdef USE_SSE42_CRC32C_WITH_RUNTIME_CHECK
-extern pg_crc32c pg_comp_crc32c_sse42(pg_crc32c crc, const void *data, size_t len);
-#endif
-#ifdef USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK
 extern pg_crc32c pg_comp_crc32c_armv8(pg_crc32c crc, const void *data, size_t len);
-#endif
 
 #else
 /*
diff --git a/src/port/meson.build b/src/port/meson.build
index cf7f07644b9..2d2bf1b70f7 100644
--- a/src/port/meson.build
+++ b/src/port/meson.build
@@ -85,6 +85,7 @@ replace_funcs_pos = [
   # x86/x64
   ['pg_crc32c_sse42', 'USE_SSE42_CRC32C'],
   ['pg_crc32c_sse42', 'USE_SSE42_CRC32C_WITH_RUNTIME_CHECK'],
+  ['pg_crc32c_sse42_choose', 'USE_SSE42_CRC32C'],
   ['pg_crc32c_sse42_choose', 'USE_SSE42_CRC32C_WITH_RUNTIME_CHECK'],
   ['pg_crc32c_sb8', 'USE_SSE42_CRC32C_WITH_RUNTIME_CHECK'],
 
diff --git a/src/port/pg_crc32c_sse42.c b/src/port/pg_crc32c_sse42.c
index 22c2137df31..4d8dec004b1 100644
--- a/src/port/pg_crc32c_sse42.c
+++ b/src/port/pg_crc32c_sse42.c
@@ -1,7 +1,7 @@
 /*-------------------------------------------------------------------------
  *
  * pg_crc32c_sse42.c
- *	  Compute CRC-32C checksum using Intel SSE 4.2 instructions.
+ *	  Compute CRC-32C checksum using Intel SSE 4.2 or AVX-512 instructions.
  *
  * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
@@ -15,6 +15,9 @@
 #include "c.h"
 
 #include <nmmintrin.h>
+#ifdef USE_AVX512_CRC32C_WITH_RUNTIME_CHECK
+#include <immintrin.h>
+#endif
 
 #include "port/pg_crc32c.h"
 
@@ -68,3 +71,92 @@ pg_comp_crc32c_sse42(pg_crc32c crc, const void *data, size_t len)
 
 	return crc;
 }
+
+#ifdef USE_AVX512_CRC32C_WITH_RUNTIME_CHECK
+
+/*
+ * Note: There is no copyright notice in the following generated code.
+ *
+ * We have modified the output to
+ *   - match our function declaration
+ *   - make whitespace match project style
+ *   - add a threshold for the alignment stanza
+ */
+
+/* Generated by https://github.com/corsix/fast-crc32/ using: */
+/* ./generate -i avx512_vpclmulqdq -p crc32c -a v1e */
+/* MIT licensed */
+
+#define clmul_lo(a, b) (_mm512_clmulepi64_epi128((a), (b), 0))
+#define clmul_hi(a, b) (_mm512_clmulepi64_epi128((a), (b), 17))
+
+pg_attribute_target("vpclmulqdq,avx512vl")
+pg_crc32c
+pg_comp_crc32c_avx512(pg_crc32c crc, const void *data, size_t length)
+{
+	/* adjust names to match generated code */
+	pg_crc32c	crc0 = crc;
+	size_t		len = length;
+	const char *buf = data;
+
+	/* Align on cacheline boundary. The threshold is somewhat arbitrary. */
+	if (unlikely(len > 256))
+	{
+		for (; len && ((uintptr_t) buf & 7); --len)
+			crc0 = _mm_crc32_u8(crc0, *buf++);
+		while (((uintptr_t) buf & 56) && len >= 8)
+		{
+			crc0 = _mm_crc32_u64(crc0, *(const uint64_t *) buf);
+			buf += 8;
+			len -= 8;
+		}
+	}
+
+	if (len >= 64)
+	{
+		const char *end = buf + len;
+		const char *limit = buf + len - 64;
+		__m128i		z0;
+
+		/* First vector chunk. */
+		__m512i		x0 = _mm512_loadu_si512((const void *) buf),
+					y0;
+		__m512i		k;
+
+		k = _mm512_broadcast_i32x4(_mm_setr_epi32(0x740eef02, 0, 0x9e4addf8, 0));
+		x0 = _mm512_xor_si512(_mm512_castsi128_si512(_mm_cvtsi32_si128(crc0)), x0);
+		buf += 64;
+
+		/* Main loop. */
+		while (buf <= limit)
+		{
+			y0 = clmul_lo(x0, k), x0 = clmul_hi(x0, k);
+			x0 = _mm512_ternarylogic_epi64(x0, y0,
+										   _mm512_loadu_si512((const void *) buf),
+										   0x96);
+			buf += 64;
+		}
+
+		/* Reduce 512 bits to 128 bits. */
+		k = _mm512_setr_epi32(0x1c291d04, 0, 0xddc0152b, 0,
+							  0x3da6d0cb, 0, 0xba4fc28e, 0,
+							  0xf20c0dfe, 0, 0x493c7d27, 0,
+							  0, 0, 0, 0);
+		y0 = clmul_lo(x0, k), k = clmul_hi(x0, k);
+		y0 = _mm512_xor_si512(y0, k);
+		z0 = _mm_ternarylogic_epi64(_mm512_castsi512_si128(y0),
+									_mm512_extracti32x4_epi32(y0, 1),
+									_mm512_extracti32x4_epi32(y0, 2),
+									0x96);
+		z0 = _mm_xor_si128(z0, _mm512_extracti32x4_epi32(x0, 3));
+
+		/* Reduce 128 bits to 32 bits, and multiply by x^32. */
+		crc0 = _mm_crc32_u64(0, _mm_extract_epi64(z0, 0));
+		crc0 = _mm_crc32_u64(crc0, _mm_extract_epi64(z0, 1));
+		len = end - buf;
+	}
+
+	return pg_comp_crc32c_sse42(crc0, buf, len);
+}
+
+#endif
diff --git a/src/port/pg_crc32c_sse42_choose.c b/src/port/pg_crc32c_sse42_choose.c
index 65dbc4d4249..74d2421ba2b 100644
--- a/src/port/pg_crc32c_sse42_choose.c
+++ b/src/port/pg_crc32c_sse42_choose.c
@@ -20,30 +20,37 @@
 
 #include "c.h"
 
-#ifdef HAVE__GET_CPUID
+#if defined(HAVE__GET_CPUID) || defined(HAVE__GET_CPUID_COUNT)
 #include <cpuid.h>
 #endif
 
-#ifdef HAVE__CPUID
+#if defined(HAVE__CPUID) || defined(HAVE__CPUIDEX)
 #include <intrin.h>
 #endif
 
+#ifdef HAVE_XSAVE_INTRINSICS
+#include <immintrin.h>
+#endif
+
 #include "port/pg_crc32c.h"
 
+/*
+ * Does XGETBV say the ZMM registers are enabled?
+ *
+ * NB: Caller is responsible for verifying that osxsave is available
+ * before calling this.
+ */
+#ifdef HAVE_XSAVE_INTRINSICS
+pg_attribute_target("xsave")
+#endif
 static bool
-pg_crc32c_sse42_available(void)
+zmm_regs_available(void)
 {
-	unsigned int exx[4] = {0, 0, 0, 0};
-
-#if defined(HAVE__GET_CPUID)
-	__get_cpuid(1, &exx[0], &exx[1], &exx[2], &exx[3]);
-#elif defined(HAVE__CPUID)
-	__cpuid(exx, 1);
+#ifdef HAVE_XSAVE_INTRINSICS
+	return (_xgetbv(0) & 0xe6) == 0xe6;
 #else
-#error cpuid instruction not available
+	return false;
 #endif
-
-	return (exx[2] & (1 << 20)) != 0;	/* SSE 4.2 */
 }
 
 /*
@@ -53,10 +60,48 @@ pg_crc32c_sse42_available(void)
 static pg_crc32c
 pg_comp_crc32c_choose(pg_crc32c crc, const void *data, size_t len)
 {
-	if (pg_crc32c_sse42_available())
+	unsigned int exx[4] = {0, 0, 0, 0};
+
+	/*
+	 * Set fallback. We must guard since slicing-by-8 is not visible
+	 * everywhere.
+	 */
+#ifdef USE_SSE42_CRC32C_WITH_RUNTIME_CHECK
+	pg_comp_crc32c = pg_comp_crc32c_sb8;
+#endif
+
+#if defined(HAVE__GET_CPUID)
+	__get_cpuid(1, &exx[0], &exx[1], &exx[2], &exx[3]);
+#elif defined(HAVE__CPUID)
+	__cpuid(exx, 1);
+#else
+#error cpuid instruction not available
+#endif
+
+	if ((exx[2] & (1 << 20)) != 0)	/* SSE 4.2 */
+	{
 		pg_comp_crc32c = pg_comp_crc32c_sse42;
-	else
-		pg_comp_crc32c = pg_comp_crc32c_sb8;
+
+		if (exx[2] & (1 << 27) &&	/* OSXSAVE */
+			zmm_regs_available())
+		{
+			/* second cpuid call on leaf 7 to check extended AVX-512 support */
+
+			memset(exx, 0, 4 * sizeof(exx[0]));
+
+#if defined(HAVE__GET_CPUID_COUNT)
+			__get_cpuid_count(7, 0, &exx[0], &exx[1], &exx[2], &exx[3]);
+#elif defined(HAVE__CPUIDEX)
+			__cpuidex(exx, 7, 0);
+#endif
+
+#ifdef USE_AVX512_CRC32C_WITH_RUNTIME_CHECK
+			if (exx[2] & (1 << 10) &&	/* VPCLMULQDQ */
+				exx[1] & (1 << 31)) /* AVX512-VL */
+				pg_comp_crc32c = pg_comp_crc32c_avx512;
+#endif
+		}
+	}
 
 	return pg_comp_crc32c(crc, data, len);
 }
diff --git a/src/test/regress/expected/strings.out b/src/test/regress/expected/strings.out
index dc485735aa4..174f0a68331 100644
--- a/src/test/regress/expected/strings.out
+++ b/src/test/regress/expected/strings.out
@@ -2330,6 +2330,30 @@ SELECT crc32c('The quick brown fox jumps over the lazy dog.');
  419469235
 (1 row)
 
+SELECT crc32c(repeat('A', 127)::bytea);
+  crc32c   
+-----------
+ 291820082
+(1 row)
+
+SELECT crc32c(repeat('A', 128)::bytea);
+  crc32c   
+-----------
+ 816091258
+(1 row)
+
+SELECT crc32c(repeat('A', 129)::bytea);
+   crc32c   
+------------
+ 4213642571
+(1 row)
+
+SELECT crc32c(repeat('A', 800)::bytea);
+   crc32c   
+------------
+ 3134039419
+(1 row)
+
 --
 -- encode/decode
 --
diff --git a/src/test/regress/sql/strings.sql b/src/test/regress/sql/strings.sql
index aeba798dac1..f7b325baadf 100644
--- a/src/test/regress/sql/strings.sql
+++ b/src/test/regress/sql/strings.sql
@@ -738,6 +738,11 @@ SELECT crc32('The quick brown fox jumps over the lazy dog.');
 SELECT crc32c('');
 SELECT crc32c('The quick brown fox jumps over the lazy dog.');
 
+SELECT crc32c(repeat('A', 127)::bytea);
+SELECT crc32c(repeat('A', 128)::bytea);
+SELECT crc32c(repeat('A', 129)::bytea);
+SELECT crc32c(repeat('A', 800)::bytea);
+
 --
 -- encode/decode
 --
-- 
2.48.1